acceptodds
Under review as a conference paper at ICLR 2027

LCKoop: Koopman-Inspired Smooth Autoregressive Neural Operators in PDE Forecasting

Abstract

Modern autoregressive neural operators evolve high-dimensional states through expressive nonlinear transition maps. However, their local first-order behavior varies rapidly along generated trajectories, leading to unstable error propagation. In contrast, Koopman-based models and dynamic mode decomposition (DMD) impose a globally or locally static linear transition, providing more stable representations of state evolution, but are inherently limited in their ability to capture high-dimensional strongly nonlinear dynamics. In this work, we introduce **LCKoop**: **L**ow-**C**urvature-is-**Koop**man, an easy-to-use, *plug-in* regularizer that bridges these two frameworks. LCKoop instantiates the model-implied first-order affine approximation and encourages it to remain consistent along generated trajectories, without modifying the forecasting architecture or inference path. We show that this regularization controls a trajectory-conditioned higher-order remainder in autoregressive error propagation, complementing conventional first-order sensitivity analysis. Across four PDE forecasting tasks and three neural-operator families, LCKoop reduces MSE by an average of 8.9% at the short forecasting horizon and 14.6% at the extended horizon, with maximum reductions of 53.7% and 42.4%, respectively.

Primary area:
learning on time series and dynamical systems
Keywords:
neural operators, PDE forecasting, autoregressive rollout, Koopman operators, dynamic mode decomposition, curvature regularization, scientific machine learning, long-horizon forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_q96uOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes to apply the "Low Curvature" Jacobian variation loss term (taken from prior work) to autoregressive PDE forecasting. Prior Jacobian regularization approaches are argued to be too conservative; on the other hand, some form of stabilisation is known to work well for PDEs.

The text markets the method as a Koopman-based method, since the low curvature penalty pulls the latent dynamics towards linear, which essentially makes it locally/approximately Koopman (with slowly changing nonlinearities, i.e. curvature).

The paper presents theoretic error bounds for their approach, and empirically compares this regularization to a Dynamic Mode Decomposition-based regularization method, alongside experiments showing gains in the LCKoop regularization vs no-regularization scenario, on several popular PDE datasets, across several architectures.

Critical strengths

Your idea is novel, and is an interesting direction to pursue in the landscape of PDE rollout forecasting methods with regularization. As regularization often helps, and traditional regularization may be viewed as too conservative, this weaker stability penalty could bring certain benefits in the cases when a system itself is inherently unstable (and therefore, intuitively, cannot be truthfully represented by a stable latent trajectory).

The Koopman connection is interesting, conceptually. And, I find it is a good idea for PDE data, to try to bound the Hessian instead of the Jacobian.

Critical weaknesses

I did not find the Koopman framing to be used in the theory or proofs. Conceptually -- I agree with you, the low curvature regularization penalty Koopman-izes a model, making its latent dynamics "more linear". But, why does one need specifically Koopman theory here? Not simply a low curvature bound (which your proofs are written around), but Koopman theory? Sure, low curvature may be Koopman, but I don't see this as anything more than an interesting observation, in the current version of your paper.

As for the baselines, you already cite (Li et al., NeurIPS 2022, McCabe et al., TMLR 2023), but do not compare to them. On the other hand, I did not find any paper which does what your DMD regularization does: specifically, add a forecasting loss penalty from a window-fitted DMD as a PDE regularizer. A comparison, where the only baseline is to your own baseline, without prior art, is useful only as an ablation. Without a fair comparison to the already-established baselines for forecasting regularization, I do not see how your motivation is supported. Why, and in which scenario would one prefer LCKoop to other methods from the PDE field? And the choice here is explicitly not among solely Koopman-style methods, it's among PDE forecasting ones.

Koopman theory also has a wider variety of baselines than the ones you presented: the Koopman models (Lusch, Xiong, Cheng), which you already cite -- these would also be a nice addition, if the paper is marketed as Koopman.

Questions for authors

I have mostly one question, I phrased in several different ways for clarity:

How do you actually use the Koopman framing? What would change in your paper, if instead of the complicated Koopman theory you would talk only about the low curvature loss penalty? "Low Curvature is Koopman" -- what does this imply?

Suggestion, which would make me change the mark. Determine a specific scenario, where your LCKoop approach improves over stricter & conservative popular PDE forecasting regularization methods. Preferably -- both in theory and in practice. That would make your finding portable, and useful for the reader.

I would be interested in a paper which would show some unstable-system failure mode for other conservative regularization methods, and demonstrate that the curvature regularization approach does not fail here.

Your Koopman framing could also be useful here, if you were to show, say, that Koopman locally-regularized methods fare better at forecasting this instability.

4 comments
AuthorsOpenReview ↗

Response to reviewer q96u

We appreciate the reviewer’s constructive assessment, which helped us clarify the role of Koopman framing and strengthen practical evaluation. The changes include Section 4.1 error-direction revision, KNO results in Table 2, Section 6.4, comparing LCKoop with a dissipativity plug-in regularizer, and Section 6.5, directly testing LCKoop on a nonlinear residual transition built on a frozen learned Koopman representation.

Question 1. The paper is viewed primarily as a low-curvature penalty. How is the Koopman framing used, what would change without it, and what does “Low Curvature is Koopman” imply?

Response. We appreciate the reviewer’s question. We clarify the framing as “Low Curvature Gives Koopman.” Section 6.5 tests the applicability of the plug-in within a Koopman autoencoder framework: we freeze a pretrained encoder, decoder, and linear latent transition KK, then refine only a nonlinear residual transition. Increasing FC reduces the directional Jacobian difference from KK from 0.022540.02254 (FC0) to 0.000230.00023 (FC1; see Figure 7), showing that LCKoop makes the nonlinear transition locally more Koopman-like. Intermediate FC retains nonlinear forecasting accuracy: FC0.1 reduces 18-step MSE by 47.3%47.3\% relative to frozen KK and 50-step MSE by 60.4%60.4\% relative to the unrestricted residual transition. Thus, LCKoop provides a plug-in interpolation between a learned linear Koopman transition and an expressive nonlinear refinement, balancing short- and long-horizon forecasting---a trade-off that also arises implicitly in neural operators. We also clarify that our curvature-based error-propagation proof reveals an implicit advantage of linear Koopman models, without relying on very specific, complex setting of full Koopman theory.

Question 2. The DMD comparison is considered an ablation rather than a fair comparison with established PDE forecasting regularizers. Why, and in which scenario, should one prefer LCKoop to other PDE forecasting regularization methods?

Response. We treat window-fitted DMD as an ablation and compare LCKoop with Li et al.'s dissipativity plug-in regularizer [1], a conservative plug-in penalty. The table reports ms9 MSE, including their combination:

Evaluation regimeBaselineLCKoopDissipativityBoth
Lorenz, near-attractor0.201500.201500.15980\mathbf{0.15980}0.786200.786200.734830.73483
Lorenz, outer sphere (5R5R)4429.614429.614432.454432.453665.163665.163649.49\mathbf{3649.49}
KS, ordinary0.030140.030140.029170.029170.029140.029140.02875\mathbf{0.02875}

Li et al.'s regularizer [1] uses additional synthetic states from a shell outside the attractor to encourage dissipative return and thereby stabilize long rollouts that leave the attractor-supported ergodic regime. Accordingly, dissipativity is effective for outer-sphere forecasts, but worsens average near-attractor Lorenz accuracy. In contrast, LCKoop improves the ordinary regime, where the relevant difficulty is error-direction curvature inside the bounded state radius. These results suggest that curvature reduction through LCKoop complements dissipativity on both KS and Lorenz, with LCKoop alone giving the largest improvement for near-attractor Lorenz forecasts.

