- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper considers the use of parallel transport in geometric machine learning, and specifically considers when such transports can be replaced by a fixed reference. Towards this, the paper introduces the Bundle-GEM methodology, which generalizes expectation maximization to the setting where sufficient statistics live in varying fibers over curved data. Roughly, this is achieved by applying GEM in each fiber and then transporting to the next fiber. The paper specifically considers this setup for the case of the tangent bundle of a sphere and the torus. The methodology is validated on weather data, roughly corresponding to a real-world version of the tangent bundle on a sphere.
Critical strengths
The paper identifies a problem in the geometric machine learning literature about when parallel transport maps are useful, and introduces a statistical approach to this problem, which seems like a partly novel approach compared to most previous empirical analyses.
Critical weaknesses
W1 The presentation of the paper is quite poor. The main body frequently references equations and theorems that are only in the appendix, and it is not often clear from the context in the main body what the theorem actually shows. Even the description of the main algorithm in the appendix refers to another appendix within the statement of the algorithm itself without explanation. Also, the paper generally reads as quite heavily AI-written and does not provide intuitive explanations for concepts or results.
W2 The theoretical basis of the work seems quite weak overall. As far as I can tell, the theorems in the main body only ever address the case of the tangent bundle on the sphere and torus. This provides an extremely narrow use case, and makes the overall theoretical grounding of the methodology being introduced incredibly limited. As a general note, it also presents several results as novel theorems, where in the mathematical literature, these would only be considered either examples or derivations due to their lack of generality. This also feels like an artifact of the paper possibly being largely AI-generated.
W3 It is not clear what the broader applications are of the methodology being introduced in the present work. The experimental use cases are quite narrow in scope and further, it is not clear how the limited theoretical analysis and experiments considered in the present work interact with the empirical results previously studied for learned vs fixed transport maps in related settings. In particular, the authors do not seem to provide a thorough comparison with existing literature on the subject. The question of whether a learned or fixed transport map is superior has been considered to some extent in several previous research works, particularly in the cellular sheaf-based machine learning literature, for instance, in [1],[2],[3],[4],[5]. Notably, the particular mathematical notation and setup used matches that of [5], but the authors only mention [3] in passing in their extended related works and do not provide a deep comparison of the findings of their methods with those of these previous works. As such, it is overall hard to judge the novelty of the methodology being introduced since it is not being placed in the context of relevant existing literature.
[1] Federico Barbero, Cristian Bodnar, Haitz Sáez de Ocáriz Borde, Michael Bronstein, Petar Veličković, and Pietro Liò. Sheaf neural networks with connection Laplacians. In Proceedings of Topological, Algebraic, and Geometric Learning Workshops 2022, volume 196 of Proceedings of Machine Learning Research, pp. 28–36, 2022.
[2] Claudio Battiloro, Zhiyang Wang, Hans Riess, Paolo Di Lorenzo, and Alejandro Ribeiro. Tangent bundle convolutional learning: from manifolds to cellular sheaves and back. IEEE Transactions on Signal Processing, 72:1892–1909, 2024
[3] Cristian Bodnar, Francesco Di Giovanni, Benjamin Chamberlain, Pietro Liò, and Michael Bronstein. Neural sheaf diffusion: A topological perspective on heterophily and oversmoothing in GNNs. In Advances in Neural Information Processing Systems, volume 35, pp. 18527–18541.
[4] Leonardo Di Nino, Gabriele D’Acunto, Sergio Barbarossa, and Paolo Di Lorenzo. Learning the structure of connection graphs. In ICASSP 2026 – 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 76–80, 2026
[5] Kartik Tandon, Julian Gould, Tanishq Bhatia, Francesa Dominici, Alejandro Ribeiro, and Claudio Battiloro. Consistent geometric deep learning via Hilbert bundles and cellular sheaves. In Advances in Neural Information Processing Systems (NeurIPS 2026), 2026.
Questions for authors
Q1 How easily can your theoretical results be extended beyond simply Levi-Civita connections on manifolds such as the sphere and torus?
Q2 Can you explain how this work would have practical implications in the setting of choosing or learning parallel transport operators in actual geometric deep learning settings, such as for sheaf neural networks?
Q3 Are there no assumptions required at the bundle or connection level for Algorithm 1 to run successfully?
6 comments
Response to Reviewer MXF9, part 2 of 2
Response to Reviewer MXF9 (2/2): W3, Q2 and Q3
W3. Relation to the sheaf literature. Thank you for these references; engaging with them has strengthened the paper. We now discuss all five works, plus Hernandez Caralt et al. (2026), in the main related work and in a comparison table (Table 11):
| work | maps | object of study |
|---|---|---|
| Hansen & Gebhart; Bodnar et al. | given or learned | expressivity, heterophily, oversmoothing |
| Barbero et al.; Singer & Wu | local PCA and Procrustes | data-driven connection Laplacians |
| Battiloro et al.; Tandon et al. | induced by a bundle | convergence and transferability under dense sampling |
| Di Nino et al. | learned orthogonal | recovering the connection from signals |
| Hernandez Caralt et al. | identity | empirical competitiveness |
These works ask whether learned maps help a downstream task, or whether a discretization converges as sampling grows. Our question is complementary: the finite-sample risk of combining fibre values under a given map, including non-isometric maps. To our knowledge none of these works gives this risk. Our contribution is therefore distinct from theirs, and our results now say something concrete about each of them, as described under Q2. On notation, we follow the standard conventions of Kobayashi and Nomizu (1963), which much of this literature shares; our setup, a reference identification compared with transport under a statistical risk, does not appear in [5].
