← All AI/ML open problems

Symbolic regression & equation discovery · Embeddings & representations · Time-series learning / Open question

Finding Coordinates Where Dynamics Become Simple

High-dimensional measurements can hide a simple dynamical law, but a learned low-dimensional picture need not preserve its evolution or give uniquely meaningful coordinates. Efficiency payoff: Replace a high-dimensional simulator with a low-dimensional symbolic model while accounting for the cost and ambiguity of learning the manifold.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-20

Seed catalog · Showing the bundled record.

Problem statement

Let latent z∈[−1,1]rz\in[-1,1]^r follow a bounded degree-qq, ss-sparse polynomial ODE z˙=f∗(z)\dot z=f_*(z), observed through an unknown C3C^3 bi-Lipschitz embedding g:[−1,1]r→RDg:[-1,1]^r\to\mathbb R^D, r≪Dr\ll D. From noisy observation trajectories, jointly learn encoder EE, decoder GG and sparse symbolic field f^\hat f. Bound E∥G(Φtf^(E(g(z0))))−g(Φtf∗(z0))∥2\mathbb E\|G(\Phi_t^{\hat f}(E(g(z_0))))-g(\Phi_t^{f_*}(z_0))\|^2 uniformly for t≤Ttestt\leq T_{\rm test}, while identifying which aspects of f∗f_* are recoverable only up to latent coordinate changes.

Assumptions and scope

  • The latent cube is forward invariant; initial latent states are independent with density bounded above and below, but are not observed. Supply derivative, inverse-Lipschitz and observation-noise bounds, sampling interval and trajectory duration. Test initial observations are noiseless here to separate learning from filtering error.
  • The global-chart assumption excludes manifolds requiring multiple charts. A low reconstruction error does not prove that a chosen bottleneck is dynamically sufficient. Coordinate transformations conjugate vector fields, and sparsity can change under them; any preferred canonical coordinates need additional normalization or intervention information.
  • Specify finite encoder/decoder capacity, training loss, optimizer and symbolic library. Charge representation learning, stored trajectories and field integration. Predictions are evaluated on the same manifold with a bounded test/train initial-density ratio; behavior away from the observed manifold is not identified.

What would settle it?

A finite-sample joint reconstruction/dynamical-prediction theorem and an identifiability statement modulo explicit conjugacies, or indistinguishable models showing which coordinate or off-manifold claims cannot be recovered.

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

SINDy autoencoders already jointly learn coordinates and equations; WgLaSDI combines weak-form latent identification with active learning, and Higher-Order-KANDy uses delay/jet embeddings. The new target is a costed finite-data guarantee with latent-coordinate ambiguity made explicit, not the first connection between manifold and equation learning. 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. Champion et al. — Data-Driven Discovery of Coordinates and Governing Equations ↗related algorithm · source checked 2026-11-20

    Primary abstract checked: autoencoder and sparse latent dynamics learned jointly. Coordinate discovery and equation discovery are already combined; latent-coordinate identifiability is a separate guarantee.

  2. 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.

  3. Slote and Fish — Higher-Order Kolmogorov-Arnold Networks for Dynamics ↗related algorithm · source checked 2026-11-20

    Primary search abstract checked: delay/jet embeddings and KAN layers for higher-order noisy dynamics without a fixed candidate library. Coarse/noisy recovery is empirical, not a general finite-data identifiability theorem.

  4. Russo and Laiu — Convergence of Weak-SINDy Surrogate Models ↗partial theory · source checked 2026-11-20

    Primary search abstract checked: projection-based surrogate-dynamics and solution-error analysis under bounded composition-operator assumptions. Basic weak-SINDy convergence is not an open existence problem.