Question 3. The paper does not identify a setting where other conservative regularization fails while local curvature regularization remains effective. Can the authors show such an unstable-system setting and demonstrate the effectiveness of LCKoop in that context?

Response. We appreciate the reviewer’s helpful suggestion. Figure 4 identifies a weakness of the ordinary forecaster within the near-attractor regime: error-direction curvature varies substantially across bounded-radius initial conditions, although dissipativity does not improve average accuracy there. Grouping these initial conditions by baseline error-direction curvature, LCKoop improves ms9 MSE by 15.3%15.3\%, 37.5%37.5\%, and 25.2%25.2\% in the three upper-curvature quartiles; the lowest-curvature quartile instead degrades by 15.6%15.6\%, with the high-curvature groups in (b) corresponding to regions that benefit from LCKoop in (a) (see Figure 4).

This trade-off indicates that LCKoop is useful where higher-order local error is pronounced. However, dissipativity regularization by Li et al. [1] substantially harms ordinary near-attractor Lorenz forecasting in this setting, and does not explicitly penalize error-direction curvature. On outer-sphere states, dissipativity is still appropriate because the problem is an out-of-distribution regime, where contraction toward the attractor is beneficial; we also stress that adding LCKoop further lowers forecasting MSE. Together with the KS result, this supports the interpretation that conservative radius control and local curvature control address distinct failure modes, and suggests complementary use of the two plug-in penalties.

[1] Z. Li et al. Learning Chaotic Dynamics in Dissipative Systems. NeurIPS, 2022.

Thanks for your answer, I'd like to reiterate on question 1 before addressing everything else. As I said, I agree that LC may Koopman-ize the model, which is what your answer was about, I did not argue with that. But, I don't see why that matters in terms of actual takeaways, which is my main concern. If you were to strip all Koopman mentions from the paper -- what would change, for an outside reader? For someone who's reading your paper, someone who's in PDEs but doesn't know anything about Koopman theory: why should they care that "low curvature is/gives Koopman", as opposed to "low curvature makes the model more linear"?

Overall, I believe your paper could be written without any mention of Koopman: it's fundamentally about low-curvature being a good regularizer, and that does not need Koopman.

AuthorsOpenReview ↗

Response to Question1

We thank the reviewer for clarifying this concern. We agree that a PDE practitioner can use LCKoop simply as a low-curvature plug-in regularizer, without learning Koopman theory. This is our contribution to PDE forecasting: a practical alternative to more conservative stability controls that preserves nonlinear expressivity. The Koopman framing also highlights a separate contribution to the Koopman community.

For the Koopman community, our contribution is to identify a useful property of an affine Koopman transition that helps explain its forecasting benefits. Our analysis identifies its vanishing higher-order remainder as one such property and makes it transferable to a nonlinear model. This converts a property of a restricted affine transition into a soft constraint that can also be applied after a Koopman representation has already been learned.

We demonstrate this connection through an application of the plug-in to learned Koopman representations. Rather than choosing between a frozen linear Koopman transition and an unconstrained nonlinear refinement, one can add the plug-in penalty to interpolate between them. Section 6.5 demonstrates this setting directly and illustrates further applicability to the Koopman community: LCKoop can build on recent developments in Koopman autoencoders, retaining their learned representations while regularizing an added nonlinear refinement. Thus, for PDE readers, the mechanism is usable as a standalone regularizer; for Koopman readers, it explains and extends a specific benefit of the affine transition.

The current introduction places greater emphasis on the Koopman contribution. To make the practical takeaways clearer for PDE practitioners, we plan to add a short guide explaining when and how to apply LCKoop, without requiring familiarity with Koopman theory.

I am still unconvinced that this paper constitutes a contribution to the Koopman community. Since the method is effectively a low-curvature plug-in and does not directly utilize Koopman theory, framing this approach as anything other than "Koopman-inspired" is misleading.

Overall, my main concerns remain unresolved, I maintain my original score.

