acceptodds
Under review as a conference paper at ICLR 2027

Disentangling Topology from Measure: Certified Koopman Spectral Learning

Abstract

Data-driven Koopman models learn finite-dimensional representations of nonlinear dynamics from trajectories sampled under a particular state-visitation distribution, yet deployment may change which regions of state space are visited while leaving the underlying dynamics fixed. Standard smooth dictionaries entangle this distribution-dependent fitting geometry with the transition structure of the dynamics inside the same Galerkin matrices. We ask which representation structures can separate these two sources of variation and remain reusable and certifiable under sampling shift. We show that measure-universal Gram diagonality is equivalent to pairwise disjoint support; on a compact connected state space, a dense continuous dictionary cannot retain this property, while indicators of refining generating partitions provide a canonical realization (Theorem 1). This representation-level separation yields the Topological Koopman Certificate (TKC), with non-asymptotic residual and pseudospectral guarantees (Theorems 2,3), a TV-linear operator-perturbation bound with sparse support-preserving reweighting under measure shift (Theorem 4), and an Krylov pseudospectral solver in place of a dense generalized eigenvalue computation. We turn the structural principle into a data-driven construction through Generative Koopman Decoupling (GKD, Algorithm 1), which uses either an analytic or learned preimage oracle to form forward–backward itinerary partitions. Proposition 1 links oracle error to partition resolution in the uniformly expanding regime. Controlled low-dimensional systems evaluate the certificates under matched assumptions, while Lorenz-63 and Hénon test the learned representation and oracle mechanism empirically beyond that regime. On Lorenz-63, the learned-itinerary representation attains the lowest in-distribution prediction MSE in our comparison using only about occupied itineraries; on Hénon, learned- and analytic-oracle constructions give statistically indistinguishable predictions.

Primary area:
learning theory
Keywords:
Koopman operator learning, symbolic dynamics, representation learning, spectral learning, distribution shift

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_a7aROpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies the problem of learning Koopman operator when the sampling distribution of the data changes under a fixed dynamics. This paper builds on symbolic EDMD, known as SEDMD which is developed by Kennedy et al., 2024; 2026. In there, Kennedy et al., 2024; 2026 uses indicator functions of the cells of a refined partition as the dictionary. The resulting Gram matrix (G) is diagonal and holds only the cell masses, and the cross-covariance (A) is nonzero only for cell pairs connected by the dynamics. In the paper, the authors describe this as separating the measure (in G) from the topology (zero pattern of A). In Theorem 1, the paper shows that a dictionary has a diagonal Gram matrix under every sampling measure when its functions have disjoint supports. The paper then proposes the Topological Koopman Certificate. Theorems 2 and 3 provides residual and pseudospectral inclusions, and Theorem 4 bounds how much the finite operator matrix changes when the sampling measure shifts. When no partition is known analytically, the paper also proposes Generative Koopman Decoupling which constructs the partition from data. In this case, Proposition 1 bounds the resulting cell size for uniformly expanding maps.

Critical strengths

The paper investigates an important problem in Koopman learning, i.e., the learned model is evaluated under a different state distribution than the one used for training. The paper frames this as a question of which parts of the learned model depend on the sampling measure, and this is a useful angle. The paper extends symbolic EDMD, known as SEDMD, to unknown partitions. The resulting models are sparse and cheap to train. The paper also provides empirical validation to justify its results.

Critical weaknesses

(1) It seems several components presented as new already appear in Kennedy et al. (2026). It seem that the new contributions appear to be the K-means seed partition, the certificate, and the per-cell affine enrichment. (2) Figure 1 attributes unstable operators and spectral pollution to changing the measure. To my knowledge, however, spectral pollution comes from a projection artifact as one is projecting the Koopman operator onto a finite-dim subspace which is not invariant under the Koopman operator. This happens regardless of whether the measure is shifted or not. (3) In page 3 C2, it says "the indicator basis of a generating partition is the unique class satisfying measure-universal Gram diagonality under... ". However, in Theorem 1 (c), it is not obvious where the uniqueness is proved. Is it an over-claim or did I miss the proof?

Minor issue: In Figure 1, Lterinary should be itinerary. In Figure 7, the caption says panels are “deferred to the journal version.” without referring where to find them.

Questions for authors

(1) Beyond the K-means seed partition and the certificate, what is new relative to Kennedy et al. (2026)? (2) Is the itinerary labeling in Algorithm 1 Step 3 causal at test time? or does it require future iterates of the test point? And how is it computed on Lorenz-63? Is the true iterates T(x)T(x) and T2(x)T^2(x) used? (3) How does the term λmin⁡((P~N−z)∗(P~N−z))\lambda_{\min} ( (\tilde{P}_N-z)^* (\tilde{P}_N - z )) bound the true residual required by Theorem 3

Reviewer_ZxP1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies Koopman/EDMD learning when the sampling distribution changes but the dynamics TT stay fixed. It shows that the EDMD Gram matrix GG is diagonal for every sampling measure μ≪λ\mu\ll\lambda iff the dictionary functions have pairwise disjoint supports (Thm. 1). It therefore uses indicators of refined generating partitions (the SEDMD estimator of Kennedy et al.). With this basis, G=diag(μ(Bi))G=\mathrm{diag}(\mu(B_i)) and the sparsity pattern of AA is the cell-transition graph. On top of this estimator the paper proposes three things: a certificate (TKC) that combines ResDMD-type residuals with a measure-distortion term ΔN\Delta_N (Thms. 2–3); a TV-linear bound on how the operator matrix changes under measure shift, with a sparse re-weighting update (Thm. 4); and a sparse Lanczos pseudospectral solver. The algorithm GKD builds cells from K-means seeds plus forward/backward itineraries, using an exact or MLP-learned inverse map. It then fits a local polynomial model in each cell. Prop. 1 links inverse-oracle error to cell diameter for expanding maps. The experiments cover 1D/2D toy maps (mostly in the appendix), one-step prediction on Lorenz-63 under zz-biased subsampling (TV ≈ 0.88) against EDMD-RBF and Deep Koopman, and a comparison of learned and exact oracles on the Hénon map.

Critical strengths
  1. Separating what a learned Koopman model inherits from the sampling distribution from what is fixed by TT is a useful angle on covariate shift with fixed dynamics. With disjointly supported dictionaries the support of AA does not depend on μ\mu. The transition graph can therefore be reused under a new sampling measure, and only the weights need re-estimating, at cost O(M′+nnz(A))O(M'+\mathrm{nnz}(A)).
  2. The estimator is block-diagonal and interpretable, and it trains in under 0.1 s on a CPU. The paper states that Lorenz-63 lies outside the theorem regime, and it reports where certification breaks down (depth 4 on the Bernoulli map). These strengths are modest compared with the weaknesses below.
Critical weaknesses
  • EDMD with indicator functions is Ulam's method (Ulam 1960; Li 1976; Ding & Li 1991, which the paper cites only for μ^N\hat\mu_N; Dellnitz & Junge 1999; Froyland; Klus et al. 2016) and the standard Markov-state-model construction, yet this literature is not discussed. Thm. 1(a) is a one-line measure-theory fact. Indicators are not the "unique" class: any disjointly supported family (e.g. hi1Bih_i\mathbf 1_{B_i}, or continuous bumps with disjoint supports, which exist on [0,1][0,1]) is Gram-diagonal for every μ\mu. The proof of Thm. 1(b) in Appendix A only uses that a nonempty open set contains two disjoint nonempty open subsets. This also holds on Cantor sets, which contradicts the claim (Thm. 1(b), L2) that connectedness is essential.

  • On a compact space a uniformly expanding map cannot be invertible, so T−1T^{-1} and ∥P−T−1∥∞\lVert P-T^{-1}\rVert_\infty are not well defined, and backward labels do not refine cells. Counterexample: take the doubling map with seeds [0,12)[0,\tfrac12) and [12,1)[\tfrac12,1). For either inverse branch, seed(Pj(x))\mathrm{seed}(P^j(x)) is constant. With ϵ=0\epsilon=0 and Nf=0N_f=0 the cell diameter therefore stays D0D_0, while the bound gives cρ−NbD0→0c\rho^{-N_b}D_0\to 0. An MSE-trained inverse learns the average of the preimages (y/2+1/4y/2+1/4 here), which only repeats seed(x)\mathrm{seed}(x). Appendix H repeats the same sketch and then uses the inverse network's test MSE in place of the sup-norm error ϵ\epsilon.

  • Only the support of AA is measure-free; the entries of K^\hat K still change with μ\mu, as is well known for Markov state models (e.g. Nüske et al. 2017). ΔN\Delta_N needs the unknown μtrue\mu_{\rm true}, and Appendix E replaces it by a self-consistency surrogate plus a concentration term with unspecified constants. Thms. 2–3 also contain O(ΔN2)O(\Delta_N^2) terms. Thm. 4's constant NK/μmin⁡+N∥A∥F/μmin⁡2\sqrt{NK}/\mu_{\min}+\sqrt{N}\lVert A\rVert_{F}/\mu_{\min}^{2} makes the bound vacuous at the TV ≈ 0.88 used in the experiments. Even for the symmetric Bernoulli shift, Table 4 needs 3×1073\times 10^{7} independent blocks to bring CNΔNC_N\Delta_N below 0.5 for just 8 cells. Thm. 4 also requires both measures to charge every cell, which excludes the hard OOD case. Finally, the "validation" of Thm. 3 in Fig. 5 checks EN−ϵ~N≥0\mathcal E_N-\tilde\epsilon_N\ge 0, which holds by the definition of EN\mathcal E_N.

  • In Table 2, EDMD-RBF is about 1000× more accurate in-distribution (5.6×10−45.6\times 10^{-4} vs. 0.61) and about 80× more accurate OOD (0.062 vs. 5.12) than cell-mean SEDMD. SEDMD is "most stable" only under the OOD/in-dist ratio, which rewards a poor in-distribution fit. EDMD-RBF's ratio (190±227190\pm 227) is too noisy to rank, and calling the gap "the empirical signature of Theorem 4" is unjustified. The abstract's claim that the learned itinerary has "the lowest in-distribution MSE" holds only because EDMD-RBF is left out of Table 3, although the data and protocol are the same (5.6×10−45.6\times 10^{-4} vs. 0.025 in-dist; 0.062 vs. 0.52 OOD). The learned itinerary also does not improve the OOD ratio (19.2 vs. 14.6 and 13.0). On Hénon (Appendix K, Fig. 9), GKD's OOD ratio is about 53, versus 4.8 for EDMD-Fourier and 5.5 for Deep Koopman, and Deep Koopman has the lowest absolute MSE. The main text reports only the agreement between the two GKD variants.

  • The Lorenz and Hénon experiments report only one-step prediction error, with no spectrum, pseudospectrum or certificate. The "certified spectral learning" claims therefore rest on toy 1D/2D maps. The gain over the uniform grid may simply be a bias–variance effect (about 120 vs. 9 training points per cell for a per-cell affine fit) rather than itinerary structure. Missing baselines include K-means (K≈25K\approx 25) plus local affine without itineraries, polynomial EDMD, and kNN. It is also not said how OOD points in cells without training data are predicted.

Questions for authors
  1. How is T−1T^{-1} defined for a uniformly expanding (hence non-invertible) map, and why should backward labels refine cells? Please check the doubling-map counterexample in the weaknesses. If the bound only holds in NfN_f, what remains of the "derived optimal depth" and of the learned backward oracle?

  2. Appendix B certifies the full residual, including the cylinder-mass terms. However, Sec. 2.2 and step 5 of Algorithm 1 compute the smallest eigenvalue of H(z)=(P~N−zI)∗(P~N−zI)H(z)=(\tilde P_N-zI)^{\ast}(\tilde P_N-zI), which is only the residual after projection onto the dictionary span. Unless every T−1BiT^{-1}B_i is a union of cells, this can be strictly smaller than the certified residual. Are the plotted pseudospectra then actually covered by Thm. 3? Also, on L2(μ)L^2(\mu) the spectrum of KK is the whole closed unit disk for the non-invertible maps used in Fig. 5, and typically the whole unit circle for invertible chaotic maps. What does the certificate tell us in these cases?

  3. Appendix E computes ΔN\Delta_N as a self-consistency surrogate plus a concentration term whose constants depend on unknown mixing parameters. How are these constants set in the experiments, and in what sense is the result a valid upper bound? Which measure is μtrue\mu_{\rm true} when both the training and test measures are non-invariant re-weightings?

  4. Why is EDMD-RBF excluded from Table 3 and from the "lowest in-distribution MSE" claim? Why are the Hénon baselines (Fig. 9) not reported in the main text? How are OOD points that land in cells without training data predicted?

  5. State exactly what is new beyond SEDMD. In particular, Kennedy et al. (2026) already show that SEDMD works without knowledge of the invariant measure, for any sampling that charges all open sets, and they propose data-driven cylinder-set estimation. Report absolute OOD MSE as the main metric, and add the missing baselines listed in the weaknesses.

  6. Besides Ulam's method and Markov state models, the authors may consider discussing the following. Nüske et al. (2017) and Wu et al. (2017) analyze and correct, by reweighting, the bias of indicator-basis (MSM) and Koopman estimates under non-equilibrium sampling; this is essentially the same "fixed dynamics, shifted sampling measure" problem. Xu et al. (2025, SDMD) study data-driven Koopman semigroup estimation, and its follow-up Reinforced SDMD (Xu et al., 2025) emphasizes that Koopman spectral estimates depend strongly on where the data are collected and learns where to sample. Korda et al. (2020) compute Koopman spectral measures, including continuous spectrum, for measure-preserving ergodic systems; all test systems here are mixing. Akindji et al. (2026; see also Slipantschuk et al., 2020) show that smooth-basis EDMD converges exponentially fast to Koopman eigenvalues for analytic full-branch maps, on a suitable function space rather than L2L^2; this is exactly the regime of Prop. 1 and the 1D experiments. Nüske et al. (2023) give finite-data error bounds for EDMD under ergodic or i.i.d. sampling, the natural comparison for the claimed computable certificate.

  7. Besides Ulam's method and Markov state models, the authors may consider citing and briefly discussing:

    • Nüske et al., "Markov state models from short non-equilibrium simulations—Analysis and correction of estimation bias," J. Chem. Phys., 2017.
    • Wu et al., "Variational Koopman models: Slow collective variables and molecular kinetics from short off-equilibrium simulations," J. Chem. Phys., 2017.
    • Xu et al., "ResKoopNet: Learning Koopman representations for complex dynamics with spectral residuals," ICML (PMLR), 2025.
    • Korda et al., "Data-driven spectral analysis of the Koopman operator," Appl. Comput. Harmon. Anal., 2020.
    • Akindji et al., "Convergence properties of dynamic mode decomposition for analytic interval maps," Comm. Pure Appl. Math., 2026.
    • Slipantschuk et al., "Dynamic mode decomposition for analytic maps," Commun. Nonlinear Sci. Numer. Simul., 2020.
    • Nüske et al., "Finite-data error bounds for Koopman-based prediction and control," J. Nonlinear Sci., 2023.
Reviewer_P7ssOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses a fundamental challenge in Koopman operator learning: the entanglement of system topology (underlying dynamics TT) and measure (data sampling distribution μ\mu) within Gram matrices, which often leads to spurious eigenvalues and out-of-distribution (OOD) instability. The authors prove that using basis functions with disjoint supports is a necessary and sufficient condition to decouple topology from measure (Theorem 1). To realize this in a data-driven manner, they propose Guided Dynamic Partitioning (GKD, Algorithm 1), which constructs fine-grained cell bases via reverse and forward itinerary labeling. Furthermore, they establish Topological Koopman Certificates (TKC, Theorems 2–4) that guarantee geometric stability and lower-bound operator approximation errors in O(KN)O(KN) time using Krylov-subspace methods. The theoretical guarantees and OOD stability are evaluated primarily on low-dimensional benchmarks, including the Lorenz-63 system (d=3d=3) and 1D/2D maps.

Critical strengths

I think there are three main strengths in this paper:

  1. Strong and Elegant Theoretical Formulation: The mathematical framework rigorously isolates topological properties from sampling density. Theorem 1 provides an insightful necessary and sufficient condition via disjoint-support cell bases, and Theorems 2–4 establish solid geometric certificates for OOD stability.

  2. Algorithmic Clarity and Computational Efficiency: The complex theoretical insights are translated into a clear, intuitive, data-driven algorithm (GKD). By employing itinerary labeling and Krylov methods, the method achieves an efficient O(KN)O(KN) complexity, making the implementation transparent and computationally lightweight.

  3. Clear Conceptual Validation on Toy Benchmarks: The paper provides clean and visually intuitive experiments on low-dimensional systems (1D/2D maps and Lorenz-63) that clearly illustrate the mechanism of cell partitioning and confirm the core theoretical assertions under controlled settings.

Critical weaknesses

If I were to point out its drawbacks, I would consider the following three points to be the main weaknesses of this paper.

  1. Lack of Robustness Analysis and Unstated Limitations: The framework heavily relies on a deterministic reverse oracle fθ≈T−1f_\theta \approx T^{-1} and identical dynamics under pure measure shift. However, the manuscript lacks a sensitivity/breakdown analysis for scenarios where these assumptions fail—such as irreversible/non-injective dynamics (e.g., convergence to fixed points or limit cycles where T−1T^{-1} is multi-valued), observation noise, or system parameter variations (concept drift). Furthermore, the fundamental limitation of cell-based partitioning regarding the "curse of dimensionality" in high-dimensional state spaces (d≫3d \gg 3) is not explicitly discussed as a Limitation in the main text.

  2. Misleading Performance Metrics in Table 2: In Table 2, the authors claim superiority in OOD stability based on a low "OOD ratio" (MSEOOD/MSEinMSE_{OOD} / MSE_{in}). However, in terms of absolute prediction error, the proposed method's OOD MSE (5.10) is significantly worse than existing baselines like EDMD-RBF (0.062) and Deep Koopman (0.44). The low ratio of the proposed method is primarily an artifact of its large In-distribution MSE (0.35), which inflates the denominator. Highlighting the ratio alone to claim superior OOD performance is misleading to readers who prioritize absolute predictive accuracy.

  3. Narrow Experimental Scope and Artificial OOD Protocol: The empirical evaluation in the main text relies heavily on Lorenz-63 (d=3d=3), omitting standard high-dimensional Koopman benchmarks (e.g., fluid dynamics or high-dimensional control systems). Additionally, the OOD evaluation in Lorenz-63 is restricted to a synthetic zz-axis sampling bias under fixed parameters (ρ=28\rho=28). While aligned with Theorem 4 (pure measure shift), this protocol is highly artificial and does not reflect realistic distribution shifts encountered in practice.

Questions for authors

To help solidify my evaluation of this paper, I would appreciate it if the authors could address the following three questions.

  1. Sensitivity to Non-Injective Dynamics and Oracle Errors: How does GKD perform when the dynamics are non-injective (many-to-one mapping), such as systems converging to fixed points or limit cycles where T−1T^{-1} is inherently multi-valued? In such cases, or when fθf_\theta suffers from approximation noise, how robust is the cell partitioning, and how gracefully do the TKC theoretical certificates degrade?

  2. Trade-off and Justification for Table 2 Metrics: Given that EDMD-RBF and Deep Koopman achieve substantially lower absolute OOD MSE than SEDMD-Local in Table 2, how should practitioners justify preferring a method with a lower relative OOD ratio over one with an order-of-magnitude lower absolute prediction error?

  3. Scalability to High Dimensions and Scope Limitations: Since spatial partitioning naturally suffers from exponential cell growth as state dimension increases (d≥10d \ge 10), how do you foresee GKD scaling to higher-dimensional physical systems? Are you willing to include an explicit "Limitations and Scope" section in the main text to clarify the curse of dimensionality and the restriction to pure measure shifts (excluding parameter drift)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.