Sharp Spectral Rates for Koopman Operator Learning

Non-linear dynamical systems can be handily described by the associated Koopman operator, whose action evolves every observable of the system forward in time. Learning the Koopman operator and its spectral decomposition from data is enabled by a number of algorithms. In this work we present for the first time non-asymptotic learning bounds for the Koopman eigenvalues and eigenfunctions. We focus on time-reversal-invariant stochastic dynamical systems, including the important example of Langevin dynamics. We analyze two popular estimators: Extended Dynamic Mode Decomposition (EDMD) and Reduced Rank Regression (RRR). Our results critically hinge on novel minimax estimation bounds for the operator norm error, that may be of independent interest. Our spectral learning bounds are driven by the simultaneous control of the operator norm error and a novel metric distortion functional of the estimated eigenfunctions. The bounds indicates that both EDMD and RRR have similar variance, but EDMD suffers from a larger bias which might be detrimental to its learning rate. Our results shed new light on the emergence of spurious eigenvalues, an issue which is well known empirically. Numerical experiments illustrate the implications of the bounds in practice.

Paper

References (49)

Scroll for more · 37 remaining

Similar papers

Peer review

Reviewer ZSn27/10 · confidence 3/52023-06-22

Summary

This paper develops bounds on how close the approximated Koopman modes and eigenvalues are to the true eigenvalues and modes, for two important classes of methods used to compute the Koopman mode decomposition. The authors find that one class, Principal Component Regression, which included extended dynamic mode decomposition (EDMD), can suffer more from poorly chosen kernels and can have larger bias than Reduced Rank Regression (RRR). They additionally provide an empirical method for determining spurious eigenvalues, which can be used for model selection.

Strengths

1. This paper is well written and easy to follow. 2. This paper provides new techniques for computing bounds on the approximated Koopman spectral objects, and the discovery that PCR can have larger bias than RRR, are important ones for the field. 3. This paper provides a new empirical method for identifying spurious eigenvalues and making model selection. Again, both of these are important topics for the field and for the application of numerical methods in applied settings. 4. The numerical examples provided in Figs. 1-3 are helpful for understanding the theory developed, and provide support for the developed claims.

Weaknesses

1. Klus et al., 2016 and Korda and Mezic, 2018, as examples, proved the convergence of EDMD to the true Koopman operator, when $M \rightarrow \infty$, where $M$ is the number of data points. This work seems to be missing from the paper, as does discussion surrounding how the paper differs from that work. I assume the primary difference is that this paper's results are not in the asymptotic limit (although, they are in the sense that the bounds for RRR, for example, in the Gaussian case converge as $1/\sqrt{n}$, which approaches $0$ as $n\rightarrow \infty$). Additionally, the results for PCR obtained by this paper would suggest that in the asymptotic limit EDMD does not converge, since it has a bias. Discussion on how this is reconciled with the work of Klus et al., 2016 and Korda and Mezic, 2018, is necessary. 2. Fig. 3 was confusing. Was the best estimator found on the test data set, and then the red line in Fig. 3 the result of applying it to the validation data? A secondary panel in that figure describing what was being done would be helpful. MINOR COMMENTS: 1. It was unclear to me how $| \lambda_i - \mu_{j(i)} | \leq || (A_\pi - \lambda_i I)^{-1}||^{-1}$ leads to observing that $||(A_\pi - S\hat{G})\hat{\psi}_i|| \leq \mathcal{E}(\hat{G})\eta(\hat{\psi})$ (lines 154-155). Adding a little more detail/comment on this would be helpful. 2. "left hand side" (line 156) should be "right hand side" no? 3. The connection between DMD and KMD should be discussed (lines 22-25) (Rowley et al., 2009). 4. The original EDMD paper (Williams et al., 2015) should be cited when discussing EDMD for the first time (line 31). 5. Very minor but both "non-linear" and "nonlinear" are written.

Questions

1. How does this work compare to previous work studying the convergence of EDMD (e.g., Klus et al., 2016; Korda and Mezic, 2018)? 2. What exactly is Fig. 2 showing (in terms of details)?

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

4 excellent

Contribution

3 good

Limitations

The authors did a good job being clear that their work was limited to self-adjoint operators.

Reviewer B3rh9/10 · confidence 3/52023-07-07

Summary

This paper studies the approximation and learning of the Koopman operator. Koopman operator is helpful in modeling a broad class of Markovian dynamical systems. In this paper, two types of approximation strategies are studied, namely, "Principal Component Regression (PCR)" and "Reduced Rank Regression (RRR)." Both types are based on general reproducing kernel Hilbert spaces. With the assumptions that the population covariance operator bounds the population cross covariance operator, the RKHS feature map being L-infinity, and the eigenvalues of the population covariance operator decay as O(i^(-1/beta)) with 0<beta<=1, the operator norm error, the eigenvalue estimation error, and the eigenfunction approximation error are bounded.

