acceptodds
Under review as a conference paper at ICLR 2027

Data-driven Random-feature Approximation of Koopman Generators of Nonlinear Continuous-time Systems

Abstract

Learning continuous-time dynamical systems from data is an active area of research. The Koopman formalism is often used here, because it is a flexible approach that is informative about the system's fundamental properties. The formalism represents dynamical systems through linear operators acting on a suitable function space. A practical difficulty arising in the numerical approximation of Koopman operators is the choice of a “dictionary”, meaning an effective finite-dimensional representation of the function space. In this paper, we focus on learning dynamical systems in continuous-time, and so we consider numerical approximations of the Koopman generator instead of the Koopman operator. Our main contribution is a data-driven random-feature method to obtain a parsimonious and effective function representation for Galerkin approximation methods such as gEDMD. The representation we construct is data-informed, which allows us to improve upon data-agnostic methods. We empirically demonstrate the suitability of our method for forecasting tasks and system analysis. The numerical experiments show that our approach is accurate, and faster than iterative optimization. Further, we prove convergence of our approximation of the vector field and the resulting trajectories. In addition to the numerical results, we also make theoretical contributions, including a general convergence criterion for the semigroup generated by the gEDMD approximation and a counterexample showing that gEDMD does not generally preserve exponential boundedness. With our contribution, we advance the frontier of efficiently learning continuous-time dynamics from data, towards interpretable and mathematically robust modelling based on the Koopman framework.

Primary area:
learning on time series and dynamical systems
Keywords:
continuous-time dynamical systems, random feature methods, Koopman operator, Koopman generator, dictionary learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Jie4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes DRAGONS for approximating the Koopman generator of continuous-time systems based on gEDMD. The key innovation is that it learns the Koopman generator on a data-driven dictionary via SWIM-sampled random features, which is gradient-free. The paper also provides convergence of the approximate vector field and trajectories. In addition, it shows LE/FTLE bounds on exploding adjoint gradients of Neural ODEs on chaotic systems.

Critical strengths

This paper studies an important problem on Koopman learning, since dictionary choice is the central bottleneck. Evaluating candidate neurons by how they act under the generator before constructing them is an interesting idea. The theoretical results on gEDMD preserving quasi-dissipativity (Lemma 4) and the resulting semigroup-level convergence via Trotter–Kato are also meaningful contributions.

Critical weaknesses

(1) The dictionary construction follows from SWIM (Bolager et al., 2023). Its combination with Koopman approximation was already proposed in Bolager et al. (2026a), and, as described in Appendix A, uses SWIM-sampled observables with EDMD and the flow-driven density. The innovation compared to Bolager et al. (2026a) could be stated more explicitly. (2) The method is formulated for deterministic systems, but is applied to stochastic systems such as the alanine dipeptide. The original EDMD algorithm does apply to both cases, yet there are several theoretical gaps in the paper. For example, Lemma 1, Theorem 2, and the Neural ODE analysis are, in their current form, stated for deterministic systems. The paper could clarify its scope more explicitly. (3) The experiments are on 2D or 3D with polynomial vector fields. As the polynomial dictionary is known to be rich enough, it is hard to tell the benefit of the proposed method. (4) As stated in Assumption 1, the samples are IID from a measure whose density is bounded away from zero. Theorem 2 additionally requires quasi-contractivity. It would be helpful to check these assumptions in the experiments.

Questions for authors

(1) Could the author clarify the contribution relative to Bolager et al. (2026a), beyond replacing EDMD with gEDMD? (2) Is the result intended for deterministic systems only, or can it be extended to stochastic systems easily as well? (3) In the Lorenz example in Table 2, DRAGONS uses 131 features against 19 for gEDMD-DL. Is the baseline also tuned, or could the author provide some explaination why the dictionary sizes differs so much.

2 comments

We sincerely thank the reviewer for their time and the provided feedback.
First, we address the questions:

  1. Although our work bears similarities with Bolager et al. (2026a), there are important differences to keep in mind: first of all, Bolager et al. only use the flow-driven density, which is an immediate application of the SWIM algorithm to their setting. We explore three different strategies: the flow-driven density an analogue of Bolager et al., for which we also provide an argument in Appendix C for why it is inadequate in our setting of irregular timesteps beyond the empirical evidence; the vector-field-driven sampling, which is the immediate application of the SWIM algorithm to the goal of regressing a vector field with a network; and the generator-driven density, which is a novel methodological contribution based on using the neuron to be constructed in the computation of the score, which has not been explored in any prior work. Furthermore, our approach does not construct a recurrent neural network; secondly, our theoretical contributions are more extensive and also applicable beyond the scope of this work; thirdly, our computational experiments include an example from molecular dynamics, and a 7D dynamical system (newly added).

  2. Our theory is developed only for deterministic systems, we now make this clear with changes in the manuscript. It is possible to extend our approach to stochastic systems if the diffusion term of the Koopman generator is approximated as well; however, we leave this to future work. In our numerical experiments we explore if our sampled dictionary choice combined with EDMD (which exists for stochastic systems) is potentially viable for a stochastic system like the Alanine Dipeptide, the text was also changed to reflect this.

  3. In Table 2 the number of features in provided in order to aid with the interpretation of the training time. The choice of features across different problems and models differs due to each model being tuned to the specific problem via a hyperparameter search, the details of which are given in Appendix F. Both methods were tuned over the same range of configurations (including the number of features). In our view, there are arguments in favor of a large dictionary (e.g. higher expressiveness, assuming independence in features), but also arguments against a large dictionary (e.g. ill-posedness of the linear system), thus is it difficult to make a statement in general about which dictionary size is favorable for a particular problem. The resulting number of features was selected according to the error on a validation set, which we consider to be in line with the conventional model selection approach in machine learning. A possible explanation for why gEDMD-DL ends up using (significantly) fewer features in this example is that DRAGONS is a random feature method which selects the features probabilistically according to a data-driven heuristic. Thus, the selected features might not be optimal, and it is also possible that essentially useless features are selected. gEDMD-DL solves an optimisation problem for each feature, meaning that the trained features will typically be reasonably close to optimality, but this is negatively reflected in the training time. It is true that a resulting larger dictionary incurs higher costs for prediction, but this could be mitigated by pruning the sampled features or compressing them using an SVD. We have, however, not explored these extensions to our method so far.

We additionally want to specifically address the concern raised in Critical Weakness 4 (the others are addressed above). Our theory assumes i.i.d. samples from a density bounded away from zero on a forward-invariant set, whereas our experiments use trajectory data. The forward invariance is a standard assumption in Koopman Theory and compatible with the conducted experiments by the presence of attractors. It is worth mentioning that we sample the initial conditions uniformly over rectangular regions, which will not typically be strictly forward invariant since some trajectories will temporarily leave them. It is, however, very difficult to describe forward invariant sets explicitly even for simple systems like Van der Pol, but we assume a slight mismatch here to be of limited relevance. The major mismatch is the use of trajectories rather than IID-sampled points. We chose this for two reasons: Firstly, we deem trajectory data to be the more realistic setting, since it more closely resembles the format of measurements from physical systems and does not require knowledge of the vector field. Secondly, we note that a sufficiently developed theory for gEDMD under dependent data, building on ergodic-sampling results, is not yet available and is thus beyond the scope of this paper; we have added this to the limitations. For example, Nüske et al. 2022 [1] provide a good foundation in this direction, but only consider the finite data approximation of generators projected on finite dictionaries, not the approximation of the true generator in the case of ergodic sampling of trajectories.
To bridge this gap, we added a second version of the Van der Pol experiment where the model is trained with i.i.d. sampled data, and provide the results in Appendix F.1.1.

All relevant changes are highlighted in blue in the revised manuscript.

[1] https://doi.org/10.1007/s00332-022-09862-1

Reviewer_PZX3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DRAGONS, a method for learning continuous-time dynamics through the Koopman generator. Instead of training a dictionary, it samples the weights and biases of a single hidden layer from pairs of data points with SWIM (Bolager et al., 2023), using a data-driven ("generator-driven") density, and then fits the generator by least squares as in gEDMD. Prediction uses the matrix exponential, with re-lifting at each step and a linear projection back to the state. On the theory side, the paper proves convergence of the learned vector field and trajectories, using existing gEDMD results and the Trotter–Kato theorem. It also gives a counterexample showing that gEDMD does not preserve exponential boundedness in general, and shows that adjoint gradients of Neural ODEs grow exponentially on chaotic data. Experiments on Van der Pol, Lorenz-63 and a mean-field cylinder-flow model report errors comparable to the best baselines, with much faster fitting than a gradient-trained dictionary. The paper also includes qualitative spectral analyses of cylinder flow and alanine dipeptide.

Critical strengths
  • A cheap, training-free, data-adapted dictionary for gEDMD, addressing the main practical bottleneck of (g)EDMD. In Table 2 it matches or beats an Adam-trained dictionary while being about 3–115× faster.
  • A clean operator-theoretic framing with consistency results and a useful negative result. Theorem 3 shows that gEDMD semigroups need not be uniformly exponentially bounded, which matters for any gEDMD-based long-time prediction.
Critical weaknesses
  • The experiments do not support the main claim that data-driven random features give a parsimonious dictionary that beats data-agnostic ones. All three forecasting systems have polynomial vector fields, where a small polynomial dictionary is already nearly exact (the paper says Poly-Dict "sets the standard"). DRAGONS uses 4–15× more functions than Poly-Dict (130 vs. 9, 131 vs. 34, 515 vs. 34). It is on par on Lorenz and slightly worse on Cylinder, where all good methods reach about 3e-7. Dictionary sizes also differ strongly between methods (Lorenz: 131 vs. 67 vs. 19), so error differences are confounded; the only size-matched win over data-agnostic features is Van der Pol. The error metric and horizon are not given in the main text, which matters a lot for chaotic Lorenz. Standard baselines are missing (e.g. SINDy, kernel gEDMD, and a Neural ODE, which Section 3.3 argues against). The two high-dimensional experiments are qualitative only, with no reference values or baselines.
  • The theory is generic and does not cover the method as used. Theorems 1–2 are existence results (a suitable sampled layer exists) without rates, built from universal approximation plus existing gEDMD and Trotter–Kato results. As the authors note (l. 270–272), they hold for any universal dictionary, so they do not explain why data-driven or generator-driven sampling should help. Their assumptions are also not met in the experiments. Assumption 1 needs i.i.d. samples from a density bounded below on a full-dimensional set, and Theorem 2 needs quasi-contractivity in L2(μ)L^2(\mu), which is shown only for the Lebesgue measure (Lemma 1). The experiments instead appear to use trajectory data concentrated on attractors, limit cycles or the MD equilibrium distribution. Algorithm 2 also re-lifts after every step, which is not the scheme analysed in Theorem 2. Finally, Theorem 4 describes a known phenomenon (e.g. Lea et al., 2000; Mikhaeil et al., 2022). It is phrased via the chaoticity of the data, although adjoint growth is driven by the learned model's tangent dynamics, and it is not used in any comparison (no Neural ODE baseline).
Questions for authors
  • How is "Error" in Table 2 defined (quantity, horizon; for Lorenz, how many Lyapunov times)? How were dictionary sizes and baseline hyperparameters chosen, especially the weight scale / input normalisation of the Gaussian features and the training budget of gEDMD-DL? Please report error vs. dictionary size (and vs. fitting time) for all methods. If the gap to Gaussian features closes at matched size with a tuned scale, my view of the main claim would change.

  • How were the training data generated (initial conditions spread over a box, or long trajectories on the attractor)? In the latter case Assumption 1 fails; how should Theorems 1–2 be read for these experiments? Also, Theorems 1–2 state that the bound holds almost surely for every M≥MNM \ge M_N with a deterministic MNM_N. This does not follow from the law of large numbers, since for any finite MM a bad sample generally has positive probability. Should MNM_N be random, or should the statement be in probability?

  • Heavy-atom MD trajectories are effectively stochastic, so the relevant generator has a diffusion term. For equilibrium (reversible) dynamics it also has a real, non-positive spectrum. How is X˙\dot X obtained, and which generator is approximated? Fig. 5a shows Im(λ)\mathrm{Im}(\lambda) of order 10−2010^{-20} (numerically zero) and ∣Re(λ)∣|\mathrm{Re}(\lambda)| below about 10−410^{-4}. This does not match "most eigenvalues are along the imaginary axis" (l. 453–454). Do the eigenvalues carry kinetic information, e.g. do implied timescales agree with reference MSM/VAMP values for this dataset?

  • The claim that spectral properties are extracted "accurately" is not quantified. How do the eigenvalues compare with the known structure (eigenvalues ikω0ik\omega_0 on the limit cycle, with ω0\omega_0 the shedding frequency) and with a baseline such as kernel gEDMD or ResDMD? Many eigenvalues have positive real part and large residual; how are reliable eigenpairs selected? Is this dataset the full flow field, while Section 4.1 uses the 3D mean-field model?

  • What is new compared with random-feature (g)EDMD (Tabish et al., 2025; Sprink et al., 2026)? What is new compared with earlier work combining data-dependent sampled hidden layers with EDMD in discrete time (Bolager et al., 2024, arXiv:2410.23467)? Is the generator-driven density the key ingredient? Suggestion: a main-text ablation over the densities of Table 1 (plus data-agnostic features at matched size), on at least one non-polynomial or higher-dimensional system without a good analytic dictionary, would directly test the central claims. A Neural ODE baseline trained on short windows or with multiple shooting would do the same.

4 comments

We thank the reviewer for their insightful comments and try to address concerns with this reply.

  • The errors given in Table 2 are MSE errors, with an exception in the case of Lorenz 63, where we estimate the empirical KL divergence between the true and predicted trajectories; details are given in Appendix F, and we have made the distinction clear in the main text. The hyperparameter search ranges were determined manually via trial and error; best combinations were determined by evaluating their performance on a held-out validation set. The search ranges are given in Appendix F. The neural network-based approaches (DRAGONS, Gaussian, gEDMD-DL) were typically given the same neuron budget. The exception is the cylinder flow experiment, where Gaussians and gEDMD-DL performed poorly in the original search range, so the search range was extended for those methods. The achieved errors versus the number of neurons are now reported for the neural-network-based approaches in the respective subsections of Appendix F in the updated manuscript.
  • The data consists of several trajectories starting from uniformly randomly sampled initial conditions throughout the domain as described in Appendix F.
    The concern regarding the number of data points as a constant in Lemma 4 and Theorems 1 and 2 is well-founded, and we thank the reviewers for highlighting it. M_N should be random, as it is in the respective results in Llamazares-Elias et al [1]. We have appropriately revised the statements and proofs of Theorems 1 and 2 and Lemma 4, where necessary. The idea and general strategy of the proofs of Theorems 1 and 2 remain unchanged; however, specific details, such as the intermediate points in the triangle inequality, have been adjusted to properly handle the convergence of the random term. The proof of Lemma 4 remains essentially unchanged.
  • We do not approximate the diffusion term and approximate the derivatives using a finite difference scheme, thus this experiment is an exploratory study of the potential for extending our method to stochastic systems. Currently our theory is limited to deterministic systems, and we have updated the manuscript to clearly reflect that. Stochastic systems remain to be explored future work, and the heavy MD example provides a glimpse of how such an extension would perform empirically. We agree also with the remarks regarding the eigenvalues and have updated the manuscript accordingly. Regarding the implied timescales, in our experiment we use a lag of 100ps, with an eigenvalue of 1.5e-4, the timescale would be 11.36ps, but according to the MSM values reported in [2] (Fig. 6d) the implied timescale is 10² to 10³ps, thus there is a difference of about one order of magnitude and we do not assume that the eigenvalues carry kinetic information.
  • The term “accurately” might be an overstatement indeed, since the conducted experiment is purely qualitative. We have appropriately revised this section. There was a mistake in the implementation of the spectral residual computation according to formula (243). We have now corrected this and updated all affected figures. For this specific experiment, one can now observe eigenvalues with lower residuals at i∗k∗ω0i*k*\omega_0 matching the known structure. Also, eigenpairs with high importance (high contribution to rollout) have lower residual values. Simply removing eigenpairs with high residuals does not, in general, provide better predictions. The dataset used is the 3D mean-field model.
  • The key distinction from Tabish et al., 2025, and Sprink et al., 2026, is indeed the use of data-driven random features with a random-feature scheme designed specifically to serve as a dictionary for gEDMD, i.e., the generator-driven density. This is also the main innovation compared to Bolager et al. 2024. We discuss the different sampling strategies and, in particular, why, for our setting of irregular time steps, the flow-driven density corresponding to Bolager et al. 2024 is potentially inadequate, in detail in Appendix C. Additionally, we make theoretical contributions that are relevant beyond our method (Lemma 4 and Theorems 2, 3, 4). An ablation over the sampling densities can be found in Appendix F.7.

The relevant changes in the revised manuscript are highlighted in blue.

[1] https://arxiv.org/abs/2405.00539

[2] http://dx.doi.org/10.1063/1.4976518

We also want to address the concern about the novelty of Theorem 4 raised in the second weakness by properly putting our contribution in context. Mikhaeil et al. (2022) theoretically relate the maximum Lyapunov Exponent (LE) of RNN trajectories to loss-gradient behavior. Their analysis, however, concerns discrete-time RNNs trained via backpropagation through time, rather than continuous-time neural ODEs trained using the adjoint sensitivity method. Lea et al. (2000) numerically demonstrated exponential growth of adjoint sensitivities in the chaotic Lorenz-63 system and associated this growth with the largest LE. Their study focused on climate sensitivity and did not establish a general theoretical relationship between LEs and parameter-gradient growth in learned neural differential equations. We have now cited this work in our manuscript.

Our contribution is therefore not the general observation that chaotic dynamics can induce gradient growth, but rather its theoretical analysis in continuous-time learned neural ODEs trained via the adjoint method. To our knowledge, this is novel and the first of its kind. Specifically, we derive finite-time bounds on neural ODE adjoint sensitivities in terms of finite-time LEs, alongside asymptotic bounds characterized by LEs.
We have intentionally excluded NeuralODEs from the main experiments because they serve a different purpose than gEDMD. Our theory also does not discuss their suitability as a vector-field model; it only concerns their training dynamics in the specific setting of long trajectories of chaotic systems. We have empirically verified the theoretical findings about the growth of the parameter gradients and how they relate to adjoint states in the experiment reported in Appendix F.9.

Regarding the first question, specifically tuned-scale Gaussians, we intentionally chose the default standard-normal weight distribution. Tuning the scale could be done in two ways, namely (a) as part of the fitting using the training data or (b) by searching over an additional hyperparameter. We decided against the first option because this would make it a data-driven method, so it wouldn’t fulfil the role of a data-agnostic baseline. Regarding option (b), since the scale is a continuously varying parameter, it is difficult to tune it as a hyperparameter. If too few values are evaluated, it doesn’t meaningfully represent the performance of a tuned scale. If many values are evaluated, this would incur high cost for evaluating the validation data many times, which is not reflected in the training time of the method and essentially turns it into a data-driven method again, only that the weights are actually computed from the validation data.
Nonetheless Gaussians with fixed scale are not a weak baseline, since ReLU is positively homogeneous with degree one, a scaling factor on the weights can be absorbed into the final coefficients, which are optimally fit by least-squares.

Finally, regarding the concern about missing baselines raised in the weaknesses, we have added SINDy results to all major experiments, except for the newly added Glycolytic Oscillator. We have tried applying SINDy to the problem, but it consistently picks up many spurious terms, leading to divergent predictions. While SINDy succeeded on this benchmark in the original work by Brunton et al., our setting is different in two regards, which we suspect to explain the performance mismatch. Brunton et al. used regularly sampled snapshots with uniform timesteps of Δt=0.001\Delta t=0.001, whereas our timesteps are sampled irregularly from a much larger range Δt∼U(0,0.1)\Delta t \sim U(0, 0.1). We suspect that the variation between the timesteps is responsible for the appearance of the spurious terms. Thus, significant tuning or modification of the iterative sparse least squares procedure appears to be necessary in our setting, which we did not have sufficient time for during the rebuttal phase.

Reviewer_8KyHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a data-driven random-feature method for constructing effective, parsimonious dictionaries for Galerkin approximation of the Koopman generator. It allows accurate, fast learning of continuous-time dynamic systems. It provides theoretical convergence guarantees and improves forecasting performance.

Critical strengths
  1. Data-informed random features offer a computationally attractive approach to dictionary selection for continuous-time Koopman modeling.

  2. The experiments demonstrate reduced fitting costs relative to iterative dictionary learning and report results across five seeds. Investigating the distinction between generator convergence and semigroup stability is also valuable.

