acceptodds
Withdrawn from ICLR 2027

Amortized Generative Modelling of Invariant Measures Across a Family of Kuramoto-Sivashinsky PDEs

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The long-run statistics of chaotic dissipative partial differential equations are described by invariant measures, and when the equation carries a parameter these measures form a family whose members can differ in kind, from steady states to sustained spatiotemporal chaos. We ask whether a single generative model can cover such a family. We train a single conditional diffusion model once, on states of the Kuramoto–Sivashinsky equation at 160 values of its viscosity, and evaluate it against held-out data at interleaved parameter values not seen in training, using a classifier two-sample test supported by spectral, geometric and tail statistics. Agreement improves as the dynamics become more chaotic. We identify a structural cause: the weakly chaotic members have invariant measures concentrated on low-dimensional sets, which a model with full-support output cannot represent, and the target approaches full dimensionality as chaos increases. Restricting training to the strongly chaotic members reduces the classifier's excess over chance on those members by roughly 40%, and this restricted conditional model matches or exceeds models trained at single parameter values. On those members, generated samples reproduce energy spectra to within a resampling floor, with a small systematic deficit at the largest scales, and match the energy balance of held-out data.

Primary area:
learning on time series and dynamical systems
Keywords:
diffusion models, invariant measures, Kuramoto–Sivashinsky equation, spatiotemporal chaos, conditional generative models, amortization, scientific machine learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3U4tOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper trains a conditional diffusion model to approximate invariant measures across a family of spatially one-dimensional Kuramoto–Sivashinsky equations, using 38 real Fourier coordinates. It reports better distributional agreement in strongly chaotic regimes and improved performance when training is restricted to those regimes. Evaluation concerns interpolation between training parameter values.

Critical strengths

The systematic parameter sweep examines both successful and unsuccessful regimes. Complementary classifier, spectral, geometric, tail and energy-balance diagnostics provide a more informative assessment than visual agreement alone. The observed variation in generative performance across dynamical regimes is potentially useful.

Critical weaknesses
  1. The proposed explanation of failure is not adequately established. The paper attributes weakly chaotic failures to incompatibility between full-support generative distributions and low-dimensional invariant measures. However, inability to represent a singular measure exactly does not establish inability to approximate it accurately. The further assertion that chaos produces full-dimensional support requires evidence beyond the reported correlation between chaos and classifier accuracy.

  2. The comparative and statistical controls are insufficient. Conditional and unconditional models receive ten epochs despite substantially different dataset sizes, leaving their optimization budgets unmatched; only three unconditional cases are examined. Additionally, randomly splitting snapshots from a single reference trajectory does not establish robustness to temporal dependence. Matched-update comparisons and independent-trajectory evaluation are needed to strengthen the amortization and distributional-fidelity conclusions.

Questions for authors
  1. Scientific contribution and generality: The experiments establish interpolation within one KS family, using a 38-dimensional representation. Which finding constitutes the main advance beyond applying conditional diffusion, and what evidence would establish that it persists with another generative architecture or PDE family?

  2. Explanation of failure in weakly chaotic regimes: How do you distinguish an inability to represent a singular measure exactly from an inability to approximate it at finite resolution? What evidence establishes that increasing chaos makes the projected invariant measure full-dimensional, and rules out model capacity, optimization or residual sampling noise as explanations for the observed failures?

  3. Fairness of the amortization comparison: Both pooled and single-parameter models are trained for ten epochs, despite substantially different dataset sizes. Does the conclusion that amortization is “free” survive matching optimizer updates or computational budgets, and expanding the comparison beyond three unconditional parameter cases?

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

In this paper, the authors investigate a method for approximating parameter-dependent invariant measures of the Kuramoto–Sivashinsky equations using a single conditional diffusion model. The experiments show that the accuracy of the invariant measure approximation improves as chaos intensifies, and that accuracy is further enhanced by limiting the training data to systems exhibiting strong chaos. The authors explain this phenomenon in terms of the relationship between the dimension of the support set of the invariant measure and the representational capacity of the diffusion model.

Critical strengths
  • The problem of approximating an invariant measure corresponding to multiple parameters using a single generative model is significant and is expected to have applications in the efficient estimation of long-run statistics.
  • The study presents interesting experimental findings regarding the relationship between the strength of chaos and generation accuracy.
  • In addition to comparing distributions using classifiers, the study evaluates generation accuracy from multiple perspectives, including energy spectra, tail probabilities, and energy balance.
Critical weaknesses
  • It has not been sufficiently demonstrated why it is necessary to learn invariant measures directly, nor what computational advantages this offers compared to numerical simulations.
  • There is no comparison with other generative models, such as Normalizing Flow, and it is unclear whether the observed phenomena are specific to diffusion models or common to generative models in general.
Questions for authors
  • Please discuss the significance of this research. Compared to numerical simulations of PDEs, what advantages does direct sampling from an invariant measure offer in terms of computation time and the accuracy of statistical estimates? It would be helpful if you could briefly mention specific application examples as well.
  • Is a similar relationship between the strength of chaos and generative accuracy also observed in generative models such as Normalizing Flows? Are the findings from this study specific to diffusion models?
  • [R1] discusses the relationship between chaos and the identifiability of dynamical systems. Please explain any connection, if any, between the relationship between chaos and generative accuracy observed in this study and that research. It seems to me that the claim that stronger chaos makes learning easier is similar.

[R1] Zakhar Shumaylov, et al., “When is a System Discoverable from Data? Discovery Requires Chaos,” arXiv, 2025.

Reviewer_v8DQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies amortized generative modeling of invariant measures across a parameterized family of Kuramoto–Sivashinsky equations. A conditional diffusion model is trained jointly on states generated at 160 hyperviscosity values and evaluated at 160 interleaved unseen parameter values. The evaluation uses a random-forest classifier two-sample test, modal energy spectra, nearest-neighbor statistics, coordinatewise tail coverage, and an energy-balance diagnostic. The paper reports that generative accuracy generally improves as the leading Lyapunov exponent increases. It further reports that restricting training to strongly chaotic members improves performance in that regime, and that conditional models sometimes outperform unconditional models trained at individual parameter values. The proposed explanation is that weakly chaotic invariant measures concentrate on low-dimensional sets that cannot be represented exactly by a diffusion model with full-support output, whereas increasingly chaotic measures become more compatible with the model.

Critical strengths
  • Multiple complementary diagnostics. The classifier test examines joint-distribution differences, while the spectral, nearest-neighbor, tail, and energy-balance statistics offer some interpretability about particular failure modes.
  • Good presentation and implementation detail. The PDE solver, state representation, diffusion architecture, training setup, classifier configuration, and data construction are described clearly.
  • Interesting problem. Learning invariant measures across a parameterized PDE family is meaningfully different from trajectory forecasting.
Critical weaknesses
  1. Evaluation does not establish invariance. The present results support the narrower claim that the model approximates several statistics of an empirical KS trajectory distribution. They do not yet fully support the stronger claim that it reproduces the invariant measure.

  2. Leading Lyapunov exponent alone does not measure dynamical complexity, thus not a good measure for performance. A strongly chaotic KS system can have multiple positive Lyapunov exponents. Observing Lyapunov spectrum would be more informative.

  3. The amortization comparison does not control training exposure. Every model is trained for ten epochs while the full conditional model uses data from 160 parameter values and each unconditional model uses data from only one. Assuming an epoch processes the respective training dataset once, the conditional model receives roughly 160 times as many optimizer steps and examples as an unconditional model. Therefore, equal epoch counts do not constitute a controlled comparison.

  4. Narrow experiment result. The experiments cover many hyperviscosity values, but all values belong to the same Kuramoto–Sivashinsky equation. This is a strong within-family study, but it does not demonstrate that the main findings generalize across chaotic dynamical systems.

  5. The experiments appear to use one model-training seed, one trajectory per parameter value, and one fixed classifier split.

Questions for authors
  1. Can the authors compute more of the Lyapunov spectrum and report the number of positive exponents?
  2. Does the generated distribution remain unchanged after its samples are evolved with the KS solver?
  3. Can the authors experiment with other chaotic systems and verify if the main claim of the paper still persists?

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

Related papers

Discussion (0)

Sign in to comment.