Pith. sign in

REVIEW 2 major objections 4 minor 58 references

Koopman-Equivariant Gaussian Processes

T0 review · 2 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that Koopman-equivariant Gaussian processes are the first universal GP models with fully tractable, closed-form confidence bounds for multi-step prediction, and that their equivariant kernel provably reduces the sample…

desk verdict Interesting GP-Koopman kernel idea, but the equivariance projection at the center does not hold for finite-window semigroup averaging, so the sample-complexity claims are unsupported. read the letter →

arxiv 2502.06645 v1 pith:V2W4E3WR submitted 2025-02-10 cs.LG cs.SYeess.SYstat.ML

Petar Bevanda , Max Beier , Armin Lederer , Alexandre Capone , Stefan Sosnowski , Sandra Hirche This is my paper · ORCID
classification cs.LGcs.SYeess.SYstat.ML
keywords KoopmanoperatorGaussianprocessesequivariancemulti-steppredictionuncertaintyquantificationinformationgainvariationalinferencedynamicalsystems
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

Gaussian processes give principled uncertainty but require approximations for multi-step prediction, while Koopman operator methods give linear forecasts without epistemic uncertainty. This paper proposes Koopman-equivariant Gaussian processes (KE-GPs), which embed the Koopman mode decomposition directly into the GP covariance function so that the posterior evolves as a linear time-invariant system in closed form. The authors aim to prove that this construction delivers universal GP models with fully tractable confidence bounds over whole time intervals and a guaranteed sample-complexity reduction, validated on predator-prey, robotic halfcheetah, and weather temperature data.

What carries the argument

The load-bearing object is the Koopman-equivariant kernel $k_y^{\mathrm{KE}}((t,\cdot),(t',\cdot')) = \sum_{j\in[D]} a_j(t,t') k_{\phi_{\lambda_j}}(\cdot,\cdot)$, built by sym-metrizing a base GP kernel through the operator $E_{\lambda} g := \mathbb{E}_{t\sim\mu([\tau_s,\tau_e])}[e^{-\lambda t} g(x(t))]$. Because $E_\lambda$ is linear, the closedness of Gaussian processes under linear operators turns the symmetrized features $\phi_{\lambda_j}$ into a GP with covariance $k_{\phi_{\lambda_j}} = E_{\lambda_j} k_g E_{\lambda_j}^*$, and the temporal factors $a_j(t,t') = e^{\lambda_j t} e^{\lambda_j^* t'}$ make the posterior evolve as a linear time-invariant system. This construction lets the standard GP posterior equations produce closed-form uncertainty over an entire time interval, and it provides the spectral structure that the information-gain theorem exploits.

What would settle it

Compute $E_\lambda^2(g)$ and $E_\lambda(g)$ numerically for a known linear system (e.g., the 2D rotation $\dot{x}_1 = -6 x_2, \dot{x}_2 = 6 x_1$ used in the paper) on the window $[\tau_s,0]$ with the trapezoidal quadrature described in the supplement; if $E_\lambda^2(g)$ differs measurably from $E_\lambda(g)$ at window boundaries, the projection property that Theorem 4.2 relies on fails to hold exactly.

Watch

Extended reading notes

Core claim

The central claim is that by folding the spectral structure of the Koopman operator into a Gaussian process prior, one obtains a model that is simultaneously a universal approximator of continuous observables, a multi-step forecaster with closed-form posterior distributions on entire time intervals, and a statistically efficient learner whose maximal information gain is provably smaller than that of any trajectory-agnostic kernel. Concretely, the Koopman-equivariant kernel (covKESD) built from the symmetrization operator $E_\lambda$ in equation (6) yields Theorem 4.2: under assumptions (HR), (WS), and (OR), the information gain satisfies $\gamma_N^{\sigma}(k_y^{\mathrm{KE}}) \in \widetilde{O}\big((\gamma_N^{\sigma}(k_x))^{1/\theta}\big)$ with $\theta \ge 1$, where $\theta$ is determined by the decay rates of the Koopman operator's eigenvalues and the Mercer eigenvalues of the base kernel. The authors claim this makes KE-GPs the first universal GP models with fully tractable and closed-form confidence bounds for multi-step prediction.

Load-bearing premise

The paper's guarantees rest on the symmetrization operator $E_\lambda$ in equation (6) being an exact projection onto Koopman-equivariant functions, a property proved via group-averaging results even though the finite-window time shifts only form a semigroup.

