When do World Models Successfully Learn Dynamical Systems?

Edmund Ross, Claudia Drygala, Leonhard Schwarz, Samir Mani Kaiser, Francesca di Mare, Tobias Breiten, Hanno GottschalkSubmission to ICLR 2026 · Withdrawn

TL;DR

World models—which compress spatial states into compact latent tokens and autoregressively predict their evolution—can accurately simulate complex physical systems over long horizons where standard neural operators fail. By grounding these architectures in classical control theory, the authors prove that observability is the critical mathematical condition determining whether a tokenized history contains enough information to reconstruct and propagate the underlying physics.

What Problem Does This Paper Address?

While neural operators like Fourier Neural Operators (FNO) and DeepONets have become popular for surrogate modeling of partial differential equations (PDEs), they often struggle with long-term rollouts. When used autoregressively as one-step predictors, errors accumulate rapidly, causing simulations to desynchronize or diverge to infinity.

Conversely, "world models" developed in computer vision and reinforcement learning compress frames into low-dimensional latent spaces, predict dynamics across a temporal history window, and then decode them back to high-resolution states. Despite their empirical success, there has been little theoretical explanation for when compressing a state and concatenating a history of tokens actually preserves physical dynamics, or how the length of the latent history relates to the true degrees of freedom in the system.

Methods

The authors formalize world models through the lens of control theory and state-space observability:

  1. System-Theoretic Formulation: Let $\dot{x} = f(x)$ be a dynamical system and $y = h(x)$ be a continuous tokenization (observation) map that reduces dimension ($m \ll n$). A system is observable over a history length $k$ if an invertibility condition holds: past observations $y_{\text{seq}}(t) = (y(t-(k-1)\Delta), \dots, y(t))$ uniquely identify the full state $x(t)$ via a continuous reconstruction mapping $G$.
  2. PAC-Learnability Guarantees: The authors prove that if observability holds, there exists a latent autoregressive dynamics map $g$ such that both $g$ and the reconstruction decoder $G$ (and thus the full time propagator $S_\Delta$) are Probably Approximately Correct (PAC) learnable using neural networks.
  3. Model Hierarchy: To isolate the impact of observability from deep learning artifacts, the authors evaluate a spectrum of models:
  4. Analytic multivariate least-squares regression.
  5. Linear autoregressive layers trained via Adam.
  6. 3D convolutional U-Nets trained in an adversarial setup (for the Kuramoto–Sivashinsky equation).
  7. Full two-stage generative video architectures (NVIDIA's LongVideoGAN) using temporal Kaiser filtering and StyleGAN-style upsamplers (for turbulent fluid flow).

Datasets

The authors benchmark across four physical systems:

  • Heat Equation (2D): Evaluated under both constant conductivity (theoretically unobservable under average pooling) and spatially varying random conductivity fields (observable).
  • Wave Equation (2D): Second-order system evaluated similarly with constant vs. spatially varying wave speed.
  • Kuramoto–Sivashinsky Equation (KSE, 2D): A classic non-linear PDE exhibiting chaotic spatiotemporal dynamics, solved via 4th-order Exponential Time Differencing Runge-Kutta (ETDRK4).
  • Flow Around a Circular Cylinder ($Re = 3900$): A 2D computational fluid dynamics (CFD) benchmark displaying a Kármán vortex street, simulated via Large Eddy Simulation (LES).

Key Insights and Ideas

  • Observability is non-negotiable: For standard patch-averaging tokenizers on a uniform grid, symmetric Fourier modes cancel out in linear PDEs with constant coefficients. As Theorem 6 shows, the system is mathematically unobservable: no amount of past history enables the model to recover the true state. Breaking this symmetry—either by introducing spatially heterogeneous coefficients or by randomizing tokenizer weights—restores observability and drops prediction error by orders of magnitude.
  • History length matches compression factor: In the linear regime, the minimum history length $k$ required to reconstruct the state corresponds directly to the spatial compression ratio $n/m$. Increasing history beyond this threshold yields rapidly diminishing returns, confirming that the latent history acts specifically to invert the rank-deficient observation operator.
  • Latent compression stabilizes long rollouts: By confining the autoregressive rollout to a compact, low-frequency latent trajectory and letting a separate decoder reconstruct high-resolution features, world models avoid feeding high-frequency errors back into the time-stepper.

Key Results

  • Long-Term Rollout Stability: On 2D Heat and Wave equations, DeepONet and FNO diverged completely by frame 1000. In contrast, the world model maintained low error (e.g., $L_1 = 0.0051$, $\text{MSE} = 4.27 \times 10^{-5}$ on Heat at frame 1000).
  • Chaotic Dynamics (KSE): The adversarial U-Net captured temporal auto-correlations significantly better than FNO and DeepONet across 200 frames (Euclidean correogram error of $5.71$ vs. $6.15$ for DeepONet and $6.62$ for FNO).
  • Turbulent Flow Inference Speed: For the cylinder wake, LongVideoGAN preserved the statistical periodicity of vortex shedding (matching the $1\sigma$ correlation envelope of real LES data) while generating frames in 0.032 seconds—a >500× speedup over the LES solver (~17 seconds per frame).

Limitations and Future Work

  • Global Observability Assumption: The mathematical proofs assume continuous invertibility of the output trajectory map, which is difficult to guarantee a priori for highly turbulent, non-linear PDEs.
  • Chaotic Horizon Limits: While world models dramatically outperform one-step neural operators, temporal correlations still degrade over very long rollouts (e.g., 1000 frames) in chaotic regimes like KSE due to extreme sensitivity to initial conditions.
  • Ill-Conditioned Observability: Observability is formally a binary theoretical condition, but practical neural networks can struggle when a system is theoretically observable yet numerically ill-conditioned ("near-unobservable"). Quantifying degree-of-observability metrics remains an open problem.

Summary updated

Abstract

In this work, we explore the use of compact latent representations with learned time dynamics ('World Models') to simulate physical systems. Drawing on concepts from control theory, we propose a theoretical framework that explains why projecting time slices into a low-dimensional space and then concatenating to form a history ('Tokenization') is effective at learning dynamical systems, and characterise when the underlying dynamics admit a reconstruction mapping from the history of previous tokenized frames to the next, with the key novel insight being observability. To validate these claims, we develop a sequence of models with increasing complexity, starting with least-squares regression and progressing through simple linear layers, shallow adversarial learners, and ultimately full-scale generative adversarial networks (GANs). We evaluate these models on a variety of datasets, including modified forms of the heat and wave equations, the chaotic regime 2D Kuramoto-Sivashinsky equation, and a challenging computational fluid dynamics (CFD) dataset of a 2D Kármán vortex street around a fixed cylinder, where our model is successfully able to recreate the flow. Comparisons to FNO and DeepONet show comparable short- and improved long-term accuracy.
BibTeX
@misc{ross2025when,
  author = {Edmund Ross and Claudia Drygala and Leonhard Schwarz and Samir Mani Kaiser and Francesca di Mare and Tobias Breiten and Hanno Gottschalk},
  title = {{When do World Models Successfully Learn Dynamical Systems?}},
  year = {2025},
  eprint = {2507.04898},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.04898},
}
Figure