- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper studies rotational drift in diffusion models and shows that, under isotropic dissipation, it is essentially a gauge freedom: it changes trajectories but not the generated distribution. The authors characterize when this invariance holds or breaks, analyze how rotation becomes visible under model misspecification, and validate the theory with analytic calculations, synthetic experiments, MNIST-based evaluations, and limited trained-model checks.
Critical strengths
-Provides a clear, exact analysis of rotational gauge freedom in a tractable class of linear diffusion processes, with extensions to a conformal class allowing anisotropic noise. -Offers a useful conceptual distinction between marginal equivalence under coordinate changes and path-space irreversibility, quantified by an explicit stationary entropy-production formula. -Provides controlled numerical checks through synthetic examples, closed-form score evaluations on the MNIST16 empirical distribution, and direct SDE simulations, while explicitly distinguishing these from trained-model experiments
Critical weaknesses
-Proposition 4(ii) appears incorrect under its stated assumptions. It claims that blocks with a common rotation speed have a common visibility-window position, whereas R4 reports different peak positions for different geometries at the same rotation strength. These geometries can be combined as independent blocks satisfying the proposition’s assumptions.The exact additive decomposition remains valid, but its claimed consequence for peak locations does not follow. This is a fixable issue, but it undermines an advertised conclusion about heterogeneous high-dimensional systems and requires additional assumptions or a weaker statement.
-The broader implications for learned diffusion models are insufficiently established. The main results concern an exactly solvable class of linear diffusion processes. While relevant to forward-process design, the coordinate-level equivalence does not by itself justify the assertion that even imperfect learners cannot detect rotation. Most experiments evaluate closed-form scores, and the reduced training pipeline removes the rotation parameter by construction. Thus, the paper provides limited evidence about finite-capacity score learning and practical sampling beyond an explicitly transformed pipeline. This gap limits its contribution to understanding learned diffusion models. The broader claims should be qualified by explicit assumptions on the learner and numerical scheme, or supported by direct training and sampling in the original coordinates.
Questions for authors
-Clarification of Proposition 4(ii). How do you reconcile the claim that blocks with a common rotation speed have a common visibility-window position with R4’s different peak positions for different geometries at the same speed? Does this conclusion require additional assumptions on the block distributions, or is it intended as an approximation rather than an exact statement? Please distinguish the rigorously established peak-location conclusions from the empirical observations.
-Clarification of learner-level invariance. Does the assertion that “even an imperfect learner cannot detect the linear rotation” apply only to the explicitly transformed co-moving-frame pipeline, or also to models trained and sampled directly in the original coordinates? For the broader interpretation, what assumptions on the hypothesis class and numerical scheme ensure invariance, and what theoretical or experimental evidence supports it beyond the reduced pipeline, where rotation does not enter training?