Editorial extensions

If this is right

  • Multi-step trajectory predictions can be made with exact, closed-form variance over an entire time interval, eliminating iterative sampling or moment-matching approximations in GP-based forecasting.
  • For systems whose Koopman spectra decay quickly, the maximal information gain of the equivariant kernel is a fractional power of that of a generic kernel, so fewer samples are needed for the same accuracy.
  • The spectral hyperprior enables principled Bayesian model selection for Koopman representations, since the eigenvalue distribution is optimized through the marginal likelihood instead of being estimated in a separate regression stage.
  • Variational inference with inducing trajectories scales the model to thousands of datapoints while avoiding the optimization of time/context-related inducing points that plagues contextual GPs.

Reading between the lines

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

  • The same symmetrization idea may carry over to stochastic or transfer-operator settings, where the semigroup acts on distributions rather than states; the paper does not discuss this extension.
  • A practical diagnostic suggested by Theorem 4.2 is to estimate the singular-value decay of the symmetrized operator from data; when the decay is fast, users should expect the equivariant kernel to dominate generic kernels, but the paper does not provide such an estimator.
  • The non-recurrence assumption (WS) means the guarantees are sharpest for transient windows; for recurrent or near-periodic dynamics the window-averaged equivariance degrades, so the closed-form uncertainty may understate error on long horizons.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes Koopman-Equivariant Gaussian Processes (KE-GPs), a GP model for forecasting dynamical systems with linear-time-invariant response in a latent Koopman representation. The main construction is a symmetrization operator E_λ defined in eq. (6) as a finite-window time average over past trajectories, which is claimed in Theorem 3.2 to project arbitrary base functions onto the set of Koopman-equivariant functions. This equivariant kernel (covKESD) is then used to derive, in Theorem 4.2 and Table 2, improved information-gain rates relative to trajectory-agnostic kernels. A variational inference scheme with inducing trajectories and experiments on predator-prey, D4RL, and weather data are also presented.

Significance. The paper targets an important gap: tractable GP uncertainty propagation for multi-step prediction of nonlinear dynamics, and it bundles this with a claimed sample-complexity reduction from Koopman equivariance. The problem statement, the spectral-decomposition kernel (covSD), and the variational inference design (Section 5) are coherent and potentially useful. However, the central theoretical contribution, Theorem 3.2, is not correct as stated, and Theorem 4.2 inherits this defect along with an internal inconsistency. Since the equivariance construction is the foundation for the kernel and the information-gain claims, the paper's main selling points are currently unsupported.

major comments (2)
  1. [Theorem 3.2, eq. (6)] The operator E_λ defined as a uniform average over t ∈ [τ_s, τ_e] is not a projection, and it does not map arbitrary functions into the set S_λ of [τ_s, τ_e]_λ-Koopman-equivariant functions. For a linear flow F_t(x) = x + vt and g(x) = e^{iξ·x}, one has E_λ g = c g with c = (1/L)∫_{τ_s}^{τ_e} e^{(iξ·v−λ)t} dt; thus E_λ^2 g = c^2 g, which equals E_λ g only if c ∈ {0,1}. In general c ∉ {0,1}, and unless λ = iξ·v, E_λ g is not equivariant at all because evaluating E_λ g at F_t(x) changes the integration limits to [τ_s+t, τ_e+t]. The proof in Appendix D.2 cites group-averaging results (Elesedy and Zaidi 2021) that apply to compact group actions with a Haar measure; the finite-window time-shift family {F_t : t ∈ [τ_s, τ_e]} is a semigroup, not a group, and composition leaves the interval. Consequently, the minimization-projection claim (7), the fixed-point characterization in Theorem D.2, and the decomposition S_λ ⊕ S_λ^⊥ are all unsubstantiated, and the kernel covKESD does not enforce the claimed Koopman equivariance.
  2. [Theorem 4.2, Appendix D.5, Assumption (OR)] There is a direct internal inconsistency in the exponential-case exponent. The main text states θ = max{2p,b}/b, while the proof in Appendix D.5 derives λ_j(A_1 ι_x ι_x^* A_1^*) ≲ e^{−j max{p,b}} and then sets θ := max{p,b}/b. These formulas differ, and since θ determines the advertised reduction (γ_N(k_x))^{1/θ}, the rate in Table 2 and the abstract depends on which exponent is correct. Independently of this inconsistency, the proof of Theorem 4.2 relies on Lemma C.1 to equate the integral operator of k_y^KE with A_1 T_{k_x} A_1^*, and that lemma presupposes that the E_{λ_j} are genuine eigenspace projections of the Koopman semigroup. Since that presupposition fails by the previous comment, the information-gain bound does not follow from the construction. Assumption (OR) postulates compact normality and a certain singular-value decay for A_1 built from the same E_{λ_j} operators, so the claimed 'reduction' is an assumed property of the construction rather than a proven consequence of Koopman equivariance.
minor comments (4)
  1. [Section 3, eq. (8)] Definition 3.1 states that φ_λ is a function on the manifold M, but eq. (8) writes φ_λ(x_{[τ_s,0]}) as a function of a trajectory segment; the domain of the GP and the domain of the equivariant features should be reconciled explicitly.
  2. [Section 4, Table 2 and surrounding text] The text contains a typo, 'Koompan' instead of 'Koopman'; likewise 'Predator-pray' in Section 6 should be 'Predator-prey'.
  3. [Section 5, eq. (12)] The variational loss in eq. (12) has inconsistent signs: the first term should contribute positively to the expected log-likelihood in an ELBO, but as printed the expression mixes negative terms in a way that does not match the standard stochastic variational inference objective.
  4. [Abstract and Section 1] The claim of being 'the first universal GP models with fully tractable and closed-form confidence bounds for multi-step prediction' is stronger than demonstrated; contextual/spatio-temporal GPs such as the C-GP baseline also provide closed-form predictive covariances, so the novelty statement should be qualified.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the claimed sample-complexity reduction is conditional on an explicit spectral-decay assumption, and the main concerns are mathematical validity rather than circular reasoning.