W3. Broader applications and the experiments. The identity applies wherever vectors in different fibres are averaged under a chosen identification. This includes message passing in gauge-equivariant and sheaf networks, local smoothing on meshes, and imputation of geophysical and imaging fields. We agree that the experiments are narrower than the theory: tangent bundles of surfaces, SPD fibres and one real dataset. The new "Scope and limitations" paragraph says so.
Q2. Practical implications for sheaf networks. Proposition 1 gives the exact finite-sample risk of one aggregation under any restriction maps:
- Orthogonal maps change only the bias. On parallel data they can match transport but cannot beat it.
- Contracting maps trade bias for variance, and they can win for small neighbourhoods.
- The Procrustes maps of Singer and Wu and of Barbero et al. are the polar factor of projection. On Sᵈ this factor equals Levi-Civita transport.
We tested these predictions on real winds (Tables 3 and 33; relative RMSE; 5 seeds at 8 months, 3 at 1 month; paired difference over 64 test fields):
| maps | 8 months | difference | 1 month |
|---|---|---|---|
| Levi-Civita (gauge net) | 0.152 | – | 0.157 |
| no transport | 0.237 | +0.085 ± 0.007 | 0.236 |
| fixed projection | 0.151 | −0.0005 ± 0.0002 | – |
| learned orthogonal | 0.153 | +0.0016 ± 0.0008 | 0.158 |
| learned general | 0.125 | −0.027 ± 0.004 | 0.131 |
- Learned orthogonal maps come close to transport but do not beat it, as Proposition 1(i) predicts.
- Fixed projection maps gain slightly, as Theorem 2 and Example 1 predict.
- Learned general maps are clearly better. They rescale messages depending on the features, which is outside our fixed-map theory, and the paper says so.
The practical guidance is as follows. With orthogonal maps, a sheaf model can at best learn the transport of the data's bundle. With fixed non-orthogonal maps, any gain comes from shrinkage, and the identity says when that helps. Proposition 1 describes one aggregation layer, so it predicts the ordering of fixed and orthogonal maps but not the size of the end-to-end gain.
Q3. Assumptions of Algorithm 1. Yes. The submission stated them in Assumption 1, and Algorithm 1 now also lists them as inputs:
- base Exp and Log;
- transport of a metric connection along minimizing geodesics in a geodesically convex region;
- an exact or Monte Carlo E-step.
The guarantees also assume unique Fréchet means, a compact likelihood superlevel set and Lipschitz gradients. Two qualifications are now in the main text:
- Deterministic ascent holds only with an exact E-step. A Monte Carlo E-step needs growing particle numbers or stochastic-approximation conditions.
- Consistency concerns the constrained global maximizer, which the nonconvex iteration need not reach.
Nothing is assumed about curvature sign, rank or topology. A new paragraph after the algorithm, "What can go wrong" (Appendix C), explains what fails when the assumptions are violated.
Thank you again. We hope the revision shows that the theory covers any metric bundle and speaks directly to the choice between learned and fixed maps, and we would welcome any further comments.
I appreciate the author’s addressing each of the concerns raised. In particular, the presentation is indeed easier to follow in the rewrite, and I will raise my presentation score to a 2 as a result.
While the revised manuscript now includes references to relevant papers such as Hernandez Caralt et al, I still do not think it meaningfully engages with the content of these works. For instance, Hernandez Caralt et al shows that using the identity map as transport, which completely ignores the geometry of the problem, can empirically outperform learned transport. How does this empirical result interact with the theory being presented here? In general, I think the point raised in W3 and Q2 still stands where it is not clear how the theory setup here would actually provide actionable insights in settings where parallel transport operators are currently used, and this is exacerbated by the limited and largely synthetic nature of the experiments presented.
This also ties into W2, where the main theoretical results being proved are purely at the level of bundles themselves, and not how learned vs fixed transport operators affect the way actual algorithms or machine learning models make use of these transports. Altogether, it is hard to place the Bundle-GEM algorithm in the larger literature wherein transport operators are currently used. For instance, as you now note in the clarified theoretical assumptions, you need a unique Frechet mean for the guarantees to hold, but these assumptions do not hold for the sphere or torus, which are the main toy examples used in the paper. Thus, while I appreciate the authors’ clarifications on the scope of the theoretical contribution, I still do not view it as substantial enough on its own right to justify acceptance. As such, I will maintain my original rating.
Response to Reviewer MXF9, part 1 of 2
Response to Reviewer MXF9 (1/2): summary, W1, W2 and Q1
Thank you for the detailed review and for recognizing that the paper takes "a partly novel approach compared to most previous empirical analyses". The question it answers is when parallel transport, the default way to combine vectors on curved spaces, actually lowers estimation error. The answer is exact and sometimes counterintuitive: an unbiased transported estimator can lose to a biased reference, and the identity says precisely when. The review prompted the largest changes in the revision. Because of the length limit, we reply in two parts; all numbers refer to the revised PDF unless we say otherwise.
On the summary. We would like to clarify one point about the method. Bundle-GEM does not apply GEM in each fibre and then transport to the next fibre. It is a single EM over bundle-valued latent variables, and transport serves two purposes in it:
The sphere and torus are test cases rather than the scope of the paper. The revised Section 5 states this directly.
W2 and Q1. Generality of the theory. We think this concern comes from how the submission was presented rather than from what it proved, and we have fixed the presentation.
We agree that a reader could easily come away thinking otherwise, because the most visible consequence of the general identity was a sphere computation. The revision makes the generality explicit:
W1. Presentation. We took each part of this point seriously.