summary: New research uses Koopman operator learning to prove that certain complex and chaotic systems have fundamental predictability limits that cannot be overcome with infinite training data. By designing an adversarial system that maps where machine learning models break down, the team explained the mathematical root causes of LLM hallucinations while introducing a highly efficient algorithm with built-in error bounds that successfully mapped hidden Arctic sea ice patterns using a standard laptop.
important facts
- The myth of infinite data has been debunked: This study shows that the technology industry’s common philosophy that “more data equals guaranteed learning” is mathematically incorrect. Certain highly complex or chaotic problems are characterized by hidden, layered patterns that are impossible to cleanly separate. This means that the absolute highest score an algorithm can score is a coin toss ($50/50$) and the problem cannot be solved mathematically, regardless of the size of the dataset.
- Why chatbots hallucinate: Mathematical instabilities that defeat long-term physical predictions explain why large-scale language models (LLMs) like ChatGPT and Claude confidently produce false information over time. In highly sensitive systems, minute changes in the starting prompt can cause compound errors, sending the model down widely disparate paths, completely losing touch with reality while maintaining short-term consistency.
- Two pillars of machine learning failure: Dr. Matthew Colbrook’s team has identified two specific structural reasons why AI modeling fails naturally when interacting with complex environments.
- Validation failure due to missing data: Machine learning algorithms have no internal mathematical mechanism to make decisions. when Enough training samples have been captured to output stable and reliable predictions.
- Obfuscation of hidden patterns: Critical tracking coordinates within dynamic architectures remain mathematically hidden or are so deeply intertwined that they cannot be distinguished by standard neural nets.
- Chaotic frequency problem: When AI analyzes chaotic systems (small changes in starting parameters cause large divergences), the Koopman operator produces a continuous spread of overlapping frequencies rather than a clean, isolated tracking variable. This explains why short-term predictions remain accurate while long-term system predictions systematically collapse.
- Proven and reliable algorithms: To address this structural weakness, the researchers designed a new mathematically rigorous algorithm that incorporates an invariant error bound. The toolkit provides researchers with a real-time certainty meter to verify exactly when AI output can be trusted without the need for multi-million dollar supercomputers.
- Laptop and supercomputer benchmarks: Stress-tested against 40 years of Arctic climate records, the team’s custom algorithm identified long-lost patterns of structural collapse in the ice sheet. It consistently outperformed the world’s leading commercial AI systems, ran entirely on basic consumer-grade standard laptops, and was a fraction of the computational cost.
sauce: cambridge university
When can we trust the results we get from AI? And when is learning impossible? Researchers have shown that there are some problems that even the most powerful AI can reliably solve, regardless of the amount of data it is given.
Researchers at the University of Cambridge and the University of California, Santa Barbara have designed an “adversarial” mathematical system aimed at fooling any AI algorithm. Similar to ethical hackers stress-testing the security of networks, these adversarial systems are designed to precisely map where and why AI predictions fail.
Many real-world systems, such as the ocean, the human brain, and robotics, are too complex to be neatly explained by equations, so researchers often use machine learning to learn how systems behave. However, these AI methods don’t always work well, returning unreliable results or incorrect predictions.
However, even with infinite data, there are times when it is fundamentally impossible to provide a reliable solution. The adversarial system developed by the researchers could help developers and users of AI systems know whether they are working on solvable or unsolvable problems, build methods that work, and avoid wasting time, effort, or AI tokens when a problem is beyond the bounds of possibility.
Their results reported in the journal nature communicationsalso helps explain why popular AI chatbots such as ChatGPT and Claude, while accurate in the short term, can drift and hallucinate over time.
“We are exploring the boundaries of what AI can and cannot do,” said lead author Dr Matthew Colbrook from the University of Cambridge’s Department of Applied Mathematics and Theoretical Physics. “It is very important to understand what problems these methods cannot solve, because otherwise you will be wasting a lot of time and money.”
Colbrook and his co-authors used an approach called Koopman operator learning, which converts complex nonlinear behavior into a linear form that is easier to analyze.
“What we were doing with these ‘adversaries’ was trying to figure out the types of systems that are difficult or impossible to predict and the types of systems that can adapt to return reliable results,” Colbrook said.
Researchers have identified two main reasons why machine learning fails when analyzing complex systems. Either the algorithm cannot determine when enough data is available to return reliable results, or the patterns in the system are hidden or difficult to distinguish.
“The general assumption in a lot of AI research is that if you collect more data, the learning will eventually get better,” Colbrook says. “However, we find that this is often wrong. Learning is often layered and requires multiple steps to be performed in the proper order.”
When a system is chaotic, that means small differences in starting conditions can lead to wildly different trajectories, like choosing your own adventure story. Koopman operators often end up with: Continuous A spread of frequencies rather than a clean, clear mode. Although short-term predictions were accurate, long-term predictions are essentially unreliable because their sensitivity to initial conditions worsens over time.
The same mathematical instability that defeats predictive algorithms may also explain why AI chatbots so confidently fabricate facts. A small change in the question can send the chatbot down a completely different path, and while it may seem plausible word for word, the longer output loses its grip on reality.
Researchers have developed a way to categorize these problems based on the number of steps required to solve them. If the data is not hierarchical enough or in the correct order, even with infinite data, the best the algorithm can do is 50/50, essentially classifying the problem as unsolvable.
The team also created a new reliable and efficient algorithm with built-in error bounds. This will allow AI researchers to know when they can trust an answer at a fraction of the cost of most supercomputers.
The researchers tested their approach based on more than 40 years of Arctic sea ice data. Using their algorithm, they were able to discover hidden patterns in how ice diminishes and outperform today’s leading AI models at a fraction of the cost on a standard laptop.
“We’re at a stage now where there are a lot of flashy examples and success stories in AI, but it’s also important to ask how robust are the models and how do we know if they’re robust,” Colbrook said. “Otherwise we are building on a very shaky foundation.”
Answers to key questions:
answer: Imagine trying to predict the path of smoke rising from a campfire. Smoke twists, loops, and breaks apart in highly complex, nonlinear ways that are nearly impossible to track with basic equations. The Koopman Operator is a mathematical technique that takes this messy, nonlinear behavior and maps it onto an alternative abstract landscape where the motion behaves like a straight line. By converting complex systems into this linear form, Dr. Matthew Colbrook’s team was able to create an “adversarial network” that stress-tests the mathematics and pinpoints where equations break down and become impossible for AI to solve.
answer: Chatbots process text the same way a chaotic weather system processes atmosphere word by word. The path depends entirely on the starting point. When a system is chaotic, small changes in the initial input cause the paths to follow very different trajectories. In chatbots, changing a single letter or word in a long prompt can deviate the AI from the factual path. Listened word for word, the response sounds perfectly plausible, but over a long period of output, the sensitivity to that small initial change adds up and the model moves away from reality, creating illusions.
answer: Technology departments need to stop blindly throwing massive computing power and unmanaged data sets at complex problems. Instead, developers can use a new algorithm from the Cambridge-UCSB team that includes error bars. Think of it like a built-in dashboard gauge that shows you exactly how reliable your AI model’s output is. Using this method, scientists can instantly identify whether complex problems are solvable or fundamentally impossible, saving millions of dollars, reducing wasted supercomputing time, and highlighting hidden patterns using basic, affordable laptops.
Editorial note:
- This article was edited by the editors of Neuroscience News.
- Journal articles were reviewed in full text.
- Additional context added by staff.
About this AI research news
author: sarah collins
sauce: cambridge university
contact: Sarah Collins – University of Cambridge
image: Image credited to Neuroscience News
Original research: Open access.
“Adversarial Dynamical Systems Characterize the Success or Failure of Data-Driven Learning” by Matthew J. Colbrook, Igor Mezic, and Alexei Stepanenko nature communications
DOI:10.1038/s41467-026-74220-8
abstract
Adversarial dynamic systems characterize whether data-driven learning succeeds or fails
Many systems resist analytical modeling, making data-driven inference of dynamics important. However, data-driven methods can fail to converge or generalize, leaving open core questions such as: When can a system’s behavior be reliably learned from data, and when is such learning not possible?
We answer this question by using an adversarial dynamics system to identify the boundaries between accessible and inaccessible regimes. Koopman operator learning, a leading framework for representing nonlinear dynamics through linear spectral objects, designs optimal data-driven spectral algorithms with convergence and authentication guarantees under conditions that commonly occur in physical systems.
This provides a convergence theory for the Koopman operator approximation and solves a long-standing open problem in Koopman spectral analysis. Conversely, by building an adversarial system, results are proven that are impossible to match. Without these conditions, a single-sequence restriction step cannot guarantee learning, regardless of the quality of the data. These results clearly characterize when data-driven spectral learning must succeed and when it must fail. We validate our framework for oscillators, chaotic fluid flows, and Arctic sea ice concentration prediction.
The latter reveals hidden modes of Arctic sea ice loss, provides long-term forecasts with geographic error bounds, outperforms state-of-the-art dynamic and deep learning models at significantly lower computational cost, and enables real-time deployment on standard CPUs.