full rationale

Walking the claimed derivation chain, I found no step in which a predicted quantity is, by the paper's own equations, identical to the input used to produce it. The covariance (covSD) is obtained from the linearity of the Koopman mode decomposition and independent GP priors on the eigenfeatures; this is a standard linear-GP construction, not a rerun of the conclusion. The Koopman-equivariant kernel (covKESD) is produced by the symmetrization operator E_lambda in eq. (6). The paper invokes external group-averaging results (Elesedy and Zaidi 2021; Elesedy 2023) to justify Theorem 3.2. Whether those results apply to the semigroup {F_t} on a finite window is a substantive validity concern, but it is not a definitional circularity: Definition 3.1 does not define Koopman-equivariance as 'fixed point of E_lambda', and the proof is not supported by a self-citation chain. Theorem 4.2 is explicitly conditional: it assumes (OR) plus explicit decays lambda_j(A1) <= j^{-p} or e^{-jp}, and then proves an information-gain inequality by standard spectral arguments (Lemma D.4 and Vakili et al. 2021). The bound is not a restatement of the assumption; it is a nontrivial functional dependence theta = max{2p,a}/a or max{p,b}/b. One can question whether Assumption (OR) is verifiable or whether the theorem is unsurprising, but formally the sample-complexity statement is derived from stated assumptions rather than from the very quantity gamma_N(kKE_y) it claims to bound. Self-citations to Bevanda et al. (2023) appear in the universality and spectral-sampling discussions, but universality also rests on external results (Korda and Mezic 2020), and the cited prior work is a separate publication; this is normal citation, not a load-bearing loop. I also note an internal inconsistency between the main-text exponential exponent (theta = max{2p,b}/b) and the appendix proof (theta = max{p,b}/b); this is a correctness bug rather than evidence of circularity. Accordingly, no circular step can be exhibited, and the paper is better assessed on mathematical-validity grounds than on circularity grounds.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central construction relies on the assumed projection and spectral properties of the symmetrization operator and on the assumed decay of the operator A1; these are not independently verified. The learned spectral and kernel hyperparameters are free parameters fitted to data.

free parameters (6)
  • number of spectral modes D = not reported per dataset; D=64 in ablation
    Model dimension balancing resolution and cost; performance saturates beyond about 64 in Fig. 5.
  • spectral distribution parameters theta (scale and bias of uniform eigenvalue prior) = init: omega scale 15, real scale 1, biases 0; optimized via marginal likelihood
    Learned locations of eigenvalues; encodes stability and frequency prior.
  • base kernel lengthscale = sqrt(nx)/2 * std(X_input) initialization, then optimized
    Standard GP hyperparameter controlling spatial smoothness.
  • observation noise variance sigma_on^2 = init to 1, optimized
    Noise level fitted to data.
  • past trajectory interval and discretization H = unit-normalized interval; H=16 or 32
    Defines the symmetrization window; chosen per dataset.
  • inducing trajectories = M initial inducing points sampled from training data, then optimized
    Variational parameters for scalable inference.
assumptions (5)
  • domain assumption The flow F_t is well-defined and the dynamics admit Koopman eigenfunctions that are dense on a non-recurrent domain (Assumption WS, Lemma B.3)
    Needed for KMD universality; excludes windows with recurrent or periodic behavior.
  • ad hoc to paper E_lambda is well-defined, self-adjoint, and idempotent (Theorem 3.2)
    Required for equivariant projection and the A1 spectral representation; proof cites group-action results not applicable to a finite-window semigroup.
  • ad hoc to paper A1 := sum_j e^{lambda_j} E_lambda_j is compact normal with eigenvalue decay j^{-p} or e^{-jp} (Assumption OR)
    Load-bearing for Theorem 4.2; unverified for any benchmark system.
  • standard math Base kernel kx is a bounded Mercer kernel with bounded eigenfunctions (Assumption HR)
    Standard assumptions for Mercer decomposition and information-gain bounds.
  • domain assumption Empirical quadrature error |hat(E_lambda) - E_lambda| is negligible
    Needed to treat discrete trajectory data as an exact equivariant operator; only informally justified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Koopman-Equivariant Gaussian Processes." pith.science (2026). https://pith.science/paper/V2W4E3WR

@misc{pith2026250206645,
  author       = {Pith},
  title        = {Pith review of: Koopman-Equivariant Gaussian Processes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V2W4E3WR}},
  note         = {Machine review of arXiv:2502.06645}
}
read the original abstract

Credible forecasting and representation learning of dynamical systems are of ever-increasing importance for reliable decision-making. To that end, we propose a family of Gaussian processes (GP) for dynamical systems with linear time-invariant responses, which are nonlinear only in initial conditions. This linearity allows us to tractably quantify forecasting and representational uncertainty, simultaneously alleviating the challenge of computing the distribution of trajectories from a GP-based dynamical system and enabling a new probabilistic treatment of learning Koopman operator representations. Using a trajectory-based equivariance -- which we refer to as \textit{Koopman equivariance} -- we obtain a GP model with enhanced generalization capabilities. To allow for large-scale regression, we equip our framework with variational inference based on suitable inducing points. Experiments demonstrate on-par and often better forecasting performance compared to kernel-based methods for learning dynamical systems.

Figures

Figures reproduced from arXiv: 2502.06645 by the authors.

Figure 1
Figure 1. Backward time equivariance interval (red) [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Multi-step mean and 2-sigma interval of the prediction for predator population from the predator-prey [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Empirical information gain γˆ for a 2D linear system scaled to remove effects of constants. The improved rates confirm our theoretical results for Koopman-equivariant GPs, leading to a lower informa￾tion gain compared to their non-equivariant counter￾part (covSD), even when a randomly sampled eigenvalue spectrum {λj} D j=1 is used instead of the true spectrum. ries are perturbed by noise as commonly encountered in p… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of the GP covariances in space and time. The spatial, KE-GP prior already strongly [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: Ablation Studies: test RMSE with varying data and spectral parameters; we report mean and [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 32 canonical work pages

  1. [10]

    doi: 10.1016/j.cnsns.2021.105833

    ISSN 10075704. doi: 10.1016/j.cnsns.2021.105833. Edwin V Bonilla, Kian Chai, and Christopher Williams. Multi-task gaussian process prediction. In Advances in Neural Information Processing Systems , volume 20,

    work page arXiv 2021

  2. [11]

    Stochastic data-driven model predictive control using Gaussian processes

    URL http://arxiv.org/abs/1908.01786. Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym,

    work page Pith review arXiv 1908

  3. [12]

    doi: 10.1146/annurev-control-042920-020211. P . Bevanda, M. Beier, A. Lederer, A. Capone, S. Sosnowski, S. Hirche Steven L. Brunton and J. Nathan Kutz. Data-Driven Science and Engineering . Cambridge University Press, 1

    doi:10.1146/annurev-control-042920-020211

  4. [15]

    doi: 10.1137/20M1338289. David R. Burt, Carl E. Rasmussen, and Mark van der Wilk. Rates of Convergence for Sparse Variational Gaussian Process Regression. 3

    work page doi:10.1137/20m1338289

  5. [16]