Strengths

The paper's originality is high, because of the new error estimation provided. This paper represents solid research results with high quality. The writing is clear and easy to follow. The paper's significance is guaranteed by the broad scope of applications of the Markovian dynamical systems, including Langevin dynamics.

Weaknesses

We do not have significant concerns about this paper. No weakness is identified.

Questions

1. Lines 116--117: What is the definition of rank-r operator? If the dimension of the image of one operator is finite, the operator is automatically Hilbert-Schmidt. If this is the definition, I guess using the notion of "rank-r Hilbert-Schmidt operator" is confusing because it implies the possibility of rank-r operators that are not Hilbert-Schmidt. 2. Which theorem does it refer to for the claim in Line 335? 3. In Line 68, the Koopman operator A_pi is assumed self-adjoint, and in Line 81, A_pi is further assumed compact. Is the second assumption made globally for the whole paper? In Theorem 3, A_pi is again assumed compact and self-adjoint, but this assumption does not appear in Theorem 4. This could not be very clear, leaving the reader wondering whether this assumption is adopted for Theorem 4.

Rating

9: Very Strong Accept: Technically flawless paper with groundbreaking impact on at least one area of AI/ML and excellent impact on multiple areas of AI/ML, with flawless evaluation, resources, and 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

4 excellent

Limitations

No limitation issues were found.

Reviewer nHa66/10 · confidence 2/52023-07-18

Summary

The authors provide bounds for the spectral decomposition of the estimated Koopman operators. The bounds are given for self-adjoint time-reversible operators in terms of two new metrics. Compared to estimation guarantees given in terms of the Hilbert Space distance norm, the proposed bounds require less restrictive assumptions. The theoretical results are specialized for two existing estimation algorithms.

Strengths

Estimating the eigenvalue of the Koopaman operator is a widespread problem. And the work seems to extend to general self-adjoint operators. The proposed new evaluation metrics are interesting, especially if they produce theoretical bounds that hold under less restrictive assumptions.

Weaknesses

It is unclear how the bounds depend on the sample size and why it is interesting to specialize them for existing estimators. The experiments show the performance of two existing estimators but have no straightforward link with the theoretical part of the paper. It would be more interesting to plot, for a given estimator, the wideness of the proposed bounds versus the (a posteriori) empirical estimation error. After assuming that the process is time-homogeneous and stationary, the learning task looks similar to standard non-parametric regression. The authors should specify what are the challenging aspects of the dynamical setup. Otherwise, the contribution of the paper is unclear. The core part of this work seems to be applying classical spectral bounds to a finite-dimensional approximation of HS operators. If the novelty is to use new metrics, the paper should focus more on explaining why these new metrics are better than the HS norm. The paper's conclusion is somehow expected. Direct learning of low-rank representations is better than projecting the data and solving an unconstrained problem. The latter option may have computational advantages. But the authors do not comment on it.

Questions

- I do not fully understand this sentence, "The Koopman operator [...], and DMD relies [...] on to, in turn, estimate its spectral decomposition." -The Koppman operator is a linear HS operator. What is peculiar about bounding the Koopman operator compared to other HS operators? - How common and restrictive are the assumptions that the operator is time-reversible and compact? - Why do you say that "the empirical PCR estimator does not minimize the empirical risk (3) under the low-rank constraint."? Is this because of the projection? Or because the optimization is unconstrained? - How can the learning method generate spurious eigenvalues? Is the standard operator norm insensitive to eigenvalue spuriousness? - Is the regularity condition called RC on page 5 used in the bounds? I am confused by the sentence on the following page, "where one chooses a universal kernel for which ...". - Intuitively, bounding the error on the estimated eigenvalue can be easier than on the estimated error functions or eigenvector. I understand the final formula may be too intricate for the main text. But it would be helpful for the reader to see an intuitive explanation of the claim, "the eigenfunction estimation bounds readily follow from (8)".

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

3 good

Limitations

The authors say that restricting the analysis to self-adjoint operators is the main limitation of their work. This is mentioned in the very last lines of the paper. It would be better to elaborate on this limitation in the introduction, where the self-adjoint assumption is made.