Reviewer_DsyjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LCKoop, an easy-to-use plug-in regularizer that bridges autoregressive neural operators and Koopman-theoretic models. It explicitly constructs the first-order affine approximation implied by the model and encourages it to remain consistent along generated trajectories, without modifying the forecasting architecture or inference path. Across multiple PDE forecasting tasks and neural-operator families, the proposed method yields consistent improvements. The theoretical part provides an analysis of higher-order error propagation.

Critical strengths
  1. The experiments are fairly broad, covering multiple domains and backbones.
  2. Hyperparameters and code are reported in detail, which facilitates reproducibility.
Critical weaknesses
  1. The paper claims that LCKoop reduces the higher-order error term by lowering curvature, yet it never reports the actual magnitudes of rollout error or curvature before and after training, nor any correlation between curvature and final error. As a result, the claim remains unverified.
  2. There is a gap between the method and its stated goal. The method is essentially a second-order smoothness regularizer on the network, and while the paper frames it as "promoting local affineness" or seeking "approximately linear evolution" in latent space, it does not actually linearize the latent dynamics. Moreover, the H in the theorem bounds the Hessian over the full rollout region, whereas training penalizes only a few sampled directions. The relation between the two is left undiscussed.
Questions for authors

My main questions are listed in the Critical Weaknesses. In addition, I have a question regarding the experimental results. In Table 13, several metrics have standard deviations comparable to their means, which weakens confidence in the results: the apparent gains may reflect unstable run-to-run variation rather than a consistent improvement.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Dsyj

We are grateful for the reviewer’s thoughtful comments. The following updates include Section 4.1, Section 6.4, comparing LCKoop with dissipativity regularization, and Section 6.5, testing its effect on a nonlinear residual transition on top of frozen learned Koopman representation.

Weakness 1. Paper claims LCKoop reduces the higher-order error term by lowering curvature, yet does not report change of curvature after training, nor any correlation. Also, the HH in the theorem bounds the Hessian over the full rollout region, whereas training penalizes only a few sampled directions.

Response. We appreciate the reviewer’s insightful question. We revised the theoretical target to Herr,tH_{\mathrm{err},t}, the supremum of curvature along the prediction-error direction over the segment between reference and predicted states. We now directly measure the curvature quantities on the error direction, which is relevant to the coefficient bounding the higher-order error term in our theoretical analysis. The reductions between unregularized (FC0) and rollout-regularized (FC10) FNO models on Shallow Water are given below:

MeasurementFC0FC10Reduction
Long-term forecasting MSE6.577×10−76.577\times10^{-7}5.386×10−75.386\times10^{-7}18.1%18.1\%
Rollout-direction curvature1.502×10−31.502\times10^{-3}3.945×10−43.945\times10^{-4}73.7%73.7\%
Error-direction curvature2.741×10−32.741\times10^{-3}1.484×10−31.484\times10^{-3}45.9%45.9\%

Thus, empirically, lowering the rollout-direction objective is accompanied by lower curvature in the error direction and lower long-horizon error. We note it is a design choice to evaluate the curvature in the error direction directly, or anticipate the error direction with rollout, avoiding arbitrary ambient directions that may be inactive or nearly null in the learned representation.

Section 6.5 provides complementary evidence in a frozen learned Koopman representation. As FC increases, the directional Jacobian difference between the nonlinear refined transition and the pretrained linear KK decreases from 0.022540.02254 to 0.000230.00023 (see Figure 7). This links error-direction curvature regularization to Koopman-like local latent dynamics, while FC0.1 improves 18-step error relative to frozen KK and 50-step error relative to the unrestricted residual transition.

We also test Lorenz. Figure 4 and the table below show that the three upper-curvature quartiles obtain ms9 MSE improvements of 15.3%15.3\%, 37.5%37.5\%, and 25.2%25.2\%, whereas the lowest-curvature quartile degrades by 15.6%15.6\% with an overall near-attractor MSE reduction of 20.7%20.7\%. These results support the interpretation that LCKoop improves long-term forecasting through reduced curvature.

Baseline error-direction curvature quartileMS9 MSE reduction with LCKoop
Q1: lowest 25%−15.6%-15.6\%
Q2: 25--50%+15.3%+15.3\%
Q3: 50--75%+37.5%+37.5\%
Q4: highest 25%+25.2%+25.2\%

Weakness 2. LCKoop appears to be a second-order smoothness regularizer on the network, but the connection to the linearization on latent dynamics is not clearly established.

Response. LCKoop first uses circular wrapping to define a closed transition in a lifted space, where affine evolution can represent nonlinear raw-state dynamics that a raw-space affine map cannot. It then penalizes the curvature of this map. In the ideal limiting case in which linearity is enforced throughout the latent region, the transition is restricted to the affine form Az+bAz+b in the lifted space.

Question 3. Several standard deviations in Table 13 are comparable to their means. Could the apparent gains reflect unstable run-to-run variation rather than a consistent improvement?

Response. Our experiments use downscaled, long-horizon forecasting regimes in which stability and robustness are important. Such challenging rollout-focused settings are common in Koopman forecasting and can exhibit greater run-to-run variation than easier one-step evaluations, especially for nonlinear models. To assess the aggregate trend, we conduct a balanced two-sided sign test in which each of the 16 domain--backbone combinations contributes four matched FC0-versus-LCKoop pairs, for 64 pairs in total.

ms9 metricImprovedWorsenedTwo-sided nominal pp
MSE48/6448/6416/6416/640.00007730.0000773
RMSE46/6446/6418/6418/640.0006170.000617
MAE42/6442/6422/6422/640.0168580.016858
L2RE41/6441/6423/6423/640.0327660.032766

For example, L2RE improves in 41/6441/64 (64.1%64.1\%) matched pairs. These two-sided nominal sign tests assume independent fair signs and suggest a consistent advantage for LCKoop. Section 6.5 further offers a representation-level explanation of how LCKoop improves forecasting: increasing FC drives the nonlinear residual transition toward the learned Koopman matrix (see Figure 7), providing a mechanistic explanation of the interpolation between nonlinear expressivity and Koopman-like stability.

I thank the authors for their response and additional experiments. However, I remain concerned about the experimental results. While the two-sided sign test provides evidence that improvements occur more often than declines, it does not account for the magnitude of those changes. Given the large standard deviations reported in Table 13, I am concerned that the gains may be driven largely by improvement in the single case with the poorest baseline performance.

Reviewer_zD4WOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LCKoop, a Koopman-inspired curvature regularizer for improving autoregressive PDE forecasting. This method controls variations in local linear approximations along generated trajectories, aiming to reduce accumulated prediction errors while preserving nonlinear expressiveness.

The authors provide a theoretical analysis of higher-order error propagation and introduce circular wrapping for applying the regularizer without modifying the inference architecture. Experiments across four PDE tasks and three neural-operator backbones demonstrate average MSE reductions of 8.9% and 14.6% at short and extended forecasting horizons.

Critical strengths

The paper identifies an important limitation of conventional first-order stability constraints. The proposed trajectory-aware curvature regularization is conceptually appealing and can be integrated into different neural operators without modifying their inference pathways.

Critical weaknesses

The theoretical connection between directional curvature and Hessian control relies on a coverage assumption. However, it remains unclear whether the sampled rollout directions sufficiently capture the important curvature directions associated with actual error propagation.

FNO, GFormer, and UNet are relatively established architectures. Evaluations on stronger and more recent neural operators, together with comparisons against existing stability and curvature regularization methods, would better establish the advantages and generality of LCKoop.

Questions for authors

Can the authors provide empirical evidence that the sampled rollout directions sufficiently cover the curvature directions relevant to autoregressive error propagation?

Does LCKoop remain effective when applied to stronger, more recent neural-operator architectures, or can the authors further explain the performance degradation observed in some domain–backbone combinations?

1 comment
AuthorsOpenReview ↗

Response to Reviewer zD4W

We thank the reviewer for the thoughtful and constructive feedback, which helped us strengthen the empirical validation of LCKoop. We have added Section 4.1, KNO results in Table 2, Section 6.4 comparing LCKoop with dissipativity regularization, and Section 6.5 studying a nonlinear residual transition on top of a frozen learned Koopman representation.

Question 1. Sampled rollout directions may not capture the curvature directions relevant to autoregressive error propagation; can the authors provide empirical evidence that they cover these relevant directions?

Response. We appreciate the reviewer’s thoughtful question. Probe direction is a hyperparameter choice: prediction-error displacement directly samples curvature relevant to the second-order bound, whereas rollout displacement anticipates future errors along learned dynamics. Although rollout probing does not ensure complete coverage, it avoids arbitrary ambient directions that may be inactive or nearly null. To provide empirical evidence for the latter, we evaluate curvature along both directions on held-out trajectories using unregularized (FC0) and rollout-regularized (FC10) FNO models on Shallow Water:

MeasurementFC0FC10Reduction
Long-term forecasting MSE6.577×10−76.577\times10^{-7}5.386×10−75.386\times10^{-7}18.1%18.1\%
Rollout-direction curvature1.502×10−31.502\times10^{-3}3.945×10−43.945\times10^{-4}73.7%73.7\%
Error-direction curvature2.741×10−32.741\times10^{-3}1.484×10−31.484\times10^{-3}45.9%45.9\%

These results suggest that rollout directions capture curvature relevant to forecasting errors. From Figure 4(a) and (b), the points that benefit from LCKoop are concentrated in high-curvature attractor regions, while points that degrade are concentrated in low-curvature regions. This supports the interpretation that LCKoop improves long-term forecasting through reduced curvature. Furthermore, Figure 7 empirically shows the evidence that increasing FC makes the learned nonlinear transition converge toward the learned Koopman matrix KK.

Question 2. Does LCKoop remain effective on stronger, more recent neural operators, or can the authors explain the performance degradation observed in some domain--backbone combinations?

Response. We additionally evaluate LCKoop on KNO, a modern Koopman neural operator, without changing its prediction architecture or inference procedure. The average MSE results are:

DomainKNOKNO+LCKoopReduction
Cavity4.30672×10−44.30672\times10^{-4}4.20014×10−44.20014\times10^{-4}2.5%2.5\%
Tube1.4666×10−21.4666\times10^{-2}1.1908×10−21.1908\times10^{-2}18.8%18.8\%
NSCH4.450×10−34.450\times10^{-3}4.084×10−34.084\times10^{-3}8.2%8.2\%
Shallow Water6.55223×10−56.55223\times10^{-5}5.78354×10−55.78354\times10^{-5}11.7%11.7\%

Thus, LCKoop improves long-horizon forecasts on a Koopman-inspired operator backbone. Section 6.5 further observes in detail how FC provides a performance interpolation between nonlinear expressivity and Koopman-like stability: low FC favors short-horizon accuracy, whereas increasing FC makes the residual transition converge to the Koopman matrix and improves longer-horizon behavior (see Figure 7). From Figure 5, LCKoop promotes stability by limiting curvature, which can improve long-term forecasts at the cost of short-term accuracy. Table 12 shows that limited model capacity can prevent such improvements. Our experiments use downscaled datasets with an emphasis on initial transient dynamics, which challenges the model and requires robustness. Since LCKoop is a plug-in loss that leaves the prediction architecture and inference path unchanged, practitioners can readily test whether it is beneficial in their setting.

Question 3. Can the authors provide comparisons against existing PDE stability regularization methods?

