- Rating
- 1
- Confidence
- 1(−1)
- Informativeness
- 1(−1)
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper studies operator transfer when the source and target share continuous symmetries but have different symmetry-breaking residuals. The authors propose NAILONG, which identifies the shared Lie-generator span from source data, aligns it to the target coordinates, and freezes it while learning a target residual. The theoretical analysis covers target OLS risk reduction with correctly calibrated shared generators, group-projection limits for jointly equivariant predictors, and error floors when the residual cannot compensate for an incorrect freeze. Experiments on noisy generator regression and spherical fields examine finite-source identification, alignment, target sample efficiency, and adaptation to symmetry-breaking targets.
Critical strengths
The paper provides a useful theoretical distinction between the variance reduction from a correctly calibrated freeze and the bias from an incorrect freeze. Under the stated matched-design OLS and residual-class assumptions, this clarifies the conditions under which transferring shared generators can be beneficial. This gives a helpful basis for assessing the proposed transfer mechanism.
Critical weaknesses
- The practical benefit of transferring the identified generators is not sufficiently isolated. Section 6.4 states that generator OLS uses analytic velocity supervision and ordered control paths, while the Wigner baseline uses endpoint supervision and control summaries. These different information conditions affect the comparison. Therefore, the reported field gap does not establish the advantage of the identify-align-freeze strategy under matched transfer conditions.
- The inspected code raises additional consistency concerns. The resolution lift uses an oracle branch and true higher-resolution generators, while the main unknown-intertwiner grid reuses target samples for alignment and residual fitting. These choices differ from the stated learned-generator and disjoint-sample protocols. The authors need to clarify the version used for the reported results and verify the corresponding experimental claims.
Questions for authors
- Can you provide one matched field comparison with a source-trained trunk plus a free residual, using the same supervision, control information, residual capacity and total target budget?
- Which code version produced the reported resolution-lift and unknown-intertwiner results? Please clarify whether the lift uses identified generators and whether alignment samples are disjoint from residual-fitting samples.
- Are the ceiling and method risks evaluated under the same Haar-symmetrized joint law? How do you verify joint equivariance of the baseline to which Theorem 2 is applied?
- Please improve the figure layout and readability. On page 7, the spacing around Figures 3–4 and Table 1 could be adjusted to enlarge Figure 3, especially its legends and heatmap labels. The x-axis tick labels in Figure 2 are also crowded and should be simplified. These changes would make the results easier to check.