    Rates of Convergence for Sparse Variational Gaussian Process Regression

    URL http://arxiv.org/abs/1903.03571. 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. arXiv preprint arXiv:2403.02811 ,

    work page Pith review arXiv 1903

  6. [18]

    doi: https://doi.org/10.1016/j.automatica.2017.12.039

    ISSN 0005-1098. doi: https://doi.org/10.1016/j.automatica.2017.12.039. Sebastian Curi, Felix Berkenkamp, and Andreas Krause. Efficient model-based reinforcement learning through optimistic policy search and planning. In Advances in Neural Information Processing Systems ,

    work page doi:10.1016/j.automatica.2017.12.039 2017

  7. [19]

    Safe reinforcement learning via confidence- based filters

    Sebastian Curi, Armin Lederer, Sandra Hirche, and Andreas Krause. Safe reinforcement learning via confidence- based filters. In 2022 IEEE 61st Conference on Decision and Control (CDC) , pages 3409–3415,

    work page 2022

  8. [20]

    P Cvitanovi´ c, R Artuso, R Mainieri, G Tanner, and G Vattay.Chaos: Classical and Quantum

    doi: 10.1109/CDC51059.2022.9992470. P Cvitanovi´ c, R Artuso, R Mainieri, G Tanner, and G Vattay.Chaos: Classical and Quantum . Niels Bohr Inst., Copenhagen,

    arXiv 2022

Show all 58 references
  1. [21]

    doi: https: //doi.org/10.1016/j.physd.2023.134044

    ISSN 0167-2789. doi: https: //doi.org/10.1016/j.physd.2023.134044. Nelson Dunford. Spectral Theory. I Convergence to Projections. Transactions of the American Mathematical Society, 54(2):185, 9

    2023

  2. [23]

    Efficiently modeling long sequences with structured state spaces

    Albert Gu, Karan Goel, and Christopher R´ e. Efficiently modeling long sequences with structured state spaces. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022,

    2022

  3. [24]

    Fr¨ ohlich, and Melanie N

    Lukas Hewing, Elena Arcari, Lukas P. Fr¨ ohlich, and Melanie N. Zeilinger. On simulation and trajectory prediction with Gaussian process dynamics. In Proceedings of the 2nd Conference on Learning for Dynamics and Control , volume 120 of Proceedings of Machine Learning Research...

  4. [26]

    doi: 10.1063/5.0094889

    ISSN 1054-1500. doi: 10.1063/5.0094889. Isao Ishikawa, Yuka Hashimoto, Masahiro Ikeda, and Yoshinobu Kawahara. Koopman operators with intrinsic observables in rigged reproducing kernel hilbert spaces. 3

    doi

  5. [27]

    Olav Kallenberg

    URL http://arxiv.org/abs/2403.02524. Olav Kallenberg. Foundations of Modern Probability. Springer-Verlag,

  6. [28]

    doi: 10.1007/ b98838

    ISBN 0-387-94957-7. doi: 10.1007/ b98838. URL http://link.springer.com/10.1007/b98838. Takahiro Kawashima and Hideitsu Hino. Gaussian Process Koopman Mode Decomposition. Neural Computation, 35(1):82–103, 01

    doi

  7. [30]

    Jonathan Ko, Daniel J Klein, Dieter Fox, and Dirk Haehnel

    doi: 10.1007/s00440-015-0619-7. Jonathan Ko, Daniel J Klein, Dieter Fox, and Dirk Haehnel. Gp-ukf: Unscented kalman filters with gaussian process prediction and observation models. In Proceedings of the International Conference on Intelligent Robots and Systems ,

    doi

  8. [32]

    doi: 10.1007/s00332-017-9423-0

    ISSN 14321467. doi: 10.1007/s00332-017-9423-0. Milan Korda and Igor Mezi´ c. Optimal Construction of Koopman Eigenfunctions for Prediction and Control.IEEE Transactions on Automatic Control, 65(12):5114–5129, 12

    doi

  9. [33]

    doi: 10.1109/TAC.2020.2978039

    ISSN 15582523. doi: 10.1109/TAC.2020.2978039. Vladimir Kostic, Pietro Novelli, Andreas Maurer, Carlo Ciliberto, Lorenzo Rosasco, and Massimiliano Pontil. Learning Dynamical Systems via Koopman Operator Regression in Reproducing Kernel Hilbert Spaces. In Advances in Neural Info...