Response. We now compare LCKoop with Li et al.'s dissipativity plug-in regularizer [1], which samples synthetic high-radius states and encourages contraction. We compare ms9 forecasts for regularization-free, LCKoop, the dissipativity penalty, and their combination on Lorenz-63 and Kuramoto--Sivashinsky (KS). Figure 4 and Table 5 show that LCKoop improves near-attractor Lorenz forecasting, with gains concentrated in the higher-curvature quartiles, whereas dissipativity benefits outer-sphere forecasts but degrades near-attractor accuracy in this setting; adding LCKoop further lowers mean MSE relative to dissipativity alone in all three evaluation regimes, and their combination achieves the lowest mean MSE on outer-sphere Lorenz and KS forecasts, demonstrating their complementary benefits.

Evaluation regimeBaselineLCKoopDissipativityBoth
Lorenz, near-attractor0.201500.201500.15980\mathbf{0.15980}0.786200.786200.734830.73483
Lorenz, outer sphere (5R5R)4429.614429.614432.454432.453665.163665.163649.49\mathbf{3649.49}
KS, ordinary0.030140.030140.029170.029170.029140.029140.02875\mathbf{0.02875}

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear Reviewers,

We thank you for your thoughtful feedback. We have revised the paper and conducted additional experiments to clarify the theoretical motivation, the connection to Koopman models, and the practical scope of LCKoop. The main updates are summarized below.

  1. Error-relevant curvature and theoretical interpretation (Section 4.1 and Table 1). We revised the curvature target to focus on the prediction-error direction along the segment between the reference and predicted states. To examine whether rollout-direction probing captures relevant curvature, we measured both directions on Shallow Water. Compared with the unregularized FNO, the rollout-regularized model reduced rollout-direction curvature by 73.7%, error-direction curvature by 45.9%, and long-term forecasting MSE by 18.1%. This provides empirical evidence of a connection between the sampled direction and error-relevant curvature.
  2. Comparison with a PDE stability regularizer (Section 6.4 and Appendix H). We added a comparison with dissipativity regularization on Lorenz-63 and Kuramoto–Sivashinsky, including the combination of both methods. LCKoop improved ordinary near-attractor Lorenz forecasting, while dissipativity was more effective for outer-sphere initial states. Combining the two reduced MSE relative to dissipativity alone in all three reported evaluation regimes, suggesting that the methods can address complementary failure modes. Appendix H provides the experimental settings and full mean and standard-deviation results.
  3. Curvature-sensitive regimes (Section 6.4, Figure 4). On near-attractor Lorenz trajectories, LCKoop improved ms9 MSE by 15.3%, 37.5%, and 25.2% in the three upper baseline error-curvature quartiles; the lowest-curvature quartile instead degraded by 15.6%. This analysis identifies a regime where curvature control is particularly helpful, while also showing that improvement is not uniform across initial conditions.
  4. Connection to learned Koopman representations (Section 6.5 and Appendix I). We applied LCKoop to a nonlinear residual transition on top of a frozen Koopman-style autoencoder and its learned linear transition. Increasing the regularization strength brought the transition's local behavior closer to the learned linear operator. An intermediate setting reduced 50-step average MSE by 60.4% relative to the unregularized residual model and reduced 18-step MSE by approximately 47% relative to the frozen linear transition. This experiment illustrates the trade-off between nonlinear expressivity and Koopman-like local dynamics, and potential plug-in application to Koopman autoencoders.
  5. Evaluation on a recent Koopman neural operator (Section 6.2 and Table 2). We added KNO results across four domains. LCKoop reduced ms9 MSE by 2.5%–18.8% across these domains without changing the prediction architecture or inference procedure. We also discuss the stability–expressivity trade-off and cases where performance can degrade.

We hope these additions address the reviewers’ concerns and clarify both the benefits and limitations of LCKoop. We appreciate the opportunity to improve the paper.

Dear reviewers,

Thank you for your reviews. The author-reviewer discussion closes on November 18, so please respond to the authors' rebuttal before then. Please say whether it addressed your concerns, and update your score if your view has changed.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.