acceptodds
Under review as a conference paper at ICLR 2027

CompKoop: Compact Observable Completion for Projected Koopman Prediction

Abstract

Observable closure presents a central challenge for long-horizon prediction with finite-dimensional Koopman models. For many nonlinear autonomous systems, a compact observable space cannot support closed linear evolution under a single finite-dimensional operator. A common approach is therefore to make the finite operator state dependent, allowing unresolved nonlinear structure to be absorbed into adaptive operator coefficients. This work investigates whether such missing structure can instead be represented through a compact learned expansion of the observable space while retaining a time-invariant projected mapping. We propose CompKoop, a Koopman framework that learns a compact product expansion of the observable space for nonlinear state prediction with a time-invariant projected readout. For state prediction, CompKoop targets the Koopman image of the physical-state observable rather than requiring closed linear evolution of the entire lifted representation, leading to a projected-closure formulation for multi-step forecasting. We further show that a class of state-dependent projected mappings can be rewritten as fixed mappings over product-expanded observables, establishing a direct connection between adaptive finite operators and observable completion. Experiments on Lorenz63 and a state-dependent Duffing system demonstrate strong long-horizon prediction with a compact parameterization. Additional experiments demonstrate accurate ten-step prediction of POD-reduced two-dimensional Navier–Stokes dynamics. Analysis of independently trained state-dependent Koopman models shows that product expansion captures their adaptive projected increments. Controlled feature probes and native-readout analysis further show that the learned product observables improve the representation of state evolution and contribute dynamical interactions to the trained predictors. These results support observable completion as a structured alternative to absorbing finite-dimensional closure deficiency through operator adaptation.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Koopman learning, nonlinear dynamical systems, observable completion, projected closure, long-horizon prediction, representation learning

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_C21FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the observable-closure problem in finite-dimensional Koopman modeling for nonlinear dynamical systems. Instead of relying on state-dependent Koopman operators to absorb unresolved nonlinearities, it proposes CompKoop, which learns a compact set of product observables while keeping the projected dynamics time-invariant. The method only requires accurate prediction of the physical-state observables, leading to a projected-closure formulation for multi-step forecasting. The paper further shows that certain state-dependent projected mappings can be interpreted as fixed operators acting on product-expanded observables, linking adaptive Koopman models to observable completion. Experiments on Lorenz63, Duffing, and reduced Navier–Stokes systems show that the learned product expansion captures important nonlinear interactions and enables strong long-horizon prediction with a compact parameterization.

Critical strengths

The motivation is very well posed, and the problem formulation is clear and easy to follow. Overall, the paper is written with strong clarity and communicates its main ideas effectively.

The progression from theory to experiments is particularly intuitive and well motivated. The experimental design feels genuinely theory-inspired rather than merely empirical, which strengthens the overall narrative.

Critical weaknesses

The scope of both the analysis and the experiments currently feels somewhat limited. If the theoretical development is made more concrete and comprehensive, the current experimental scale may be sufficient. Otherwise, I think the empirical evaluation would need to be expanded substantially. See the questions below for more specific suggestions.

Questions for authors

Q1. Subsection 4.2 seems to be the core methodological component of the paper, but I found the discussion around this design somewhat underdeveloped. In particular, it is not yet clear why the proposed product structure should be expected to work. The gating term ss is straightforward, and the structure (Ac(x))⊙(Bc(x))(Ac(x)) \odot (Bc(x)) is intuitive as a compact way to introduce multiplicative interactions. However, the paper would benefit from a more explicit explanation of how this construction relates to, or approximates, the preceding qj(x)c(x)q_j(x)c(x)-type terms. Why is this the right structural surrogate, and under what assumptions should it be expressive enough?

Q2. More concretely, each learned observable has the form χi(x)=sid(ai⊤c(x))(bi⊤c(x)),\chi_i(x)=\frac{s_i}{\sqrt{d}}(a_i^\top c(x))(b_i^\top c(x)), which can be interpreted as a learned quadratic/product feature that captures interactions among components of c(x)c(x). This suggests that explicitly modeling interaction terms is an important source of the method's empirical gain. However, it is not yet clear why these interactions are the relevant missing structure for observable completion. I think the paper would be substantially strengthened by either a theoretical argument or a more targeted empirical analysis explaining why this particular interaction structure is an appropriate inductive bias.

Reviewer_Tg9gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CompKoop augments a neural observable dictionary with learned factorized products and predicts physical states through a fixed linear readout with recursive re-embedding. The paper distinguishes projected closure from full Koopman invariance and evaluates the method on two ODE systems and POD-reduced Navier–Stokes dynamics. The distinction between state prediction and full lifted invariance is clearly stated. The reported ODE improvements are promising, particularly on Duffing. Five-seed evaluations, retrained ablations, random-product probes, and native-readout interventions provide useful evidence that the product branch contributes to prediction. Nevertheless, I have several major concerns about the paper, as articulated below. Overall, I lean toward rejection because the methodological advance and the source of the empirical gains remain insufficiently established. The formulation appears sound and the analyses are valuable. My main reservations concern novelty and controlled empirical validation.

Critical strengths

The paper clearly distinguishes projected state prediction from full Koopman invariance and develops a simple, computationally efficient factorized product expansion for enriching learned observables. The ODE results are promising, particularly on Duffing, and the evaluation includes multiple random seeds, retrained ablations, random-product controls, and branch-level analyses. These experiments provide useful evidence that the learned product features participate meaningfully in the state update rather than serving merely as unused additional coordinates.

Critical weaknesses
  1. The novelty beyond established prediction strategies and dictionary expansion is insufficiently established. The paper is motivated by the difficulty of achieving closed linear evolution in a finite-dimensional observable space. However, the main step that circumvents this difficulty is the relaxation to projected closure, together with re-embedding after each state prediction. Closely related prediction strategies already exist in projected EDMD-DL and neural flow-map learning [1]. The paper should clarify what is new in this formulation beyond these precedents.

    The remaining contribution is the learned factorized product expansion. Adding observables enlarges the available representation space, so improved prediction over a smaller dictionary alone provides limited evidence of a distinctive contribution. Although increased expressivity does not guarantee better generalization, the reported improvements need to be distinguished from the benefits of additional capacity. Given prior work on dictionary learning and multiplicative representations [2, 3], what new insight or advantage does this particular construction provide? Comparisons with alternative expansions at comparable capacity, together with accuracy–cost comparisons against full quadratic expansion, would help establish whether the contribution is a useful architectural bias or a computational advantage.

  2. The experiments do not sufficiently isolate the benefit of the multiplicative architecture.

    Capacity-controlled ablations. Removing the product branch in Table 5 simultaneously reduces the feature dimension and parameter count. The resulting performance degradation shows that the branch is useful in the reported model, but does not establish that multiplicative features are preferable to other ways of increasing representational capacity. A more informative comparison would replace the product branch with an equally wide nonmultiplicative feature expansion, while matching the parameter budget, training objective, and optimization setting. The equal-width random-product probe in Table 2 is helpful, but it freezes an encoder that was jointly trained with the learned products. An end-to-end control in which fixed random products, the encoder, and the readout are trained jointly would provide stronger evidence for the value of learning the product directions.

    Comparability of the PKDL baseline. The implemented PKDL baseline is trained using a one-step loss over the full lifted dictionary, whereas CompKoop is trained using a multi-step physical-state prediction objective and only requires projected closure. These objectives impose different requirements and may substantially affect recursive prediction. The paper should also clarify whether PKDL re-embeds the predicted physical state at every step or recursively propagates the predicted lifted coordinates. Therefore, PKDL is a useful overall baseline, but the current comparison does not by itself isolate the benefit of the proposed product architecture. A comparison under the same state-prediction objective and rollout procedure is needed for that claim.

  3. The value of the compact expansion is not established through accuracy–cost comparisons. For the same base dictionary, a full quadratic expansion can represent every proposed factorized product. The relevant question is therefore whether learning a compact subset of quadratic combinations provides comparable or better prediction at lower computational cost. The paper provides complexity expressions, but lacks a direct comparison against full quadratic expansion over the same learned base dictionary, including prediction accuracy, parameter count, training time, and rollout cost. Such evidence is needed to substantiate the practical advantage of the compact construction.

  4. The fluid experiment does not isolate the product branch’s contribution. The POD-DMD and POD-polyEDMD comparisons demonstrate gains over the selected reduced-order baselines. Without a no-product neural predictor operating on the same POD coefficients, however, it remains unclear how much of the improvement comes from the proposed products rather than general neural nonlinear modeling. This control is particularly important because the fluid task is the main evidence beyond the two low-dimensional ODE systems.

References

[1] Jake Buzhardt, C. Ricardo Constante-Amores, and Michael D. Graham. On the relationship between Koopman operator approximations and neural ordinary differential equations for data-driven time-evolution predictions. Chaos, 35(4):043131, 2025.

[2] Qianxiao Li, Felix Dietrich, Erik M. Bollt, and Ioannis G. Kevrekidis. Extended dynamic mode decomposition with dictionary learning: A data-driven adaptive spectral decomposition of the Koopman operator. Chaos, 27(10):103111, 2017.

[3] Grigorios G. Chrysos et al. Π\Pi-nets: Deep Polynomial Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020.

Questions for authors
  1. What is the precise advance over projected EDMD-DL/neural flow-map learning and existing factorized multiplicative networks? Is the main claim a new representation principle, a useful architectural bias, or improved computational efficiency?

  2. Under the same multistep state-prediction objective and comparable parameter budgets, how does CompKoop compare with a residual MLP and a model whose random product factors remain fixed while its encoder and readout are trained jointly?

  3. How does full quadratic expansion over the same learned base dictionary compare in prediction accuracy, parameter count, training time, and rollout cost? At which dimensions and product widths does factorization offer a clear advantage?

Reviewer_QeJhOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CompKoop, a Koopman-based framework that uses learned product observables for nonlinear dynamical prediction. The method relates state-dependent Koopman operators to fixed linear readouts in an expanded observable space, while seeking a compact representation. Experiments on Lorenz63, Duffing, and POD-reduced Navier–Stokes systems evaluate its predictive performance.

Critical strengths

The main strength of this paper is its compact representation with promising predictive performance. CompKoop captures nonlinear dynamics using a small number of learned product features rather than a full polynomial feature library. The model is parameter-efficient and achieves good long-horizon results on Lorenz63 and Duffing. The ablation studies further support the usefulness of the product features.

Critical weaknesses
  1. Theoretical novelty and justification of the Koopman formulation.

The theoretical contribution appears limited. Proposition 1 essentially follows from set inclusion, while Proposition 2 expresses a state-dependent matrix multiplication as a fixed linear mapping over expanded features. The error bound ϵC also follows from a basic matrix norm inequality. These results do not provide meaningful guarantees on the number of product features required, the compactness of the representation, or long-horizon prediction errors. Overall, the analysis appears to reformulate existing ideas rather than establish substantial new theoretical results.

The methodological distinction is also insufficiently explained. CompKoop recomputes nonlinear observables at every prediction step instead of evolving the latent state through a fixed linear operator. The resulting model therefore resembles a nonlinear residual network with factorized quadratic features. Similar product features have been widely explored in polynomial EDMD and related methods. The paper does not sufficiently clarify the difference from conventional bilinear neural networks or the specific benefit provided by the Koopman formulation.

  1. Experimental scope and adequacy of the comparisons.

The experiments mainly consider the low-dimensional Lorenz63 and Duffing systems. Although a two-dimensional Navier–Stokes example is included, forecasting is performed in a 64-dimensional POD space for only 10 steps, and the comparisons are restricted to POD-DMD and POD-polyEDMD. The absence of stronger nonlinear forecasting baselines limits the evidence for the proposed method. Moreover, Lorenz63 already contains quadratic interaction terms, so the effectiveness of product features on this system is not unexpected. The current experiments do not sufficiently establish an advantage on more complex dynamics. Comparisons with stronger nonlinear models and longer-horizon Navier–Stokes forecasts would help assess whether the proposed feature structure provides benefits beyond these relatively favorable settings.

Questions for authors
  1. Could the authors clarify the theoretical insights of Propositions 1 and 2 beyond existing Koopman dictionary learning and polynomial feature expansions? In particular, can the analysis provide more informative guarantees on feature complexity, representation capacity, or long-horizon prediction errors?

  2. Could the authors clarify what distinguishes CompKoop from a residual neural network with factorized quadratic features, given that nonlinear observables are recomputed at every prediction step rather than propagated through closed linear Koopman dynamics? What specific advantage does the Koopman formulation provide?

  3. Could the authors provide comparisons with parameter-matched MLPs, ResNets, or other nonlinear dynamics prediction models? Longer-horizon Navier–Stokes forecasts with stronger baselines would also help assess the advantages of the product-feature structure.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.