    2020

  10. [35]

    doi: 10.1007/s00233-018-9958-x

    ISSN 00371912. doi: 10.1007/s00233-018-9958-x. Pijush K Kundu, Ira M Cohen, and David R Dowling. Fluid mechanics. Academic press,

    doi

  11. [36]

    Maud Lemercier, Cristopher Salvi, Thomas Cass, Edwin V Bonilla, Theodoros Damoulas, and Terry J Lyons

    URL http://arxiv.org/abs/2305.04625. Maud Lemercier, Cristopher Salvi, Thomas Cass, Edwin V Bonilla, Theodoros Damoulas, and Terry J Lyons. SigGPDE: Scaling Sparse Gaussian Processes on Sequential Data. In Proceedings of the 38th International Conference on Machine Learning , ...

    arXiv

  12. [37]

    J. Li, M. Zagorowska, G. De Pasquale, A. Rupenyan, and John Lygeros. Safe time-varying optimization based on gaussian processes with spatio-temporal kernel. In Proceedings of NeurIPS 2024,

    2024

  13. [38]

    Yingzhao Lian and Colin N

    URL http://arxiv.org/abs/2208.01711. Yingzhao Lian and Colin N. Jones. On Gaussian process based koopman operators. In IF AC-PapersOnLine, volume 53, pages 449–455. Elsevier B.V.,

    arXiv

  14. [39]

    Bryan Lim and Stefan Zohren

    doi: 10.1016/j.ifacol.2020.12.217. Bryan Lim and Stefan Zohren. Time-series forecasting with deep learning: a survey. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences , 379:20200209, 4

    2020 doi

  15. [40]

    Kartik Loya, Jake Buzhardt, and Phanindra Tallapragada

    doi: 10.1126/science.1208351. Kartik Loya, Jake Buzhardt, and Phanindra Tallapragada. Koopman Operator Based Predictive Control With a Data Archive of Observables. ASME Letters in Dynamic Systems and Control , 3(3):031009, 10

    doi

  16. [41]

    Alexandre Mauroy, Igor Mezi´ c, and Yoshihiko Susuki.The Koopman Operator in Systems and Control , volume 484 of Lecture Notes in Control and Information Sciences

    URL http: //arxiv.org/abs/2405.15945. Alexandre Mauroy, Igor Mezi´ c, and Yoshihiko Susuki.The Koopman Operator in Systems and Control , volume 484 of Lecture Notes in Control and Information Sciences . Springer International Publishing, Cham,

  17. [43]

    Igor Mezi´ c

    URL http://arxiv.org/abs/2405.14778. Igor Mezi´ c. Spectrum of the Koopman Operator, Spectral Expansions in Functional Spaces, and State-Space Geometry. Journal of Nonlinear Science , 30(5):2091–2145,

    arXiv

  18. [44]

    Samuel E Otto and Clarence W Rowley

    URL http://arxiv.org/abs/2311.07143. Samuel E Otto and Clarence W Rowley. Koopman Operators for Estimation and Control of Dynamical Systems. Annual Review of Control, Robotics, and Autonomous Systems , 4:2021,

    2021 arXiv

  19. [45]

    Exact Multiple-Step Predictions in Gaussian Process- based Model Predictive Control: Observations, Possibilities, and Challenges

    Maik Pfefferkorn, Michael Maiworm, and Rolf Findeisen. Exact Multiple-Step Predictions in Gaussian Process- based Model Predictive Control: Observations, Possibilities, and Challenges. In 2022 American Control Conference (ACC), pages 2829–2836,

    2022

  20. [46]

    Thomas Pinder and Daniel Dodd

    doi: 10.23919/ACC53348.2022.9867259. Thomas Pinder and Daniel Dodd. Gpjax: A gaussian process framework in jax. Journal of Open Source Software , 7(75):4455,

    2022

  21. [47]

    URL https://doi.org/10.21105/joss.04455

    doi: 10.21105/joss.04455. URL https://doi.org/10.21105/joss.04455. Kyriakos Polymenakos, Luca Laurenti, Andrea Patane, Jan-Peter Calliess, Luca Cardelli, Marta Kwiatkowska, Alessandro Abate, and Stephen Roberts. Safety guarantees for iterative predictions with gaussian process...

    doi

  22. [48]

    Joaquin Qui˜ nonero-Candela and Carl Edward Rasmussen

    ISBN 9781728174471. Joaquin Qui˜ nonero-Candela and Carl Edward Rasmussen. A unifying view of sparse approximate gaussian process regression. Journal of Machine Learning Research , 6(65):1939–1959,

    1939

  23. [49]

