Pith. sign in

REVIEW 2 major objections 6 minor 54 references

Sharp Concentration Bounds for Bundle-Valued Statistics on Manifolds

T0 review · 2 major / 6 minor · reviewed 2026-07-14 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Averaging vectors living on different points of a curved manifold leaves a curvature-driven error floor that more data cannot remove.

desk verdict Clean, usable concentration theory for transported bundle means that correctly isolates a curvature-driven holonomy floor with matching lower bounds and sphere validation. read the letter →

arxiv 2607.10592 v1 pith:C7HFPGJI submitted 2026-07-12 cs.LG

Swagatam Das , Vaclav Snasel This is my paper · ORCID
classification cs.LG MSC 60E1562H1153C29
keywords concentrationinequalitiesRiemannianstatisticsvectorbundlesparalleltransportholonomygeometricmachinelearningbias-variancedecomposition
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Many geometric machine-learning pipelines produce observations that live in different local vector spaces attached to different points of a manifold. To average them one must first parallel-transport them into a single reference space; that transport is path-dependent once geodesics are no longer unique. The paper proves that the resulting empirical mean obeys a clean bias-variance split: ordinary sampling noise that decays like 1 over square-root of sample size, plus a deterministic holonomy bias controlled by the curvature of the bundle and the diameter of the data support. Both pieces are shown to be unavoidable for any transport-based estimator, and the theory supplies explicit finite-sample concentration radii, a robust median-of-means variant, and a central-limit theorem once the geometric floor becomes negligible. Controlled experiments on the sphere confirm that the predicted error floor appears exactly where the formulas say it should.

What carries the argument

The bias-variance decomposition ||Ȳ_n - m⋆|| ≲ B/√n + Δ_hol, where the stochastic piece is controlled by dimension-free Hilbert-space Hoeffding/Bernstein inequalities and the geometric floor Δ_hol is bounded by the operator norm of the bundle curvature times the squared diameter of the support (with a sharp closed-form expression on the round sphere).

What would settle it

On the unit sphere, draw samples from a geodesic ball of fixed radius ρ, form the transported mean under two different transport rules, and check whether the gap between the two means remains constant (and matches 2 sin(π ho^{2}/2)) as sample size grows from hundreds to tens of thousands; if the gap shrinks or systematically deviates from the formula, the claimed error floor is false.

Watch

Extended reading notes

Core claim

Once bundle-valued observations are parallel-transported to a fixed reference fiber, the transported empirical mean concentrates about its expectation at the classical Euclidean rate, but any residual path-dependence of the transport injects an irreducible, curvature-controlled bias Δ_hol that is independent of sample size; the two terms together form a minimax-optimal bias-variance decomposition for every transport-based estimator.

Load-bearing premise

The argument needs a measurable, essentially unique way to choose the transport paths from every data point to the reference point; outside a normal ball or a non-positively curved cover that uniqueness fails and the holonomy term becomes a modeling choice rather than a pure geometric constant.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper develops a non-asymptotic concentration theory for empirical means of bundle-valued observations on Riemannian manifolds. Observations live in fibers of a vector bundle and are reduced to a fixed reference fiber by parallel transport; the resulting estimator is analyzed via sharp Hilbert-space inequalities. The main results are dimension-free Hoeffding and Bernstein tail bounds (Theorems 1–2), an explicit bias–variance decomposition that isolates a curvature/holonomy-driven deterministic floor Δ_hol (Theorem 3, Eq. 3, Proposition 1), matching minimax lower bounds for the class of transport-based estimators (Theorem 4), a median-of-means estimator under second-moment assumptions (Corollary 2), and a CLT in the reference fiber (Theorem 5). Controlled experiments on the tangent bundle of the round sphere confirm both the n^{-1/2} stochastic decay and the holonomy floor, matching the sharp area formula to within a few percent.

Significance. If the claims hold—and the derivations appear sound—the work fills a genuine gap between classical Hilbert/Banach concentration and manifold statistics for data that live in varying fibers rather than a single vector space. Geometric ML pipelines (gauge-equivariant message passing, intrinsic regression residuals, diffusion-tensor averaging) routinely perform exactly this transport-and-average step; having finite-sample radii that separate sampling noise from an irreducible geometric floor is practically useful and theoretically clean. Strengths that should be credited include: complete appendix proofs of the main theorems via Pinelis inequalities and Ambrose–Singer holonomy; an exact closed-form holonomy formula on TS^2_r; minimax lower bounds that match the upper bounds up to universal constants within the natural estimator class; a robust MoM extension; and reproducible sphere experiments that validate the predicted floor quantitatively. The contribution is more geometric isolation and statistical packaging than new concentration technology, but that packaging is the right object for the applications.

major comments (2)
  1. Appendix H / Section 6.4: The holonomy-floor experiments implement Rule B as a synthetic post-composition of Rule A with a fixed rotation of angle θ=πρ², rather than as parallel transport along a genuinely distinct minimizing geodesic. Inside the normal-ball regime ρ<πr/2 used throughout the experiments, minimizing geodesics to x0 are unique, so the construction correctly validates the operator-norm formula of Proposition 1 but does not exercise cut-locus multi-geodesic ambiguity. The manuscript should state this limitation explicitly and indicate whether the same quantitative agreement is expected (or how the theory changes) when support reaches the cut locus and the transport rule becomes a genuine modeling choice.
  2. Theorem 4 and Remark 1: The minimax lower bounds are correctly scoped to transport-based estimators and require either pinched positive curvature or constant positive curvature for the geometric term. The upper bound (Eq. 3) is stated for general bundle curvature ζ. The paper should make more prominent, already in the main-body statement of Theorem 4, that the matching lower bound on the holonomy floor is a positive-curvature phenomenon and that under the Hadamard branch of Assumption 1 one has Δ_hol=0 by uniqueness, so the two-term rate collapses to the pure stochastic term. This is implicit but easy to miss.
minor comments (6)
  1. The three forms of the holonomy discrepancy (canonical section-dependent Δ(P,ẽP;s), section-uniform Δ^unif_hol, and the per-sample operator-norm form used in Theorem 11) are carefully distinguished in the appendix but appear somewhat abruptly in the main text. A short paragraph in Section 5 collecting the three definitions and their relative tightness would help readers.
  2. Section 6.2 applications (gauge GNNs, DTI, Wasserstein tangent spaces) are useful recipes but contain no new numerical checks beyond the sphere validation. Even a small synthetic gauge-GNN aggregation example illustrating discrete cycle holonomy would strengthen the claim that the bias–variance decomposition is immediately actionable.
  3. Corollary 1 notes that the Bernstein linear term 2B can be improved to B under exact recentering; this is correct but easy to overlook. Flagging the improved constant in the main display of the confidence radius would be helpful for practitioners.
  4. Notation: main text uses Assumptions 1–3 while the appendix re-labels them A1–A3 for self-containment. A single sentence at the start of Appendix B noting the correspondence would reduce friction.
  5. Table 1 is a clear comparison; the entry for this work correctly lists the holonomy floor as irreducible by data alone when Δ_hol>0. Consider adding a column or footnote indicating which frameworks handle heavy tails, since Corollary 2 is a genuine extension of Lugosi–Mendelson to the bundle setting.
  6. Minor typographical/consistency items: the abstract and introduction both use “n^{-1/2}” and “n−1/2” interchangeably in places; standardize. In Appendix H the theoretical row of Table 4 is italicized as non-experimental—good—but the caption could state this more explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: concentration follows from Pinelis after geometric reduction; holonomy and minimax are independent geometric/probabilistic arguments, not fitted or self-defined.

