acceptodds
Under review as a conference paper at ICLR 2027

Transport-Aware Inference on Vector Bundles: When Is Parallel Transport Statistically Beneficial?

Abstract

Many learning problems attach a vector to every point of a curved space: a velocity at each point of a sphere, a feature at each vertex of a mesh. The vectors at each location live in their own vector space, the fibre, and together the fibres form a vector bundle. Vectors in different fibres cannot be averaged directly; one must first decide how to carry a vector from one location to another. Geometric machine learning usually uses parallel transport, on the grounds that it is the geometrically canonical choice. We ask when it is also the statistically better one. We show that the error of replacing transport by a fixed reference rule depends on the rule, so it is not an intrinsic curvature effect: two standard rules on the sphere have biases of opposite sign. Curvature enters invariantly through holonomy, the rotation of a vector carried around a closed loop. An exact mean-squared-error identity shows that at low curvature a biased reference that contracts noise beats transport, and that for varying fields the answer depends on curvature relative to the field’s own variation. To make the comparison operational for incomplete data, we develop Bundle-GEM, a generalized EM procedure whose parameters live in moving fibres. In controlled experiments the predicted crossover matches the exact theory, a plug-in estimate of the identity picks the better estimator in 94% of cases, full Bundle-GEM on data with up to 80% missing fibre values comes within 0.006 of the true-parameter error where fixed references and global frames fail, and a surface without a global frame shows the predicted breakdown of coordinate baselines. On real global wind fields (NCEP/NCAR reanalysis), a gauge-equivariant network beats all non-transport baselines on every test field, and the plug-in identity picks the better local estimator at 98% of locations.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Geometric Machine Learning, Vector Bundles, Parallel Transport, Statistical Inference, Expectation Maximization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MXF9OpenReview ↗
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
AuthorsOpenReview ↗

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:

  • it carries posterior fibre values into a common fibre, so that the sufficient statistics can be summed;
  • it moves the fibre parameters when an anchor moves.

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.

  • The central result, the exact MSE identity, was already in the main body of the submission as eq. (5). It holds for any rank-r vector bundle with a metric connection and any bounded linear reference, at every sample size.
  • The submission's main body contained two numbered results. One is the topology proposition (Proposition 1 in the submission), which holds for any vector bundle. The other is the sphere crossover, which was already labelled a corollary.
  • The single-path lemma, the holonomy and radius results, the path and connection bounds, and all Bundle-GEM guarantees were also general. Only the closed-form constants κ* and Ξ > 4 were evaluated on the sphere.

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:

  • Lemma 1 and Theorem 1 are now stated in the main text. The holonomy result (Theorem 6), the radius trade-off (Theorem 7), the path and connection bounds (Propositions 4 and 5) and the Bundle-GEM guarantees (Theorems 3 to 5) hold on any metric bundle.
  • Theorem 2 is new and holds for any bundle. It gives the leading-order gap ‖Mv*‖² + (2σ²/n) tr M, and an exact 1/n law for how the gap depends on the sample size. It assumes the reference agrees with transport to first order, and the paper says so in the theorem and in the abstract.
  • Examples versus theorems. We agree with the principle. The sphere-specific statements are now Example 1 and Corollary 1, and the Ξ condition is stated for any bundle.
  • Beyond Levi-Civita (Q1). Nothing in the theory uses the Levi-Civita connection or a particular manifold. It applies to any metric connection, for instance on normal bundles, the feature bundles of gauge-equivariant networks, or discrete connection graphs. A new remark covers non-metric connections; the identity then compares two patterns of shrinkage.

W1. Presentation. We took each part of this point seriously.

