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

The theory of Koopman operators allows to deploy non-parametric machine learning algorithms to predict and analyze complex dynamical systems. Estimators such as principal component regression (PCR) or reduced rank regression (RRR) in kernel spaces can be shown to provably learn Koopman operators from finite empirical observations of the system's time evolution. Scaling these approaches to very long trajectories is a challenge and requires introducing suitable approximations to make computations feasible. In this paper, we boost the efficiency of different kernel-based Koopman operator estimators using random projections (sketching). We derive, implement and test the new"sketched"estimators with extensive experiments on synthetic and large-scale molecular dynamics datasets. Further, we establish non asymptotic error bounds giving a sharp characterization of the trade-offs between statistical learning rates and computational efficiency. Our empirical and theoretical analysis shows that the proposed estimators provide a sound and efficient way to learn large scale dynamical systems. In particular our experiments indicate that the proposed estimators retain the same accuracy of PCR or RRR, while being much faster.

Paper

Similar papers

Peer review

Reviewer 8Gvs7/10 · confidence 3/52023-06-20

Summary

This paper utilizes theory developed around random projections to develop fast/efficient numerical methods for approximating the Koopman mode decomposition of data sets with long-time trajectories, that additionally come with learning rates that are optimal.

Strengths

1. This paper developed new fast/efficient algorithms for computing the Koopman spectra, the effectiveness of which was demonstrated on very large data sets with long-time trajectories. These include a traditional benchmark (Lorenz '63), as well as molecular data sets, which are a growing area of interest to the machine learning and Koopman communities. Additionally, the molecular data sets were of a size that could not have been studied previously. 2. This paper provided learning rates for their new algorithms, which they show to be optimal. 3. This paper discusses how different methods for computing the Koopman mode decomposition (e.g., Nystrom RRR vs. PCR) differ in their learning rate dependencies. 4. The paper was well-written.

Weaknesses

1. Motivation for why very long-time trajectories of data would be needed for some systems was missing from the Introduction. This was partially discussed at the end in Sec. 5 "Molecular dynamics datasets", but developing this more is important. Additionally, the title and some places in the text describe the sketching approach as being useful for "large scale dynamical systems". As I understand it, the methods are beneficial for "long-time trajectory dynamical systems". Of course, the molecular examples shown have both large dimension and long-time trajectories, but the title and text should be corrected to accurately emphasize that it is long-time trajectory dynamical systems that this method is useful for. 2. The discussion on how Nystrom PCR and RRR differ in their learning rates was very interesting. Including more on this (maybe in table form) would be nice. Additionally, it would be helpful to see all 3 methods compared, so KRR should be included in Figure 1. Additionally, the reason for using just RRR in Figs. 3 and 4 should be noted. 3. I found it a little confusing that going from Eq. 8 to Eq. 9, the P_Y \hat{C}_{YX} P_X became K^\dagger_{\tilde{Y}, \tilde{Y}} K_{\tilde{Y}, Y} K_{X, \tilde{X}}, but that going from Eq 10 to Eq 11 the P_Y \hat{C}_{YX} also led to K^\dagger_{\tilde{Y}, \tilde{Y}} K_{\tilde{Y}, Y} K_{X, \tilde{X}}. I assume this has to do with the way the \dagger was absorbed into the [ \cdot ]_r? 4. How the random $\tilde{\Phi}$ are generated should be explained in the text. Minor comments: 1. Mezic, 2005 should be cited when discussing the Koopman mode decomposition and Budisic et al. 2012 should be cited when discussing how Koopman mode decomposition has been leveraged in the past (line 95). 2. Abbreviations should be explained before using them (DMD, tICA, VAC, etc.). 3. The box plot denoting the sKAF data in Fig. 2 has a strange shape in its lower box (an inverted U). Why is this? 4. It would be helpful and appropriate to have citations when noting that "An important application of Koopman operator theory is in the analysis of molecular dynamics". 5. The colors of Fig. 3 should be explained in more detail. 6. It might be helpful to have a line across the 3 states in Fig. 4 to show how Eigenfunction 1 has linear separation between State 0 and the others, and Eigenfunction 2 has linear separation between State 1 and the others. 7. There are a few typos: i. "allows to deploy" (line 1) ii. "as theirs exact" (line 68) iii. "to study large class" (line 90) iv. "]0, 1]" (Assumption 4.3) v. "]0, \tau]" (Assumption 4.4)

Questions

To summarize, the Weakness section above, 1. Why would one want to use very long-time trajectory data when computing the Koopman mode decomposition? (You explain this some, but making it more explicit would be helpful) 2. What are the differences between KRR, PCR, and RRR in their performance on the different data sets studied (at least the Lorenz '63)? 3. How are the random $\tilde{\Phi}$ are generated?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

The authors do a sufficient job explaining their limitations (the assumptions they make) and what kinds of data sets their proposed methods work best on. The only exception to this is that it should be clarified that the proposed methods improve Koopman mode decomposition estimates of long-time trajectory data and not large dimension data.

Reviewer jbb37/10 · confidence 3/52023-07-05

Summary

This paper proposes sketching algorithms for approximating three different kernel-based estimators of Koopman operators from large scale temporal data. New non-asymptotic error bounds are derived.

