acceptodds
Under review as a conference paper at ICLR 2027

Unpaired measure transport with dynamical alignment

Abstract

For decades, numerical models have been central to forecasting and understand- ing physical systems. Yet unresolved scales, simplified physics, and uncertain parameters make them imperfect. Over long trajectories, model states and ob- servations therefore lie on different attractors, described by different invariant measures. Relating these measures is a transport problem for which generative models provide expressive maps. Standard transport objectives, however, rely on geometric proximity and may pair states whose futures are dynamically incompati- ble. We introduce a framework for transporting invariant measures while aligning their dynamics. We define a discrepancy that quantifies how well two systems can be aligned up to an admissible change of coordinates. We then bound the map-dependent commutation residual by a pairwise ground cost independent of the transport map, reducing dynamical alignment to a linear optimal transport problem. From unpaired trajectories, intertwining Extended Dynamic Mode Decomposition (iEDMD) learns this cost by aligning reduced representations of the two dynam- ics. The resulting coupling supervises a Flow Matching model without requiring integration of the transport during training. The learned map can then be wrapped around an unchanged numerical solver in a transport–solve–transport procedure. Experiments on chaotic systems and fluid flows show that the dynamical cost improves conjugation when geometry and dynamics disagree; when the systems already share their coordinates, geometric costs or simpler maps remain preferable.

Primary area:
learning on time series and dynamical systems
Keywords:
Generative modeling, Optimal transport, dynamical systems, Koopman operators

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_RDmeOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies how to relate an imperfect numerical model F to the system it describes, M, when the two are observed only through unpaired trajectories and settle on different invariant measures µ_X and µ_Y. The authors seek a map T that transports µ_X onto µ_Y and approximately commutes with the dynamics, T ∘ F ≈ M ∘ T. When the commutation is exact and T is invertible, T is a conjugacy.

They define a discrepancy D_Ω: the smallest commutation residual achievable by a measure-transporting map in an admissible class Ω. Ω = {Id} recovers the direct forecast error, and richer classes forgive coordinate changes. Their main technical step (Prop. 1) bounds the map-dependent residual by a pairwise cost that does not depend on T: c(x, y) = ‖Ψ(y) − GΨ(x)‖ + ‖Ψ(My) − GΨ(Fx)‖, where Ψ is a feature dictionary and G is a feature-space alignment. Minimizing this bound over couplings is a linear OT problem. G is learned from unpaired transitions with "intertwining EDMD" (iEDMD), which sets K_U = G⁻¹K_V G and fits each system to its own transitions, plus an MMD penalty. Minibatch Sinkhorn plans under this cost supervise simulation-free Flow Matching. The learned map is used for transport–solve–transport forecasting, T(F_t(T⁻¹y₀)), around an unchanged solver.

Experiments cover Lorenz-63, Chen, Rössler and Lorenz-96 pairs under parameter shifts and/or constructed coordinate deformations, cylinder-wake and fluidic-pinball flows with separate POD bases, and a full-state to delay-embedding transport. The iEDMD coupling reduces conjugation error by 43–62% relative to Euclidean OT on deformed ODE pairs and by 32–74% on fluids. It gives little or no gain, and sometimes harm, when coordinates are shared.

Critical strengths
  1. Removing the cost of enforcing dynamical consistency. Enforcing T ∘ F ≈ M ∘ T directly requires integrating and differentiating through the learned transport. Proposition 1 replaces this with a pairwise cost evaluated on recorded states and their successors (or reduced-model predictions, in the operator form), so that minimizing an upper bound on the residual becomes a linear OT problem and Flow Matching stays simulation-free. The two-sided result in Prop. 3 (cross-moment lower bound, successor-cost upper bound) is a helpful certificate. The reported diagnostic places learned-map residuals between the two bounds on all thirteen ODE pairs.

  2. A useful conceptual reframing. The discrepancy D_Ω, with Ω = {Id} recovering ordinary forecast error, makes precise the idea that a model with poor direct forecasts can still be close to the observed system once its states are correctly identified. Figure 1 illustrates this cleanly: under an exact nonlinear conjugacy, direct error and W₂ grow while the residual over residual MLPs stays at the fitting floor. A genuine change of dynamics (ρ from 28 to 40) raises it.

  3. Learning the alignment without pairs. Defining K_U := G⁻¹K_V G builds the intertwining relation into the parameterization, so G can be learned from each system's own transitions without any correspondence. This is a sensible extension of EDMD to two systems.

  4. The paper explicitly reports cases where transport hurts (Lorenz-96, native Rössler, longer-horizon parametric wake) and where simpler maps win (affine on Chen b, GW + Procrustes on pure rotation).

  5. On the structural wake, where the source misplaces the cylinder by two diameters and the POD bases are unregistered, conjugated forecasting reduces quarter-period error from 0.325 (true state in the wrong model) to 0.074. Euclidean coupling visibly fails: its conjugation error is 0.441 vs. 0.114 for iEDMD (Fig. 2), and its pushforward contracts inside the target cycle (Fig. 4).

Critical weaknesses

1. The contribution of the dynamical (successor) term is not isolated. The paper's central claim is that pairing by dynamical compatibility, not proximity, drives the gains. The cost has two parts: an endpoint term ‖Ψ(y) − GΨ(x)‖ (OT in G-aligned coordinates) and a successor term ‖Ψ(My) − GΨ(Fx)‖. The paper's own evidence suggests much of the gain may come from the learned alignment G rather than from the successor term: i) On ring-deformed Lorenz-96, K = 1 orbit coupling (no successor term) gives 3.892 vs. 3.883 for the full statewise cost. The authors state there is no measurable contribution from the successor term on this pair. ii) On the structural wake, adding raw successors to the Euclidean cost barely changes the result (0.441 → 0.429), while a learned linear intertwiner recovers about a third of the gap (0.305).

An ablation of endpoint-only cost vs. full cost, with the same G, on all ODE and fluid pairs is necessary. Without it, the method may be mostly "learn a nonlinear alignment, then do Euclidean OT in aligned space." That is a substantially weaker claim than the one made.

2. Missing related work and baselines on OT with transformation invariances. Aligning unpaired point clouds up to a transformation class is an established problem (https://arxiv.org/abs/1806.09277, https://arxiv.org/abs/1805.11222, https://proceedings.mlr.press/v97/bunne19a.html). Structurally, D_Ω is the same kind of object, a minimum over transformations in a class jointly with a marginal-respecting coupling (cf. Eq. 21), but with a dynamical residual in place of a geometric cost. None of these works are cited or compared. The paper does show that geometric invariance is insufficient under nonlinear deformations: GW recovers only 14% of the known correspondence, and FM trained on GW plans reaches 9.02 vs. 0.87 for oracle pairs. However, the only invariance-based baselines are GW and rigid Procrustes alignment, which are mismatched to nonlinear deformations by construction.

3. Baselines are weakest where the method wins. The headline gains (43–62% on deformed ODEs, 32–74% on fluids) are relative to Euclidean OT in unregistered coordinates, where proximity is meaningless by construction. Euclidean coupling sometimes performs worse than random pairing. The dynamical-systems literature the paper cites offers stronger natural comparisons that are not evaluated (https://epubs.siam.org/doi/10.1137/17M116207X, https://link.springer.com/chapter/10.1007/978-3-030-01228-1_8).

In addition, on the wakes the Euclidean baseline uses ε = 0.5 × median cost, while iEDMD uses ε = 0.02 × mean cost. The authors partly control for this (iEDMD at the matched scale gives 0.161 vs. 0.441), but the main table should use a common protocol.

4. Coupling-only training is not the strongest variant, and the simulation-free benefit is not quantified. On deformed Lorenz-63, random pairing with the cost used as a generated-state penalty beats coupling-only iEDMD (2.06 vs. 2.995). The solver-in-the-loop reference is better still on most pairs (Table 11). The case for the coupling therefore rests on the training-cost savings of staying simulation-free, but beyond a qualitative count of integrations per regularized batch, the paper does not compare wall-clock or compute costs of coupling-only, penalized, and solver-in-the-loop training.

5. Evaluation is restricted to constructed, noise-free settings. The motivation is a model compared with real observations: noisy, partial, short records, and unknown observation operators (Gulf Stream separation, initialization shocks). All experiments are noise-free. The limitations section acknowledges this, but at least one experiment with observation noise or short records would substantially strengthen the claims.

6. The forecasting comparison omits standard alternatives. The only non-transport baseline is the true state fed into the wrong model, plus nearest-neighbour and known-translation controls on one wake. Learned model-error correction (https://rmets.onlinelibrary.wiley.com/doi/full/10.1002/qj.4116) and simple statistical bias correction are natural competitors for "making an imperfect model more useful," and neither is evaluated. As the authors acknowledge, transport also degrades accurate models (Lorenz-96: 0.488 direct vs. 1.230 conjugated; native Rössler a: 0.070 vs. 0.131; parametric wake: 0.118 vs. 0.204 at 1T).

Questions for authors

Please see "Critical Weaknesses" section of the review above. The paper is interesting and approaches the problem in a different way. I am keen to see the authors' response and am eager to revise my rating if the response addresses all my concerns.

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

The paper considers unpaired transport between dynamical systems, where matching states based only on geometric distance can give the wrong correspondence. The main idea is to use dynamics to define the OT cost that iEDMD learns an alignment from unpaired trajectories, this alignment gives a dynamics-aware coupling, and Flow Matching then learns the final transport. The learned map can also be wrapped around an existing simulator for forecasting. The method works particularly well when the two systems are expressed in misaligned coordinates or have structural mismatch, while the gains are smaller when the coordinates are already aligned.

Critical strengths

Using dynamics rather than Euclidean distance to decide which unpaired states should correspond makes sense, and the connection to the commutation error gives the proposed cost a clear justification rather than making it just another heuristic matching loss.

Also the paper does not claim that dynamics-aware transport always helps: when the systems already share coordinates, simpler mappings can be just as good or better. I found this useful for understanding when the method is actually needed.

Critical weaknesses

The advantage is very clear after coordinate deformation or structural changes, but much less consistent when the two systems are already expressed in comparable coordinates. I am therefore somewhat unsure how often the harder setting considered here arises in practice. Also since the final OT cost depends on the learned G, I would like to better understand whether different equally good iEDMD solutions can lead to different transport maps.

Questions for authors

When should one choose not to apply the transport? Some shared-coordinate cases get little benefit or even degrade.

How sensitive is the method to trajectory quality and amount of data, or say noisy data? I think a experiment here might help.

Why the primary iEDMD coupling uses epsilon=0.02 times the mean cost, while the Euclidean baseline uses 0.5 times the median cost?

Have the authors considered Gromov Dynamic Time Warping (Cohen et al., AISTATS 2021) as a baseline? It seems particularly relevant since it also uses temporal structure to align trajectories living in incomparable spaces.

A comparison with dynamical similarity analysis measure will be better. Also a ablation separating the two terms in the iEDMD cost, learned alignment G itself versus the successor term.

Reviewer_L58MOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for modeling observed (target) dynamical systems by transporting known model (source) dynamics via a learnable invertible map TT, based on the notion of conjugacy. Under exact conjugacy, the target and source systems are dynamically equivalent up to a coordinate transformation, allowing the target dynamics to be reliably predicted by transferring the source dynamics through the learned map.

To bypass the computational cost of directly minimizing the conjugacy commutation residual, the authors formulate the optimal transport (OT)-based surrogate objective; rather than learning the conjugacy in the state space, the authors first introduce a feature-space alignment map GG, learned through the iEDMD method inspired by Koopman operator theory. This alignment is then used to construct the transport cost that provides an upper bound on the conjugacy residual and enables a Kantorovich OT formulation without explicitly evaluating TT. After obtaining the probabilistic coupling between the source and target distributions by solving the OT problem, the authors employ flow matching to construct a deterministic, invertible transport map TT.

The authors evaluate their framework on various dynamical systems, including chaotic oscillators and fluid dynamics. The results demonstrate that the proposed approach can effectively align the target dynamics with the source model, particularly when the two systems are related by a coordinate transformation. Additional experiments compare the proposed iEDMD-based alignment with random coupling and Euclidean OT, demonstrating the benefits of incorporating dynamical information into the transport cost.

Critical strengths

Conjugacy-based learning provides a promising approach to dynamical systems modeling, as it allows the topological/dynamical structures of a source system to be transferred to a target system straightforwardly via a learnable diffeomorphism. While learning a sufficiently expressive diffeomorphism (such as a complex neural ODE) can be computationally expensive, the authors address this challenge through a motivated surrogate formulation combining OT and Koopman operator theory. Overall, I find the problem well-motivated and the proposed combination of operator theory, OT, and flow matching technically interesting.

The authors also conduct extensive experiments on relatively complex dynamical systems, including chaotic oscillators and fluid dynamics, demonstrating the empirical effectiveness of the proposed approach.

Critical weaknesses

W1. The proposed algorithm involves several steps in constructing a surrogate for conjugacy learning. In particular, the successor bound establishes that the proposed OT-based objective provides an upper bound on the conjugacy commutation residual. Although the authors discuss the tightness of this bound, the overall approach remains an indirect surrogate for learning conjugacy.

Therefore, I believe that a direct comparison against baseline methods that explicitly optimize conjugacy residuals, in terms of predictive accuracy, conjugation error, and computational cost, would provide important evidence supporting the central motivation of this work. I could not find such a comparison in the main text, making it difficult to assess the practical advantages and trade-offs of the proposed surrogate formulation. I have not checked the Appendix, so I would appreciate clarification if such experiments have already been conducted.

In particular, several previous studies have explored learning conjugacy diffeomorphisms by directly minimizing trajectory-level conjugacy residuals [1] or enforcing vector-field-level pushforward relations [2]. Although these studies do not appear to benchmark systems as complex as those considered in the present paper, they offer closely related methodological alternatives. Thus, a comparison with a baseline that directly minimizes the conjugacy residual would help demonstrate the advantages of the proposed surrogate and better distinguish its methodological contributions from prior work.

[1] Sagodi, A., & Park, I. M. (2025). Dynamical archetype analysis: Autonomous computation. arXiv preprint arXiv:2507.05505.

[2] Friedman, R., Moriel, N., Ricci, M., Pelc, G., Weiss, Y., & Nitzan, M. (2025). Characterizing nonlinear dynamics via smooth prototype equivalences. arXiv preprint arXiv:2503.10336.


W2. A central motivation of the paper is to learn transport maps that preserve invariant measures with respecting dynamical compatibility. The authors specifically emphasize that "Standard transport objectives, however, rely on geometric proximity and may pair states whose futures are dynamically incompatible."

However, much of the empirical evaluation focuses on numerical errors in predicted states, such as normalized field RMSE and conjugation error. Although these metrics are important for assessing forecasting accuracy and approximate commutation, they do not necessarily establish whether the underlying topological or dynamical structures are preserved. The attractor-distance analysis in Figure 2 provides some relevant evidence, but a more systematic investigation of dynamical invariants would strengthen the paper.

For example, evaluating the preservation of equilibrium structures and their stability, invariant sets, or basins of attractors could provide more direct evidence of the claimed dynamical compatibility. This seems important because a small yet nonzero conjugacy error does not generally guarantee the persistence of invariant structures without additional stability assumptions.

Questions for authors

Q1. Would it be possible to compare the proposed method against direct commutation residual minimization? Since the framework assumes access to observed target trajectories and the transport map TT is an invertible diffeomorphism, direct trajectory-level residual minimization appears feasible. For example, given consecutive target observations (y,y+)(y,y^{+}), one could directly minimize E(y,y+)∼DY[∥T∘F∘T−1(y)−y+∥2]\mathbb{E}_{(y,y^+) \sim \mathcal{D}_Y}[\Vert T \circ \mathcal{F} \circ T^{-1} (y) - y^+ \Vert^2]. This formulation does not require paired source-target trajectories or explicit access to the target system. A comparison in terms of conjugacy, forecasting performance, and cost would help quantify the practical advantage of the proposed surrogate formulation.


Q2. Could the authors provide additional evaluations of the preservation of qualitative dynamical structures, e.g., the correspondence of equilibria, stability, invariant sets, basins of attractors, ..., when applicable?


Very minor comment: The title listed on OpenReview differs slightly from that in the submitted PDF.

Comments on the paper

3 comments

We thank the three reviewers for their time and their comments. the revised manuscript is now uploaded and we answered each of you in your own thread, so this message is only a short summary.

what the paper does. to our knowledge it is the first work that turns the dynamical alignment of two systems, known only through unpaired trajectories, into a pairwise optimal transport cost. compared with a standard Flow Matching training it adds a single stage: we train an iEDMD model, and its cost gives the coupling through a linear transport problem. Training stays simulation-free, and since the dynamics sit in the coupling and not in the loss, Flow Matching keeps its original objective (Lipman et al., 2023), so the marginals are transported as usual.

positioning. the paper is first of all about measure transport: mapping the invariant measure of a model onto that of the observed system, without pairs, as in unpaired domain translation. generative transport methods (Flow Matching, OT-CFM) and OT with invariances do transport the measure, but they pair states by geometric proximity. the conjugacy methods cited in the reviews go the other way: they align the dynamics with a prototype system, from a trajectory residual or from velocities, and their objective is that residual, not the transport of a measure. similarity scores such as DSA compare dynamics but return no map. we do both at once: the measure is transported exactly as in Flow Matching, and the dynamical alignment enters through the transport cost.

update of the manuscript. most of your points led to new experiments, now in appendices J to O (direct minimisation of the residual, invariant structures, ablation of the cost, noise and short records, compute cost). several answers were already in the submitted appendices and are now pointed to from the main text. as announced to reviewers Mdmh and RDme we also added a realistic experiment (Sec. 4.3): a quasi-geostrophic model of the north atlantic against a finer one observed only through its sea surface height, where the iEDMD coupling still improves on a euclidean coupling defined on the same physical field.

we believe the revised manuscript addresses the main concerns, and we invite you to have a look at it and to reconsider your assessment. thanks again.

Reviewer_MdmhOpenReview ↗

Reviewer Followup

Thank you for the detailed response and additional experiments. Most of my original concerns have been addressed.

I have a remaining question: Appendix H.6 uses the same learned gg and KVK_V to select among identity, affine, Euclidean, and iEDMD maps. Would this include the bias for the metrics?

Reviewer_MdmhOpenReview ↗

Reviewer Comment

Thanks for your response. I decide to keep my positive rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.