Pith. sign in

REVIEW 1 major objections 6 minor 1 cited by

Self-Supervised Evolution Operator Learning for High-Dimensional Dynamical Systems

T0 review · 1 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Minimizing a self-supervised contrastive loss is equivalent to learning the evolution operator and its spectral decomposition.

desk verdict Clean algebra and a genuinely useful bridge between contrastive learning and evolution operators, but the deterministic experiments sit outside the paper's own Hilbert-Schmidt theory. read the letter →

arxiv 2505.18671 v1 pith:S4G7SETK submitted 2025-05-24 cs.LG math.DS

Giacomo Turri , Luigi Bonati , Kai Zhu , Massimiliano Pontil , Pietro Novelli This is my paper
classification cs.LGmath.DS
keywords evolutionoperatorsself-supervisedlearningcontrastiverepresentationspectraldecompositionKoopmanoperatormoleculardynamicsclimatevariabilityVAMP-2score
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 sets out to show that a single self-supervised contrastive objective can learn the evolution operator of a Markovian dynamical system and its spectral decomposition, replacing hand-built dictionaries and reconstruction decoders. The objective is the squared error between the true transition density ratio $r(x,y)=p(y|x)/p(y)$ and a bilinear model $\langle \varphi(x), P\varphi(y)\rangle$; for any fixed encoder $\varphi$ the optimal predictor is $P^* = C_X^{-1}C_{XY}C_Y^{-1}$, so the implied operator $E_\varphi = P^* C_Y$ coincides with the classical least-squares estimator (with zero regularization) and the minimized loss equals the negative VAMP-2 score. When the evolution operator is Hilbert-Schmidt, the same loss is exactly the squared error of a finite-rank operator regression, which is why training the encoder makes its span approximate the leading singular modes. The paper demonstrates the approach on protein folding, host-guest binding, and global climate data, where the learned modes match known physical coordinates and climatic indices. If the claim is right, spectral analysis of large dynamical systems becomes a scalable, encoder-only pipeline that still produces interpretable modes.

What carries the argument

The central object is the density-ratio estimator $r(x,y)=p(y|x)/p(y)$ modeled by the bilinear form $\langle \varphi(x), P\varphi(y)\rangle$, trained with the contrastive loss (8): $\hat{\varepsilon}(\varphi,P)=\frac{1}{N(N-1)}\sum_{i\neq j}\langle \varphi(x_i),P\varphi(y_j)\rangle^2 - \frac{2}{N}\sum_i \langle \varphi(x_i),P\varphi(y_i)\rangle$. The argument is carried by three identities: the closed-form minimizer $P^*=C_X^{-1}C_{XY}C_Y^{-1}$; the least-squares identity $P^*C_Y = C_X^{-1}C_{XY}$; and the Hilbert-Schmidt equivalence $\varepsilon(\varphi,P)=\|E-\Phi_\nu^*P\Phi_\mu\|_{HS}^2$, where $\Phi_\mu$ is the evaluation map into the span of the encoder. Together these identities show the loss is minimized exactly when the encoder's span best approximates the leading singular space of the evolution operator.

What would settle it

Run the method on a deterministic dynamical system with an analytically known Koopman spectrum (for instance, a linear rotation on the circle, whose eigenfunctions are the Fourier modes), and compare the learned eigenfunctions with the analytic ones as the amount of data grows; since the rotation's evolution operator is not Hilbert-Schmidt, this directly tests whether the spectral-decomposition claim survives where the paper's Assumption 1 fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the density-ratio loss (8) — the squared error between the true transition density ratio $r(x,y)=p(y|x)/p(y)$ and its bilinear model $\langle \varphi(x), P\varphi(y)\rangle$ — is not merely a heuristic contrastive objective but an evolution-operator regression loss. For any fixed encoder $\varphi$, the loss is convex in the predictor $P$, and the minimizer is $P^* = C_X^{-1}C_{XY}C_Y^{-1}$, so the implied finite-dimensional operator $E_\varphi = P^* C_Y = C_X^{-1}C_{XY}$ is exactly the least-squares estimator (3) with $\lambda=0$. At this optimum the loss equals $-\mathrm{VAMP}_2(\varphi)$, and when the true evolution operator $E$ is Hilbert-Schmidt the loss equals the squared Hilbert-Schmidt error $\|E - \Phi_\nu^* P \Phi_\mu\|_{HS}^2$. Consequently, optimizing $\varphi$ makes the span of the encoder approximate the leading singular space of $E$, and the eigenvectors of $E_\varphi$ give coherent dynamical modes. The paper supports this claim by showing that modes learned from raw trajectory data recover folding and binding coordinates in molecular simulations and the El Niño–Southern Oscillation pattern in global sea-surface-temperature data.

Load-bearing premise

The load-bearing premise is that the evolution operator $E$ is Hilbert-Schmidt, so that the density ratio $r(x,y)=p(y|x)/p(y)$ is a finite square-integrable function and the contrastive loss is a genuine operator-regression error; deterministic systems, including the paper's own Lorenz '63 experiment, have point-mass transitions and violate this premise.

Editorial extensions

If this is right

  • An encoder-only network can be trained with no reconstruction decoder and no hand-built dictionary, then frozen; the operator and its modes are obtained by forming covariance matrices and diagonalizing $E_\varphi$.
  • Because the loss requires only matrix multiplications, it runs efficiently on GPUs and scales to very high-dimensional representations, such as all-heavy-atom protein models and global climate fields.
  • The learned representations transfer across related molecular systems: an encoder trained on two ligands recovers the binding modes of a third ligand without retraining.
  • The resulting spectral modes are interpretable: the leading eigenfunctions correlate with known physical coordinates (RMSD during folding, host-guest distance and angle during binding) and with the Oceanic Niño Index in climate data.
  • Since the minimized loss equals the negative VAMP-2 score, the method inherits the variational principle for Markov processes, so higher implied timescales indicate a better approximation of the slow dynamics.

Reading between the lines

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

  • The Hilbert-Schmidt assumption is load-bearing but the paper's own Lorenz '63 experiment is deterministic, where $p(y|x)$ is a point mass and $r(x,y)$ is not a finite function; a useful follow-up would be to test whether the objective still tracks the true Koopman modes of a deterministic system, or to develop a regularized variant with a well-defined limit.
  • The identity $E_\varphi=C_X^{-1}C_{XY}$ suggests the method could be used as a pretraining step for any downstream linear spectral estimator, such as time-lagged linear analysis or Markov-state modeling, by providing a nonlinear encoder that makes the linear operator informative.
  • The same contrastive objective could be extended to multiple time lags or to non-Markovian contexts by augmenting the state with a history context, as the paper notes; a multi-step variant would yield a family of operators and could expose state-dependent timescale separation.
  • The climate experiment's reported training instability suggests the objective may need architectural stabilization, such as spectral normalization or gradient clipping, in high-dimensional settings; this is an engineering consequence of applying the method, not a limitation of the spectral identity itself.
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

1 major / 6 minor

Summary. The paper proposes an encoder-only, self-supervised method for learning evolution operators (Koopman/transfer operators) of high-dimensional dynamical systems. The method optimizes a contrastive loss (Eq. 8) over an encoder φ and a linear predictor P, with a theoretical argument that, for fixed φ, the optimal predictor gives E_φ = C_X^{-1} C_XY (the standard EDMD estimator) and that the loss at the optimum equals -VAMP2(φ). Under a Hilbert-Schmidt assumption on the evolution operator, the loss is shown to equal an operator-regression error in Hilbert-Schmidt norm, so that training φ is claimed to make its span approximate the leading singular space of the operator. The method is demonstrated on Lorenz '63, protein folding (Trp-Cage), ligand binding, and climate (ENSO) data.

Significance. If the theoretical claims are established, the paper offers a scalable and interpretable approach to spectral analysis of high-dimensional dynamical systems, bridging self-supervised representation learning and operator theory. The central identities are correct: the fixed-encoder predictor coincides with EDMD, and the loss matches negative VAMP-2. The experiments are compelling and the released code supports reproducibility. However, the theoretical justification hinges on a Hilbert-Schmidt assumption that is violated in the deterministic systems used as warm-up and climate demonstrations, so the scope of the contribution is narrower than the abstract suggests.

major comments (1)
  1. [Supp. A, Sec. 4.1, Sec. 4.4] The proof of Lemma 1 in Supplementary Section A establishes ε(φ,P) = ||E - Φ*_ν P Φ_μ||²_HS only under Assumption 1 that E is Hilbert-Schmidt. For deterministic continuous-state systems, the Koopman operator is typically not Hilbert-Schmidt on L2 spaces, so the 'remarkable connection' between contrastive learning and operator regression is not valid for the very systems (Lorenz, climate) that motivate the method. The paper should state more explicitly the class of stochastic processes for which the theory applies and explain why the method can still be expected to work heuristically in deterministic cases.
minor comments (6)
  1. [Sec. 3, Eq. (7)] The notation for expectations over the product measure X_t ⊗ X_{t+1} versus the joint measure (X_t, X_{t+1}) is easy to confuse; a brief explanation after Eq. (7) would improve readability.
  2. [Table 1] The caption 'RMSE values are scaled by 10^{-2}' is ambiguous: it is unclear whether the reported numbers are the true RMSE multiplied by 10^{-2} or should be read as percentages. Please clarify.
  3. [Sec. 4.4] The authors select the 11th eigenfunction as the ENSO mode but do not justify why this particular index is 'leading nontrivial' or how it was chosen among the other modes. A clearer selection criterion or a discussion of the spectrum's structure would strengthen the claim.
  4. [Supplementary Material B.4] The training instability mentioned for the climate experiment (difficulty in reproducing runs with strong validation performance) is an important reproducibility concern; consider reporting the variance across seeds or a robustness statement in the main text.
  5. [Sec. 3.1] The final computation of E_φ via Eq. (3) still requires inverting C_X + λI, despite the earlier claim that the loss avoids matrix inversions; please clarify that the inversion is needed for the estimator, not for the training loss.
  6. [Algorithm 1] The algorithm line 'for all i do' is slightly informal; specifying the batch index set explicitly would make the pseudocode more precise.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core equivalence is derived in-paper; the Hilbert-Schmidt limitation is a scope gap, not a circular reduction.

full rationale

The claimed derivation chain is: Eq. (7) is an L2 density-ratio regression; Lemma 1 (Supp. A) proves ε(φ,P)=||E−Φ*_ν P Φ_µ||^2_HS under Assumption 1 via direct trace/cyclic computations; optimizing P yields P*=C_X^{-1}C_XY C_Y^{-1}, hence E_φ=C_X^{-1}C_XY, exactly the λ→0 least-squares estimator (Eq. 9); substituting P* gives ε=−VAMP2(φ). Each step is an algebraic identity proved inside the paper, not an equivalence imposed by definition. The spectral-mode conclusion is the standard Eckart–Young fact for Hilbert-Schmidt operators once Lemma 1 is granted; it does not depend on a fitted parameter being relabeled as a prediction. The only potentially load-bearing self-citation is [54] (overlapping authors), cited for the Hilbert-Schmidt equivalence; however, the paper states the proof is formal and self-contained in the Supplementary Material, and Lemma 1 reproduces the calculation, so the citation is not the sole support. Experiments are external benchmarks (Lorenz, Trp-Cage, host-guest, ORAS5/ChaosBench climate data) with comparisons to baselines and physical descriptors; no experiment uses a quantity that was itself fitted from the reported target. The Limitations paragraph candidly states the Hilbert-Schmidt assumption is 'often violated by deterministic dynamical systems,' which covers the Lorenz warm-up and largely deterministic climate experiment; this is an assumption/scope gap (the formal equivalence may not hold there), not a circular reduction—the paper does not use the violated assumption to define the experimental outputs. Therefore no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central derivation rests on regularity conditions (Hilbert-Schmidt operator, existence of transition densities, covariance invertibility) and on the authors' prior theorem from [54]. No invented entities are introduced. The hand-set numbers are standard hyperparameters, but the climate grid search combined with instability makes the selected configuration load-bearing for the ENSO claim.

free parameters (4)
  • Ridge regularization λ in E_φ = (C_X + λI)^{-1} C_XY = not stated in main text
    Chosen by hand; required for stable inversion of the covariance buffers; affects the estimated operator and hence all reported modes.
  • Lag time per experiment = 10 steps (Lorenz), 1 month (climate), MD lag not stated
    Defines which transition E_φ approximates; changing it changes the implied timescales and the recovered modes.
  • Climate model selection configuration = latent dim 128, history length 1, simplicial dim 4, gradient clipping 0.2
    Selected by grid search over Table 3 ranges using validation score in an unstable regime; the ENSO mode is reported for this selected run.
  • EMA decay for covariance buffers = unspecified
    The online covariance estimate in Sec 3.1 uses an exponentially moving average whose decay rate is not reported; this affects the final E_φ computation.
assumptions (6)
  • domain assumption Assumption 1: the evolution operator E is Hilbert-Schmidt
    Required for the identity ε(φ,P) = ||E - Φ*_ν P Φ_µ||²_HS (Supplementary A). The authors state this is 'often violated by deterministic dynamical systems', which covers their own Lorenz experiment.
  • domain assumption The transition kernel p(y|x) admits a density with respect to p(y), so the ratio r(x,y) = p(y|x)/p(y) is a finite function
    Needed for the importance-sampling rewrite in Eqs (5)-(7); fails for deterministic flows where the transition is a point mass.
  • domain assumption The observed sequence is Markovian
    Assumed at the start of Sec 2; the paper notes the history-augmentation trick when it fails, and the climate experiment appends the previous snapshot as an extra channel.
  • standard math Encoder components are square-integrable with respect to both μ and ν
    Needed for Φ_μ and Φ_ν to be well-defined operators in Supplementary A.
  • standard math C_X and C_Y are invertible, or the ridge term makes them so
    Needed for the closed-form minimizer P* = C_X^{-1} C_XY C_Y^{-1} in Eq (9) and Supplementary A.
  • domain assumption The main equivalence lemma is taken from prior work: [54, Theorem 1], with [56, Lemma 3.2] and [53, Lemma 4.1] as sources
    Supplementary A states the material is 'based on the results of' these papers; the paper does not re-prove the equivalence from scratch, and [54] shares four of five authors with this work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Evolution Operator Learning for High-Dimensional Dynamical Systems." pith.science (2026). https://pith.science/paper/S4G7SETK

@misc{pith2026250518671,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Evolution Operator Learning for High-Dimensional Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4G7SETK}},
  note         = {Machine review of arXiv:2505.18671}
}
read the original abstract

We introduce an encoder-only approach to learn the evolution operators of large-scale non-linear dynamical systems, such as those describing complex natural phenomena. Evolution operators are particularly well-suited for analyzing systems that exhibit complex spatio-temporal patterns and have become a key analytical tool across various scientific communities. As terabyte-scale weather datasets and simulation tools capable of running millions of molecular dynamics steps per day are becoming commodities, our approach provides an effective tool to make sense of them from a data-driven perspective. The core of it lies in a remarkable connection between self-supervised representation learning methods and the recently established learning theory of evolution operators. To show the usefulness of the proposed method, we test it across multiple scientific domains: explaining the folding dynamics of small proteins, the binding process of drug-like molecules in host sites, and autonomously finding patterns in climate data. Code and data to reproduce the experiments are made available open source.

Figures

Figures reproduced from arXiv: 2505.18671 by the authors.

Figure 1
Figure 1. 2nd leading eigenfunction. possible, we didn’t concatenate additional projection heads to the encoder φ, as suggested in [63, 64, 65, 66]. Furthermore, because of the identity (9) we kept P linear, despite tiny MLPs being usually employed for the predictors. While experimenting with these architectural variants is out of the scope of this work, we believe it is an exciting direction for future work. To get the least… view at source ↗
Figure 2
Figure 2. Slow dynamical modes of biomolecular processes. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. ENSO mode from an encoder trained with Alg. 1. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Leading eigenfunctions computed by our and baseline approaches. Each row corresponds [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 6
Figure 6. Figure 6: MSE of approximating Ψ1 by LASSO regression on meaningful physical de￾scriptors. For Trp-cage we constructed a library of hydrogen-bond contact functions. The selected descriptors are reported in [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Normalized hydrogen-bond coefficients selected by the LASSO model (left) and represen [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Analysis of the leading eigenfunctions in the space of the host-guest distance [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: The three different molecules studied in the ligand-binding experiment. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Outcome-Aware Spectral Feature Learning for Instrumental Variable Regression

    stat.ML 2025-11 conditional novelty 6.0 of 10

    Augmented Spectral Feature Learning biases the learned treatment–instrument features toward the outcome, recovering NPIV structural functions that spectrum-only SpecIV features miss.

Reference graph

Works this paper leans on

108 extracted references · 60 canonical work pages · cited by 1 Pith paper

  1. [54]

    Neural conditional probability for uncertainty quantification.Advances in Neural Information Processing Systems, 37:60999–61039, 2024

    Vladimir Kostic, Grégoire Pacreau, Giacomo Turri, Pietro Novelli, Karim Lounici, and Massimiliano Pontil. Neural conditional probability for uncertainty quantification.Advances in Neural Information Processing Systems, 37:60999–61039, 2024

    work page 2024

  2. [1]

    Hans Hersbach, Bill Bell, Paul Berrisford, Shoji Hirahara, András Horányi, Joaquín Muñoz- Sabater, Julien Nicolas, Carole Peubey, Raluca Radu, Dinand Schepers, Adrian Simmons, Cornel Soci, Saleh Abdalla, Xavier Abellan, Gianpaolo Balsamo, Peter Bechtold, Gionata Biavati, Jean Bidlot, Massimo Bonavita, Giovanna De Chiara, Per Dahlgren, Dick Dee, Michail Di...

    1999

  3. [2]

    Lawrence Zitnick, and Zachary Ulissi

    Lowik Chanussot*, Abhishek Das*, Siddharth Goyal*, Thibaut Lavril*, Muhammed Shuaibi*, Morgane Riviere, Kevin Tran, Javier Heras-Domingo, Caleb Ho, Weihua Hu, Aini Palizhati, Anuroop Sriram, Brandon Wood, Junwoong Yoon, Devi Parikh, C. Lawrence Zitnick, and Zachary Ulissi. Open catalyst 2020 (oc20) dataset and community challenges.ACS Catalysis, 2021

    2020

  4. [3]

    Acemd: accelerating biomolecular dynamics in the microsecond time scale.Journal of Chemical Theory and Computation, 5(6):1632–1639, 2009

    Matt J Harvey, Giovanni Giupponi, and G De Fabritiis. Acemd: accelerating biomolecular dynamics in the microsecond time scale.Journal of Chemical Theory and Computation, 5(6):1632–1639, 2009

    2009

  5. [4]

    Smith, Berk Hess, and Erik Lindahl

    Mark James Abraham, Teemu Murtola, Roland Schulz, Szilárd Páll, Jeremy C. Smith, Berk Hess, and Erik Lindahl. Gromacs: High performance molecular simulations through multi- level parallelism from laptops to supercomputers.SoftwareX, 1–2:19–25, September 2015

    2015

  6. [5]

    Chodera, Robert T

    Peter Eastman, Jason Swails, John D. Chodera, Robert T. McGibbon, Yutong Zhao, Kyle A. Beauchamp, Lee-Ping Wang, Andrew C. Simmonett, Matthew P. Harrigan, Chaya D. Stern, Rafal P. Wiewiora, Bernard R. Brooks, and Vijay S. Pande. Openmm 7: Rapid development of high performance algorithms for molecular dynamics.PLOS Computational Biology, 13(7):e1005659, 2017

    2017

  7. [6]

    The quiet revolution of numerical weather prediction.Nature, 525(7567):47–55, 2015

    Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction.Nature, 525(7567):47–55, 2015

    2015

  8. [7]

    Cambridge University Press, 2014

    Shai Shalev-Shwartz and Shai Ben-David.Understanding Machine Learning: From Theory to Algorithms. Cambridge University Press, 2014

    2014

Show all 108 references
  1. [8]

    Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural operators.arXiv preprint arXiv:2202.11214, 2022

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopad- hyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high-resolution weather model using adaptive fourier neural op...

    2022 arXiv

  2. [9]

    Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast.arXiv preprint arXiv:2211.02556, 2022

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast.arXiv preprint arXiv:2211.02556, 2022

    2022 arXiv

  3. [10]

    Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    Remi Lam, Alvaro Sanchez-Gonzalez, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Suman Ravuri, Timo Ewalds, Zach Eaton-Rosen, Weihua Hu, et al. Learning skillful medium-range global weather forecasting.Science, 382(6677):1416–1421, 2023

    2023

  4. [11]

    Neural general circulation models for weather and climate.Nature, 632(8027):1060–1066, 2024

    Dmitrii Kochkov, Janni Yuval, Ian Langmore, Peter Norgaard, Jamie Smith, Griffin Mooers, Milan Klöwer, James Lottes, Stephan Rasp, Peter Düben, et al. Neural general circulation models for weather and climate.Nature, 632(8027):1060–1066, 2024

    2024

  5. [12]

    MIT Press, 1998

    Richard S Sutton and Andrew G Barto.Reinforcement learning: An introduction, volume 1. MIT Press, 1998

    1998

  6. [13]

    Mackey.Chaos, Fractals, and Noise, volume 97 ofApplied Mathematical Sciences

    Andrzej Lasota and Michael C. Mackey.Chaos, Fractals, and Noise, volume 97 ofApplied Mathematical Sciences. Springer New York, 1994

    1994

  7. [14]

    Cambridge University Press, April 2009

    David Applebaum.Lévy Processes and Stochastic Calculus. Cambridge University Press, April 2009

    2009

  8. [15]

    Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

    2023

  9. [16]

    Academic Press, 1972

    Michael Reed and Barry Simon.Methods of Modern Mathematical Physics. Academic Press, 1972

    1972

  10. [17]

    Molgedey and H

    L. Molgedey and H. G. Schuster. Separation of a mixture of independent signals using time delayed correlations.Physical Review Letters, 72(23):3634–3637, June 1994

    1994

  11. [18]

    Springer, 2001

    Ch Schütte, Wilhelm Huisinga, and Peter Deuflhard.Transfer operator approach to conforma- tional dynamics in biomolecular systems. Springer, 2001

    2001

  12. [19]

    Dynamic mode decomposition of numerical and experimental data.Journal of Fluid Mechanics, 656:5–28, 2010

    Peter J Schmid. Dynamic mode decomposition of numerical and experimental data.Journal of Fluid Mechanics, 656:5–28, 2010

    2010

  13. [20]

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

    1931

  14. [21]

    Identification of slow molecular order parameters for markov model construction.The Journal of Chemical Physics, 139(1), 2013

    Guillermo Pérez-Hernández, Fabian Paul, Toni Giorgino, Gianni De Fabritiis, and Frank Noé. Identification of slow molecular order parameters for markov model construction.The Journal of Chemical Physics, 139(1), 2013

    2013

  15. [22]

    Deep learning the slow modes for rare events sampling.Proceedings of the National Academy of Sciences, 118(44):e2113533118, 2021

    Luigi Bonati, GiovanniMaria Piccini, and Michele Parrinello. Deep learning the slow modes for rare events sampling.Proceedings of the National Academy of Sciences, 118(44):e2113533118, 2021

    2021

  16. [23]

    Peter J. Schmid. Dynamic mode decomposition of numerical and experimental data.Journal of Fluid Mechanics, 656:5–28, July 2010

    2010

  17. [24]

    Nathan Kutz, Steven L

    J. Nathan Kutz, Steven L. Brunton, Bingni W. Brunton, and Joshua L. Proctor.Dynamic Mode Decomposition. Society for Industrial and Applied Mathematics, 2016

    2016

  18. [25]

    On convergence of extended dynamic mode decomposition to the Koopman operator.Journal of Nonlinear Science, 28:687–710, 2018

    Milan Korda and Igor Mezi´c. On convergence of extended dynamic mode decomposition to the Koopman operator.Journal of Nonlinear Science, 28:687–710, 2018

    2018

  19. [26]

    Learning dynamical systems via Koopman operator regression in reproducing kernel Hilbert spaces.Advances in Neural Information Processing Systems, 35:4017–4031, 2022

    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.Advances in Neural Information Processing Systems, 35:4017–4031, 2022

    2022

  20. [27]

    Sharp spectral rates for Koopman operator learning.Advances in Neural Information Processing Systems, 36:32328–32339, 2023

    Vladimir Kostic, Karim Lounici, Pietro Novelli, and Massimiliano Pontil. Sharp spectral rates for Koopman operator learning.Advances in Neural Information Processing Systems, 36:32328–32339, 2023. 11

    2023

  21. [28]

    Finite-data error bounds for Koopman-based prediction and control.Journal of Nonlinear Science, 33(1):14, 2023

    Feliks Nüske, Sebastian Peitz, Friedrich Philipp, Manuel Schaller, and Karl Worthmann. Finite-data error bounds for Koopman-based prediction and control.Journal of Nonlinear Science, 33(1):14, 2023

    2023

  22. [29]

    A kernel-based approach to data-driven Koopman spectral analysis.arXiv preprint arXiv:1411.2260, 2014

    Matthew O Williams, Clarence W Rowley, and Ioannis G Kevrekidis. A kernel-based approach to data-driven Koopman spectral analysis.arXiv preprint arXiv:1411.2260, 2014

    2014 arXiv

  23. [30]

    Dynamic Mode Decomposition with Reproducing Kernels for Koopman Spectral Analysis

    Yoshinobu Kawahara. Dynamic Mode Decomposition with Reproducing Kernels for Koopman Spectral Analysis. InAdvances in Neural Information Processing Systems, volume 29, 2016

    2016

  24. [31]

    Eigendecompositions of transfer operators in reproducing kernel Hilbert spaces.Journal of Nonlinear Science, 30(1):283–315, 2019

    Stefan Klus, Ingmar Schuster, and Krikamol Muandet. Eigendecompositions of transfer operators in reproducing kernel Hilbert spaces.Journal of Nonlinear Science, 30(1):283–315, 2019

    2019

  25. [32]

    Koopman spectra in reproducing kernel Hilbert spaces.Applied and Computational Harmonic Analysis, 49(2):573–607, 2020

    Suddhasattwa Das and Dimitrios Giannakis. Koopman spectra in reproducing kernel Hilbert spaces.Applied and Computational Harmonic Analysis, 49(2):573–607, 2020

    2020

  26. [33]

    Operator-theoretic framework for forecasting nonlinear time series with kernel analog techniques.Physica D: Nonlinear Phenomena, 409:132520, 2020

    Romeo Alexander and Dimitrios Giannakis. Operator-theoretic framework for forecasting nonlinear time series with kernel analog techniques.Physica D: Nonlinear Phenomena, 409:132520, 2020

    2020

  27. [34]

    Estimating koopman operators with sketching to provably learn large scale dynamical systems

    Giacomo Meanti, Antoine Chatalic, Vladimir Kostic, Pietro Novelli, Massimiliano Pontil, and Lorenzo Rosasco. Estimating koopman operators with sketching to provably learn large scale dynamical systems. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, edi...

    2023

  28. [35]

    Deep learning for universal linear embeddings of nonlinear dynamics.Nature Communications, 9(1):4950, 2018

    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

  29. [36]

    Forecasting sequential data using consistent koopman autoencoders

    Omri Azencot, N Benjamin Erichson, Vanessa Lin, and Michael Mahoney. Forecasting sequential data using consistent koopman autoencoders. InInternational Conference on Machine Learning, pages 475–485. PMLR, 2020

    2020

  30. [37]

    Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics.The Journal of Chemical Physics, 148(24), 2018

    Christoph Wehmeyer and Frank Noé. Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics.The Journal of Chemical Physics, 148(24), 2018

    2018

  31. [38]

    Neural Koopman prior for data assimilation.IEEE Transactions on Signal Processing, 2024

    Anthony Frion, Lucas Drumetz, Mauro Dalla Mura, Guillaume Tochon, and Abdeldjalil Aissa El Bey. Neural Koopman prior for data assimilation.IEEE Transactions on Signal Processing, 2024

    2024

  32. [39]

    Vampnets for deep learning of molecular kinetics.Nature communications, 9(1):5, 2018

    Andreas Mardt, Luca Pasquali, Hao Wu, and Frank Noé. Vampnets for deep learning of molecular kinetics.Nature communications, 9(1):5, 2018

    2018

  33. [40]

    Learning invariant representations of time-homogeneous stochastic dynamical systems.arXiv preprint arXiv:2307.09912, 2023

    Vladimir R Kostic, Pietro Novelli, Riccardo Grazzi, Karim Lounici, and Massimiliano Pontil. Learning invariant representations of time-homogeneous stochastic dynamical systems.arXiv preprint arXiv:2307.09912, 2023

    2023 arXiv

  34. [41]

    Latent representation and simulation of markov processes via time-lagged information bottleneck.arXiv preprint arXiv:2309.07200, 2023

    Marco Federici, Patrick Forré, Ryota Tomioka, and Bastiaan S Veeling. Latent representation and simulation of markov processes via time-lagged information bottleneck.arXiv preprint arXiv:2309.07200, 2023

    2023 arXiv

  35. [42]

    Task-oriented Koopman-based control with contrastive encoder

    Xubo Lyu, Hanyang Hu, Seth Siriya, Ye Pu, and Mo Chen. Task-oriented Koopman-based control with contrastive encoder. In7th Annual Conference on Robot Learning, 2023

    2023

  36. [43]

    Data-efficient reinforcement learning with self-predictive representations.arXiv preprint arXiv:2007.05929, 2020

    Max Schwarzer, Ankesh Anand, Rishab Goel, R Devon Hjelm, Aaron Courville, and Philip Bachman. Data-efficient reinforcement learning with self-predictive representations.arXiv preprint arXiv:2007.05929, 2020

    2007 arXiv

  37. [44]

    Learning mesh-based simulation with graph networks

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learning mesh-based simulation with graph networks. InInternational Conference on Learning Repre- sentations, 2020. 12

    2020

  38. [45]

    Learning to simulate complex physics with graph networks

    Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. InInternational Conference on Machine Learning, pages 8459–8468. PMLR, 2020

    2020

  39. [46]

    Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

    2010 arXiv

  40. [47]

    Koopman-assisted reinforcement learning.arXiv preprint arXiv:2403.02290, 2024

    Preston Rozwood, Edward Mehrez, Ludger Paehler, Wen Sun, and Steven L Brunton. Koopman-assisted reinforcement learning.arXiv preprint arXiv:2403.02290, 2024

    2024 arXiv

  41. [48]

    Operator world models for reinforcement learning.arXiv preprint arXiv:2406.19861, 2024

    Pietro Novelli, Marco Pratticò, Massimiliano Pontil, and Carlo Ciliberto. Operator world models for reinforcement learning.arXiv preprint arXiv:2406.19861, 2024

    2024 arXiv

  42. [49]

    Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  43. [50]

    On mutual information maximization for representation learning

    M Tschannen, J Djolonga, P Rubenstein, S Gelly, and M Lucic. On mutual information maximization for representation learning. InEighth International Conference on Learning Representations. OpenReview. net, 2020

    2020

  44. [51]

    Noise-contrastive estimation: A new estimation principle for unnormalized statistical models

    Michael Gutmann and Aapo Hyvärinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. InProceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pages 297–304. JMLR Workshop and Conference Pro...

    2010

  45. [52]

    On vari- ational bounds of mutual information

    Ben Poole, Sherjil Ozair, Aaron Van Den Oord, Alex Alemi, and George Tucker. On vari- ational bounds of mutual information. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors,Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of...

    2019

  46. [53]

    Spectral representation learning for conditional moment models.arXiv preprint arXiv:2210.16525, 2022

    Ziyu Wang, Yucen Luo, Yueru Li, Jun Zhu, and Bernhard Schölkopf. Spectral representation learning for conditional moment models.arXiv preprint arXiv:2210.16525, 2022

    2022 arXiv

  47. [55]

    A class of statistics with asymptotically normal distribution.Breakthroughs in Statistics: Foundations and Basic Theory, pages 308–334, 1992

    Wassily Hoeffding. A class of statistics with asymptotically normal distribution.Breakthroughs in Statistics: Foundations and Basic Theory, pages 308–334, 1992

    1992

  48. [56]

    Provable guarantees for self- supervised deep learning with spectral contrastive loss.Advances in neural information processing systems, 34:5000–5011, 2021

    Jeff Z HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma. Provable guarantees for self- supervised deep learning with spectral contrastive loss.Advances in neural information processing systems, 34:5000–5011, 2021

    2021

  49. [57]

    Jeff Z HaoChen, Colin Wei, Ananya Kumar, and Tengyu Ma. Beyond separability: Analyzing the linear transferability of contrastive representations to related subpopulations.Advances in Neural Information Processing Systems, 35:26889–26902, 2022

    2022

  50. [58]

    Spectral decomposition representation for reinforcement learning.arXiv preprint arXiv:2208.09515, 2022

    Tongzheng Ren, Tianjun Zhang, Lisa Lee, Joseph E Gonzalez, Dale Schuurmans, and Bo Dai. Spectral decomposition representation for reinforcement learning.arXiv preprint arXiv:2208.09515, 2022

    2022 arXiv

  51. [59]

    Spectral repre- sentation for causal estimation with hidden confounders.arXiv preprint arXiv:2407.10448, 2024

    Haotian Sun, Antoine Moulin, Tongzheng Ren, Arthur Gretton, and Bo Dai. Spectral repre- sentation for causal estimation with hidden confounders.arXiv preprint arXiv:2407.10448, 2024

    2024 arXiv

  52. [60]

    Yiwei Lu, Guojun Zhang, Sun Sun, Hongyu Guo, and Yaoliang Yu.f-micl: Understanding and generalizing infonce-based contrastive learning.arXiv preprint arXiv:2402.10150, 2024

    2024 arXiv

  53. [61]

    Variational approach for learning markov processes from time series data.Journal of Nonlinear Science, 30(1):23–66, 2020

    Hao Wu and Frank Noé. Variational approach for learning markov processes from time series data.Journal of Nonlinear Science, 30(1):23–66, 2020. 13

    2020

  54. [62]

    G. H. Golub and V . Pereyra. The differentiation of pseudo-inverses and nonlinear least squares problems whose variables separate.SIAM Journal on Numerical Analysis, 10(2):413–432, April 1973

    1973

  55. [63]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational Conference on Machine Learning, pages 1597–1607. PmLR, 2020

    2020

  56. [64]

    Bootstrap your own latent-a new approach to self-supervised learning.Advances in Neural Information Processing Systems, 33:21271–21284, 2020

    Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning.Advances in Neural In...

    2020

  57. [65]

    Barlow twins: Self- supervised learning via redundancy reduction

    Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self- supervised learning via redundancy reduction. InInternational Conference on Machine Learning, pages 12310–12320. PMLR, 2021

    2021

  58. [66]

    Exploring simple siamese representation learning

    Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15750–15758, 2021

    2021

  59. [67]

    Simplicial embeddings in self-supervised learning and downstream classification.arXiv preprint arXiv:2204.00616, 2022

    Samuel Lavoie, Christos Tsirigotis, Max Schwarzer, Ankit Vani, Michael Noukhovitch, Kenji Kawaguchi, and Aaron Courville. Simplicial embeddings in self-supervised learning and downstream classification.arXiv preprint arXiv:2204.00616, 2022

    2022 arXiv

  60. [68]

    Deterministic nonperiodic flow.Journal of Atmospheric Sciences, 20(2):130– 148, 1963

    Edward Lorenz. Deterministic nonperiodic flow.Journal of Atmospheric Sciences, 20(2):130– 148, 1963

    1963

  61. [69]

    How fast-folding proteins fold.Science, 334(6055):517–520, 2011

    Kresten Lindorff-Larsen, Stefano Piana, Ron O Dror, and David E Shaw. How fast-folding proteins fold.Science, 334(6055):517–520, 2011

    2011

  62. [70]

    Hythem Sidky, Wei Chen, and Andrew L Ferguson. High-resolution markov state models for the dynamics of trp-cage miniprotein constructed over slow folding modes identified by state-free reversible vampnets.The Journal of Physical Chemistry B, 123(38):7999–8009, 2019

    2019

  63. [71]

    Graphvampnet, using graph neural networks and variational approach to markov processes for dynamical modeling of biomolecules.The Journal of Chemical Physics, 156(18), 2022

    Mahdi Ghorbani, Samarjeet Prasad, Jeffery B Klauda, and Bernard R Brooks. Graphvampnet, using graph neural networks and variational approach to markov processes for dynamical modeling of biomolecules.The Journal of Chemical Physics, 156(18), 2022

    2022

  64. [72]

    Schnet: A continuous-filter convolutional neural network for modeling quantum interactions.Advances in Neural Information Processing Systems, 30, 2017

    Kristof Schütt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert Müller. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions.Advances in Neural Information Processing Systems, 30, 2017

    2017

  65. [73]

    Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

    Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems.Proceedings of the national academy of sciences, 113(15):3932–3937, 2016

    2016

  66. [74]

    Descriptor-free collective variables from geometric graph neural networks.Journal of Chemical Theory and Computation, 20(24):10787–10797, 2024

    Jintu Zhang, Luigi Bonati, Enrico Trizio, Odin Zhang, Yu Kang, TingJun Hou, and Michele Parrinello. Descriptor-free collective variables from geometric graph neural networks.Journal of Chemical Theory and Computation, 20(24):10787–10797, 2024

    2024

  67. [75]

    Characterizing metastable states with the help of machine learning.Journal of Chemical Theory and Compu- tation, 18(9):5195–5202, 2022

    Pietro Novelli, Luigi Bonati, Massimiliano Pontil, and Michele Parrinello. Characterizing metastable states with the help of machine learning.Journal of Chemical Theory and Compu- tation, 18(9):5195–5202, 2022

    2022

  68. [76]

    A variational approach to modeling slow processes in stochastic dynamical systems.Multiscale Modeling & Simulation, 11(2):635–655, 2013

    Frank Noé and Feliks Nuske. A variational approach to modeling slow processes in stochastic dynamical systems.Multiscale Modeling & Simulation, 11(2):635–655, 2013

    2013

  69. [77]

    Overview of the sampl5 host–guest challenge: Are we doing better?Journal of Computer-Aided Molecular Design, 31:1–19, 2017

    Jian Yin, Niel M Henriksen, David R Slochower, Michael R Shirts, Michael W Chiu, David L Mobley, and Michael K Gilson. Overview of the sampl5 host–guest challenge: Are we doing better?Journal of Computer-Aided Molecular Design, 31:1–19, 2017. 14

    2017

  70. [78]

    The role of water in host-guest interaction.Nature Communications, 12(1):93, 2021

    Valerio Rizzi, Luigi Bonati, Narjes Ansari, and Michele Parrinello. The role of water in host-guest interaction.Nature Communications, 12(1):93, 2021

    2021

  71. [79]

    Cambridge University Press, 2000

    Henry F Diaz and Vera Markgraf.El Niño and the Southern Oscillation: multiscale variability and global and regional impacts. Cambridge University Press, 2000

    2000

  72. [80]

    Reviewing the oceanic niño index (oni) to enhance societal readiness for el niño’s impacts.International Journal of Disaster Risk Science, 11:394–403, 2020

    Michael H Glantz and Ivan J Ramirez. Reviewing the oceanic niño index (oni) to enhance societal readiness for el niño’s impacts.International Journal of Disaster Risk Science, 11:394–403, 2020

    2020

  73. [81]

    Persistent effect of el niño on global economic growth.Science, 380(6649):1064–1069, 2023

    Christopher W Callahan and Justin S Mankin. Persistent effect of el niño on global economic growth.Science, 380(6649):1064–1069, 2023

    2023

  74. [82]

    Atmospheric teleconnections from the equatorial pacific.Monthly Weather Review, 97(3):163–172, 1969

    Jakob Bjerknes. Atmospheric teleconnections from the equatorial pacific.Monthly Weather Review, 97(3):163–172, 1969

    1969

  75. [83]

    El nino southern oscillation phenomena.Nature, 302(5906):295–301, 1983

    S George H Philander. El nino southern oscillation phenomena.Nature, 302(5906):295–301, 1983

    1983

  76. [84]

    Enso as an integrating concept in earth science.Science, 314(5806):1740–1745, 2006

    Michael J McPhaden, Stephen E Zebiak, and Michael H Glantz. Enso as an integrating concept in earth science.Science, 314(5806):1740–1745, 2006

    2006

  77. [85]

    Changing el niño–southern oscillation in a warming climate.Nature Reviews Earth & Environment, 2(9):628–644, 2021

    Wenju Cai, Agus Santoso, Matthew Collins, Boris Dewitte, Christina Karamperidou, Jong- Seong Kug, Matthieu Lengaigne, Michael J McPhaden, Malte F Stuecker, Andréa S Taschetto, et al. Changing el niño–southern oscillation in a warming climate.Nature Reviews Earth & Environment,...

    2021

  78. [86]

    El niño and southern oscillation (enso): a review

    Chunzai Wang, Clara Deser, Jin-Yi Yu, Pedro DiNezio, and Amy Clement. El niño and southern oscillation (enso): a review. InCoral Reefs of the Eastern Tropical Pacific: Persistence and Loss in a Dynamic Environment, pages 85–106. Springer, 2017

    2017

  79. [87]

    El niño–southern oscillation complexity.Nature, 559(7715):535–545, 2018

    Axel Timmermann, Soon-Il An, Jong-Seong Kug, Fei-Fei Jin, Wenju Cai, Antonietta Capo- tondi, Kim M Cobb, Matthieu Lengaigne, Michael J McPhaden, Malte F Stuecker, et al. El niño–southern oscillation complexity.Nature, 559(7715):535–545, 2018

    2018

  80. [88]

    Spectral analysis of climate dynamics with operator-theoretic approaches.Nature Communications, 12(1):6570, 2021

    Gary Froyland, Dimitrios Giannakis, Benjamin R Lintner, Maxwell Pike, and Joanna Slaw- inska. Spectral analysis of climate dynamics with operator-theoretic approaches.Nature Communications, 12(1):6570, 2021

    2021

  81. [89]

    Deep learning for multi-year enso forecasts.Nature, 573(7775):568–572, 2019

    Yoo-Geun Ham, Jeong-Hwan Kim, and Jing-Jia Luo. Deep learning for multi-year enso forecasts.Nature, 573(7775):568–572, 2019

    2019

  82. [90]

    Climate Prediction Center - ONI — ori- gin.cpc.ncep.noaa.gov

    NOAA’s climate prediction center NCP Center. Climate Prediction Center - ONI — ori- gin.cpc.ncep.noaa.gov. https://origin.cpc.ncep.noaa.gov/products/analysis_ monitoring/ensostuff/ONI_v5.php. [Accessed 09-05-2025]

    2025

  83. [91]

    The ecmwf operational ensemble reanalysis–analysis system for ocean and sea ice: a description of the system and assessment.Ocean Science, 15(3):779–808, 2019

    Hao Zuo, Magdalena Alonso Balmaseda, Steffen Tietsche, Kristian Mogensen, and Michael Mayer. The ecmwf operational ensemble reanalysis–analysis system for ocean and sea ice: a description of the system and assessment.Ocean Science, 15(3):779–808, 2019

    2019

  84. [92]

    Chaosbench: A multi-channel, physics-based benchmark for subseasonal- to-seasonal climate prediction.arXiv preprint arXiv:2402.00712, 2024

    Juan Nathaniel, Yongquan Qu, Tung Nguyen, Sungduk Yu, Julius Busecke, Aditya Grover, and Pierre Gentine. Chaosbench: A multi-channel, physics-based benchmark for subseasonal- to-seasonal climate prediction.arXiv preprint arXiv:2402.00712, 2024

    2024 arXiv

  85. [93]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  86. [94]

    A method for unsupervised learning of coherent spatiotemporal patterns in multiscale data.Proceedings of the National Academy of Sciences, 122(7):e2415786122, 2025

    Karl Lapo, Sara M Ichinaga, and J Nathan Kutz. A method for unsupervised learning of coherent spatiotemporal patterns in multiscale data.Proceedings of the National Academy of Sciences, 122(7):e2415786122, 2025

    2025

  87. [95]

    Old and new matrix algebra useful for statistics.See www

    Thomas P Minka. Old and new matrix algebra useful for statistics.See www. stat. cmu. edu/minka/papers/matrix. html, 4, 2000. 15

    2000

  88. [96]

    [Accessed 19-05-2025]

    Kooplearn.https://kooplearn.readthedocs.io/. [Accessed 19-05-2025]

    2025

  89. [97]

    The lorenz attractor exists.Comptes Rendus de l’Académie des Sciences- Series I-Mathematics, 328(12):1197–1202, 1999

    Warwick Tucker. The lorenz attractor exists.Comptes Rendus de l’Académie des Sciences- Series I-Mathematics, 328(12):1197–1202, 1999

    1999

  90. [98]

    Large sample analysis of the median heuristic.arXiv preprint arXiv:1707.07269, 2017

    Damien Garreau, Wittawat Jitkrittum, and Motonobu Kanagawa. Large sample analysis of the median heuristic.arXiv preprint arXiv:1707.07269, 2017

    2017 arXiv

  91. [99]

    A unified framework for machine learning collective variables for enhanced sampling simulations: mlcolvar.The Journal of Chemical Physics, 159(1), 2023

    Luigi Bonati, Enrico Trizio, Andrea Rizzi, and Michele Parrinello. A unified framework for machine learning collective variables for enhanced sampling simulations: mlcolvar.The Journal of Chemical Physics, 159(1), 2023

    2023

  92. [100]

    Blind prediction of cyclohexane–water distribution coefficients from the sampl5 challenge.Journal of Computer-Aided Molecular Design, 30:927–944, 2016

    Caitlin C Bannan, Kalistyn H Burley, Michael Chiu, Michael R Shirts, Michael K Gilson, and David L Mobley. Blind prediction of cyclohexane–water distribution coefficients from the sampl5 challenge.Journal of Computer-Aided Molecular Design, 30:927–944, 2016

    2016

  93. [101]

    Plumed 2: New feathers for an old bird.Computer Physics Communications, 185(2):604– 613, 2014

    Gareth A Tribello, Massimiliano Bonomi, Davide Branduardi, Carlo Camilloni, and Giovanni Bussi. Plumed 2: New feathers for an old bird.Computer Physics Communications, 185(2):604– 613, 2014

    2014

  94. [102]

    Development and testing of a general amber force field.Journal of Computational Chemistry, 25(9):1157–1174, 2004

    Junmei Wang, Romain M Wolf, James W Caldwell, Peter A Kollman, and David A Case. Development and testing of a general amber force field.Journal of Computational Chemistry, 25(9):1157–1174, 2004

    2004

  95. [103]

    A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the resp model.The Journal of Physical Chemistry, 97(40):10269–10280, 1993

    Christopher I Bayly, Piotr Cieplak, Wendy Cornell, and Peter A Kollman. A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the resp model.The Journal of Physical Chemistry, 97(40):10269–10280, 1993

    1993

  96. [104]

    Comparison of simple potential functions for simulating liquid water.The Journal of Chemical Physics, 79(2):926–935, 1983

    William L Jorgensen, Jayaraman Chandrasekhar, Jeffry D Madura, Roger W Impey, and Michael L Klein. Comparison of simple potential functions for simulating liquid water.The Journal of Chemical Physics, 79(2):926–935, 1983

    1983

  97. [105]

    Canonical sampling through velocity rescaling.The Journal of Chemical Physics, 126(1), 2007

    Giovanni Bussi, Davide Donadio, and Michele Parrinello. Canonical sampling through velocity rescaling.The Journal of Chemical Physics, 126(1), 2007

    2007

  98. [106]

    Funnel metadynamics as accurate binding free-energy method.Proceedings of the National Academy of Sciences, 110(16):6358–6363, 2013

    Vittorio Limongelli, Massimiliano Bonomi, and Michele Parrinello. Funnel metadynamics as accurate binding free-energy method.Proceedings of the National Academy of Sciences, 110(16):6358–6363, 2013

    2013

  99. [107]

    Soumendranath Bhakat and Pär Söderhjelm. Resolving the problem of trapped water in binding cavities: prediction of host–guest binding free energies in the sampl5 challenge by funnel metadynamics.Journal of Computer-Aided Molecular Design, 31:119–132, 2017

    2017

  100. [108]

    Spectral normaliza- tion for generative adversarial networks.arXiv preprint arXiv:1802.05957, 2018

    Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normaliza- tion for generative adversarial networks.arXiv preprint arXiv:1802.05957, 2018. 16 Supplementary Material A Theory: connection between operator learning and self-supervised learning The ma...

    2018 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.