    Cristopher Salvi, Thomas Cass, James Foster, Terry Lyons, and Weixin Yang

    URL http://arxiv.org/abs/2211.11103. Cristopher Salvi, Thomas Cass, James Foster, Terry Lyons, and Weixin Yang. The Signature Kernel Is the Solution of a Goursat PDE. SIAM Journal on Mathematics of Data Science , 3(3):873–899, 1

    arXiv

  24. [50]

    doi: 10.1137/20M1366794

    ISSN 2577-0187. doi: 10.1137/20M1366794. Simo Sarkka, Arno Solin, and Jouni Hartikainen. Spatiotemporal learning via infinite-dimensional bayesian filtering and smoothing: A look at gaussian process regression through kalman filtering. IEEE Signal Processing Magazine, 30(4):51–61,

    doi

  25. [52]

    Financial time series forecasting with deep learning : A systematic literature review: 2005–2019

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. Financial time series forecasting with deep learning : A systematic literature review: 2005–2019. Applied Soft Computing , 90:106181,

    2005

  26. [53]

    doi: 10.1109/TIT.2011.2182033

    ISSN 0018-9448. doi: 10.1109/TIT.2011.2182033. Ingo Steinwart and Clint Scovel. Mercer’s theorem on general domains: On the interaction between measures, kernels, and rkhss. Constructive Approximation, 35:363–417,

    2011

  27. [54]

    doi: 10.1214/009053607000000613

    ISSN 0090-5364. doi: 10.1214/009053607000000613. Mark van der Wilk, Matthias Bauer, ST John, and James Hensman. Learning invariances using the marginal likelihood. In Advances in Neural Information Processing Systems , volume 31,

    doi

  28. [55]

    Jack Wang, Aaron Hertzmann, and David J Fleet

    doi: 10.1038/s41592-019-0686-2. Jack Wang, Aaron Hertzmann, and David J Fleet. Gaussian process dynamical models. In Advances in Neural Information Processing Systems, volume

    doi

  29. [56]

    doi: 10.1007/s00332-019-09567-y

    ISSN 0938-8974. doi: 10.1007/s00332-019-09567-y. Houssam Zenati, Alberto Bietti, Eustache Diemert, Julien Mairal, Matthieu Martin, and Pierre Gaillard. Efficient kernelized ucb for contextual bandits. In Proceedings of The 25th International Conference on Artificial Intelligen...

    doi

  30. [57]

    Zhizhen Zhao and Dimitrios Giannakis

    URL http://arxiv.org/abs/2303.01927. Zhizhen Zhao and Dimitrios Giannakis. Analog forecasting with dynamics-adapted kernels. Nonlinearity, 29(9): 2888, aug

    arXiv

  31. [58]

    Mauricio ´Alvarez, David Luengo, and Neil D

    doi: 10.1088/0951-7715/29/9/2888. Mauricio ´Alvarez, David Luengo, and Neil D. Lawrence. Latent force models. In Proceedings of the Twelth International Conference on Artificial Intelligence and Statistics , volume 5 of Proceedings of Machine Learning Research, pages 9–16. PML...

    doi

  32. [1909]

    doi: 10.1098/rsta.1909.0016

    ISSN 0264-3952, 2053-9258. doi: 10.1098/rsta.1909.0016. Dimitri Meunier, Zikai Shen, Mattes Mollenhauer, Arthur Gretton, and Zhu Li. Optimal rates for vector-valued spectral regularization learning algorithms. 5

    1909

  33. [1931]

    doi: 10.1073/pnas.17.5.315

    ISSN 0027-8424. doi: 10.1073/pnas.17.5.315. 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, 4

    doi

  34. [1943]

    doi: 10.2307/1990329

    ISSN 00029947. doi: 10.2307/1990329. David Kristjanson Duvenaud. Automatic Model Construction with Gaussian Processes . PhD thesis, University of Cambridge,

    doi

  35. [1997]

    doi: 10.1007/978-1-4612-0653-8

    ISBN 978-1-4612-6857-4. doi: 10.1007/978-1-4612-0653-8. A. Billard, S. Mirrazavi, and N. Figueroa. Learning for Adaptive and Reactive Robot Control: A Dynamical Systems Approach. MIT Press,

    doi

  36. [2008]

    Tianshi Chen

    doi: 10.1088/0266-5611/24/3/034004. Tianshi Chen. On kernel design for regularized LTI system identification. Automatica, 90:109–122,

    doi

  37. [2011]

    Henrik Kreidler

    URL https://proceedings.neurips.cc/paper_files/ paper/2011/file/f3f1b7fc5a8779a9e618e1f23a7b7860-Paper.pdf. Henrik Kreidler. Compact operator semigroups applied to dynamical systems. Semigroup Forum, 97(3):523–547, 12

