acceptodds
Under review as a conference paper at ICLR 2027

Right States, Wrong Paths: Controlling Temporal Dependence in Chaotic Forecasting

Abstract

In chaotic systems, small forecast errors grow over time, setting a predictability horizon beyond which forecasts lose track of the reference trajectory. Past this horizon, the goal shifts from tracking a realization to reproducing the correct statistics. Yet matching the distribution of visited states (state-space fidelity) is not enough: those states must also be organized correctly in time (path-space fidelity). We show that the Energy Score (ES), a standard probabilistic objective whose expectation over whole trajectories is uniquely minimized by the true joint distribution, becomes increasingly insensitive to errors in temporal dependence as the effective number of spatial dimensions grows, while remaining sensitive to marginal-variance errors. Similar ES values can therefore hide very different temporal dynamics. We propose a post-hoc calibration method that makes temporal dependence an explicit, adjustable target. For generators with per-time stochastic inputs, such as the coarse-to-fine JUMP–COMPLETE forecaster we use, correlating these inputs across forecast times changes temporal dependence while keeping the learned weights fixed and the one-time latent marginals unchanged. The coupling is selected with a temporal Variogram Score; we show that its standard finite-ensemble estimator is biased and can move temporal correlations away from the reference, and use an unbiased version instead. Across three chaotic systems, calibration substantially improves temporal correlations and, for two of them, temporal spectra, while joint ES remains nearly unchanged.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Chaotic dynamics, neural PDE surrogates, probabilistic forecasting, proper scoring rules, temporal dependence

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Zne3OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper attempts to distinguish tracking, state-space, and path-space fidelity in chaotic forecasting. From Energy Score, this paper proposes a post-hoc calibration for generators with per-time stochastic inputs, in which a single coupling parameter correlates these inputs across forecast times with the learned weights frozen, selected with an unbiased estimator of a temporal Variogram Score. Experiments on Kolmogorov flow, 2D Kuramoto-Sivashinsky, and Hasegawa-Wakatani show substantial changes in temporal correlations, and spectral improvements on two systems, with small changes in joint Energy Score.

Critical strengths
  1. The method is straightforward and easy to understand. Temporal dependence is adjusted through a single coupling parameter in the stochastic inputs, while the trained network remains fixed. This provides a simple calibration procedure without retraining.

  2. The experimental results support the usefulness of the approach. Across three chaotic systems, calibration improves the reported within-window temporal correlations. Temporal spectra also improve on Kolmogorov flow and HW2D, while the joint Energy Score remains nearly unchanged.

Critical weaknesses
  1. Sensitivity to endpoint prediction errors remains unclear. Since Jump–Complete conditions the intermediate trajectory on a predicted distant endpoint, the paper should clarify whether, and to what extent, inaccuracies in that endpoint degrade the quality and temporal fidelity of the completed path. Although Appendix F.6 examines the conditioning mismatch, it does not quantify how completion quality changes as endpoint prediction errors increase.

  2. The empirical results do not establish a consistent advantage over baselines. Calibration improves several diagnostics relative to the uncalibrated model, but the resulting temporal fidelity is not consistently better than competing methods. Moreover, the comparison is limited to DYffusion, PDE-Refiner, and an endpoint-free Joint-Forecast control. Additional comparisons with approaches that explicitly model joint trajectories or temporal coherence would strengthen the evidence for the method’s comparative advantage.

  3. Limited discussion of the underlying chaotic dynamics. The experiments evaluate state distributions and temporal dependence, but the paper does not sufficiently explain how these design and results relate to characteristic properties(i.e., the Lyapunov Exponent) of chaotic systems.

  4. The applicability of the ES analysis to chaotic dynamics is insufficiently discussed. The main scaling result is established for zero-mean Gaussian trajectories with separable covariance. Although Appendix D provides a distribution-free decomposition, the stated high-dimensional decay rate relies on Gaussian moment bounds to control the remainder. The manuscript does not adequately examine whether the required concentration and covariance assumptions approximately hold for the chaotic trajectory distributions considered.

Questions for authors
  1. Inconsistent table formatting. Please standardize the use of boldface for the best results and clearly indicate metric preferences across tables. For example, Table 1 neither highlights the best results in bold nor indicates the preferred direction for ES, whereas Tables 12–14 do both. For signed correlation errors, consistently specify that values closer to zero are better.

  2. How does the proposed method relate to existing approaches that jointly address trajectory consistency and distributional fidelity in chaotic forecasting, and what are its main methodological differences?

  3. What’s the key contribution of this work in terms of the methodology or theoretical perspectives for chaotic forecasting?

5 comments
AuthorsOpenReview ↗

Response to Reviewer Zne3 (1/2): W1-W4

We thank the reviewer for their feedback, which motivated seven new experiments, additional comparisons, and clearer contributions (in blue in the revision of the pdf document).

Our central contribution links evaluation and correction: we quantify why the joint Energy Score (ES), despite being proper, can be insensitive to temporal-dependence errors, and address this through post-hoc calibration. We correlate stochastic inputs across forecast times with a single parameter ρ\rho, selected with an unbiased temporal Variogram Score (VSt\mathrm{VS}_t).

W1: Sensitivity to endpoint prediction errors.

Following the reviewer's suggestion, we quantify endpoint-error sensitivity and show that calibration remains effective under perturbations. Adding Gaussian endpoint noise degrades completion quality for both JUMP-COMPLETE and DYffusion, but JUMP-COMPLETE consistently performs better. Without retuning, calibration maintains near-zero within-window correlation error and reduces VSt\mathrm{VS}_t relative to the native coupling ( ρ=0\rho=0) (new App. F.7, Fig. 15).

W2: Comparisons with temporally coherent forecasting methods.

Competitive temporal fidelity at substantially lower cost. We added two new baselines, stochastic interpolants (SI) [1] and Rolling Diffusion [2], the latter jointly denoising future frames (Table 1, new App. C.6). Calibrated JUMP-COMPLETE achieves the lowest within-window correlation error on KS2D and HW2D, and ranks second on Kolmogorov, narrowly behind PDE-Refiner (0.069 vs. 0.060). On Kolmogorov, it requires 23-800$\times$ fewer network evaluations than DYffusion, PDE-Refiner, SI, and Rolling (Table 7). Thus calibration provides competitive temporal accuracy at a fraction of the computational cost.

The new baselines reinforce the limitation of joint ES. On Kolmogorov, Rolling and SI achieve lower joint ES despite 1.7$\times$ and 1.6$\times$ larger within-window correlation errors than calibrated JUMP-COMPLETE (new results in Table 1). ES-based selection also favors the uncalibrated coupling (ρ=0\rho=0) on all three systems despite worse temporal correlations. Hence, a better joint ES does not necessarily indicate more faithful temporal dynamics.

Calibration is not architecture-specific. We clarified JOINT-FORECAST as an endpoint-free temporal-bundling baseline [3] (new App. C.6). Calibration also improves this baseline (App. F.8), demonstrating that the mechanism is not restricted to endpoint-conditioned JUMP-COMPLETE.

W3: Connection to chaotic dynamics.

We added a Lyapunov-time analysis linking predictability to completion, evaluation, and calibration difficulty (new App. C.2). Estimated Lyapunov times are 17.6, 16.5, and 8.4 frames respectively for Kolmogorov, KS2D, and HW2D (new Table 3). Completion windows span 0.91, 0.97, and 1.43 Lyapunov times, respectively, making endpoints sensitive to initial-condition errors. The 496-frame rollouts span tens of Lyapunov times across all three systems, where tracking loses relevance and temporal fidelity matters. HW2D has the longest window in Lyapunov units, lowest selected coupling, and smallest calibration gain, consistent with greater calibration difficulty over longer relative windows.

W4: Applicability of the ES analysis.

On Kolmogorov, KS2D, and HW2D, both Gaussian concentration and covariance separability assumptions are substantially violated (as assessed through the measures in the new Table 8). We performed new tests showing that ES sensitivity analysis remains predictive on the three equations. The theoretical approximation predicts temporal-dependence ES gaps within 88%. New Figure 7 in the appendix shows that both the loss of ES sensitivity to temporal-dependence errors and its consequences for detectability persist in the nonlinear 2D systems. ES sensitivity to temporal-dependence errors decreases approximately inversely with effective spatial dimension, while sensitivity to variance errors remains nearly constant. Consequently, detecting temporal-dependence errors at full resolution requires an estimated 10410^4 to 5×1045\times10^4 forecast cases, compared with only 50 available (new Fig. 7).

References are listed in the following comment.

AuthorsOpenReview ↗

Response to Reviewer Zne3 (2/2): Q1-Q3 / References

Q1: Table formatting.

Tables now bold best means, indicate lower ES is better, and specify zero as optimal for signed correlation errors.

Q2: Relationship to existing approaches.

Our method explicitly controls temporal dependence after training, rather than modifying the learned model or generated states. We clarified this in a new Positioning paragraph (new App. A). We adjust only the joint law of per-time stochastic inputs, keeping weights, conditioning, and one-time latent marginals fixed. Unlike approaches implicitly constraining temporal coherence, we explicitly calibrate temporal dependence using a dependence-sensitive validation criterion and assess it with lag-resolved diagnostics.

Q3: Central contribution.

Our contribution is the diagnosis-to-remedy connection: we quantify why joint ES poorly resolves temporal dependence and use this insight for post-hoc calibration. Beyond known qualitative limitations of ES, we derive effective-dimensionality scaling and finite-sample detection budgets, empirically validated in chaotic systems. The novelty lies in this diagnosis-guided calibration, not the coupling family alone.

References

[1] Yifan Chen, Mark Goldstein, Mengjian Hua, Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. Probabilistic forecasting with stochastic interpolants and Föllmer processes. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research. PMLR, 2024.

[2] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research. PMLR, 2024.

[3] Johannes Brandstetter, Daniel E. Worrall, and Max Welling. Message passing neural PDE solvers. In International Conference on Learning Representations, 2022.

As the discussion period is coming to an end, we wanted to ask whether our rebuttal has adequately addressed your concerns or whether any points would benefit from further clarification. We would be happy to respond promptly to any remaining questions before the deadline.

Reviewer_Dw7uOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper argues that, past the predictability horizon of a chaotic system, matching the distribution of visited states is not the same as organizing those states correctly in time. The Energy Score is strictly proper for the joint trajectory law, but for Gaussian trajectories its gap for a fixed temporal-covariance error shrinks like the inverse of the effective spatial dimension, while its sensitivity to marginal-variance errors grows with that dimension. Similar Energy Scores can therefore hide very different temporal dynamics. The proposed fix is post-hoc. For a generator with one stochastic input per forecast time, here the coarse-to-fine JUMP–COMPLETE forecaster, those inputs are coupled by a scalar ρ\rho through z′=ρz+1−ρ2ϵz' = \rho z + \sqrt{1 - \rho^2}\epsilon. This changes temporal dependence while the weights stay fixed and each one-time latent marginal remains standard normal. The value of ρ\rho is chosen with a temporal Variogram Score. The usual finite-ensemble estimator of that score is biased and can push correlations away from the reference, so the paper uses an unbiased U-statistic instead. On Kolmogorov flow, Kuramoto–Sivashinsky, and Hasegawa–Wakatani, a coupling selected on short windows improves long-rollout lag-one correlations, and it improves temporal spectra substantially on two of the three systems, while the joint Energy Score changes by less than about 0.2%0.2\%. The control requires per-time stochastic inputs, and a single ρ\rho represents only one temporal scale.

Critical strengths

The paper makes a useful distinction for long-horizon chaotic forecasting among tracking fidelity, state-space fidelity, and path-space fidelity. Past the predictability horizon, a forecast can visit the right states while organizing those states incorrectly in time, and the re-anchoring example makes that gap concrete. The coarse-to-fine generator is also well matched to this distinction. The Jumper draws a distant endpoint from the current context, and the Completer fills the interior conditioned on that endpoint, with one stochastic input per forecast time. Temporal dependence can then be adjusted after training while the learned weights stay fixed.

Critical weaknesses

The experiments are limited to Kolmogorov flow, Kuramoto–Sivashinsky, and Hasegawa–Wakatani. The coupling itself is taken from a noise prior introduced for video diffusion (Ge et al., 2023), but the paper never tests the calibration on video generation, where temporal dependence across frames is the quantity being judged.

  • Songwei Ge, Seungjun Nah, Guilin Liu, Tyler Poon, Andrew Tao, Bryan Catanzaro, David Jacobs, Jia-Bin Huang, Ming-Yu Liu, and Yogesh Balaji. "Preserve your own correlation: A noise prior for video diffusion models." IEEE/CVF International Conference on Computer Vision (ICCV), pp. 22873–22884, 2023.
Questions for authors

The Completer conditions on a future endpoint and on noise that is correlated across the window, while the stated context is the current state. How does the method behave for a Markovian system, where the transition depends on the present state, and for a non-Markovian system, where the relevant context is a history? The latent coupling raises the same question for stationarity. It is a time-homogeneous recursion with one coefficient, so each input has the same marginal and a correlation that depends only on the lag. Does this remain appropriate when temporal dependence changes with time, and do the authors see a difference in feasibility between stationary and non-stationary dynamics, or between Markovian and non-Markovian dynamics? For the Kolmogorov-flow forecasts, can the authors also compare the spatial energy spectrum of the long rollouts with the reference spectrum of the same simulation? Figure 5 reports temporal spectra, and the spatial energy spectrum would check whether the generated states carry the right energy across wavenumbers. Finally, does the model compare with the probabilistic forecasting method based on stochastic interpolants and Föllmer processes in Chen et al. (2024)?

  • Yifan Chen, Mark Goldstein, Mengjian Hua, Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. "Probabilistic forecasting with stochastic interpolants and Föllmer processes." Proceedings of the 41st International Conference on Machine Learning, PMLR 235:6728–6756, 2024.
3 comments
AuthorsOpenReview ↗

Response to Reviewer Dw7u

We thank the reviewer for the positive assessment and questions motivating new experiments (marked in blue in the revision of the pdf document).

W1: Video generation.

Thanks for the suggestion. While we focus on physical dynamics, the distinction between state and temporal fidelity also matters in video generation. We test whether our post-hoc calibration, which adjusts temporal dependence via a latent coupling ρ\rho selected with the temporal Variogram Score (VSt\mathrm{VS}_t), transfers to video. We added new experiments on the KTH and CLEVRER video datasets (new App. G.3), using the same method as for PDEs, without video-specific tuning.

Calibration reduces the normalized lag-one temporal correlation error ∣ΔC~1∣|\widetilde{\Delta C}_1| from 0.620.62 to 0.060.06 on CLEVRER and from 0.510.51 to 0.070.07 on KTH (new Fig. 24). Fréchet Video Distance (FVD), a video-quality metric which reflects both appearance and motion, also drops (lower is better) from about 520520 to 300300 and from 340340 to 205205, respectively in these datasets. VSt\mathrm{VS}_t improves while joint ES changes little. As FVD is not used to select ρ\rho, these gains independently support transfer of the method beyond PDEs.

Q1: Markovianity and temporal context.

JUMP-COMPLETE can condition on multiple past states (App. C.3). In order to evaluate the method on non-Markovian environments, we added two spectrally coarse-grained Kolmogorov flows (new App. G.1): near-Markovian with more spatial modes, and strongly non-Markovian with fewer, each using one or two context frames.

Temporal context and stochastic coupling play complementary roles, even with non-Markovian observations. In all four configurations (the two regimes and the two context lengths), calibration reduces absolute normalized within-window lag-one correlation error ∣ΔC~1∣|\widetilde{\Delta C}_1| and temporal Variogram Score VSt\mathrm{VS}_t with little effect on joint Energy Score (ES) (new Fig. 22). In the strongly non-Markovian single-frame setting, ∣ΔC~1∣|\widetilde{\Delta C}_1| drops from about 0.390.39 to 0.040.04. A second context frame improves uncalibrated forecasts, especially with stronger coarse-graining, while calibration provides further gains at either context length.

Q2: Non-stationary dynamics.

We added freely decaying 2D turbulence (new App. G.2), separating early, intermediate, and late phases as turbulence weakens and temporal correlations evolve. We compare native sampling (ρ=0\rho=0), a global coupling selected across phases, and phase-specific couplings selected separately.

A single global coupling improves temporal fidelity throughout the decay: it substantially reduces ∣ΔC~1∣|\widetilde{\Delta C}_1| in all three phases (new Fig. 23). Phase-specific tuning yields no consistent further improvement and only marginal VSt\mathrm{VS}_t gains.

Q3: Spatial energy spectra.

Following the reviewer's suggestion, we added spatial kinetic-energy spectra for 496-frame Kolmogorov rollouts, compared with the reference and PDE-Refiner, a high-wavenumber correction baseline (new Fig. 9, App. F.3).

Calibration improves temporal dependence while largely preserving spatial energy spectra. Native and calibrated JUMP-COMPLETE match the reference closely at low and intermediate spatial frequencies, with similar excess energy at high frequencies; PDE-Refiner deviates earlier. This matches Table 10, where spatial statistics barely change with the coupling, and complements the temporal results in Figs. 4 and 5.

Q4: Stochastic interpolants.

We added the autoregressive one-step stochastic-interpolant (SI) forecaster of [1] as a baseline (new results in Tables 1, 14-16; App. C.6). Results are reported for a single seed due to the method's high inference-time computational cost. SI requires 100 network evaluations per frame versus 0.125 for JUMP-COMPLETE (800-fold; added to Table 7).

The comparison shows that calibrated JUMP-COMPLETE generally achieves better temporal fidelity than SI, except at HW2D window boundaries. On Kolmogorov, SI improves upon native JUMP-COMPLETE but remains worse than calibrated JUMP-COMPLETE in within-window temporal correlation, highlighting the benefit of post-hoc calibration.

Importantly, joint ES does not necessarily reflect temporal fidelity. On Kolmogorov flow, SI achieves a lower joint ES than calibrated JUMP-COMPLETE and PDE-Refiner, despite having a within-window correlation error 1.6×1.6\times and 1.8×1.8\times larger, respectively.

We hope these experiments address the reviewer's concerns and would be grateful if they reconsidered their assessment.

References

[1] Yifan Chen, Mark Goldstein, Mengjian Hua, Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. Probabilistic forecasting with stochastic interpolants and Föllmer processes. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research. PMLR, 2024.

As the discussion period is coming to an end, we wanted to ask whether our rebuttal has adequately addressed your concerns or whether any points would benefit from further clarification. We would be happy to respond promptly to any remaining questions before the deadline.

I would like to thank the authors for the extensive additional experiments they have done over the discussion period. I think these have addressed my concerns and therefore I raise my evaluation on the submission.

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

This paper investigates temporal dependence fidelity in probabilistic forecasting of chaotic systems. The authors show that Energy Score (ES), despite being strictly proper for joint trajectory distributions, becomes less sensitive to temporal covariance errors as spatial dimensionality increases. To address this issue, they propose a post-hoc calibration method that adjusts the temporal correlation of stochastic inputs without retraining the model. The approach is implemented with a coarse-to-fine JUMP–COMPLETE architecture, and the coupling parameter is selected using an unbiased temporal Variogram Score (VS). Experiments on three chaotic systems demonstrate improved temporal correlations and spectra while largely preserving joint ES.

Critical strengths
  1. The paper highlights the often-overlooked distinction between state-space and path-space fidelity in chaotic forecasting. The theoretical analysis provides a quantitative explanation for why ES may fail to distinguish forecasts with different temporal dependencies, particularly in high-dimensional settings.

  2. Adjusting temporal dependence through correlated stochastic inputs is an intuitive and computationally efficient approach that requires no retraining. The derivation of the unbiased VS estimator is also valuable, as it addresses the finite-ensemble bias that can otherwise lead to undesirable temporal correlations.

  3. The authors evaluate the approach on three chaotic systems and provide extensive ablations on coupling strength, ensemble size, noise amplitude, and training objectives. The results demonstrate that temporal fidelity can be substantially improved while joint ES remains nearly unchanged.

Critical weaknesses
  1. The main theoretical results rely on separable Gaussian covariance assumptions, which may not adequately characterize complex nonlinear chaotic systems. Although empirical evidence is provided for Kuramoto–Sivashinsky dynamics, the applicability of the theoretical conclusions to more general non-Gaussian and non-separable settings remains unclear.

  2. The proposed method uses a single AR(1) coupling parameter, which may be insufficient to capture complex temporal dependencies involving multiple timescales. Moreover, Appendix F.5 shows that different temporal metrics prefer different coupling values, raising questions about whether VS-based calibration consistently improves overall path-space fidelity.

  3. While calibration improves within-window temporal correlations, Table 1 shows that cross-window boundary errors increase for Kolmogorov flow and HW2D. This suggests that optimizing temporal dependence on short windows does not necessarily improve global trajectory consistency, which is particularly important for long-horizon chaotic forecasting.

Questions for authors
  1. How well do the ES sensitivity and detection-budget results extend to non-Gaussian and non-separable spatiotemporal distributions commonly encountered in chaotic systems?

  2. Given that different temporal metrics favor different coupling values (Appendix F.5), how robust is the VS-selected coupling in capturing temporal dependencies across multiple timescales?

  3. Table 1 shows increased cross-window boundary correlation errors after calibration on Kolmogorov flow and HW2D. What causes this degradation, and could a boundary-aware calibration objective help address it?

4 comments
AuthorsOpenReview ↗

Response to Reviewer UwgF (1/2): W1-W2 / Q1-Q2

We thank the reviewer for the careful reading and constructive comments. Following the reviewer's suggestions, we have added several experiments (marked in blue in the revision of the pdf document).

W1 / Q1: Applicability to non-Gaussian, non-separable chaotic systems.

Our main theoretical result shows that as the effective spatial dimension DD grows, the joint Energy Score (ES) becomes less sensitive to temporal-dependence errors (roughly 1/D1/D), making them costly to detect, while remaining sensitive to variance errors. New experiments confirm this persists in non-Gaussian, non-separable chaotic systems, beyond Proposition 1’s assumptions (we summarize below the main findings of our additional experiments; full experimental details and results are provided in the new App. E.2).

On Kolmogorov, KS2D, and HW2D, both assumptions are substantially violated (as assessed through the measures in the new Table 8). Nevertheless, the theoretical approximation predicts temporal-dependence ES gaps within 8%. New Figure 7 in the appendix shows that both the loss of ES sensitivity to temporal-dependence errors and its consequences for detectability persist in our nonlinear 2D systems. More precisely, ES temporal-dependence sensitivity decreases approximately inversely with the effective spatial dimension, while sensitivity to variance errors remains nearly constant. Consequently, detecting temporal-dependence errors at full resolution requires an estimated 10410^4 to 5×1045\times10^4 cases, compared with our 50 test trajectories (new Fig. 7). Hence, the dilution predicted by Proposition 1 holds in all three chaotic 2D systems used for the experiments, even though they violate both of its assumptions.

W2 / Q2: Robustness across timescales, metrics, and coupling complexity.

Our post-hoc calibration controls temporal dependence through a latent coupling ρ\rho, selected using the temporal Variogram Score (VSt\mathrm{VS}_t). Following your suggestions, we performed new experiments. They are detailed in Appendix F while the main findings are summarized below. These experiments quantitatively assess the robustness of the control mechanism.

Multiple timescales. VSt\mathrm{VS}_t is sensitive to both fast and slow temporal dependencies within the completion window. A new two-timescale experiment confirms that it detects errors in both rapidly and slowly varying temporal components (new App. F.6, Fig. 13).

Metric-dependent optima. The three diagnostics prefer ρ=0.9\rho=0.9, 0.950.95, and 11, but validation VSt\mathrm{VS}_t is flat across [0.9,1.0][0.9,1.0] within seed variability, and every value there improves strongly over ρ=0\rho=0 (Table 11 and new Fig. 11).

Coupling complexity. Couplings with multiple parameters provide no consistent improvement over a single-parameter coupling in our new experiments. We compared couplings where each noise input depends on the previous pp noise inputs, with p∈1,2,4,8p \in \\{1,2,4,8\\}, on Kolmogorov and KS2D (new Fig. 14). One possible explanation is that the nonlinear Completer allows even simple latent noise coupling to induce richer temporal dependencies in physical space. Additionally, a single global coupling remains effective across phases of non-stationary decaying turbulence (see the exp. in new App. G.2).

AuthorsOpenReview ↗

Response to Reviewer UwgF (2/2): W3 / Q3

W3: Global trajectory consistency.

Calibration improves temporal consistency over long rollouts, despite some degradation at window boundaries (Kolmogorov and HW2D; addressed in Q3 below). Temporal power-spectrum errors also decrease (Fig. 5 in the main text), indicating improvements beyond short-term dependence.

New evaluation. We performed new experiments over full 496-frame rollouts, boundaries included. The VSt\mathrm{VS_t}-selected coupling reduces the measured errors across the temporal diagnostics over the native coupling. More precisely, the results show that calibration improves state correlations, increment correlations, and second-difference statistics on all three systems (new Fig. 12).

Q3: Boundary degradation and boundary-aware calibration.

While calibration improves temporal dependence overall, as you observed, it can increase correlation errors at the boundaries between consecutive generation windows, particularly on Kolmogorov and HW2D. We analyzed the cause of the boundary errors and proposed a boundary-aware calibration objective. Our new experiments show that this can reduce the boundary degradation, but reveal a trade-off between boundary calibration and spectral fidelity.

Cause of degradation. Our new analysis of temporal correlations at different positions within each window shows that correcting interior under-correlation can over-correlate the window ends on Kolmogorov and HW2D (new App. F.12, Fig. 19).

Separate boundary coupling. We tested a separate coupling parameter at transitions between consecutive generation windows to correct boundary errors. However, this approach succeeds on Kolmogorov but not on HW2D (new Fig. 20), showing that adjusting the coupling at boundaries alone is not consistently effective.

Chained-window calibration. Since boundary errors also depend on the chaining scheme, we tested selecting ρ\rho using VSt\mathrm{VS}_t on chained rather than isolated validation windows. This reduces boundary errors on both Kolmogorov and HW2D, below the uncalibrated model on Kolmogorov but not on HW2D (new Table 17). However, it adds back some high-frequency power, revealing a trade-off between boundary and spectral fidelity. Importantly, the original calibration already improves global temporal diagnostics over complete long rollouts.

We hope these additional analyses clarify the generality of the ES sensitivity result, the robustness of the calibration procedure, and its remaining limitations. We thank the reviewer for considering the revised manuscript.

Thank you for the detailed response and the substantial additional experiments. I appreciate the authors' efforts to address my concerns. In particular, the new results in Appendix E.2 provide convincing empirical support for the ES sensitivity analysis beyond the Gaussian and separable assumptions. I also acknowledge that the additional long-rollout experiments demonstrate improvements across multiple temporal diagnostics.

However, I remain concerned about the robustness of the calibration strategy in achieving overall path-space fidelity. As shown in Table 17, chained-window calibration reduces boundary errors compared with single-window calibration, but the boundary error on HW2D remains larger than that of the uncalibrated model. Moreover, this improvement comes at the expense of spectral fidelity relative to single-window calibration. Together with the metric-dependent optimal coupling values in Table 11, these results suggest that the proposed calibration does not consistently reconcile different aspects of temporal fidelity.

While the additional experiments strengthen the paper and clarify its limitations, I believe the remaining trade-offs are important for reliable long-horizon chaotic forecasting. Therefore, I intend to maintain my score.

AuthorsOpenReview ↗

Response to Reviewer UwgF’s follow-up

Thank you for the follow-up. The three observations are correct and are reported in the paper (Section 5.4; Conclusion; App. F.5, F.12). Where we differ is on what they imply.

The paper makes three contributions: the ES dilution result, which the reviewer now finds convincingly supported (App. E.2); the unbiased VSt\mathrm{VS}_t estimator, listed as a strength in the original review; and post-hoc calibration, for which the reviewer acknowledges long-rollout improvements across multiple temporal diagnostics. The remaining concern bears on how completely the calibration succeeds. Our claim there is that, relative to the same model sampled with independent inputs (ρ=0\rho=0), calibration improves within-window correlations on all three systems and temporal spectra on two, with joint ES nearly unchanged; it is not that one coupling is optimal for every diagnostic: ρ⋆\rho^\star is “a calibration parameter for the targeted temporal property rather than a universally optimal setting” (Section 5.4).

Size of the trade-off (Table 17, three-seed means; signed correlation errors and spectrum error):

nativesingle-window ρ⋆\rho^\starchained ρ2⋆\rho^\star_2
Kolmogorov, within-window−0.328+0.069−0.030
Kolmogorov, boundary−0.037+0.051+0.027
Kolmogorov, spectrum0.8900.1720.381
KS, within-window−0.026−0.004−0.002
KS, boundary−0.018−0.006−0.005
KS, spectrum0.0380.0270.039
HW2D, within-window−0.097+0.056−0.012
HW2D, boundary+0.133+0.238+0.190
HW2D, spectrum0.3910.2210.300

Both couplings are selected with VSt\mathrm{VS}_t alone; only the validation windows differ: single windows start from a true state, chained windows from a generated one, as every window after the first does in a rollout. Under chained selection, seven of the nine means improve over native sampling, five of them beyond seed variability; the KS spectrum is unchanged within seed variability; and one, the HW2D boundary, worsens. Under single-window selection the HW2D boundary is likewise the only entry that worsens beyond seed variability. Both selections reduce the spectrum error on Kolmogorov and HW2D: by 81% and 43% with single-window selection, by 57% and 23% with chained.

Long rollouts. The spectrum error is computed over full 496-frame rollouts, boundaries included. Over the same rollouts, the single-window coupling reduces increment-correlation and second-difference errors on all three systems, and lag-one state-correlation error on Kolmogorov and KS, with non-overlapping error bars; on HW2D the lag-one reduction is within seed variability (Fig. 12). Joint ES differs by less than 0.2% across the three settings.

HW2D boundary. The position-resolved analysis added for Q3 (App. F.12) separates two effects. About three quarters of the native error (+0.133) comes from the hand-off pair between windows. It is present before calibration and remains nearly unchanged under every latent coupling we tested, within the window or across the seam (Figs. 19–20): it is set by how windows are chained, not by the coupling. The increase under calibration (+0.105 single-window, +0.057 chained) comes mainly from the endpoint-approach pair: the native error profile is U-shaped and a uniform coupling shifts it as a whole, so removing the interior under-correlation over-correlates the unanchored end of the window. Position-dependent coupling could target this increase, while correcting the native hand-off error may require changes to the chaining scheme. Both are identified as future work in the conclusion.

Metric-dependent optima. The three preferred couplings in Table 11 (0.9, 0.95, 1.0) lie in a range where validation VSt\mathrm{VS}_t varies by less than its seed variability (Fig. 11), so the choice within it is immaterial. Their disagreement is small next to the effect of coupling: across [0.9,1.0][0.9,1.0] the three errors stay below 0.004, 0.04 and 0.014, against 0.042, 0.49 and 0.056 for the same weights at ρ=0\rho=0. The same flat region appears for checkpoints trained at other couplings (Fig. 4).

In summary. Under chained selection, the HW2D boundary is the only entry whose error increases beyond seed variability relative to native sampling, and its cause is identified. This limits how far the calibration reaches; the ES analysis and the unbiased estimator do not depend on it. The calibration requires a one-parameter sweep and no retraining.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Summary of Revisions

We thank the reviewers for their constructive feedback. We have substantially strengthened the manuscript with new experiments examining the generality of our theoretical results, the robustness and transferability of temporal calibration, and comparisons with additional baselines. Changes are marked in blue.

Our central message is that generating plausible states does not guarantee generating plausible trajectories. We establish that the joint Energy Score (ES), despite being a proper scoring rule, can have very limited practical sensitivity to temporal-dependence errors in high-dimensional chaotic systems. This motivates explicitly evaluating and controlling temporal dependence in probabilistic forecasting. We therefore introduce a post-hoc calibration mechanism: correlating stochastic inputs across forecast times through a parameter ρ\rho, selected using an unbiased temporal Variogram Score (VSt\mathrm{VS}_t), without retraining or modifying model weights.

The new results strengthen both our theoretical and methodological contributions. They demonstrate that the limited sensitivity of ES has important practical consequences beyond our theoretical assumptions, and that temporal dependence can be substantially improved through post-hoc calibration across diverse forecasting settings. They also clarify the limitations of calibration at generation-window boundaries.

The main additions are:

  1. Generality and practical implications of the ES analysis. New experiments on non-Gaussian, non-separable chaotic systems confirm that ES sensitivity to temporal errors decreases with effective spatial dimension. The theoretical approximation predicts the observed ES gaps within 8% without fitted parameters. At full resolution, detecting these errors would require an estimated 10410^4 to 5×1045\times10^4 forecast cases (new App. E.2). Thus, ES dilution is not merely a theoretical phenomenon: it can make substantial temporal errors practically undetectable.

  2. Robustness and generality of calibration. New experiments show that calibration improves temporal fidelity beyond chaotic PDE forecasting, including on the KTH and CLEVRER video datasets. FVD decreases by approximately 40%, although this metric is not used to select ρ\rho (new App. G.3). Gains also extend to non-Markovian observations (new App. G.1), non-stationary turbulence (new App. G.2), and perturbed endpoints (new App. F.7). More complex temporal couplings provide no consistent additional benefit (new App. F.6). These results support the effectiveness of a simple, transferable calibration procedure without retraining.

  3. New baselines and evidence of ES limitations. We added two new baselines, Stochastic Interpolants (SI) and Rolling Diffusion (new results in Table 1). On Kolmogorov, both improve within-window temporal correlations over native (ρ=0\rho=0) JUMP-COMPLETE, but calibrated JUMP-COMPLETE performs better at substantially lower computational cost. SI and Rolling Diffusion achieve lower joint ES despite 1.6-1.7$\times$ larger temporal-correlation errors. This directly demonstrates that better joint ES does not necessarily imply better temporal fidelity.

  4. Temporal fidelity across timescales and long horizons. New experiments investigate multiple temporal resolutions and 496-frame free-running rollouts. Calibration improves aggregate temporal diagnostics, including state correlations, increment correlations, and second-difference Wasserstein discrepancies. These gains extend beyond individual generation windows. However, boundary errors can increase on some systems; new boundary-aware calibration experiments reduce them at the expense of some spectral fidelity (new App. F.12). We now explicitly acknowledge this trade-off.

  5. Further analyses and clarifications. We added new Lyapunov-time analysis, new spatial energy spectra for Kolmogorov flow, clearer positioning relative to prior work, and improved presentation of experimental results.

Taken together, these results establish that temporal fidelity cannot be reliably assessed through joint ES alone, and demonstrate that it can be substantially improved through simple post-hoc calibration, while identifying the remaining challenge of global trajectory consistency.

We address each reviewer's concerns in detail below.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.