When is a System Discoverable from Data? Discovery Requires Chaos
summary
The gist
The information presented in both sections is highly technical, focusing on the mathematical conditions under which dynamical systems can be uniquely identified (discovered) from observational data,
In short
The research investigates when a dynamical system can be uniquely identified from observational data. The core finding is that chaos is often necessary for this discovery to be provable, as chaotic systems provide enough complexity for unique identification. Conversely, conservative systems are inherently non-discoverable from data alone.
Key concepts
- Chaos
- Chaos describes a system's behavior where small changes in initial conditions lead to wildly different long-term outcomes. The paper argues that this inherent unpredictability is actually what makes the system's trajectory rich enough to uniquely identify the governing equations.
- Discoverability
- This refers to the mathematical ability to uniquely determine a system's governing equations from observing its behavior (trajectories). The study proves that for continuous systems, chaos ensures this uniqueness, a property that conservative systems often lack.
- First Integrals
- A first integral is a conserved quantity in physics—a value that remains constant along the system's trajectory. The paper shows that if an analytic system possesses such a non-constant global first integral, it cannot be uniquely discovered from data alone.
- Analytic Discoverability
- This is the strict mathematical requirement for finding equations using analytic functions (functions that can be represented by power series). The research shows this is impossible if a system has first integrals, but it can still be achieved under specific geometric conditions, such as those met by the Lorenz attractor.
Terminology used across episodes
This episode discusses
- When is a System Discoverable from Data? Discovery Requires Chaos · Paper Radio
- Pangu-Weather: A 3D High-Resolution Model for Fast and Accurate Global Weather Forecast
- Topological chaos: what may this mean ?
- Canonical Bayesian Linear System Identification
- Symplectic Neural Flows for Modeling and Discovery
- Identifiability Challenges in Sparse Linear Ordinary Differential Equations
- Scientific Machine Learning of Chaotic Systems Learns Reduced-Order Equations for Neural Populations
- Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl
- Lagrangian Neural Networks
- Learning Symbolic Physics with Graph Networks
- Lorenz like flows: exponential decay of correlations for the Poincar'e map, logarithm law, quantitative recurrence
- Robust identifiability for symbolic recovery of differential equations
- An example of a compact non-C-analytic real subvariety of R cubed
- Estimation of Lyapunov dimension for the Chen and Lu systems
- FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators
- A Survey of Deep Learning for Scientific Discovery
- Forecasting subcritical cylinder wakes with Fourier Neural Operators
- Extracting Sparse High-Dimensional Dynamics from Limited Data
- Symbolic Recovery of Differential Equations: The Identifiability Problem
- Deep Learning of the Evolution Operator Enables Forecasting of Out-of-Training Dynamics in Chaotic Systems
- Bridging the Gap Between Scientific Laws Derived by AI Systems and Canonical Knowledge via Abductive Inference with AI-Noether
The paper
When is a System Discoverable from Data? Discovery Requires Chaos · Read on arXiv
Zakhar Shumaylov, Peter Zaika, Philipp Scholl, Gitta Kutyniok, Lior Horesh, Carola-Bibiane Schönlieb
University of Cambridge · IBM Research · LMU Munich and Munich Center for Machine Learning (MCML)
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: Today's paper: "When is a System Discoverable from Data? Discovery Requires Chaos".
Jane: The information presented in both sections is highly technical, focusing on the mathematical conditions under which dynamical systems can be uniquely identified (discovered) from observational data,
Tom: First, who's behind it and why it matters.
Paper summary: Tom: So if we look at the title again, "When is a System Discoverable from Data? Discovery Requires Chaos", it really boils down to that idea that uniqueness in discovery hinges on chaotic dynamics, not just any kind of complex movement >
Jane: Right. And the authors really drive home that chaos isn't just some random property; it’s what provides the necessary complexity for the data to uniquely specify the model >
Meng: From a deployment standpoint, this tells us that if we want robust AI models for modeling physical systems, we need to look at whether those systems possess chaotic behavior or if they have those simple conservation laws that make discovery inherently non-unique >
Lu: The implication is that for many engineering applications where the system is well-behaved and predictable, you might actually struggle to find a unique governing equation just from data >
Lalam: They’re suggesting that when chaos isn't present, you probably can't rely on pure trajectory data to find the right model without adding external physical knowledge about conservation laws >
Jane: So it’s a call for hybrid approaches where we use the AI for pattern recognition but we use our physics knowledge to guide the search when the data alone doesn't provide enough certainty >
Conclusion: Tom: So we’re talking about this paper, “When is a System Discoverable from Data? Discovery Requires Chaos.” It’s by some folks in dynamical systems and they’re making a big claim about how we find governing equations from observational data.
Jane: Okay, so the main idea is that chaos isn't just some messy behavior; it seems to be what makes a system actually discoverable, which I know sounds counter-intuitive because chaos usually means unpredictability.
Lu: Exactly. They’re proving that without chaos—without that certain type of complexity—you can't uniquely identify the system from just the data you’ve collected, whether it’s continuous or analytic functions.
Meng: From a deployment standpoint, so if I have a system that doesn't exhibit this chaotic structure, I can’t rely on my AI to find the right model just by looking at trajectories.
Lalam: And they show that conservative systems with those first integrals—those conservation laws—are inherently non-discoverable from data alone; you need external knowledge to solve them.
Tom: So the authors are saying that for many physical systems we encounter, the chaos is actually a prerequisite for us to be able to find the right AI model in the first place.
Jane: It suggests that when AI excels at things like weather forecasting, it’s because those systems have this chaotic structure that makes them "discoverable" in a way other models aren't.
Lu: They even verified this experimentally with the classical Lorenz system, showing it is analytically discoverable, which is a pretty strong result for analytic functions.
Meng: But that doesn't mean they are easy to control; chaotic systems are often the hardest to control precisely once you’ve found the model.
Tom: So what this means for us in AI development is that we need to be careful about which systems we try to model with pure data discovery versus hybrid approaches.
Jane: It really shifts the focus from just fitting curves to understanding the underlying topological structure of the dynamics itself before we even start training a model.
Lu: And the next step is figuring out how to use those chaotic properties effectively in our representation learning frameworks, which is where my work fits in.
More episodes
- 2607.06772-Efficient Long-Horizon Learning for Learned Optimization
- 2509.20645-Anticipatory Evaluation of Language Models: Predicting LLM Benchmark Scores from Descriptions
- 2510.07231-EconCausal: A Context-Aware Economic Reasoning Benchmark for Large Language Models
- 2510.06225-Generalized Multi-agent Social Simulation Framework
- 2511.09568-VEDA: Generation of 3D Molecules via Variance-Exploding Diffusion with Annealing
- 2606.11245-Position: Hippocampal Explicit Memory Is the Cornerstone for AGI
- 2311.04902-Beyond Size: How Gradients Shape Pruning Decisions in Large Language Models
- 2505.20085-Explanation User Interfaces: A Systematic Literature Review
- 2507.12145-PRISM: Distributed Inference for Foundation Models at Edge
- 2507.14843-The Invisible Leash: Why RLVR May or May Not Escape Its Origin