← All AI/ML open problems

Symbolic regression & equation discovery · Embeddings & representations · Neural architecture & complexity / Open question

When Does a KAN Reveal a Real Formula?

An attractive edge plot or low test error need not identify the generating equation. Equivalent decompositions and near-indistinguishable formulas can make symbolic conclusions unstable. Efficiency payoff: Turn a trained surrogate into a compact auditable formula while reducing expensive symbolic search and preserving predictive behavior.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-20

Seed catalog · Showing the bundled record.

Problem statement

Let Gs,b\mathcal G_{s,b} be expressions on [0,1]d[0,1]^d with at most ss nodes using +,×,sin⁡+,\times,\sin and rational constants of at most bb bits, with all intermediate values bounded by RR. Observe Yi=g∗(Xi)+ξiY_i=g_*(X_i)+\xi_i for uniform XiX_i and Gaussian noise. Assume distinct represented functions in Gs,b\mathcal G_{s,b} are separated by at least Δ\Delta in L2L^2. Determine when spline-KAN training followed by joint symbolic replacement recovers g∗g_* up to functional equivalence with probability 1−δ1-\delta, without exhaustive grammar enumeration.

Assumptions and scope

  • The grammar, size/bit/range bounds and noise variance are fixed in advance. Functional equivalence identifies expressions with the same values throughout the cube; hidden-node rescalings, redundant paths and algebraically identical syntax do not count as different recovered laws.
  • Charge learned-spline accuracy, operator-library search, numerical fitting, validation and any active queries. An independent-edge fit can alter downstream inputs, so a theorem must control the complete substituted network.
  • The separation promise makes statistical identification meaningful. A finite-class testing bound and exhaustive search are baselines; seek polynomial dependence on d,s,b,1/Delta,log(1/delta) or a conditional hardness result. Observational formula recovery is not causal identification or an automatic extrapolation theorem.

What would settle it?

A finite-noise identification and computational guarantee for a concrete KAN-assisted extraction algorithm, or a hardness/ambiguity result explaining the additional structural assumptions required. Approximate function fitting alone does not settle exact expression-class identification.

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

KAN 2.0 already compiles and extracts formulas; context-aware symbolic regression addresses coupled replacements, and PH-Symb preserves dynamical structure by construction. The new target is an identifiability/sample/computation guarantee for a finite grammar, not first symbolic extraction or a claim that learned coordinates are uniquely meaningful. 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. Liu et al. — KAN 2.0: Kolmogorov-Arnold Networks Meet Science ↗related algorithm · source checked 2026-11-20

    Primary abstract checked: multiplication nodes, symbolic-to-KAN compilation, modular structure and formula extraction already exist. Scientific-law recovery requires identifiability beyond a visually simple network.

  2. Sovrano et al. — In-Context Symbolic Regression for Robustness-Improved Kolmogorov-Arnold Networks ↗related algorithm · source checked 2026-11-20

    Primary search abstract checked: greedy context-aware operator replacement and gated matching pursuit mitigate independent-edge extraction errors empirically. Global operator selection is already proposed; stable finite-noise recovery remains a stronger target.

  3. El Messaoudi et al. — PH-Symb: Sparse Symbolic Identification of Nonlinear Port-Hamiltonian Systems Using Kolmogorov-Arnold Networks ↗related algorithm · source checked 2026-11-20

    Primary search abstract checked: structure-preserving symbolic identification with KANs and sparse regression. Preserving port-Hamiltonian form and empirical recovery do not alone establish general identifiability of arbitrary extracted formulas.

  4. Liu et al. — KAN: Kolmogorov-Arnold Networks ↗partial theory · source checked 2026-11-20

    HTML Sections 2.1–2.5 and Theorem 2.1 checked: smooth-edge spline approximation with representation-dependent constants, grid extension, pruning and empirical training. The representation theorem does not provide smooth low-cost decompositions for every target or a gradient-training guarantee.