Reviewer GGn47/10 · confidence 3/52023-07-24

Summary

This paper analyses two approximation techniques for self-adjoint compact Koopman operators. Both estimators are constructed based on regularized least-squares with a certain rank condition in mind. The Principal Component regression (PCR) performs unconstrained Tikhonov estimation of data projected on a low-dimensional subspace. The Reduced Rank Regression (RRR) minimizes the Tikhonov objective function subject to a rank constraint on the optimizer. Both estimators admit closed form solutions. The paper study spectral properties of these estimators, specifically deviation of eigenvalues and eigenvectors. The novelty of this study resides on the error measures: operator norm error of the Koopman operator estimation (instead of Hilbert-Schmidt) and the "metric distortion" that compares the two Hilbert space, the RKHS that approximates the process, and the ambient L^2 Hilbert space where the Koopman operator is initially defined. The conclusion of this study is that both estimators have a similar variance, but the PCR may have a potentially larger bias, particularly for badly chosen kernels.

Strengths

The authors employ existing state-of-the-art bounds in spectral theory of compact operators. The study sheds light on the phenomenon of "spurious eigenvalues". The asymptotic rates of convergence are shown to be optimal. Overall it seems a solid paper.

Weaknesses

The rates of convergence and error bounds are tight, but only asymptotically. Since both PCR and RRR have similar asymptotic rates (for variance), the constants become important. A more careful analysis of the constants would strengthen the paper. However it is understandable that such a study might be analytically too complex. The authors mention that results are limited to self-adjoint Koopman operators. This is true, however, compact operators admit a SVD factorization with similar spectral properties (control of singular values) as self-adjoint operators. At a cursory reading, the results obtained in this study seem extendable to non-self-adjoint compact Koopman operators.

Questions

1. The paper is well-written. Just a few typos that can be easily fixed. I suppose in Example 3, the definition of the permutation Pi, the first case is i<= r, right? 2. I understand that condition (RC) is weaker than Im (A_pi S)\subset Im(S). Is it weaker also than Im (A_pi S)\subset closure(Im(S)), with closure w.r.t. L^2-norm ? 3. It would be useful to indicate the Hilbert space norms throughout the paper. Particularly in equation (10), but also elsewhere.

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

4 excellent

Contribution

3 good

Limitations

Nothing to be reported here.

Reviewer 1FxQ7/10 · confidence 4/52023-07-26

Summary

This paper proves sharp upper bounds for eigenvalue estimation of a Koopman operator for a time-homogeneous Markovian dynamical system using either reduced rank regression or principal component regression. The bounds include operator norm error and metric distortion. These results are illustrated on simple models and some discussion of a higher dimensional molecular example is included. The error estimate also yields a design principle for kernels, based on spectral bias.

Strengths

The paper is clearly written and articulates the both the theoretical results and their consequences on practical examples in a lucid manner. The results on metric distortion are novel for this problem. The paper synthesizes a number of existing arguments in a compelling way to generate a clear estimate on the spectral learning rate.

Weaknesses

My impression is that the argument in Sec. 5 is not particularly new in the case that the HS norm error is used. But it is not very clear why to perfer the operator norm error.

Questions

The estimator of metric distortion is introduced in the main text and some reference to its bearing on the experiments is made, but it is not clear from the experiments what role it plays. Can this be better explained? If I understand correctly, the "ugly" kernel is chosen to make large the bias. Is there a way of making the deformation of the metric structure large in this example to illustrate the contribution of that term in the error? In the appendix, it would be better to say in the alanine dipeptide example more explicitly how the RMSE is estimated. I assume the authors are forecasting the structure, but it could be some observable.

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

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Yes, the conclusion captures limitations well. Negative impacts are not very relevant to this paper.

Reviewer ZSn22023-08-11

Response to reviewers

Thank you for your detailed rebuttal. 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 GGn42023-08-11

I thank the authors for addressing my questions the comments. I keep my rating and recommendation.

Reviewer nHa62023-08-11

I agree with the other reviewers and raise my score to Accept.

Many thanks to the authors for the detailed rebuttal. Probably, I should have read the appendix more carefully. Thank you for explaining the difference between Koopman and HS operators. The authors' answers also address my concern about the connection between theory and experiments.

Reviewer 1FxQ2023-08-14

Thanks for the clarification concerning the operator norm. I will keep my rating as accept.

Reviewer B3rh2023-08-16

Including key theorem into main body

For 2, since it is claimed that "We established minimax optimal rates for the operator norm error in the Koopman regression problem" in line 335, I think it makes sense to include the relative theorem(s) to the main body. Of course, the proofs could be placed in the appendix.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC