acceptodds
Under review as a conference paper at ICLR 2027

PadéKoop: Learning Rational Observables for Long-Horizon Dynamics

Abstract

Finite-dimensional Koopman models depend critically on the observables used to represent nonlinear dynamics. Generic neural encoders can fit in-distribution trajectories, but provide no structural preference for inverse-power, fractional, or saturation dynamics and can extrapolate poorly beyond the training region. We introduce **PadéKoop**, a Koopman framework that learns observables as ratios of independent numerator and strictly positive denominator networks. This Padé-style parameterization represents denominator-driven structure while keeping every observable well-defined; the resulting features augment the physical state and evolve through a linear Koopman operator. To separate representation error from latent amplification, PadéKoop combines rational observables with optional spectral control of the Koopman matrix. We derive well-posedness, rational-approximation, and rollout-error properties that clarify these complementary roles. Across four equation-generated physical and biochemical benchmarks, PadéKoop achieves the lowest mean horizon-100 out-of-distribution error among controlled MLP- and KAN-observable Koopman baselines. A parameter-matched enzyme-kinetics ablation shows 69% lower error than an MLP and 54% lower error than a constant-denominator control. Spectral-control experiments further show that hard projection can suppress catastrophic latent amplification, but does not consistently improve predictive accuracy. Together, these results support rational observables as an effective inductive bias for denominator-driven dynamics and spectral control as a complementary robustness mechanism, rather than a guarantee of universally accurate long-horizon prediction.

Primary area:
learning on time series and dynamical systems
Keywords:
Koopman operator learning, rational neural networks, Padé approximation, nonlinear dynamical systems, long-horizon prediction, out-of-distribution generalization, structured inductive bias, spectral regularization.

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_KEs2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PadéKoop, a finite-dimensional Koopman learning framework in which nonlinear observables are parameterized as elementwise quotients of independently learned numerator and positive denominator networks. The resulting rational features are concatenated with the physical state and propagated using a learned linear Koopman operator. The method additionally considers denominator regularization and spectral/operator control to address numerical instability in the quotient representation and amplification during latent rollouts.

The paper provides several basic analytical results concerning well-posedness and Lipschitz continuity of the quotient observables, approximation error when the target admits a rational representation, and accumulation of one-step latent errors under repeated Koopman propagation. Empirically, the method is evaluated on four equation-generated physical and biochemical dynamical systems and compared against Koopman models using MLP- and KAN-based observables. The reported results show improved long-horizon out-of-distribution prediction on the considered benchmarks, including a parameter-matched ablation against a standard MLP and a constant-denominator variant. The paper also investigates spectral control of the learned Koopman operator and finds that it can prevent severe latent amplification, although it does not consistently improve prediction accuracy.

Overall, the paper presents a simple and empirically effective rational-observable architecture, while the central question is whether the Padé-inspired structure itself has a sufficiently principled advantage over generic neural observable parameterizations.

Critical strengths

The proposed quotient parameterization is simple, interpretable at the architectural level, and easy to incorporate into existing neural Koopman models. The paper also makes a commendable effort to connect its regularization choices to explicit failure modes: the denominator-margin analysis addresses instability caused by small denominators, while the rollout analysis separates representation error from amplification by the learned linear dynamics. The distinction between latent and re-encoded rollouts is also useful for assessing approximate Koopman closure. Finally, the empirical results indicate that the method is practically viable across several dynamical systems.

Critical weaknesses

The main conceptual motivation for the Padé-inspired parameterization is currently unclear. The analysis establishes that the quotient representation is well-defined and that its approximation and rollout errors can be controlled, but this does not answer the more important question: why should this structure be preferred over a standard neural observable? Good empirical performance alone is not sufficient to establish this point. There are already many competing Koopman lifting architectures, and without a very broad comparison it is difficult to attribute the observed gains specifically to the quotient structure. The paper should therefore provide a more principled argument for why the Padé form is expected to yield a better or more efficient representation. In this regard, the lack of discussion of the classical Padé/rational approximation literature is surprising. Padé approximants are often motivated by approximation properties that can be superior to truncated Taylor or polynomial expansions. If these properties motivate the present architecture, the paper should explain which of them carry over to Koopman observables and how they are reflected in the proposed theory. As written, the current propositions justify the stability of the parameterization, but not the choice of the parameterization itself.

Questions for authors

Q1. What is the strongest principled reason for preferring the proposed Padé-inspired quotient observable over a standard neural observable? The current propositions establish well-posedness and error control. Still, they do not appear to show that the quotient structure is inherently more expressive, more data-efficient, or more suitable for Koopman closure. Is there a theoretical property of the rational parameterization that the authors believe is central to the method?

Q2. How should we understand the connection to classical Padé/rational approximation? Which specific approximation properties of Padé approximants motivate the present architecture, and do any of those properties carry over to the learned Koopman setting considered here? It would be useful to clarify whether “Padé-inspired” is primarily an architectural analogy or whether a deeper approximation-theoretic connection is intended.

2 comments
AuthorsOpenReview ↗

Rationale for the Padé-Inspired Observable

Response to Reviewer 3

We thank the reviewer for noting that well-posedness alone does not justify a quotient observable. We revised the manuscript to explain its Koopman-specific rationale and distinguish it from classical Pade theorems.

Question 1: Why prefer the quotient observable?

Our strongest principled reason is not universal expressivity or guaranteed sample efficiency. It is algebraic compatibility with denominator-driven dynamics under Koopman composition. Inverse-power and saturating mechanisms have forms such as 1/r21/r^2, 1/r7−1/r131/r^7-1/r^{13}, and Vmax⁡s/(Km+s)V_{\max}s/(K_m+s). The quotient makes numerator variation and denominator-driven scaling explicit. A generic MLP can approximate these functions on a compact domain, but does not encode this structure as an inductive bias. The Van der Pol boundary experiment, where this mechanism is absent, shows that we do not expect the same benefit for arbitrary dynamics.

Let Ω\Omega be forward-invariant under a rational transition F:Ω→ΩF:\Omega\to\Omega, and let hh be rational with all denominators nonzero on Ω\Omega. Then

UFh=h∘F∈R(Ω),UFR(Ω)⊆R(Ω).\mathcal{U}_F h=h\circ F\in\mathcal{R}(\Omega),\qquad \mathcal{U}_F\mathcal{R}(\Omega)\subseteq\mathcal{R}(\Omega).

Indeed, if Fi=pi/qiF_i=p_i/q_i and h=ph/qhh=p_h/q_h, substitution and clearing finitely many denominators expresses h∘Fh\circ F as a ratio of polynomials. Likewise, for a rational vector field ff,

Lfh(x)=∇h(x)⊤f(x)∈R(Ω).\mathcal{L}_f h(x)=\nabla h(x)^\top f(x)\in\mathcal{R}(\Omega).

Thus rational functions are algebraically compatible with both the Koopman operator and its generator for rational dynamics. A polynomial dictionary is generally not preserved by composition with a genuinely rational transition.

This motivates the function class before finite-dimensional truncation; it does not imply a finite-dimensional invariant subspace or a smaller closure residual in every problem. We added a conditional closure-transfer bound. If g⋆(F(x))=K⋆g⋆(x)g_\star(F(x))=K_\star g_\star(x), sup⁡x∣g−g⋆∣2≤ε\sup_x\\|g-g_\star\\|_2\leq\varepsilon, ∣K−K⋆∣2≤κ\\|K-K_\star\\|_2\leq\kappa, and sup⁡x∣g(x)∣2≤B\sup_x\\|g(x)\\|_2\leq B, then

sup⁡x∣g(F(x))−Kg(x)∣2≤(1+∣K⋆∣2)ε+κB.\sup_x\\|g(F(x))-Kg(x)\\|_2\leq(1+\\|K_\star\\|_2)\varepsilon+\kappa B.

This exposes how observable approximation error and operator-estimation error enter closure. It does not claim that the rational family always has smaller approximation error than an MLP; the direct OOD closure diagnostics show mixed, system-dependent outcomes.

Question 2: Relation to classical Pade approximation

Classical [m/n][m/n] Pade approximation constructs a ratio of degree-mm and degree-nn polynomials whose local series matches the target to the highest order allowed by those degrees. Under suitable analytic assumptions, rational approximants can represent singularity structure and may be useful beyond the disk where a truncated Taylor series is effective. These facts motivate explicit numerator-denominator representations for inverse-power, fractional, and saturating behavior.

The properties that carry over are conditional:

  1. Function-class compatibility. Pole-free rational transitions map pole-free rational observables to rational observables under Koopman composition.
  2. Quotient approximation control. If f=p/qf=p/q, q(x)≥q0>0q(x)\geq q_0>0, ∣N−p∣∞≤δp\\|N-p\\|_\infty\leq\delta_p, and ∣Q−q∣∞≤δq<q0\\|Q-q\\|_\infty\leq\delta_q<q_0, then

\left\\|\frac{N}{Q}-\frac{p}{q}\right\\|_\infty\leq\frac{\delta_p}{q_0-\delta_q}+\frac{\\|p\\|_\infty\delta_q}{q_0(q_0-\delta_q)}.

  1. Conditional closure transfer. The preceding approximation error enters the closure bound through (1+∣K⋆∣2)ε+κB(1+\\|K_\star\\|_2)\varepsilon+\kappa B.

These results connect rational representation to Koopman composition and quantify quotient and closure errors under explicit assumptions. At the same time, we do not claim that classical coefficient matching, convergence rates, pole recovery, or analytic-continuation guarantees transfer automatically. Our branches are neural networks, no Taylor coefficients are matched, and the denominator is constrained positive,

Qψ(x)=ϵ+exp⁡(clip⁡(Gψ(x),−c,c))>0,Q_\psi(x)=\epsilon+\exp(\operatorname{clip}(G_\psi(x),-c,c))>0,

to avoid real encoder poles. Therefore, the connection is strong at the representation and function-class level, but the model is not a classical coefficient-matched Pade approximant.

Revised scope

The denominator-matched enzyme ablation supports the learned denominator in the intended regime. Mixed closure diagnostics, the polynomial Van der Pol mismatch experiment, the long-horizon reversal, and the limited robot stress test bound this conclusion. We describe PadeKoop as a conditional structural inductive bias for denominator-driven dynamics, not a universally superior replacement for standard neural observables.

I understand your explanation and agree with your reasoning, particularly as it reflects a general property of Pade approximations. However, it does not directly address my original concern.

I believe the paper needs a clearer survey of the existing literature and a more thorough comparison with established methods. More importantly, I would expect the proposed method to be evaluated on benchmarks where existing approaches struggle, demonstrating that it can overcome their limitations for the reasons you have outlined.

In this regard, I do not consider my concern to be sufficiently addressed. Therefore, I will maintain my original score.

Reviewer_Zo5ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PadéKoop, a Koopman-based framework that learns rational observables using independent numerator and positive-denominator neural networks for out-of-distribution (OOD) prediction of nonlinear dynamics. The learned observables augment the physical state and are propagated by a linear Koopman operator, with optional spectral control to mitigate error amplification. The paper provides theoretical results on observable well-posedness, approximation error, and rollout-error propagation. Experiments on synthetic physical and biochemical systems report improved OOD prediction over MLP- and KAN-based Koopman baselines, while a small robotic pushing experiment evaluates the applicability of rational parameterizations to measured dynamics. The results also reveal limitations in noise robustness and long-horizon prediction.

Critical strengths
  1. The paper reports improvements over MLP-based and KAN-based observables on several synthetic OOD prediction tasks. The comparisons are conducted within the same Koopman framework, making the results relatively informative.
  2. The authors report unfavorable results in the long-horizon and noise experiments, which helps clarify the limitations of the proposed method.
Critical weaknesses
  1. The methodological novelty is limited. The proposed method mainly replaces conventional neural observables with ratios of two neural networks, while retaining the standard Koopman architecture and training objectives. The distinction from existing Koopman learning and rational neural network methods is not sufficiently established.
  2. The theoretical analysis provides limited support for the proposed method. The three propositions mainly establish basic approximation properties and error bounds, without explaining why rational observables should yield better Koopman-invariant representations or OOD generalization. No direct analysis of Koopman closure errors is provided.
  3. The ablation studies do not fully isolate the contribution of the learnable denominator. In Table 2, the constant-denominator baseline has 4,994 parameters, compared with 9,866 for PadéKoop. Although a parameter-matched MLP is included, the effect of the denominator structure remains insufficiently distinguished from architectural and capacity differences.
  4. The experimental validation is limited. Most experiments use synthetic systems with predefined dynamics, and OOD evaluation mainly considers state-range extrapolation. The robotic experiment contains only two trajectories, and its best-performing model is PadéODE rather than PadéKoop, providing limited evidence for the generalizability of the proposed Koopman framework.
Questions for authors
  1. Can the authors provide a comparison of Koopman closure errors between PadéKoop and the MLP/KAN baselines, particularly in OOD regions? This would help clarify whether the improvements come from better Koopman representations or simply better nonlinear function approximation.
  2. In Table 2, why does the constant-denominator baseline have substantially fewer parameters than PadéKoop? Would the performance advantage remain with a parameter-matched numerator-only baseline?
  3. Table 1 reports only mean values despite mentioning standard deviations. Moreover, the relative improvements differ considerably from those in Figures 2 and 3 under nominally similar settings. Could the authors clarify the evaluation protocols and provide the seed-level results?
  4. In the robotic experiment, PadéODE achieves the best performance, while the evaluation contains only two trajectories. Can the authors provide a more directly matched comparison between MLP- and Padé-observable controlled Koopman models to better support the claimed benefit of PadéKoop on measured dynamics?
3 comments
AuthorsOpenReview ↗

Methodological Scope and Additional Evaluations

Response to Reviewer 2

We thank the reviewer for identifying concerns about methodological novelty, theoretical motivation, baseline coverage, and experimental scope. We added controlled and diagnostic experiments, revised related work, and narrowed our claims.

1--2. Novelty and theoretical support

The linear Koopman transition, decoder, and training objectives are standard. Our contribution is specifically the learned strictly positive denominator inside the observable map,

ϕ(x)=Nθ(x)⊘Qψ(x),Qψ(x)=ϵ+exp⁡(clip⁡(Gψ(x),−c,c)).\phi(x)=N_\theta(x)\oslash Q_\psi(x),\qquad Q_\psi(x)=\epsilon+\exp(\operatorname{clip}(G_\psi(x),-c,c)).

The numerator represents signed feature variation, while the denominator provides state-dependent scaling and saturation without real encoder poles. We claim a conditional structural bias for inverse-power, fractional, and saturating mechanisms, not universal superiority over neural observables.

The Koopman-specific rationale is algebraic compatibility. Let Ω\Omega be forward-invariant, and let F:Ω→ΩF:\Omega\to\Omega and h:Ω→Rh:\Omega\to\mathbb{R} be rational and pole-free on Ω\Omega. Then

UFh=h∘F∈R(Ω),UFR(Ω)⊆R(Ω).\mathcal{U}_F h=h\circ F\in\mathcal{R}(\Omega),\qquad \mathcal{U}_F\mathcal{R}(\Omega)\subseteq\mathcal{R}(\Omega).

For a rational vector field ff, the generator also preserves this class:

Lfh(x)=∇h(x)⊤f(x)∈R(Ω).\mathcal{L}_f h(x)=\nabla h(x)^\top f(x)\in\mathcal{R}(\Omega).

This gives a function-class reason for rational observables before finite-dimensional truncation, but does not imply a finite-dimensional invariant subspace. We also added the conditional closure-transfer bound. If g⋆(F(x))=K⋆g⋆(x)g_\star(F(x))=K_\star g_\star(x), sup⁡x∣g−g⋆∣2≤ε\sup_x\\|g-g_\star\\|_2\leq\varepsilon, ∣K−K⋆∣2≤κ\\|K-K_\star\\|_2\leq\kappa, and sup⁡x∣g(x)∣2≤B\sup_x\\|g(x)\\|_2\leq B, then

sup⁡x∣g(F(x))−Kg(x)∣2≤(1+∣K⋆∣2)ε+κB.\sup_x\\|g(F(x))-Kg(x)\\|_2\leq(1+\\|K_\star\\|_2)\varepsilon+\kappa B.

Thus better observable approximation can reduce closure error conditional on operator estimation; it does not prove that the rational family always has smaller approximation error. Our direct closure results are mixed and system-dependent.

3. Baseline scope, SINDy-PI, and Mechanistic Neural Networks

We added a supplementary model-class diagnostic covering DMD, polynomial and kernel EDMD, NeuralODE, Deep Koopman, Koopman and consistent Koopman autoencoders, and a simplified polynomial-library SINDy implementation. This is not a universal ranking because the methods differ in representation, objective, solver, and parameter count. It includes unfavorable outcomes: NeuralODE is strongest on Coulomb and Michaelis-Menten, and MLP is marginally strongest on LJ. The matched MLP/KAN comparison remains the primary test of observable choice because the rest of the pipeline is fixed.

We now discuss SINDy-PI and Mechanistic Neural Networks explicitly. SINDy-PI targets implicit and rational governing-equation discovery; Mechanistic Neural Networks target explicit mechanistic ODE representations. PadeKoop instead learns observables constrained to evolve linearly in a finite-dimensional latent space. The supplied SINDy diagnostic is a simplified polynomial-library implementation, not an official SINDy-PI reproduction. We label it accordingly and do not claim to have reproduced either method.

4. Synthetic scope, structural mismatch, and robot data

We added a structural-mismatch experiment on the polynomial Van der Pol oscillator,

r˙=v,v˙=(1−r2)v−r,\dot r=v,\qquad \dot v=(1-r^2)v-r,

which lacks the denominator-driven structure motivating our bias. Pooled OOD MSE is 0.200±0.1290.200\pm0.129 for matched MLP, 0.265±0.1710.265\pm0.171 for KAN, and 0.252±0.1170.252\pm0.117 for PadeKoop. PadeKoop wins only 2/5 paired seeds against MLP (p=0.546p=0.546) and 3/5 against KAN (p=0.918p=0.918). This negative result bounds the claim: the quotient is not a reliable advantage when the target lacks the motivating structure.

For measured dynamics, we added a parameter-matched controlled-Koopman comparison and leave-one-trajectory-out stress test, detailed in the response to Reviewer 1. The matched test favors PadeKoop in mean and median but is highly variable and not statistically conclusive. Transfer is asymmetric and both models can fail catastrophically in one direction. We agree that two demonstrations cannot establish broad robotic generalization or closed-loop improvement; the manuscript now states this limitation.

Revised scope

The denominator-matched and parameter-matched controls support the learned denominator in the intended enzyme concentration-OOD regime. The Van der Pol mismatch result, mixed closure diagnostics, long-horizon reversal, and robot stress test bound this conclusion. We therefore present PadeKoop as a conditional inductive bias for denominator-driven dynamics rather than a universally superior Koopman architecture.

Thank you for the rebuttal and additional experiments. Several experimental concerns have been addressed, but two main concerns remain.

  1. Methodological novelty and theoretical justification.

The methodological novelty remains limited. The main modification is the introduction of neural quotient observables into an otherwise standard Koopman framework. The new algebraic compatibility argument motivates rational observables at the function-class level, but does not explain why the proposed parameterization should achieve better Koopman closure than conventional neural observables.

  1. Applicability of the inductive bias.

The authors now characterize PadéKoop as a conditional inductive bias for denominator-driven dynamics, but the applicable class remains insufficiently defined. Inverse-power, fractional, and saturating dynamics cover broad classes of systems, without specifying which structural properties make the proposed parameterization beneficial. The negative Van der Pol results make this distinction particularly important.

What specific structural properties determine the applicability of this inductive bias, and how can they be identified from physical prior knowledge or training data rather than OOD test results? I therefore maintain my rating of 2.

We thank the reviewer for the clarification. We agree that the algebraic compatibility result alone does not prove that a learned quotient has lower Koopman closure error than an MLP. Our claim is conditional: the quotient is a candidate bias when the target dynamics contain a state-dependent scale or rate factor that is naturally represented by a nonconstant, pole-free denominator. We have added a training-only applicability screening and revised the claim accordingly.

1. What structural properties define the applicable regime?

The relevant property is not merely nonlinearity or the existence of a rational formula. The bias is most appropriate when an observable or update law has the form

h(x)≈a(x)/b(x),inf⁡x∈Ωb(x)≥b0>0,h(x)\approx a(x)/b(x),\qquad \inf_{x\in\Omega}b(x)\ge b_0>0,

where the denominator varies with the state and controls a scale, rate, normalization, or saturation. This includes inverse-power factors, normalized or saturating rates such as s/(K+s)s/(K+s), fractional/Hill responses, and bounded domains where the physical denominator stays positive and away from zero. It is less appropriate for numerator-only polynomial dynamics, nearly constant denominators, or denominators approaching zero. The negative Van der Pol result is consistent with this boundary and argues against universal superiority.

2. Why can this affect Koopman closure?

For a rational transition FF and a pole-free rational observable hh, composition keeps h∘Fh\circ F in the same rational function class. This is a compatibility statement, not a strict closure theorem. A closure benefit is expected only when the target is closer to a low-dimensional rational subspace than to the selected numerator-only subspace. In our bound,

sup⁡x∣g(F(x))−Kg(x)∣2≤(1+∣K⋆∣2)ε+κB,\sup_x\\|g(F(x))-Kg(x)\\|_2\le(1+\\|K_\star\\|_2)\varepsilon+\kappa B,

the quotient can help by reducing observable approximation error ε\varepsilon, but it need not reduce ε\varepsilon for every system and does not remove operator-estimation error κ\kappa.

3. Training-only applicability screening.

We evaluated five ID validation splits (32 trajectories of horizon 100 per seed) for Pade, constant-denominator, and parameter-matched MLP models on the enzyme system, using the same states for all three models. These validation trajectories were also used for checkpoint selection, so we treat this as a screening diagnostic rather than an independent confirmatory test. The Pade denominator remained positive and state-dependent: its minimum was 0.533±0.0800.533\pm0.080, first percentile 0.611±0.0810.611\pm0.081, mean 1.529±0.1231.529\pm0.123, coefficient of variation 0.433±0.1310.433\pm0.131, and relative range 3.356±1.0303.356\pm1.030. The ID one-step MSEs were 5.786×10−75.786\times10^{-7} (Pade), 5.818×10−75.818\times10^{-7} (constant denominator), and 5.385×10−75.385\times10^{-7} (matched MLP); decoded closure MSEs were 5.811×10−75.811\times10^{-7}, 5.788×10−75.788\times10^{-7}, and 5.307×10−75.307\times10^{-7}, respectively. Thus, the screen confirms that the learned denominator is nonconstant and well-conditioned, but it does not show that Pade has the lowest ID error. This negative evidence is important: training-only errors cannot be presented as a universal predictor of OOD gains.

Accordingly, our model-selection rule is modest: use the quotient when physical prior knowledge suggests a positive, state-varying scale factor and the training-only screen shows that the denominator is used without approaching zero; otherwise prefer the simpler numerator-only observable. The screening is a diagnostic, not a theorem and not an OOD-label-dependent criterion. The five-seed values above, together with the corresponding seed-level rows, are included in the supplementary material.

We therefore accept the reviewer's distinction. The theory establishes well-posedness, rational-class compatibility, approximation control, and conditional closure transfer, but not uniformly better closure. PadeKoop should be understood as a testable, mechanism-matched inductive bias whose benefit is expected only in the structural regime above; the mixed screening results and the negative Van der Pol result delimit that claim.

Reviewer_Z1dNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method to model dynamical models that follow rational dynamical equations. The paper positions against the Koopman dynamical perspective.

Koopman theory models nonlinear systems through linear evolution with a number of practical methods: DMD, EDMD and learned latent models. Long horizon prediction only requires repeated powers of a linear operator. However, the exact Koopman operator is infinite dimensional; finite dimensional approximation has residuals that propagate making choice of observable important outside training.

MLPs can approximate the functional form of dynamics but it is not clear how the observables should extrapolate.

The idea of the paper is: observables should expose the rational structure of the systems (coulomb, lennard jones etc).

Current learned Koopman models address the observable problem with general encoders or fixed dictionaries. difficulty with denominators in these terms.

The final model Padekoop is a Koopman model with Pade rational observables. It learns a numerator and denominator and the ratio is concatenated with the physical state. The force law is not explicitly specified: instead the observable family is restricted toward behaviour that includes inverse-power structure.

Critical strengths

The model is well described and the paper is clearly written. There are ablation experiments over model size and regularization penalties and for transfer experiments to contact dynamics.

Critical weaknesses

It appears to me that relevant prior work on rational function and more general dynamical modeling is not considered in prior work or compared against; Models such as SINDy-PI or Mechanistic neural networks.

Evaluation is also limited to small scale systems with explicitly known dynamics that follow the rational structure the model is optimized for. There aren’t experiments to show what happens when the data doesn’t follow such a rational structure.

The baselines are also limited: with only MLP and KAN compared against.

Questions for authors

I am missing part of the motivation for the model. why should the observable expose the rational structure of the system? What happens if it doesn’t? How is the observable exposing the rational structure distinct from the dynamics including the rational structure?

Can you explain what you mean by neural encoders not explicitly preferring denominator driven behaviour (lines 58,59)?

Do you have more baselines from the dynamics literature that are relevant and could be compared with? What motivates the choice of the current baseline comparisons?

How does the performance and experiments generalize when the data does not follow the specific rational structure the model is design for?

Others Table 1 doesn’t have standard deviation

1 comment
AuthorsOpenReview ↗

Additional Experiments and Clarifications

Response to Reviewer 1

We thank the reviewer for the careful assessment. We agree that the original version did not sufficiently separate representation quality from rollout error, did not fully isolate the denominator contribution, and did not make the measured-data limitation explicit enough. We added the following analyses and revised the claims accordingly.

1. Koopman closure errors in OOD regions

Rollout MSE alone cannot identify the source of an improvement. We therefore added direct one-step OOD closure diagnostics for PadeKoop, a parameter-matched MLP, and KAN observables. The lifted dimension, Koopman operator, decoder, losses, data splits, and training protocol are fixed; only the observable family changes. Seed-level values, 95% Student-tt intervals, and paired tests are included in the supplementary artifact.

The normalized lifted closure error and decoded closure error are

Elift=∣Pdyn(KE(xt)−E(xt+1))∣22∣PdynE(xt+1)∣22+10−12,Edec=∣D(KE(xt))−D(E(xt+1))∣22.E_{\mathrm{lift}}=\frac{\\|P_{\mathrm{dyn}}(K E(x_t)-E(x_{t+1}))\\|_2^2}{\\|P_{\mathrm{dyn}}E(x_{t+1})\\|_2^2+10^{-12}},\qquad E_{\mathrm{dec}}=\\|D(K E(x_t))-D(E(x_{t+1}))\\|_2^2.

PadeKoop has lower lifted closure error than the matched MLP in 5/9 system/OOD comparisons and lower decoded closure error in 6/9. Relative to KAN, the counts are 9/9 and 7/9. The result is system-dependent: PadeKoop improves both diagnostics on Coulomb and the enzyme concentration shift, while the matched MLP is better in the Argon-LJ near-OOD region and LJ is approximately tied under the decoded metric. We therefore do not claim uniformly smaller closure error.

2. Denominator contribution versus parameter count

The constant-denominator model has fewer parameters because it removes the trainable denominator branch. We added a numerator-only MLP with matched total parameter count:

ModelParametersHorizon-100 MSEPadeKoop wins
Constant denominator4,9940.256±0.1320.256\pm0.1324/5
Matched numerator-only MLP9,8950.377±0.3000.377\pm0.3004/5
PadeKoop9,8660.117±0.054\mathbf{0.117\pm0.054}--

On the enzyme concentration-OOD benchmark, PadeKoop reduces mean error by 54.3% versus the constant-denominator control and by 69.0% versus the parameter-matched MLP. This supports a benchmark-specific contribution from the learned denominator, while not ruling out every optimization interaction. We retain the unfavorable horizon-1000 result, where the matched MLP is better under pure latent rollout without spectral control.

3. Table 1, Figures 2--3, and seed-level results

We corrected an aggregation inconsistency. Table 1 and Figures 2--3 were regenerated directly from archived per-seed CSV files. Table 1 now reports mean and sample standard deviation over five seeds; seed values, 95% confidence intervals, and paired tests are provided in the artifact. Table 1 uses 1,000 training transitions and horizon 100. Figure 2 varies training-set size at horizon 100, while Figure 3 varies horizon at 1,000 transitions. The corresponding cells therefore use the same protocol, but Figures 2--3 show within-system ratios and percentage reductions whereas Table 1 reports raw MSE.

The corrected mean reductions relative to MLP are 97.7%, 31.4%, 46.2%, and 37.4% for Coulomb, LJ, Argon-LJ, and Michaelis-Menten. Paired unadjusted pp-values are 0.298, 0.332, 0.105, and 0.260; all paired 95% intervals include zero. We now describe these as mean trends rather than statistically conclusive advantages. The unfavorable Figure 2 cells and the LJ reversal at horizons 200 and 500 in Figure 3 are retained.

4. Robot evidence and two trajectories

We added a parameter- and learning-rate-matched controlled-Koopman comparison:

ModelParametersRollout MSE, mean ±\pm SDMedian
Matched MLP-observable Koopman1,52718.388±37.85818.388\pm37.8580.221
Pade-observable Koopman1,5320.161±0.1660.161\pm0.1660.084

Both use the same temporal split, learning rate 3×10−43\times10^{-4}, and five seeds. Pade wins 3/5 seeds; the paired tt-test gives p=0.343p=0.343 and the Wilcoxon test gives p=0.438p=0.438. One MLP seed diverges strongly, so we present this as exploratory feasibility evidence, not a statistically conclusive advantage.

We also added a leave-one-trajectory-out test. Training on shoving and testing on pushing gives median MSE 0.193 for the matched MLP and 0.056 for PadeKoop. In the reverse direction, both models can fail catastrophically. PadeKoop is lower in 4/5 fold-averaged seeds, but the paired result is inconclusive (p=0.351p=0.351). We agree that Pad茅ODE leads the broader heterogeneous robot comparison and that two demonstrations cannot establish broad measured-data generalization or closed-loop improvement. The revised paper states this limitation explicitly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.