Back to search
2402.18377

Out-of-Domain Generalization in Dynamical Systems Reconstruction

Niclas Göring, Florian Hess, Manuel Brenner, Zahra Monfared, Daniel Durstewitz

correctmedium confidence
Category
Not specified
Journal tier
Strong Field
Processed
Sep 28, 2025, 12:56 AM

Audit review

The paper’s Theorem 4.2 constructs an infinite family of C^1 vector fields that match the ground-truth vector field on the training basin but differ on the test basin, guaranteeing strictly positive test-domain error by appealing to their general error bound (Theorem 3.3) and a smooth bump-function replacement on a compact subset of the test basin; see theorem statement and proof sketch in the main text and Appx. E.3 , together with the definitions of E_stat and E_top in Defs. 3.1–3.2 . The candidate solution reaches the same conclusion via a more constructive modification (local hyperbolic sink plus a transport corridor) that yields explicit uniform lower bounds for both the topological and statistical (SW1) errors on M_test. Minor fixable technical gaps appear in the model’s Step 3 (where cancellation of f should be guaranteed on the whole contraction region), and in the paper’s phrasing that the whole new basin lies inside V; neither affects the main conclusion. Net: both arguments are valid and consistent, with different proof techniques.

Referee report (LaTeX)

\textbf{Recommendation:} minor revisions

\textbf{Journal Tier:} strong field

\textbf{Justification:}

The work provides a rigorous, conceptually clear negative result for OOD learnability in DSR under multistability and non-transitivity, aligning well with empirical observations. The proof strategy in Appx. E.3 is standard and sound; the main text frames the result appropriately. A few phrasing and completeness improvements (e.g., basin containment statement; explicit constants for the error lower bound) would enhance precision but do not undermine correctness. The candidate solution offers a complementary constructive approach that reinforces the result’s intuition.