← All problems

Petersen–Raslan–Voigtlaender ReLU realization-space closedness conjecture

For every fixed ReLU-network architecture A\mathcal{A} and input domain Ω\Omega, the realization space RA(Ω)\mathcal{R}_{\mathcal{A}}(\Omega) of functions represented by networks with architecture A\mathcal{A} is closed in C(Ω)C(\Omega) with respect to the uniform topology; equivalently, if a sequence (fn)⊆RA(Ω)(f_n)\subseteq\mathcal{R}_{\mathcal{A}}(\Omega) converges uniformly on Ω\Omega to f∈C(Ω)f\in C(\Omega), then f∈RA(Ω)f\in\mathcal{R}_{\mathcal{A}}(\Omega).

References

Source: arXiv

Progress summary

Refreshed
Open · solution reported

A new preprint claims to refute the general conjecture with a concrete counterexample, while closedness still holds for some simpler network architectures.

Petersen, Raslan, and Voigtlaender conjectured that fixed-complexity ReLU realization spaces are closed in the uniform topology C(Ω)C(\Omega). Earlier work established important special cases but left the general question open.

Known results

  • Fixed-architecture ReLU realization spaces can be non-closed in LpL^p norms (Petersen, Raslan, and Voigtlaender, 2020).
  • Two-layer ReLU realization spaces are closed in the uniform norm, including on compact boxes.
  • Closedness was also proved for architectures with bounded scaling weights.

October 2026 counterexample

Kathryn A. Lindsey's preprint The Geometry of Limits in ReLU Function Spaces claims closedness for fixed-width shallow scalar architectures in one norm setting, but gives an explicit (2,3,2∣1)(2,3,2\mid 1) counterexample under weighted global and compact-uniform convergence. If correct, this refutes the broad closedness conjecture; the result is presently unverified.

Current status (as of October 2026): The general conjecture is claimed refuted by Lindsey's explicit counterexample, but that claim is unverified; closedness questions for restricted architectures and topologies remain open.

Sources

Solutions 0

No solutions have been posted yet.