    2011

  38. [2012]

    doi: 10.1063/1.4772195

    ISSN 10541500. doi: 10.1063/1.4772195. Dmitry Burov, Dimitrios Giannakis, Krithika Manohar, and Andrew Stuart. Kernel analog forecasting: Multiscale test problems. Multiscale Modeling & Simulation , 19(2):1011–1040,

    doi

  39. [2013]

    V ´ ıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling

    doi: 10.1109/MSP.2013.2246292. V ´ ıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) equivariant graph neural networks. In Proceedings of the 38th International Conference on Machine Learning , volume 139 of Proceedings of Machine Learning Research, pages 9323–9332...

    2013

  40. [2014]

    doi: 10.1146/annurev-economics-080213-041213

    ISSN 1941-1383. doi: 10.1146/annurev-economics-080213-041213. Carl Hvarfner, Erik Orm Hellsten, and Luigi Nardi. Vanilla Bayesian optimization performs great in high dimensions. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarl...

    1941 doi

  41. [2015]

    Felix Berkenkamp, Matteo Turchetta, Angela Schoellig, and Andreas Krause

    doi: 10.1109/ECC.2015.7330913. Felix Berkenkamp, Matteo Turchetta, Angela Schoellig, and Andreas Krause. Safe model-based reinforcement learning with stability guarantees. In Advances in Neural Information Processing Systems , volume 30,

    2015

  42. [2016]

    doi: https://doi.org/10.1016/j.acha.2015.03.002

    ISSN 1063-5203. doi: https://doi.org/10.1016/j.acha.2015.03.002. Petar Bevanda, Stefan Sosnowski, and Sandra Hirche. Koopman operator dynamical models: Learning, analysis and control. Annual Reviews in Control , 52:197–212,

    2015 doi

  43. [2017]

    Gosafe: Globally optimal safe robot learning

    Dominik Baumann, Alonso Marco, Matteo Turchetta, and Sebastian Trimpe. Gosafe: Globally optimal safe robot learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA) , pages 4452–4458,

    2021

  44. [2018]

    Schoellig

    Felix Berkenkamp and Angela P. Schoellig. Safe and robust learning control with Gaussian processes. In 2015 European Control Conference (ECC), pages 2496–2501,

    2015

  45. [2019]

    doi: 10.1017/9781108380690

    ISBN 9781108380690. doi: 10.1017/9781108380690. Steven L Brunton, Marko Budiˇ si´ c, Eurika Kaiser, and J Nathan Kutz. Modern Koopman Theory for Dynamical Systems. SIAM Review, 64(2):229–340,

    doi

  46. [2020]

    doi: https://doi.org/10.1016/j.physd.2020.132520

    ISSN 0167-2789. doi: https://doi.org/10.1016/j.physd.2020.132520. Joshua Angrist and Jorn-Steffen Pischke. Mostly Harmless Econometrics: An Empiricist’s Companion . Princeton University Press, 1 edition,

    2020

  47. [2021]

    Thomas Beckers and Sandra Hirche

    doi: 10.1109/ICRA48506.2021.9560738. Thomas Beckers and Sandra Hirche. Prediction with Approximated Gaussian Process Dynamical Models. IEEE Transactions on Automatic Control , 67(12):6460–6473,

    2021

  48. [2022]

    Mikhail Belkin

    doi: 10.1109/TAC.2021.3131988. Mikhail Belkin. Approximation beats concentration? an approximation view on inference with smooth radial kernels. In Proceedings of the 31st Conference On Learning Theory , volume 75 of Proceedings of Machine Learning Research, pages 1348–1361. P...

    2021

  49. [2023]

    Jinwoo Kim, Dat Nguyen, Ayhan Suleymanzade, Hyeokjun An, and Seunghoon Hong

    doi: 10.1109/TIE.2022.3158007. Jinwoo Kim, Dat Nguyen, Ayhan Suleymanzade, Hyeokjun An, and Seunghoon Hong. Learning probabilistic symmetrization for architecture agnostic equivariance. 36:18582–18612,

    2022

  50. [2024]

    Ralf Banisch and P´ eter Koltai

    URL http://arxiv.org/abs/2405.05890. Ralf Banisch and P´ eter Koltai. Understanding the geometry of transport: Diffusion maps for Lagrangian trajectory data unravel coherent sets. Chaos: An Interdisciplinary Journal of Nonlinear Science , 27(3):035804, 02

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.