  • Main text relying on appendix results. Lemma 1, Theorems 1 and 2, Example 1, Corollary 1 and Propositions 1 to 3 are now stated in the main text, each followed by an explanation of what it says and why. Theorem 1 also has a proof idea.
  • Algorithm referring to another appendix. Algorithm 1 is now self-contained. It lists its inputs and assumptions and writes out every update in place.
  • Intuition. Section 2 opens with a running example: on the sphere, projection shortens a vector by cos(s/R), which adds bias but also shrinks noise. Each section says what it does and why, a worked numerical example follows Corollary 1, and the experiments are framed as five questions, each answered first in words.
  • Writing style. We have revised the prose throughout and agree that the exposition should be clear, precise and checkable. All proofs are in the appendix and all code is public. As ICLR requires, the paper discloses our use of an AI assistant, and we take full responsibility for every statement.
AuthorsOpenReview ↗

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):

workmapsobject of study
Hansen & Gebhart; Bodnar et al.given or learnedexpressivity, heterophily, oversmoothing
Barbero et al.; Singer & Wulocal PCA and Procrustesdata-driven connection Laplacians
Battiloro et al.; Tandon et al.induced by a bundleconvergence and transferability under dense sampling
Di Nino et al.learned orthogonalrecovering the connection from signals
Hernandez Caralt et al.identityempirical 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):

maps8 monthsdifference1 month
Levi-Civita (gauge net)0.152–0.157
no transport0.237+0.085 ± 0.0070.236
fixed projection0.151−0.0005 ± 0.0002–
learned orthogonal0.153+0.0016 ± 0.00080.158
learned general0.125−0.027 ± 0.0040.131
  • 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.

Reviewer_CiuLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In computational Riemannian geometry, using parallel transport to move tangent vectors is a standard practice. However, the paper notes that parallel transport is not always the statistically preferable to approaches based on references, where there are some bias-variance tradeoff. The paper derives an analytical MSE identity for comparing the different methods, which allows identifying which method is better for the considered scenario. The authors further developed Bundle-GEM, an algorithm for inference based on incomplete observations. Numerical results are provided, verifying the predictive power of the proposed MSE identity and the utility of the proposed Bundle-GEM algorithm.

Critical strengths

The paper looks into the usage of parallel transport, a seemingly obvious choice, and shows that it might not be statistically optimal. The analytical MSE identity allows predicting which method behaves better in practice. This perspective is interesting and potentially useful for geometric machine learning.

Critical weaknesses
  1. I find the presentation hard to follow at times, and I think the presentation can be improved. For instance, the EM algorithm is introduced under paragraph titled Geometry in Section 2, even though EM is a general probabilistic machine learning method that does not need to be geometric. The practical method and theoretical results are presented densely, without much motivation and introduction to allow readers not familiar with the field to follow.

  2. It is unclear, at least according to the current manuscript, how the performance of the algorithms as captured using the analytical MSE identity relates to the practical Bundle-GEM algorithm, as these seem to be semi-disjoint contributions.

Questions for authors

Please refer to Critical Weaknesses.

  1. Can the authors comment on how the analytical MSE identity relates to the practical Bundle-GEM algorithm? The MSE formula allows compare which method is better, while the Bundle-GEM algorithm is a practical algorithm for inference under incomplete data.

  2. How does the analytical MSE identity behave under different numbers of available data, e.g. when there are few data points available?

3 comments
AuthorsOpenReview ↗

Response to Reviewer CiuL

Thank you for the constructive review. Your first question refers to the weaknesses, so we answer W1–W3 and then Q2 and Q3 in turn. All numbers refer to the revised PDF.

W1. EM introduced under "Geometry". We agree, and this is fixed. EM now has its own paragraph ("EM in brief", Section 5), separate from the geometric setup in Section 2. It presents EM as a general latent-variable method and explains that geometry enters only in the M-step: posterior means of fibre values lie in different fibres and must be carried to a common fibre before they can be summed.

W2. Dense presentation with little motivation for non-specialists. We restructured the main text so that it runs from question to answer:

  • Section 2 now opens with a running example in plain terms. On the sphere, projecting a vector shortens it by cos(s/R); this creates a bias but also shrinks the noise, which can win with few samples.
  • The exact theory for complete data comes first (Sections 3 and 4). Bundle-GEM follows in Section 5 and opens by explaining why a likelihood method is needed.
  • Each section opens with a short statement of what it does and how it follows from the previous one.
  • Each main result is stated in the main text with an explanation:
    • a proof idea for Theorem 1;
    • a plain reading of Theorem 2;
    • a worked numerical scale after Corollary 1 (with n = 40, transport pays off only when the neighbourhood covers more than about a seventh of the radius of curvature);
    • a paragraph on when to use the exact identity, its local form, or the plug-in rule.
  • The experiments are framed as five questions in the order of the theory, and each paragraph starts with its answer.

