acceptodds
Under review as a conference paper at ICLR 2027

Representational Collapse and Spectral Recovery in Koopman-Structured Time-Series JEPAs

Abstract

Joint-embedding predictive architectures (JEPAs) are an effective framework for self-supervised learning on sequential data, but their latent representations remain difficult to interpret, particularly when a dynamical system generates the data. We study predictive self-supervised learning through Koopman operator theory and delay embeddings. Extending prior analyses of the eigenvalue-one case, we show that, under standard assumptions, training a linear latent predictor amounts to learning a finite-dimensional approximation to a Koopman-invariant subspace. Absent variance and covariance constraints on the latent space, both unconstrained and norm-preserving predictors converge toward degenerate representations rather than well-organized ones; this collapse, rather than the loss of oscillatory structure, is the dominant failure mode of predictive self-supervised learning on dynamical data. Standard variance-covariance regularization suffices to prevent it, and a Koopman-structured predictor adds interpretability: its learned rotation angles can be compared directly against the spectral content of the underlying system. We relate context length, latent dimension, and prediction horizon to observability, spectral capacity, and predictability. On synthetic dynamical systems, the learned representations recover the true oscillation frequencies, including combination frequencies not explicitly targeted by the model, while exhibiting the collapse and capacity trade-offs predicted by our analysis.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Joint-embedding predictive architectures (JEPA), Self-supervised representation learning, Koopman operator theory, Delay embeddings, Time-series dynamical systems

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

This paper looks at JEPA-style predictive self-supervised learning on dynamical data using Koopman operator theory and delay embeddings. It shows three things: a linear latent predictor with zero loss spans a Koopman-invariant subspace; training without regularisation collapses to a constant representation regardless of how the predictor is built; and VICReg-style regularisation prevents that collapse. It then proposes Direct Koopman-JEPA, a block-diagonal orthogonal predictor whose learned rotation angles can be compared against the system’s Koopman eigenfrequencies. Experiments on five synthetic systems test collapse, frequency recovery, and the effect of context length, latent dimension, and prediction horizon.

Critical strengths
  1. The paper connects JEPA training to Koopman subspace identification and delay embedding clearly.
  2. The structured predictor gives a direct, readable link between its learned parameters and the system’s frequencies.
  3. Recovering combination frequencies is a useful demonstration.
Critical weaknesses
  1. The experiments only check whether the predictor’s rotation angles match known frequencies. They do not show that the encoder’s latent coordinates track physical quantities such as oscillation phase, and the identity block seems to mostly encode which system a window came from.
  2. Since the claim is about the learned representation, please show that the latent coordinates follow the system’s dynamics, for example by relating each rotation block’s phase to the true phase of the matching mode.
  3. The paper reports that an unconstrained predictor trained with the same regulariser learns eigenvalues close to the unit circle. Does this mean rotation frequencies could also be recovered from a standard JEPA by eigendecomposing its predictor after training?
  4. A comparison between those frequencies and the learned angles of Direct Koopman-JEPA would show whether the block-diagonal structure is needed for interpretability, or whether an unconstrained JEPA already gives comparable frequency information.
Questions for authors

See weakness

1 comment

We thank the reviewer for these comments. Its main point is fair: our experiments check the predictor's rotation angles, not the encoder's coordinates, and a claim about the representation should be tested on the representation. We focused on the angles because each one has an exact theoretical value that also shifts in a known way with the prediction horizon. This gave us a direct test against ground truth, including for frequencies that do not appear in the observed signal. A correct spectrum in the predictor does not, however, show by itself that the coordinates carry the phase.

There is a theoretical link between the two. If a rotation block predicted its own future exactly, its two coordinates would form a Koopman eigenfunction, and for a torus or a limit cycle that eigenfunction is the physical phase up to an offset and a scale. Matching frequencies together with a small prediction error inside the block would therefore imply matching phases. We did not measure that error block by block. Comparing each block's phase with the true phase, as the reviewer suggests, is the test that could settle the question.

The identity block does mostly encode which system a window comes from. This is expected when several systems are pooled: the quantities that stay constant over time are the system labels, so the non-rotating block learns them. It is also why we did not use regime classification as evidence of within-system dynamics.

On the unconstrained JEPA, we agree this is the right control, and the paper does not include it. We reported how close its eigenvalues are to the unit circle but did not compare their frequencies with the targets, so it may well carry similar information. We chose the block form because it assigns each frequency to a fixed pair of coordinates, which makes the interpretation more direct. With a dense predictor, the frequencies would first have to be extracted from its eigendecomposition and then matched to coordinates. In the comparison the paper does contain, the block structure gives similar regime separability and roughly a third of the prediction error. The structure helps, then, but the paper fails to show that it is necessary.

The reviewer's two suggestions are to compare block phases with true phases, and to compare the unconstrained predictor's eigenvalues with the targets. These are what this work needs to go from a correct spectrum in the predictor to a representation that demonstrably tracks the dynamics. We are grateful to the reviewer for making that gap so clear.

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

The paper interprets JEPA-style predictive self-supervised learning on time series through Koopman operator theory and delay embeddings. With a linear latent predictor M, it shows the following:

  • Zero predictive loss holds exactly when the encoder's coordinates span a Koopman-invariant subspace and M represents the Koopman operator on it. This extends Ruiz-Morales et al. (2026) beyond the eigenvalue-one case.
  • The context length must exceed the attractor's embedding dimension (Takens/Sauer).
  • The latent dimension must be at least 1 + 2c to represent c oscillatory modes.
  • As the prediction horizon grows, mixing (continuous-spectrum) components become unpredictable.

The paper argues that without variance/covariance regularization, training collapses to constant representations regardless of the predictor's structure. It proposes "Direct Koopman-JEPA": a block-diagonal orthogonal predictor (an identity block plus 2×2 rotations) combined with a VICReg-style regularizer, so that the learned rotation angles can be read off as Koopman eigenfrequencies.

Experiments use five synthetic systems (Ornstein–Uhlenbeck, Stuart–Landau, quasi-periodic torus, Lorenz-63, hyperchaotic Rössler), each observed through a scalar projection, with an MLP encoder. They show collapse without the regularizer, partial recovery of base and combination frequencies, and the effects of context length and latent dimension on regime k-NN accuracy, false-nearest-neighbor rate, and prediction error.

Critical strengths
  1. Clear, interpretable framing. Linking a linear JEPA predictor to Koopman invariant subspaces and delay embedding is a sensible lens, and the rotation-angle readout gives a verifiable check of what the latent dynamics encode. The recovery of combination frequencies (ω1 ± ω2, 2ω2) that are not explicitly targeted is a nice demonstration.
  2. Careful and transparent reporting. The paper reports per-seed results, failure cases (Rössler-locked blocks, seed-dependent recovery, the identity block acting as a regime label), and limitations honestly.
Critical weaknesses
  1. The theory is largely known or follows directly from definitions.

    • Theorem 3.3 restates the definition of a finite-dimensional Koopman-invariant subspace. The JEPA-as-Deep-EDMD connection is essentially the setup of existing deep Koopman work (Takeishi et al., 2017; Lusch et al., 2018; Otto & Rowley, 2019).
    • Theorem 3.5 restates the Takens/Sauer delay-embedding results.
    • Theorem 3.8 is dimension counting (k ≥ dimension of the target subspace).
    • Theorem 3.11 follows from decay of correlations, under a strong disjoint-coordinates assumption that is not checked.
    • Proposition 4.1 is a one-line integral.
    • The theory therefore adds little beyond a reinterpretation of existing results.
  2. The collapse finding is not new. Collapse of joint-embedding predictive objectives without regularization, and its prevention by variance/covariance terms (VICReg, Barlow Twins), is well established. It is also expected that the orthogonal predictor collapses completely, since its identity block admits the trivial constant solution with M = I. Calling this "the dominant failure mode" is not a new insight.

  3. Limited architectural novelty and missing baselines. Block-diagonal rotation and unitary/orthogonal Koopman predictors already appear in Koopman autoencoder work (e.g., Lusch et al., 2018, and consistent/unitary Koopman autoencoders). The comparison is only against an unconstrained linear predictor. There are no baselines from:

    • Koopman learning: DMD/EDMD, Koopman autoencoders;
    • time-series SSL: TS2Vec, TNC, TS-TCC, PatchTST, all of which the related work cites;
    • classical spectral estimation for frequency recovery, e.g., a periodogram, which already finds the same peaks (Appendix D.3).
  4. Purely synthetic, small-scale experiments with a weak downstream task.

    • All data are 1D projections of five low-dimensional systems, and the encoder is a small MLP.
    • The main downstream metric, regime classification among five very different systems, is nearly trivial: accuracy is 0.96–0.997, and the authors show the identity block alone acts as a regime label.
    • The claimed relevance to climate, fluids, and ECG/EEG (broader impacts) is untested; no real data are used.
  5. Frequency recovery is partial and loosely evaluated.

    • At Δ = 10, ω2 is recovered in 0/6 seeds. The combination frequencies are recovered in at most 2/6 seeds at any horizon.
    • Several blocks lock onto a Rössler frequency, which is explained post hoc as a "genuine frequency".
    • Angles are folded modulo 2π into [0, π], with several targets only 0.2–0.46 rad apart and a 0.05 rad tolerance. There is no significance test against chance or against target aliasing.
    • Which modes are recovered depends on the seed.
  6. Gap between theory and experiments. The theorems assume:

    • deterministic dynamics with a single ergodic measure;
    • an ideal target network equal to the online encoder;
    • linearly independent coordinates.

    The experiments instead use stochastic regimes, a pooled mixture of five ergodic systems, and an EMA target. The context-length "threshold" agreement is qualitative ("flattens near n_c ≳ 6"), the unconstrained predictor's later transition is explained ad hoc, and the continuous-spectrum/horizon prediction is admittedly not tested. How the variance regularizer interacts with the theory (e.g., Remark A.9's 2‖F_cont‖² penalty) is not analyzed.

  7. Arbitrary design choices are not ablated. The fixed rule assigning at least half of the latent dimensions to the identity block is not justified or varied. Gains over the unconstrained predictor are modest in k-NN accuracy (0.05–0.08) and inconsistent in FNN.

  8. Minor issues.

    • The estimated dimension d_A ≈ 2.5 for a 4D hyperchaotic Rössler system seems low (hyperchaotic attractors typically have Kaplan–Yorke dimension above 3) and deserves justification, since it sets the n_c threshold.
    • The AI-use statement saying generative AI was used to generate the synthetic datasets is unclear and should be clarified.
Questions for authors

see weakness

1 comment

We thank the reviewer for a detailed report and for recognizing the care taken in the experiments. We agree that much of the theory builds on known results. We wrote it out for JEPAs to make explicit what these results mean for this training objective. This is, mainly, that a constant representation is a perfect solution of the predictive loss, which is the starting point of the collapse analysis. The part that goes beyond the classical embedding theorems is the impossibility result. Below a certain window length, no observation can separate the system's states, so the loss cannot reach zero. The horizon result relies on an assumption we did not verify, and it is not tested, as the paper notes.

Three points about the experiments may be worth clarifying. First, the impossibility result is visible in the data: with one-sample windows, every model shows the same aliasing as the raw input. Second, the periodogram was meant as a reference, not as a competitor. Our observation is linear in the two base oscillations, so the combination frequencies do not appear in the observed signal, yet some learned rotations recover them. Finally, recovery is well above chance. We did not report a significance test, and we thank the reviewer for asking, since one follows from existing data. Folding only merges equivalent rotations, and the tolerance is narrower than the gap between any two targets. A random angle would therefore hit a target about 16% of the time, whereas 15 of 17 learned angles do so at the intermediate horizon. The variation across seeds is expected, as three randomly initialized rotations cannot cover an infinite set of frequencies. The angle that locks onto the Rössler system follows an independent estimate of that system's dominant frequency.

On collapse and novelty we largely agree. Collapse without regularization is well known, and we tested it as a prediction. What we found interesting is that a predictor unable to contract still collapses. "Dominant failure mode" was too strong, and rotational and unitary predictors predate this work. Our interest was in what such a predictor reveals inside a standard JEPA.

Most other gaps follow from the paper's question. The data are synthetic because the true dimensions and frequencies must be known, and the encoder is small to avoid architectural confounds. Pooled training with a moving-average target mirrors how JEPAs are used, while the noisy systems lie outside the theory. Regime classification is informative mainly as the context grows, rising from near chance before it saturates. The capacity split is fixed so that the frequency readout needs no tuning, and results change little with the number of rotations. We agree that the classification gains are modest, the case for the structure being its readout and its lower prediction error.

The reviewer is right on both minor points. Our Rössler estimate was a correlation dimension, and since the system is hyperchaotic, its dimension exceeds 3, which moves the required window up by one step. No data were generated by AI: AI tools wrote the data-simulation script, which we reviewed and tested, and every trajectory is its output.

The reviewer's suggestions outline what a stronger version of this work would need: baselines from Koopman learning and time-series self-supervised learning, a significance test built into the frequency evaluation, an ablation of the capacity split, a direct test of the horizon effect, and real data, without which claims about applications are not justified.

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

This paper is focused on JEPA-style predictive self-supervised learning on time series, and it studies it through the Koopman operator theory using delay embeddings. The authors show that with a linear latent predictor and an ideal target network, zero predictive loss is characterized as the encoder spanning a Koopman-invariant subspace, and they provide three further results that relate: (i) the context length to the delay-embedding injectivity, (ii) the latent dimension to the number of targeted spectral components, and (iii) the prediction horizon to the decay of mixing components. Empirically, it is shown that unregularized predictors collapse and that VICReg variant of variance/covariance regularization prevents this. Additionally to that, authors propose Direct Koopman-JEPA, whose block-diagonal orthogonal predictor has rotation angles that can be compared with analytic Koopman frequencies. Experiments use five synthetic regimes trained jointly (Ornstein–Uhlenbeck, Stuart–Landau, a quasi-periodic torus, Lorenz-63 and hyperchaotic Rossler) and report learned angles that track base and combination frequencies across a sweep of horizons. Importantly, The theory assumes noiseless observations of deterministic, measure-preserving systems.

Critical strengths

Honest and verifiable empirical study.

  • I appreciate that each experiment is well presented, and that negative or partial outcomes are reported plainly. In particular I like the seed-dependent frequency recovery, the Rössler lock-in, the identity block acting as a regime classifier, and the finding that the structured predictor is not needed to prevent collapse. This transparency is not (unfrotunaly) typical, and it makes the empirical claims good to read and easy to check.

  • Analytic targets that move with Δ\Delta make a sharper test than a fixed-frequency check. Also, I wish to note that recovering combination frequencies (neither simulation parameters nor present in the linear observation), is a non-trivial.

  • Unregularized and regularized training are compared well to isolate the regularizer (and not the predictor structure) as what prevents the collapse. I find this being a useful practical message for JEPA practitioners working on dynamical data.

Critical weaknesses

Scope

While abstract and introduction suggest much broader scope, every theoretical result assumes noiseless observations of a deterministic, invertible, measure-preserving system, which is essentially a study of unitary Koopman group (very idealized setting usually far from practical use cases of JEPA). The authors say “under standard assumptions” and “dynamical data”, and introduction motivates the work with observations that are “a noisy, partial, and often nonlinear view” of the system. But then the theory excludes noise, studies a linear observation map, and unitarity is only mentioned in the appendix notation. This is not a minor shift in scope, since under process or observation noise, Lem A.1 fails (the window is not anymore a function of the state), the loss acquires an irreducible variance term so that main strategy of this paper (zero loss collapse in Thm 3.3, delay-embedding result of Thm 3.5, and point/continuous dichotomy of Thm 3.11) is failing completely. I would also like to note that (see bellow), even within the deterministic scope, the non-degenerate case of Thm 3.3 is attained only by the torus regime, since for mixing systems no non-constant finite-dimensional invariant subspace exists. So, in the empirical part of the paper: (1) the two stochastic regimes are therefore not “stress tests” of the presented theory but fully lie outside of it, and (2) all regimes are trained jointly on a non-ergodic mixture. To conclude, the authors should either restrict the claims accordingly and state the correct scope explicitly in the abstract and introduction, or extend the analysis to Markov transfer operators carefully reviewing novelties w.r.t. the operator-learning literature of the next paragraph (which I believe is a significant revision of the work).

Positioning and missing literature

The paper frames several questions as new, despite the fact that they have been already studied Koopman and spectral representation-learning literature. This includes what a linear latent predictor learns, how to avoid collapse, how to constrain the predictor for measure-preserving systems, and how to verify learned spectra. While notion of JEPA became recently popularized (and the name was coined recently), the work that predates it is significant and is largely not cited here. So the contribution is hard to assess and in my opinion the novelty is overstated. To help the authors position their work better, I tried to group the missing work by topic and briefly explain why each group matters.

  • Learning Koopman/transfer-operator representations without collapse. The paper’s central observation is that a linear latent predictor learns an approximately invariant subspace and that an unnormalized loss collapses. This has been discussed in deep Koopman representation learning literature, see e.g. VAMPnets [1, 2] and DPNets [3] where collapse is excluded by construction and does not need an add-on regularizer, or operator-regression formulations [4] that make the “Deep EDMD” of Remark A.4 precise, including the projection error. In my opinion these methods are the natural baselines for Sec 5, and they should be discussed when presenting VICReg as the remedy. Beyond that, non-contrastive SSL with linear predictors has been analyzed in [5], and VICReg-type objectives have been related to spectral embeddings in [6], and I believe that the paper should explain how its collapse analysis (Remark A.3, Section 5.1) relates to these results. Finally, Lusch et al. is cited, but the authors do not say that it already parameterizes the operator with rotation-scaling blocks.

  • Low-rank-approximation and contrastive objectives as spectral learning. Another relevant line of work is the one that learns the leading singular functions of a conditional expectation operator (stochastic version of Koopman operator called also transfer operator) with neural networks. It minimizes a low-rank approximation loss that needs no whitening or matrix inversion. See e.g. NestedLoRA [7, 8], work [9] that shows how the same objective underlies SSL with a spectral decomposition as its minimizer, NCP [10] that discusses covariance type regularizations and proves with guarantees for inference and uncertainty quantification, and finally [11] that frames SSL explicitly as evolution operator learning for high-dimensional dynamics, which is also this paper’s framing. So, together these works directly impact the paper’s question “what should a non-collapsing predictive SSL objective recover?” but also come with operator-theoretic guarantees covering the stochastic and non-normal setting (in my opinion more typical in practice), which this paper’s theory excludes.

  • Unitary predictors and verification of learned spectra. Constraining the predictor to be unitary for measure-preserving systems is the idea of mpEDMD [12], and ResDMD [13] certifies learned eigenpairs through residuals and handles continuous spectrum. On the other hand spectral learning rates and spurious eigenvalues of stochastic systems are also analyzed in [4]. These works bear directly on the claims of frequency recovery. I think they are needed to tell true eigenvalues from pseudo-eigenvalues and to show which part of the block-orthogonal design is new.

Theoretical issues (measure related) I find that following issues significantly impact the paper and that they need to be addressed correctly:

(1) Proving Thm 3.3, step “using the ergodic theorem to equate the expectation over the data distribution with the spatial integral over the invariant measure” in Thm A.2 is not done correctly. It is unclear which theorem from Ergodic theory and assumptions under which this can be done are missing. According to my understanding, the only case that satisfies all the hypotheses of Thm 3.3 in a non-trivial way is torous.

(2) Ass 3.1 says “an ergodic invariant probability measure”, but Lorenz and Rossler have infinitely many, and depending on which one is ment the intended the content cis affected: (i) for a fixed point, L2(μ)L^2(\mu) is one-dimensional, so the linear-independence hypothesis fails for even two functions, (ii) for a periodic orbit, L2(μ)L^2(\mu) is finite-dimensional, and the content becomes trivial. Since “Spatial averages equal empirical averages” holds for Lebesgue-a.e. initial condition in the basin only when the measure is physical, it is my understanding that the authors intended the SRB measure for these systems, But in that case, this needs to be assumed explicitly, and note that its existence is not established for the hyperchaotic Rossler.

(3) While Def 3.4 refers to “the attractor supporting μ\mu”, the proof of Thm 3.5 sets A=suppμA = supp \mu, which differ in general, and importantly mAm_A and dAd_A change with the choice of μ\mu (e.g. for a periodic-orbit measure they are 0). Because the loss is a μ\mu-expectation, what matters is injectivity on a set of full μ\mu-measure, not on all of AA. Moreover, Thm 3.5 I think has issues in use of Sauer–Yorke–Casdagli, handling of prevalent family of observables, and mistaking correlation dimension with the box dimension.

To conclude

My current score reflects the main tension within the paper: (A) keep the theoretical setting (I believe that current technical issues can be resolved in minor revision) and significantly reduce the scope and impact of the paper, or (B) aim intended impact and do a proper theory that studies the representation collapse in practical setting (noisy and stochastic data from multiple systems) which in my opinion would need significant rewriting (in a sense is quite different paper to write). In both cases I see the possibility of paper meeting the acceptance bar unlikely. That said, I acknowledge the authors path and do believe that this topic is interested and that it deserves a good in-depth research that can bring benefits to JEPA community.

References:

  • [1] Mardt, Pasquali, Wu, Noé. VAMPnets for deep learning of molecular kinetics. Nat. Commun. 2018.
  • [2] Wu, Noé. Variational approach for learning Markov processes from time series data. J. Nonlinear Sci. 2020.
  • [3] Kostic et al. Learning invariant representations of time-homogeneous stochastic dynamical systems. ICLR 2024.
  • [4] Kostic et al. Sharp spectral rates for Koopman operator learning. NeurIPS 2023.
  • [5] Tian, Chen, Ganguli. Understanding self-supervised learning dynamics without contrastive pairs. ICML 2021.
  • [6] Balestriero, LeCun. Contrastive and non-contrastive self-supervised learning recover global and local spectral embedding methods. NeurIPS 2022.
  • [7] Ryu et al. Operator SVD with neural networks via nested low-rank approximation. ICML 2024.
  • [8] Jeong et al.. Efficient parametric SVD of Koopman operator for stochastic dynamical systems. NeurIPS 2025.
  • [9] HaoChen et al. Provable guarantees for self-supervised deep learning with spectral contrastive loss. NeurIPS 2021.
  • [10] Kostic et al. Neural conditional probability for uncertainty quantification. NeurIPS 2024.
  • [11] Turri et al.. Self-supervised evolution operator learning for high-dimensional dynamical systems. ICLR 2026.
  • [12] Colbrook. The mpEDMD algorithm for data-driven computations of measure-preserving dynamical systems. SIAM J. Numer. Anal. 2023.
  • [13] Colbrook and Townsend. Rigorous data-driven computation of spectral properties of Koopman operators for dynamical systems. Comm. Pure Appl. Math. 2024.
Questions for authors
  1. Can you please report per-block residuals ∣∣KΔφ−eiωφ‖|| K^\Delta \varphi − e^{i\omega} \varphi‖ (φ\varphi being an eigenfunction), as in ResDMD, so that true eigenfunctions can be told apart from approximate ones (Rössler, Lorenz)?

  2. What Δ\Delta you used in Figs 1 and 3?

  3. What if the systems are trained separately rather than pooled, do the conclusions change? This is interesting since the theory is developed for a single deterministic ergodic system, while in all experiments the sampling measure is the mixture of five regimes (invariant, but not ergodic, and also including two stochastic components). So, it is not clear for me how do the conclusions would change when each system is trained separately: (i) are the same frequencies recovered per regime (noting the some are shared by systems), (ii) does the identity block still absorb most of the capacity, and (iii) do the reported context-length transition and MSE gains persist?

  4. Since for the stochastic regimes, the Koopman eigenvalues generally lie inside the unit disk, does a radius-scaled rotation changes the results?

2 comments

We thank the reviewer for a very careful report. We think that it presents the choice well: keep the idealized setting and narrow the claims, or develop the theory for noisy, stochastic data. The paper was written for the first option, although we see that the abstract reads as if it had taken the second.

We chose the idealized setting because our question was whether the parameters of a JEPA predictor can be checked against the true dynamics, and that is only possible when the Koopman spectrum is known exactly. We agree with the reviewer's point on mixing systems: for them no non-degenerate encoder can reach zero loss, which matches the error floor we observe. Our mistake was in how we described this scope. "Standard assumptions" does not convey how restrictive it is. The two noisy systems also lie outside the theory rather than stress-testing it, and we kept them to see how the model behaves beyond its assumptions. Training on a pooled mixture is closer to how JEPAs are used, and it remains compatible with the main result, which only needs an invariant measure.

We too agree on the measure-theoretic comments, and our statements may be looser than what the arguments require. The proofs rely on the training samples following the system's invariant measure, which is what our data generation aims for, not on the ergodic theorem. The intended measure is the physical (SRB) one, which for the hyperchaotic Rössler system is assumed, not established. The embedding result also needs the full hypotheses of Sauer, Yorke and Casdagli. With one fixed observation map, the part of our context result that the data can test is the impossibility statement, and the results agree with it: with one-sample windows, every model shows the same aliasing as the raw input. Finally, our Rössler dimension was a correlation dimension. Since the system is hyperchaotic, its dimension exceeds 3, which moves the required window up by one step, to where our models already perform best.

We are also grateful for the overview of related work, which we should have engaged with more explicitly. We framed the paper around JEPAs because we wanted to understand what the recipe used in practice learns from dynamical data, instead of designing a collapse-free operator-learning method, and that is why our comparisons stay within that recipe. The methods the reviewer lists, from VAMPnets and DPNets to low-rank operator learning, take the second approach and handle stochastic dynamics. In several cases they also anticipate our analysis and our rotational predictor. They are the right context and the right baselines for this work.

On the questions, Figures 1 and 3 use a prediction horizon of 20 samples, and Figure 2 varies it. We did not compute residuals, so learned frequencies are only checked against analytic values, and for the chaotic systems a match should be read as approximate. We did not train each system separately, nor did we test contracting rotations. For a single system, the theory only predicts that the identity block would no longer encode which system a window comes from.

We share and appreciate the reviewer's view that the topic deserves deeper study. A stronger version of this work would treat noise and contraction directly, certify the learned frequencies, train on single systems as well as mixtures, and compare with the operator-learning methods above.

I thank the Authors for their honest reply and acknowledgment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.