← All AI/ML open problems

Symbolic regression & equation discovery · Data-efficient learning · Control & robotics / Open question

Which Experiment Reveals the Equation?

Different equations can fit the same observed trajectory. The right initial condition or forcing can separate them far more effectively than simply collecting more points along that trajectory. Efficiency payoff: Reduce the simulations or measurements required for reliable equation discovery by selecting informative trajectories.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-20

Seed catalog · Showing the bundled record.

Problem statement

For x˙=fθ(x)+Gu\dot x=f_\theta(x)+Gu, let fθf_\theta be a degree-qq polynomial vector field with at most ss nonzero bounded coefficients and minimum nonzero magnitude βmin⁡\beta_{\min}. Choose each reset state from a known compact set and controls from bounded piecewise-constant signals with at most JJ switches. Given only noisy state samples, determine the least expected observation-and-duration cost for support recovery and ∥θ^−θ∥2≤ε\|\hat\theta-\theta\|_2\leq\varepsilon with probability 1−δ1-\delta, allowing experiments to depend on previous data.

Assumptions and scope

  • G, the coefficient bounds, control amplitude, reset set and maximum experiment horizon are known. Candidate dynamics admit unique solutions in a common bounded operating region. State sampling and reset/duration prices are explicit positive inputs; exact derivatives and zero-cost counterfactual simulations are unavailable.
  • Use independent Gaussian measurement noise and a minimum admissible sampling interval. Distinguish models that all permitted experiments leave observationally equivalent; no design can identify such a pair. Information or excitation conditions must be quantified over achievable experiments.
  • Charge planning/search over reset states and control signals, not just the samples acquired by an optimal design oracle. A proposed lower bound must apply to adaptive experiments; a positive algorithm must propagate uncertainty into its design and final support decision.

What would settle it?

Instance-dependent or minimax adaptive acquisition bounds with an implementable design algorithm and identification lower bounds, separating the cost of informative data from the computation of finding it.

Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.

Status and evidence

Ensemble-SINDy, active sparse model discovery and active latent-model learning already demonstrate adaptive acquisition. The added target is a sharp, costed fixed-confidence identification frontier for controlled polynomial dynamics, not the first active equation-learning method. PDE experiment design requires its own input/solution model and is not silently covered by this ODE theorem. This is an editorial, source-motivated research target. The preliminary review through 2026-11-20 does not certify novelty or exclude an unlocated resolution.

Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.

Sources and further reading

  1. Larrañaga, Fasel and Brunton — How Low Can You Go? Active Learning for Sparse Model Discovery in the Ultra-Low-Data Limit ↗related algorithm · source checked 2026-11-20

    Primary search abstract checked: ensemble uncertainty guides ODE/PDE sampling with empirical sample savings. An optimal adaptive experiment-complexity theorem is a different target.

  2. Fasel et al. — Ensemble-SINDy: Robust Sparse Model Discovery in the Low-Data, High-Noise Limit, with Active Learning and Control ↗related algorithm · source checked 2026-11-20

    Primary search abstract checked: bagged inclusion probabilities, forecasts, active learning and control. Inclusion frequency is not automatically a calibrated confidence statement about the true symbolic structure.

  3. He et al. — Physics-Informed Active Learning with Simultaneous Weak-Form Latent Space Dynamics Identification ↗related algorithm · source checked 2026-11-20

    Primary search abstract checked: joint autoencoder/WENDy training and active parametric sampling for reduced ODE/PDE models. Low-dimensional modeling and active selection already coexist in an implemented method.

  4. Brunton, Proctor and Kutz — Discovering Governing Equations from Data: Sparse Identification of Nonlinear Dynamical Systems ↗related algorithm · source checked 2026-11-20

    Primary abstract checked: sparse equation discovery, parameterized/time-varying/forced examples and interpretable models are established. A fixed candidate library is an inductive assumption, not automatic discovery of every physical law.