Strengths

- The paper leverages the concept of sketching to speed up the estimation of Koopman operators without sacrificing accuracy. The newly proposed algorithms can potentially have a high impact in several scientific disciplines where the analysis of dynamical systems is of interest. - The authors explain and contextualize very well their contributions and the related work is comprehensive. - The paper is very well written and organized. - New theoretical results are rigorously derived to provide robust guarantees for the proposed methodology. - The empirical analysis is sound and well-conducted. To showcase various aspects of the method, two distinct applications are considered. The first experiment effectively demonstrates the tradeoff between time and accuracy. The second experiment is particularly interesting as it provides a practical example where it is necessary to consider full, non-subsampled long trajectories. Furthermore, it focuses on the interpretation of dynamics and scientific discovery. It also highlights the scalability with respect to the state space dimension, which serves as motivation for the study of Koopman operator methods.

Weaknesses

- While the idea of using sketching is not original, this contribution has the potential to significantly advance dynamical systems analysis by providing efficient learning algorithms for Koopman operators. Sketches for operators mapping between Hilbert spaces have recently been studied in [3]. However, the authors outline the unique aspects of their contribution and provide a clear motivation for deriving new results.

Questions

- What is the complexity with respect to the dimension of the state space? Is it linear? - Does the Markovian assumption in line 84 present a limitation for real-world applications? Is there a way to adapt Koopman operator learning (not necessarily sketched) to the non-Markovian case?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

Limitations are discussed.

Reviewer Hh1S6/10 · confidence 3/52023-07-06

Summary

This paper improves the computational efficiency of existing kernel method for Koopman operator approximation, including KRR, PCR, RRR based on the Nystrom estimator. In this way, the authors reduce the original computational complexity from $\mathcal{O}(n^3)$ to $\mathcal{O}(n\sqrt{n})$. They provide theoretical learning bounds and several numerical studies to demonstrate the efficacy of the proposed method.

Strengths

[1] The theoretical estimation of the learning bounds guarantee the performance of the proposed methods in practical scenarios. [2] The experiments validate the efficacy of the proposed methods can scale to large dynamical systems.

Weaknesses

[1] The idea of introducing random projections to the existing kernel methods is not novel enough, since the similar idea has been used in the deep learning community, such as generative models [R1] and neural ODEs [R2]. [2] The last two experiments seem to be interesting but the authors do not describe them clearly. [3] In line 62, the authors have typos about 'decomposition'. __References__ [R1] Song, Y., Garg, S., Shi, J., & Ermon, S. (2020, August). Sliced score matching: A scalable approach to density and score estimation. In Uncertainty in Artificial Intelligence (pp. 574-584). PMLR. [R2] Grathwohl, W., Chen, R. T., Bettencourt, J., Sutskever, I., & Duvenaud, D. (2018). Ffjord: Free-form continuous dynamics for scalable reversible generative models. arXiv preprint arXiv:1810.01367.

Questions

[1] In the experiment of Molecular dynamics datasets, do the $\phi,\psi$ correspond to the first two eigenfunctions with largest eigenvalues of the approximated Koopman matrix? And why not demonstrate the predictive performance of the proposed methods in the original 45-D state space in this task?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

The authors have not provided experiments to demonstrate the efficacy of the proposed methods with the non-i.i.d. data, but the current work is complete anyway.

Reviewer FXhE6/10 · confidence 3/52023-07-06

Summary

The paper talks about estimating Koopman operators. Sketching (random projections), specifically the Nyström method is used as a computational tool to make things tractable. Furthermore, it is shown that these approximates have the same convergence rate as their slower counter parts. Further error bounds are provided in terms of the operator norm.

Strengths

The authors do a good job in positing the current work within the literature which is very wide otherwise. The flow of the paper is very good and reads well. Assumptions in Section 4 are clearly mentioned before stating the main result. I appreciate authors commenting on the non iid case, apart from detailed analysis of the iid scenario.

Weaknesses

I view the papers contributions as two fold. The first part is the use of sketching for approximation of the three scenarios of KRR, PCR, RRR considered. In this case (as authors have acknowledged, references [3] and [6] therein), there are other works who have pointed this out. Again authors mention the special case of their sketched PCR as kernel DMD. These make the contributions of the current work from this point of view a bit restricted. Having said that, I greatly appreciate authors mentioning them in the paper! The second is in terms of learning bounds via operator norm, where I think the contribution is solid. However, it is hard to have empirical evidence to support this and is understandable. It would be good to have some discussion around this, as Koopman operators are heavily used in practical settings and having some evidence would greatly improve the impact of the paper over it's current form.

Questions

1) What is $A^\ast$ is line 104? The equation 5 just above has only $A$, are they the same? 2) In experimental evaluation, only PCR and RRR are considered, what about KRR ?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

I don't see a direct potential negative social impact.

Reviewer 8Gvs2023-08-12

Response to authors

Thank you for your detailed rebuttal. Thank you for clarifying the confusion about large scale vs. long time. I see how this is an unavoidable issue, but I think it was explained clearly. These responses, as well as the reviews from the other reviewers (and the authors' responses to those reviewers) make me confident that this is a strong paper with good contributions. The changes/clarifications the authors propose to make in the revised version of the manuscript will further increase its quality. I will therefore increase my score.

Reviewer FXhE2023-08-18

Thank you for the rebuttal. Please clarify the adjoint notation, it will be helpful. I appreciate the experiments added and am increasing my score.

Reviewer Hh1S2023-08-19

Review to the rebuttal

Thanks for the detailed rebuttal, after reading the general response as well as the other reviews, I am happy to amend my score of the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC