Summary: A new study uses Koopman operator learning to show that some complex, chaotic systems have fundamental limits to predictability that cannot be overcome simply by adding more training data. By constructing adversarial dynamical systems that expose where data-driven models must fail, the research team revealed mathematical causes of long-horizon errors — including the hallucinations seen in large language models — and introduced a computationally efficient algorithm with provable error bounds that uncovered hidden Arctic sea ice patterns while running on an ordinary laptop.
Key Facts
- The infinite-data myth challenged: The study proves that the common assumption “more data always yields better learning” is false for a class of highly complex or chaotic problems. When essential patterns are layered or intertwined in certain ways, no algorithm can outperform random guessing (50/50), no matter how much data is available. These problems are mathematically unsolvable by data-driven methods alone.
- Why chatbots hallucinate: The same mathematical instabilities that prevent reliable long-term physical prediction help explain why large language models (LLMs) like ChatGPT or Claude can produce confident but false statements over extended outputs. In systems with extreme sensitivity to initial conditions, tiny differences in a prompt amplify into divergent trajectories: short responses may appear coherent, yet longer outputs can lose contact with factual reality.
- Two structural reasons machine learning breaks down: Dr. Matthew Colbrook and colleagues identified two fundamental failure modes that make AI modeling unreliable in complex environments:
- Inability to verify data sufficiency: Standard learning algorithms lack an internal, mathematically rigorous test to determine when they have seen enough samples to produce a provably stable prediction.
- Hidden or entangled patterns: Key coordinates or modes that determine system behavior can be hidden or deeply mixed, so conventional neural networks cannot separate them into reliable predictors.
- The chaos-frequency phenomenon: For chaotic systems, Koopman operator spectra form a continuous band of overlapping frequencies instead of discrete tracking modes. This structure explains why short-term forecasts remain usable while long-term forecasts systematically fail as sensitivity to initial conditions compounds.
- Provably reliable algorithm with error bounds: The researchers developed a new, mathematically rigorous algorithm that delivers built-in error bounds and convergence guarantees under realistic conditions. This provides a practical “certainty gauge” so users can know when model outputs are trustworthy, without requiring massive supercomputing resources.
- High performance on ordinary hardware: Applied to more than four decades of Arctic sea ice data, the algorithm discovered previously hidden modes of ice decline, produced long-range forecasts with geographic error bounds, and outperformed leading commercial and research models while running on a standard laptop at a fraction of the computational cost.
Source: University of Cambridge
When is AI trustworthy, and when is learning fundamentally impossible? The researchers show clear mathematical criteria that separate problems a data-driven method can learn from those it cannot, even with unlimited data.
Researchers at the University of Cambridge and the University of California, Santa Barbara designed adversarial dynamical systems to probe the limits of data-driven learning. Just as security researchers craft test cases to reveal weaknesses, these adversarial systems demonstrate where and why learning algorithms must fail, allowing practitioners to recognize unsolvable tasks before wasting resources.
Many real-world phenomena — from turbulent fluid flows to neural dynamics and climate processes — resist simple analytic models, so scientists turn to machine learning to infer behavior from observational data. However, that approach can break down: sometimes it produces unreliable or unstable predictions, and sometimes the failure is provably unavoidable.
The team used Koopman operator learning, a framework that represents nonlinear dynamics through linear spectral objects, converting messy nonlinear behavior into a form that can be analyzed with linear tools. By testing this representation against purpose-built adversarial systems, the researchers classified when Koopman-based and related spectral methods converge and when they cannot.
Their analysis reveals that many learning tasks are inherently layered: they require information to be revealed in a particular sequence for algorithms to succeed. If the necessary layers are missing or scrambled, no algorithm can assemble the hidden structure from observations alone. In those cases the best possible performance is equivalent to random choice.
This same sensitivity underlies long-horizon errors in language models. In chaotic-like settings, small changes in the input prompt can redirect generation along a different trajectory. The model’s token-by-token fluency can hide a diverging path that, over many steps, produces fabricated or incorrect content.
To address these issues, the authors designed optimal data-driven spectral algorithms with provable convergence under broadly applicable conditions. The method provides certified error bounds and a real-time measure of reliability, enabling researchers and engineers to determine when a model’s output should be trusted.
On Arctic sea ice concentration forecasting, the algorithm uncovered hidden modes of decline and delivered long-range forecasts with quantified geographic error bounds. The approach outperformed state-of-the-art dynamical and deep learning models while running in real time on standard central processing units, demonstrating that provable, efficient learning is achievable without massive hardware.
“It’s crucial to understand the limits of data-driven methods,” said lead author Dr Matthew Colbrook. “Knowing which problems are unsolvable prevents wasted effort and helps us build methods that provide reliable, certified predictions.”
Key Questions Answered:
A: A Koopman operator maps nonlinear system behavior into a linear spectral representation. By analyzing that linearized form, the team could design adversarial examples and optimal spectral algorithms that reveal when data-driven learning converges and when it provably cannot. This approach exposes the mathematical boundary between solvable and unsolvable inference tasks.
A: Chatbot generation can behave like a chaotic dynamical system: tiny changes in an initial prompt can steer the output along very different paths. Over many steps, sensitivity to those early differences compounds, producing fluent but factually incorrect continuations. The mathematics of Koopman spectra helps describe why such drift occurs.
A: Rather than indiscriminately scaling data and compute, developers should adopt methods that provide certification and error bounds. The Cambridge–UCSB algorithm offers a built-in certainty indicator so teams can determine whether a problem is learnable, avoid wasted computation, and focus effort where reliable predictions are possible.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context added by staff.
About this AI research news
Author: Sarah Collins
Source: University of Cambridge
Contact: Sarah Collins – University of Cambridge
Image: The image is credited to Neuroscience News
Original Research: Open access. “Adversarial dynamical systems characterize when data-driven learning succeeds or fails” by Matthew J. Colbrook, Igor Mezić & Alexei Stepanenko. DOI: 10.1038/s41467-026-74220-8
Abstract
Adversarial dynamical systems characterize when data-driven learning succeeds or fails
Many physical and engineering systems resist analytic modeling, making data-driven inference of their dynamics essential. Yet data-driven methods can fail to converge or generalize, leaving a fundamental question: when can system behavior be learned reliably from data, and when is such learning impossible?
This work answers that question by using adversarial dynamical systems to identify the boundary between accessible and inaccessible regimes. In the Koopman operator learning framework — which represents nonlinear dynamics through linear spectral objects — the authors design optimal data-driven spectral algorithms with convergence and certification guarantees under conditions that commonly arise in real systems.
The analysis produces a convergence theory for Koopman-operator approximations and resolves a longstanding open problem in Koopman spectral analysis. Conversely, by constructing adversarial systems, the authors prove matching impossibility results: if the required structural conditions are absent, no single-sequence limiting procedure can guarantee learning, regardless of data quality. These results precisely characterize when data-driven spectral learning can succeed and when it must fail. The framework is validated on oscillators, chaotic fluid flows and Arctic sea ice forecasting, where hidden modes are revealed and efficient, certified forecasts are achieved on standard CPUs.