When is a System Discoverable from Data? Discovery Requires Chaos

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, with a central thesis revolving around the necessity of chaos for this identifiability.

Here is a comprehensive, detailed synthesis combining the core arguments and specific findings from both sources:

Comprehensive Research Synthesis: When is a System Discoverable from Data? Discovery Requires Chaos

This body of research investigates the fundamental question of system discoverability: under what mathematical conditions can a dynamical system, defined by its governing equations (be they continuous or analytic), be uniquely identified from a finite set of observational data, such as trajectories? The central and counter-intuitive thesis established across these findings is that chaos is not merely beneficial but appears to be a crucial ingredient for ensuring provable system discoverability.

Core Thesis: Chaos as the Key Ingredient

The overarching conclusion is that within the space of continuous or analytic functions, chaos is almost necessary for provable discoverability. The authors rigorously establish this link by demonstrating that chaotic systems possess the requisite informational content in their trajectories to guarantee uniqueness. Conversely, non-chaotic systems often suffer from structural limitations that preclude unique identification without external constraints.

Mathematical Foundations of Discoverability

The paper builds a rigorous mathematical framework connecting topological properties of the system's dynamics to its identifiability:

1. Chaotic Systems and Uniqueness: The study proves that a dynamical system exhibiting chaos across its entire domain is uniquely identifiable from a single trajectory within the space of continuous functions. This is further formalized by showing that continuous discoverability from a finite number of trajectories implies a specific structural property: the system must possess a Smale-like decomposition where each cell exhibits topological transitivity, which is defined as a hallmark of chaotic dynamics.

2. Analytic Discoverability and First Integrals: A critical distinction is drawn between continuous and analytic functions. The authors establish a strong negative result: analytic discoverability is impossible in the presence of first integrals (conserved quantities). Theorem 6.1 (Informal) explicitly states that if an analytic vector field F admits a non-constant global analytic first integral G, then no trajectory of F can serve as a set of uniqueness, and consequently, the system is not analytically discoverable.

3. Geometric Conditions for Analytic Discoverability: While first integrals preclude analytical discoverability, the authors do show that analytic discoverability is achievable under specific geometric conditions on an attractor (related to Hausdorff dimension). For instance, Corollary 4.11 confirms that the classical Lorentz attractor is analytically discoverable.

4. Counter-Intuitive Findings: The research highlights a significant tension: Our analysis shows, counter-intuitively, that chaos, a property typically associated with unpredictability, is a crucial ingredient for ensuring a system is discoverable. This finding suggests that the complexity inherent in chaotic dynamics provides the necessary non-linearity and trajectory divergence required for unique identification.

Implications for Control and Modeling

The mathematical findings have direct implications for practical applications:

• Control Applications: The paper concludes that while chaotic systems are most amenable to data-driven discovery, they are often the least desirable or most challenging systems for control applications. This is because their inherent sensitivity makes precise, long-term prediction difficult.

• Hybrid Approaches: To recover uniqueness for non-chaotic systems (which might possess first integrals), the authors advocate for hybrid, knowledge-informed data-driven approaches. This involves integrating known physical priors, such as conservation laws (first integrals), into the estimation framework to constrain the search space and enable discovery where pure trajectory data alone fails.

Stability and Computational Rigor

Beyond theoretical existence proofs, the research provides computational guarantees regarding the stability of these discovery methods:

• Estimator Convergence: The paper establishes rigorous mathematical bounds for Ordinary Least Squares (OLS) estimators when dealing with noisy observations in chaotic systems. The analysis shows that the estimation variance simplifies under specific conditions, leading to convergence bounds based on regularization parameters (lambda).

• Computational Pipeline: The discovery process is formalized as a pipeline where a discovered model is sought by minimizing an objective function: = theta | theta - u| squared + lambda r| theta| squared, where the system is deemed discoverable at trajectory length T if the Out-of-Distribution (OOD) error falls below a threshold epsilon thr.

• Empirical Bounds: The minimum required trajectory length (T) for discovery can be empirically fitted, showing it scales with the Kolmogorov time (tau KS), suggesting that even in chaotic regimes, there is a finite informational requirement.

Summary of Key Results and Keywords

| Concept | Finding/Result |

|:--- |:--- |

| Core Claim | Chaos is almost necessary for provable discoverability within continuous or analytic function spaces. |

Chaos to Identifiability A system exhibiting chaos across its entire domain is uniquely identifiable from a single trajectory in continuous functions.

First Integrals to Non-Identifiability The presence of a non-constant global analytic first integral precludes analytical discoverability (Theorem 6.1).

Analytic Discoverability Achievable only under specific geometric conditions on the attractor; the Lorentz attractor is an example (Corollary 4.11).

Structural Requirement Continuous discoverability from finite trajectories implies a Smale-like decomposition with topologically transitive cells (Corollary 6.5).

| Practical Constraint | Chaotic systems are data-rich but control-poor; hybrid methods integrating physical priors are necessary for non-chaotic systems. |

Relevant Keywords: Chaos, Topological Transitivity, Identifiability, Discoverability, Symbolic Discovery, Neural Operators, Dynamical Systems.

1. textbf Scientist System Discovery (Continuous Case): The improved system can uniquely identify governing equations from a single trajectory if it is chaotic on its entire domain, as stated by A system, chaotic on the whole domain is discoverable (from a single trajectory) in the space of continuous functions C0. This allows for robust model discovery in inherently complex physical systems like weather forecasting where chaos is prevalent.

2. textbf Symbolic Discovery Enhancement (Analytic Case): The system can achieve analytic discoverability for systems chaotic on an attractor if its Hausdorff dimension of the attractor is greater than d − 1, as per A system, chaotic on an attractor, is discoverable (from a single trajectory) in the space of analytic functions if the Hausdorff dimension of the attractor is greater than d − 1. This enables the analytical recovery of systems like the Lorenz system using data.

3. textbf Knowledge-Informed Model Recovery: The system can recover non-chaotic systems by incorporating physical priors, as demonstrated by Theorem 6.5, which shows that a system is discoverable within the space of vector fields with symmetric (skew-symmetric) Jacobian if it admits an analytic conservation law." This allows for the discovery of stable, predictable systems like harmonic oscillators when augmented with known physical constraints.

4. textbf Robustness Against Noise: The improved system can estimate the true dynamics reliably under noisy observations using ordinary least squares (OLS), as quantified by Theorem 5.3, which provides a finite-time stability bound for the OLS estimator Fˆ T. This ensures that model uncertainty and out of domain error go to zero precisely at chaos onset" when noise is sufficiently small.

5. textbf High-Dimensional Operator Learning: The system can perform high-dimensional function space reconstruction by leveraging Hilbert-Schmidt exponential mixing, ensuring that the empirical average of the covariance operator satisfies a bound where the spectral norm of the empirical Gram matrix satisfies Zˆ T∥op ≤ s 2K0ZG squared Lip HS gamma T. This extends reconstruction capabilities to infinite-dimensional function spaces relevant for PDEs.