W3 and Q2. How does the identity relate to Bundle-GEM? The two meet in the M-step, where the mean update averages transported posterior means. New Proposition 3 makes this exact in the simplest setting: one component, a fixed anchor and fully observed fibres with Gaussian noise. At every EM fixed point,

μ̂ = (Σᵢ Wᵢ)⁻¹ Σᵢ Wᵢ τᵢ yᵢ, where Wᵢ = (Σ̂ + σ² τᵢ τᵢ*)⁻¹.

  • For transport, and for any isometric reference, the weights are equal. The Bundle-EM mean is then exactly the transported average, so Theorem 1 gives its risk.
  • With missing fibres, a missing fibre contributes only the prior, and a partially observed fibre counts as part of an observation. We therefore propose that the identity holds approximately with an effective sample size n_eff. This is an approximation we test, not a theorem; we do not claim exact inheritance for the full multi-component algorithm.

Table 24 tests both statements:

  • With full data, the measured gap matches the identity within two standard errors for n = 4 to 64, and changes sign between n = 8 and 16.
  • With 50% missing data, the identity at the nominal n gets the sign wrong at n = 16, while the identity at n_eff tracks the measurement.

The two contributions are now presented as one argument. The identity says which identification to use, and Bundle-GEM applies it to incomplete data.

Q3. How does the identity behave with few data points? Theorem 1 is exact for every n ≥ 1, so it is not an asymptotic statement. Theorem 2(iii) adds that the gap equals ‖Bias‖² + c/n with c independent of n, so its sign changes at most once in n. Figure 2 (main text) and Table 23 check this for n from 2 to 320 at κ ∈ {0.05, 0.2, 1}, with 2 to 4 × 10⁵ paired trials per cell:

  • The identity agrees with simulation within two standard errors in 26 of 27 cells, including n = 2.
  • The exact crossovers are n* = 38.3 and 8.3; the leading-order formula gives 40 and 10.
  • At κ = 1 transport is better for every n.
  • Only the leading-order form degrades, at very small n with high curvature.

In short, a shrinking reference is more likely to win with few data points, and the identity says exactly up to which n.

We hope these changes make the paper easier to follow and show how its parts connect. Thank you again for the helpful review.

I thank the authors for their response and for improving the presentations. I am not that familiar with the research topic, and can not form a strong opinion for this paper. I am hesitant between 2 and 3.

AuthorsOpenReview ↗

Responding to Further Comment by Reviewer CiuL

Thank you for reading the revision and for your candid comment. We appreciate that you engaged with the paper outside your main area. In case it helps you settle on a view, here is a short guide to what the paper claims and to the parts that can be checked without a specialist background in differential geometry.

The question and the main answer. When vectors live at different points of a curved space, they have to be carried to a common point before they can be averaged. The standard choice is parallel transport. The paper asks when transport actually gives a more accurate average than a simpler fixed rule. Theorem 1 answers this with an exact formula for the difference in mean squared error, made of three interpretable terms: a bias, a sampling variance, and a change in noise. The surprising consequence (Theorem 2) is that a biased rule which shrinks vectors can beat the unbiased transported average when data are scarce, because shrinking also removes noise.

What can be checked directly.

  • The formula matches simulation at every sample size. Figure 2 in the main text plots the predicted and measured gap for n from 2 to 320 at three curvatures. The two agree within two standard errors in 26 of 27 settings, including n = 2. This also answers your earlier question about few data points: the formula is exact for every n, and the gap changes sign at most once as n grows.
  • Your question on Bundle-GEM is answered by a proof and a test. Proposition 3 shows that, with one component and complete data, the mean computed by Bundle-GEM is exactly the transported average, so Theorem 1 gives its error. With missing data the link holds approximately, through an effective sample size. Table 24 confirms both statements, and shows that ignoring the missing data gives the wrong sign while the effective sample size tracks the measurement.
  • The practical rule works on real data. Estimated from training data only, the formula predicts which of two estimators is better at 95–98% of locations on real global wind fields. It stays near 96% when the training data themselves are masked and noisy (Table 32).
  • Learned and fixed maps behave as predicted. In networks trained on real winds, learned orthogonal maps come close to transport but do not beat it, and fixed contracting maps gain slightly. Both match the theory (Table 3).
  • Everything is reproducible. All proofs are in the appendix, and every reported number is produced by a script in the released code.

What changed in response to your review. EM now has its own paragraph, separate from the geometry. Section 2 opens with a plain-language example on the sphere. Each main result is followed by an explanation of what it means. Bundle-GEM now comes after the theory, in Section 5, and opens by explaining why a likelihood method is needed.

Thank you again for your time and for comments that made the paper clearer. We appreciate your consideration, whichever score you settle on.

Reviewer_WYemOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates when parallel transport, a default choice for geometric learning, is beneficial for inference on vector bundles. The authors distinguish reference-dependent single-path errors from gauge-invariant loop holonomy, and derive an exact MSE identity to characterize the bias–variance trade-off between transport-aware and reference-based estimators. To support likelihood-based inference on partially observed vector bundles, the paper proposes Bundle-GEM, a generalized EM procedure that transports sufficient statistics into common fibres and updates moving anchors and fibre parameters through a transported retraction. Experiments on synthetic vector fields, curved meshes, and real wind data validate the theoretical results and reveal the strengths and weaknesses of parallel transport in different settings.

Critical strengths
  • The paper brings up an interesting and novel research questions about whether parallel transport, a default choice for geometric learning, is universally beneficial. A thorough investigation challenges common sense in geometric learning and provides new insights that broadly benefit research in this area.
  • The proposed Bundle-GEM provides a unified framework to likelihood-based inference on partially observed vector bundles. I find such unified framework a clever design that enables principled comparisons between parallel transport and other reference identifications, thereby answering the key question of whether parallel transport is universally beneficial.
  • The paper provides detailed theoretical analysis regarding the bias–variance–holonomy trade-off, including an exact MSE identity, a curvature crossover threshold, etc., revealing when parallel transport is preferable. The empirical evaluation seems closely align with these theoretical results, which also provides a consistent validation of the theoretical results.
Critical weaknesses
  • It seems that the proposed plug-in MSE criterion requires a lot of calibration data for selecting transport-aware and reference-based estimator, which can limit its practicality when available calibration data is scarce. According to Appendix V.2, the reported 94% accuracy requires 8,000 independent calibration samples, while 2,000 samples are explicitly reported as insufficient, despite the fact that only 40 samples are used to estimate the target vector.
Questions for authors
  • Appendix V.2 reports that 8,000 independent calibration samples achieve 94% accuracy while 2,000 samples are too noisy to be useful. Can you include an additional sensitivity study of the MSE criterion against the number of calibration data used for selecting an estimator? This should be a particular important calibration data is scarce.
1 comment
AuthorsOpenReview ↗

Response to Reviewer WYem

Thank you for the positive and accurate summary. Your weakness and your question raise the same issue, how much calibration data the plug-in rule needs, so we answer them together. All numbers refer to the revised PDF.

Weakness and Q1. Sensitivity of the plug-in criterion to the amount of calibration data.

Synthetic study (Table 25). We used the same 105 fields and varied the calibration sample from 50 to 8,000 points, with 20 independent calibration samples per setting. Because you rightly contrasted the calibration size with the 40 samples used for the target, we also varied the target size, n = 10, 40 and 160. Results for n = 40:

calibration pointscorrectmean excess MSE
5072.6%12.2%
20079.8%4.4%
40083.3%2.0%
1,60089.6%0.6%
8,00093.3%0.15%
always transport70.6%18.4%
always reference29.4%52.5%

Three findings:

  • Accuracy rises steadily with calibration size and depends little on the target size n.
  • Excess MSE is the more informative measure. At every calibration size and every n, the rule has lower excess MSE than both fixed rules, because its mistakes fall mostly where the two estimators are close.
  • With 50 points the accuracy is only modest (73%). We therefore do not claim the rule is reliable with very little data. A few hundred points are needed for a dependable choice, and the conclusion of the paper now says that the rule needs enough calibration data.

The first version said 2,000 points were "too noisy to be useful". The new study shows that this was too pessimistic: with 1,600 points the rule is right in about 90% of cases. We have removed the sentence.

Real data with realistic calibration (Table 32). In the first version the real-wind plug-in was calibrated on complete, noise-free months. In the new version, each calibration month is seen only through masked, noisy copies, like the test inputs, and the variance term is corrected for the noise.

calibration monthscleanmasked and noisy
174.9%74.7%
489.9%89.1%
897.7%95.8%

Always transporting is right at 66% of locations. Noise costs about two points of accuracy, and a single month already beats every fixed rule.

Why moderate calibration suffices. The decision depends only on the sign of one scalar. The noise term needs no data, because it depends only on the geometry of the reference and on σ². The other two terms need a local quadratic fit with 6 coefficients per fibre component.

Thank you again for a question that led to a clearer and more useful result.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response (to all reviewers)

We thank the reviewers for their careful reading. The reviewers agree on the core question. WYem calls it "interesting and novel" and says it "challenges common sense in geometric learning". CiuL finds it "interesting and potentially useful". MXF9 calls the statistical approach novel "compared to most previous empirical analyses". We have uploaded a revision with major changes in blue. The main text uses the 10 pages allowed for the revision; the extra page goes to exposition, a running example, and a new figure.

What the paper contributes. Parallel transport is the default way to combine vectors on curved spaces, but it has not been known when it actually lowers estimation error. The paper gives an exact answer.

  1. Theorem 1 is an exact risk identity for any vector bundle with a metric connection and any bounded linear reference, valid for every n ≥ 1. It splits the gap into a change in noise variance, a sampling-variance term, and a squared bias. The submission already stated it in this generality, as eq. (5) of the main text.
  2. Theorem 2 shows that the unbiased transported estimator can lose to a biased reference that shrinks vectors, because shrinking also removes noise. For references that agree with transport to first order, it gives the leading-order crossover sample size n* in closed form. The exact gap is affine in 1/n, so its sign changes at most once.
  3. Example 1 shows that ambient projection on Sᵈ is Levi-Civita transport followed by a symmetric shrinkage. Proposition 1 extends the identity to sheaf restriction maps.
  4. A plug-in estimate of the identity picks the better estimator in 94% of controlled cases and at 98% of real-wind locations.
  5. Bundle-GEM brings the same choice to incomplete data. It comes with ascent, consistency and rate guarantees, each under stated conditions (exact E-step; identifiable observed model). Its link to the identity is exact for one component with complete fibres, and approximate under missing data.

What the revision adds.

  • Generality (MXF9). Theorem 2, a general form of the Ξ condition, and a remark on non-metric connections. The sphere computations are now labelled Example 1 and Corollary 1.
  • Sheaf networks (MXF9).
    • Proposition 1.
    • A positioning table (Table 11) against all five cited works and Hernandez Caralt et al. (2026).
    • A real-wind experiment with learned and fixed restriction maps.
  • Identity and Bundle-GEM (CiuL). Proposition 3 and a supporting experiment (Table 24).
  • Small n (CiuL). A sample-size sweep (Figure 2 in the main text, Table 23).
  • Calibration data (WYem). Sensitivity studies on synthetic and real data (Tables 25 and 32).
  • Presentation (CiuL, MXF9).
    • EM has its own paragraph.
    • All main results are stated in the main text with intuition.
    • A paragraph explains how the exact identity, its local form and the plug-in rule differ.
    • The paper now runs from question to answer: the exact theory for complete data (Sections 3 and 4), then Bundle-GEM for incomplete data (Section 5), which opens by explaining why a likelihood method is needed.
    • Algorithm 1 lists its assumptions.
    • A new "Scope and limitations" paragraph says where the theory stops.

Scope. We want to state clearly what is and is not claimed.

  • Theorem 2 is local and assumes first-order alignment.
  • Proposition 3 is exact for one component, a fixed anchor, and complete fibres; beyond that, the effective-sample-size rule is an approximation, which we test empirically.
  • The theory holds on any metric bundle. The experiments cover tangent bundles of surfaces, SPD fibres and one real dataset.

On writing style. We are grateful for the comments on presentation. We have taken them seriously and revised the exposition throughout: the text now gives plainer explanations, more intuition, and clearer links between the sections. We share the reviewers' view that a theoretical paper should be clear, precise, and easy to check. To make checking easy, every proof is in the appendix, the code is public, and every reported number can be reproduced with a released script. We welcome any further suggestions to improve the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.