full rationale

The load-bearing chain is: Assumption 1 supplies a measurable transport rule so that Yi live i.i.d. in the fixed Hilbert space Ex0; Theorems 1–2 then invoke Pinelis’ dimension-free martingale inequalities (external, classical) on the centered summands bounded by 2B (or variance proxy σ^{2}). The holonomy term Δ(P,ẽP;s) is defined as the operator-norm discrepancy of two transports and bounded by bundle curvature via the Ambrose–Singer theorem and the explicit sphere formula (Prop. 1), neither of which is fitted to data nor defined in terms of the concentration radius. Theorem 4’s minimax lower bounds are obtained by a standard Le Cam two-point construction that produces a geometric separation of order κD^{2}B while keeping total variation small; the construction does not recycle the upper-bound constants. Sphere experiments (Appendix H) are pure validation: theoretical Δhol = 2 sin(πρ^{2}/2) is compared to observed Err(B) with no free parameters adjusted to force the match. There are no self-citations of prior work by the present authors that underwrite uniqueness, no ansatz smuggled via citation, and no fitted parameter later called a prediction. The derivation is therefore self-contained against external classical inequalities and differential-geometric identities.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central claims rest on standard Riemannian geometry (Levi-Civita connection, Ambrose-Singer theorem, injectivity radius) and on classical Hilbert-space concentration (Pinelis martingale inequalities). The only modeling choices are the three explicit assumptions that guarantee a measurable transport rule and bounded moments; no free parameters are fitted and no new physical or mathematical entities are postulated.

assumptions (4)
  • standard math Pinelis' sharp martingale inequalities for Hilbert-valued sums yield dimension-free Hoeffding and Bernstein tails once observations are reduced to a fixed fiber.
    Invoked directly in the proofs of Theorems 1-2 / 6-7; the paper does not re-derive them.
  • standard math Ambrose-Singer theorem: holonomy around a loop is generated by the curvature 2-form integrated over a spanning surface.
    Used to convert curvature bounds into operator-norm bounds on transport discrepancy (Lemma 3, Theorems 9-10).
  • domain assumption Assumption 1: either the base is Cartan-Hadamard or the support lies inside a normal ball of radius less than the injectivity radius, guaranteeing a unique measurable minimizing geodesic.
    Stated in Section 3.1; without it the transport map is path-dependent by definition and the reduction to i.i.d. Hilbert-space variables fails.
  • domain assumption Uniform bound ||s(x)|| ≤ B on the support of µ (Assumption 2).
    Needed for the almost-sure bound that feeds the Hoeffding/Bernstein constants; relaxed later by the median-of-means estimator.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sharp Concentration Bounds for Bundle-Valued Statistics on Manifolds." pith.science (2026). https://pith.science/paper/C7HFPGJI

@misc{pith2026260710592,
  author       = {Pith},
  title        = {Pith review of: Sharp Concentration Bounds for Bundle-Valued Statistics on Manifolds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C7HFPGJI}},
  note         = {Machine review of arXiv:2607.10592}
}
abstract

Many geometric statistics and manifold learning pipelines routinely produce observations -- such as tangent vectors or local frames -- whose natural home is a varying family of fibers attached to different points of a base manifold, rather than a single shared vector space. Forming empirical averages requires transporting these observations to a common reference fiber, thereby introducing curvature- and holonomy-driven effects that are absent from classical concentration theory. We develop a non-asymptotic concentration theory for such transported empirical means, deriving finite-sample, dimension-free Hoeffding- and Bernstein-type bounds via sharp Hilbert-space inequalities. When shortest paths to the reference point are non-unique, transport becomes path-dependent and introduces a deterministic holonomy bias; we isolate and quantify this bias through bundle curvature and loop geometry, with sharp closed-form formulas for the tangent bundle of a round sphere. The resulting bias-variance decomposition separates the stochastic fluctuation decaying at the classical $n^{-1/2}$ rate in sample size $n$, from a curvature-driven error floor that no amount of additional data can eliminate; minimax lower bounds confirm both terms are unavoidable. We further establish a robust median-of-means estimator achieving optimal rates under heavy tails and the central limit theorem in the reference fiber. Controlled experiments on the sphere validate all theoretical predictions.

Figures

Figures reproduced from arXiv: 2607.10592 by the authors.

Figure 1
Figure 1. Exponential and logarithm maps: z = logy (x) ∈ TyM and x = expy (z) ∈ M. paths (e.g., minimizing geodesics) yields deterministic ambiguity in the transported estimator; Appendix C provides full holonomy control via curvature. Metric compatibility is crucial: because each Px→x0 is an isometry, moment bounds on s(X) transfer directly to moment bounds on Y = PX→x0 s(X). In particular, Assumption 2 implies ∥Y ∥ ≤ B almo… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 1 canonical work pages

  1. [1]

    Riemannian Lp center of mass: existence, uniqueness, and convexity

    Bijan Afsari. Riemannian Lp center of mass: existence, uniqueness, and convexity. Proceedings of the American Mathematical Society, 139(2):655–673, 2011

    2011

  2. [2]

    Ambrose and I

    W. Ambrose and I. M. Singer. A theorem on holonomy.Trans. Amer. Math. Soc., 75: 428–443, 1953

    1953

  3. [3]

    Birkh¨auser, 2008

    Luigi Ambrosio, Nicola Gigli, and Giuseppe Savar´e.Gradient Flows: In Metric Spaces and in the Space of Probability Measures. Birkh¨auser, 2008

    2008

  4. [4]

    Machine learning line bundle connections.Physics Letters B, 827:136972, 2022

    Anthony Ashmore, Rehan Deen, Yang-Hui He, and Burt A Ovrut. Machine learning line bundle connections.Physics Letters B, 827:136972, 2022

    2022

  5. [5]

    Birkh ¨auser, 2006

    Werner Ballmann, Mikhael Gromov, and Viktor Schroeder.Manifolds of Nonpositive Curvature, volume 61 ofProgress in Mathematics. Birkh ¨auser, 2006. Originally published 1985

    2006

  6. [6]

    A Geometric Unification of Generative AI with Manifold-Probabilistic Projection Models, 2025

    Leah Bar, Liron Mor Yosef, Shai Zucker, Neta Shoham, Inbar Seroussi, and Nir Sochen. A Geometric Unification of Generative AI with Manifold-Probabilistic Projection Models, 2025

    2025

  7. [7]

    Springer, 2003

    Marcel Berger.A Panoramic View of Riemannian Geometry. Springer, 2003

    2003

  8. [8]

    Bhattacharya and A

    R. Bhattacharya and A. Bhattacharya.Nonparametric Inference on Manifolds: With Applications to Shape Spaces. Cambridge University Press, 2012

    2012

Show all 54 references
  1. [9]

    Large sample theory of intrinsic and extrinsic sample means on manifolds—I.The Annals of Statistics, 31(1):1–29, 2003

    Rabi Bhattacharya and Vic Patrangenaru. Large sample theory of intrinsic and extrinsic sample means on manifolds—I.The Annals of Statistics, 31(1):1–29, 2003. doi: 10.1214/ aos/1046294456. URLhttps://doi.org/10.1214/aos/1046294456

    2003 doi

  2. [10]

    Large sample theory of intrinsic and ex- trinsic sample means on manifolds—II.The Annals of Statistics, 33(3):1225–1259,

    Rabi Bhattacharya and Vic Patrangenaru. Large sample theory of intrinsic and ex- trinsic sample means on manifolds—II.The Annals of Statistics, 33(3):1225–1259,

  3. [11]

    URL https://doi.org/10.1214/ 009053605000000093

    doi: 10.1214/009053605000000093. URL https://doi.org/10.1214/ 009053605000000093

    doi

  4. [12]

    Cambridge University Press, 2018

    Jean-Daniel Boissonnat, Fr´ed´eric Chazal, and Mariette Yvinec.Geometric and Topological Inference, volume 57. Cambridge University Press, 2018

    2018

  5. [13]

    Sliced-Wasserstein Distances and Flows on Cartan–Hadamard Manifolds.Journal of Machine Learning Research, 26:1–76, 2025

    Cl´ement Bonet, Lucas Drumetz, and Nicolas Courty. Sliced-Wasserstein Distances and Flows on Cartan–Hadamard Manifolds.Journal of Machine Learning Research, 26:1–76, 2025

    2025

  6. [14]

    Oxford University Press, Oxford, 2013

    S´ebastien Boucheron, G´abor Lugosi, and Pascal Massart.Concentration Inequalities: A Nonasymptotic Theory of Independence. Oxford University Press, Oxford, 2013. ISBN 978-0-19-953525-5

    2013

  7. [15]

    Bridson and Andr´e H¨afliger.Metric Spaces of Non-Positive Curvature, volume 319 ofGrundlehren der Mathematischen Wissenschaften

    Martin R. Bridson and Andr´e H¨afliger.Metric Spaces of Non-Positive Curvature, volume 319 ofGrundlehren der Mathematischen Wissenschaften. Springer, 1999. 16

    1999

  8. [16]

    Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c

    Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veliˇckovi´c. Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges, 2021. URL https://arxiv. org/abs/2104.13478

    2021 arXiv

  9. [17]

    Caffarelli

    Luis A. Caffarelli. The regularity of mappings with a convex potential.Journal of the American Mathematical Society, 5(1):99–104, 1992

    1992

  10. [18]

    Cohen, Mario Geiger, Jonas Koehler, and Max Welling

    Taco S. Cohen, Mario Geiger, Jonas Koehler, and Max Welling. Gauge equivariant convolutional networks and the icosahedral CNN. InProceedings of the 36th Interna- tional Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Research, pages 1321–1330. PMLR, 2019

    2019

  11. [19]

    do Carmo.Riemannian Geometry

    Manfredo P. do Carmo.Riemannian Geometry. Birkh ¨auser, 1992

    1992

  12. [20]

    P. T. Fletcher, S. Joshi, C. Lu, and S. Pizer. Principal geodesic analysis for the study of nonlinear statistics of shape.IEEE Transactions on Medical Imaging, 23(8):995–1005,

  13. [21]

    doi: 10.1109/TMI.2004.831793

    2004 doi

  14. [22]

    Thomas Fletcher and Sarang Joshi

    P. Thomas Fletcher and Sarang Joshi. Principal geodesic analysis on symmetric spaces: Statistics of diffusion tensors. pages 87–98, 2004

    2004

  15. [23]

    Universitext

    Sylvestre Gallot, Dominique Hulin, and Jacques Lafontaine.Riemannian Geometry. Universitext. Springer, Berlin, 3 edition, 2004. ISBN 978-3-540-20493-0. doi: 10.1007/ 978-3-642-97242-3

    2004

  16. [24]

    Thomas Hotz, Huiling Le, and Andrew T. A. Wood. Central limit theorem for intrinsic Fr´echet means in smooth compact Riemannian manifolds.Probability Theory and Related Fields, 189(3):1219–1246, 2024

    2024

  17. [25]

    A Lower Bound for Estimating Fr´echet Means, 2024

    Shayan Hundrieser. A Lower Bound for Estimating Fr´echet Means, 2024. URL https: //arxiv.org/abs/2402.12290

    2024 arXiv

  18. [26]

    Shoshichi Kobayashi and Katsumi Nomizu.Foundations of Differential Geometry, Vol. I. Wiley, 1963

    1963

  19. [27]

    Springer Series in Statistics

    Lucien Le Cam.Asymptotic Methods in Statistical Decision Theory. Springer Series in Statistics. Springer, 1986

    1986

  20. [28]

    American Mathematical Society, 2001

    Michel Ledoux.The Concentration of Measure Phenomenon, volume 89 ofMathematical Surveys and Monographs. American Mathematical Society, 2001

    2001

  21. [29]

    A Series of Modern Surveys in Mathematics

    Michel Ledoux and Michel Talagrand.Probability in Banach Spaces: Isoperimetry and Processes. A Series of Modern Surveys in Mathematics. Springer, Berlin, 1991. doi: 10.1007/978-3-642-20212-4

    1991 doi

  22. [30]

    Lee.Riemannian Manifolds: An Introduction to Curvature, volume 176 of Graduate Texts in Mathematics

    John M. Lee.Riemannian Manifolds: An Introduction to Curvature, volume 176 of Graduate Texts in Mathematics. Springer, 1997

    1997

  23. [31]

    Mean estimation and regression under heavy-tailed distributions: A survey.Foundations of Computational Mathematics, 19(5):1145–1190, 2019

    G´abor Lugosi and Shahar Mendelson. Mean estimation and regression under heavy-tailed distributions: A survey.Foundations of Computational Mathematics, 19(5):1145–1190, 2019. 17

    2019

  24. [32]

    A Riemannian framework for tensor computing.International Journal of Computer Vision, 66(1):41–66, 2006

    Xavier Pennec, Pierre Fillard, and Nicholas Ayache. A Riemannian framework for tensor computing.International Journal of Computer Vision, 66(1):41–66, 2006

    2006

  25. [33]

    Optimum bounds for the distributions of martingales in Banach spaces.The Annals of Probability, 22(4):1679–1706, 1994

    Iosif Pinelis. Optimum bounds for the distributions of martingales in Banach spaces.The Annals of Probability, 22(4):1679–1706, 1994

    1994

  26. [34]

    M. S. Pinsker.Information and Information Stability of Random Variables and Processes. Holden-Day, 1964

    1964

  27. [35]

    Pullback Bundles and the Geometry of Learning.Entropy, 25(10): 1450, 2023

    St´ephane Puechmorel. Pullback Bundles and the Geometry of Learning.Entropy, 25(10): 1450, 2023

    2023

  28. [36]

    Wasserstein Parallel Transport for Predicting the Dynamics of Statistical Systems, 2026

    Tristan Luca Saidi, Gonzalo Mena, Larry Wasserman, and Florian Gunsilius. Wasserstein Parallel Transport for Predicting the Dynamics of Statistical Systems, 2026. URL https: //arxiv.org/abs/2603.23736

    2026

  29. [37]

    Probability measures on metric spaces of nonpositive curvature

    Karl-Theodor Sturm. Probability measures on metric spaces of nonpositive curvature. InHeat Kernels and Analysis on Manifolds, Graphs, and Metric Spaces, volume 338 of Contemporary Mathematics, pages 357–390. American Mathematical Society, Providence, RI, 2003. doi: 10.1090/con...

    2003 doi

  30. [38]

    Joel A. Tropp. An Introduction to Matrix Concentration Inequalities.Foundations and Trends in Machine Learning, 8(1–2):1–230, 2015. ISSN 1935-8237. doi: 10.1561/ 2200000048. URLhttp://dx.doi.org/10.1561/2200000048

    2015 doi

  31. [39]

    Tsybakov.Introduction to Nonparametric Estimation

    Alexandre B. Tsybakov.Introduction to Nonparametric Estimation. Springer Series in Statistics. Springer, 2009

    2009

  32. [40]

    A Wasserstein-Type Distance for Gaussian Mixtures on Vector Bundles with Applications to Shape Analysis.SIAM Journal on Imaging Sciences, 17(3):1433–1466, 2024

    Michael Wilson, Tom Needham, Chiwoo Park, Suparteek Kundu, and Anuj Srivastava. A Wasserstein-Type Distance for Gaussian Mixtures on Vector Bundles with Applications to Shape Analysis.SIAM Journal on Imaging Sciences, 17(3):1433–1466, 2024. doi: 10.1137/23M1620363. URLhttps://...

    2024 doi

  33. [41]

    Exponential concentration for geometric-median-of-means in non-positive curvature spaces.Bernoulli, 29(4):2927–2960, 2023

    Ho Yun and Byeong U Park. Exponential concentration for geometric-median-of-means in non-positive curvature spaces.Bernoulli, 29(4):2927–2960, 2023. 18 Appendix Roadmap.The appendix is organized as follows. Section A establishes geometric notation, reviews Riemannian and vecto...