Critical weaknesses
  1. Theorem 7’s parameter-gradient lower bound is false as stated. An upper bound on the parameter Jacobian and growth of the adjoint do not imply growth of their integrated product. Consider fθ(x,y)=(x(1−x2),θ1−y21+y2), z(0)=0, J=12(x(T)−1)2f_\theta(x,y) = \left( x(1-x^2), \theta \frac{1-y^2}{1+y^2} \right), \, z(0)=0, \, J=\tfrac12(x(T)-1)^2. At θ=0\theta=0, the maximal FTLE is 11, aT=(−1,0)a_T=(-1,0) satisfies the non-orthogonality condition, and ∣∂θf∣≤1|\partial_\theta f|\le1. However, ∇θJ=0\nabla_\theta J=0, whereas Eq. (213) requires a lower bound of eT−1>0e^T-1>0.

  2. Lemma 3 requires a missing core argument. Absolute continuity does not imply D(L)=W1,2D(\mathcal L)=W^{1,2}. The required graph-norm density therefore remains unjustified. This is a consequential proof gap, rather than a counterexample to Theorems 1–2 themselves.

  3. Generator-driven sampling can be undefined. For pure rotation F(z)=ΩzF(z)=\Omega z, with Ω⊤=−Ω\Omega^\top=-\Omega, every weight in Eq. (37) vanishes.

Questions for authors

how does the method performance for high-dimensional dynamic systems?

1 comment

We are grateful for the insightful feedback from the reviewer.
First of all, we address the question regarding performance for higher-dimensional dynamical systems by introducing an experiment where the dynamics evolve in 7D, and furthermore, cannot be described solely using polynomial terms; we call this experiment “Glycolytic oscillator” and the results have been incorporated into the manuscript. Our method is able to capture the dynamics of this system with an accuracy comparable to that of existing state-of-the-art methods.

The weaknesses mentioned by the reviewer also contain valid points, which we have re-evaluated now.
Weakness 1: We thank the reviewer for correctly identifying this gap, which we have closed by adding an explicit nondegeneracy assumption under which the lower bound holds. Accordingly, we have revised Theorem 7 and its proof. Now, regarding the reviewer's example at θ=0\theta = 0:

  • z(t)≡0z(t) \equiv 0, ∂θf=(0,1)⊤\partial_\theta f = (0,1)^\top, Φ(T,t)=diag(eT−t,1)\Phi(T,t) = \mathrm{diag}(e^{T-t},1), aT=(−1,0)a_T = (-1,0), β(0,T)=−1\beta(0,T) = -1.
  • Hence a(t)=(−eT−t,0)a(t) = (-e^{T-t},0) and ∇θJ=0\nabla_\theta J = 0.
  • KT=(0,T)⊤K_T = (0,T)^\top and a(0)=(−eT,0)a(0) = (-e^{T},0), so a(0)KT=0a(0)K_T = 0 and γT=0\gamma_T = 0.

The corrected finite-time two-sided bound gives ∣β(0,T)∣ γT eT=0≤∥∇θJ∥≤∥aT∥ γT eT=0|\beta(0,T)|\,\gamma_T\,e^{T} = 0 \le \|\nabla_\theta J\| \le \|a_T\|\,\gamma_T\,e^{T} = 0, which is exact. The theorem also keeps a priori integral bound ∥∇θJ∥≤Cθ∥aT∥∫0TeT−t dt=eT−1\|\nabla_\theta J\| \le C_\theta\|a_T\|\int_0^T e^{T-t}\,dt = e^T - 1 which is valid but not tight. Likewise, the corrected asymptotic bounds give 0≤∥∇θJ∥≤00 \le \|\nabla_\theta J\| \le 0. All of these are consistent with ∇θJ=0\nabla_\theta J = 0.

Weakness 2: We thank the reviewers for bringing this mistake to our attention. It has been addressed by appropriately revising Lemmas 1 and 3. We no longer require or assume that D=W1,2\mathcal{D}=W^{1,2}. In Lemma 1, we show that W1,2W^{1,2} is a core of the generator (and thus dense in the Domain, rather than equal to it). Roughly, this follows from invariance under the Koopman operator, which in turn follows from exponential boundedness (see Lemma 1 for details). In Lemma 3, we then extend the approximation of Sobolev functions to functions in the Domain by bounding the graph norm in terms of the Sobolev norm.

Weakness 3: Yes, this is a valid objection and can in fact happen for all three sampling strategies in degenerate cases (e.g. the zero vector field). We have added a clarifying remark in the main text explaining that we follow Bolager et al. 2023 [1] and fall back to uniform pair sampling in these cases.

For convenience, relevant changes are highlighted in blue in the updated manuscript.

[1] https://arxiv.org/abs/2306.16830

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.