    2023

  34. [42]

    (Hadamard case) M is a Cartan–Hadamard manifold (complete, simply connected, non- positive sectional curvature), so every pair of points is joined by a unique minimizing geodesic

  35. [43]

    Assumption A2(Uniform boundedness; cf

    (Normal-ball case) S := supp(µ)⊂B(x 0, r) for some r <Inj(x 0), so each x∈ S is joined tox 0 by a unique minimizing geodesic depending smoothly onx. Assumption A2(Uniform boundedness; cf. Assumption 2).There exists B >0 such that ∥s(x)∥Ex ≤Bfor allx∈ S. Assumption A3(Variance ...

  36. [44]

    , Yn are independent copies of Y and we set ξi :=Y i −E[Y i], then ∥ξi∥Ex0 ≤2B almost surely

    If Y1, . . . , Yn are independent copies of Y and we set ξi :=Y i −E[Y i], then ∥ξi∥Ex0 ≤2B almost surely. 27 Proof. 1)Because the connection is metric compatible, parallel transport preserves the fiber inner product. In particular, for anyx∈Mand anyv∈E x, ∥Px→x0v∥Ex0 =∥v∥ Ex....

  37. [45]

    This yields a valid concentration bound, but with an additional multiplicative factor that depends on the net cardinality (hence, dimension-dependent constants)

    (Elementary net argument) For finite dimensionk, one can fix a δ-net N on the unit sphere S(Ex0) and apply scalar Hoeffding to each direction in the net, then take a union bound. This yields a valid concentration bound, but with an additional multiplicative factor that depends...

  38. [46]

    Theorem 6 achieves a dimension-free prefactor by invoking the Hilbert-space inequality of Pinelis [31]

    (Pinelis optimal bound) The optimal dimension-independent constants are obtained by applying Pinelis’ martingale inequalities for Banach-space-valued sums; these give the precise factor2and the denominator2 P b2 i in the exponent (see [31]). Theorem 6 achieves a dimension-free...

  39. [47]

    Fixx, x 0 ∈U

    the connection curvature is bounded onB(p, R): Λ0 := sup y∈B(p,R) ∥Ωy∥op <∞, Λ1 := sup y∈B(p,R) ∥∇Ωy∥op <∞. Fixx, x 0 ∈U. Letγ 1, γ2 be piecewiseC 1 curves inUjoiningxtox 0. Let Γ :=γ 1 ◦γ −1 2 37 be the resulting loop based atx, and define L := Length(Γ) = Length(γ1) + Length...

  40. [48]

    This path leads to functional-analytic subtleties (the measurability of sections and operator- valued kernels)

    (Intrinsic L2-section approach) An alternative to parallel-transporting to a reference fiber is to regard sections as elements of the Hilbert space L2(M, µ;E) such as square-integrable sections and develop empirical-process concentration directly in that Hilbert space. This pa...

  41. [49]

    (Non-i.i.d. data) The martingale (Azuma/Pinelis) machinery used to obtain Banach-space concentration extends to certain dependent-data regimes (martingale difference arrays, mixing sequences) under additional assumptions; see [13, 31] for avenues

  42. [50]

    For dimension-free optimal constants, one should use the full strength of Pinelis’ inequalities for martingales in Banach spaces as in [31]

    (Sharper constants) For finite-dimensional fibers Ex0 ∼= Rk the elementary net argument gives explicit dimension-dependent constants. For dimension-free optimal constants, one should use the full strength of Pinelis’ inequalities for martingales in Banach spaces as in [31]. We...

  43. [51]

    Stochastic regime.Construct s0, s1 ∈ S B differing only in mean by δsto :=B/ √n, e.g., take s1(x) =s 0(x) + (δsto/B)P x0→xw for a fixed unit vector w∈E x0 chosen orthogonal to v

    Le Cam’s inequality then gives inf ˆmn sup s Es∥ˆmn −m ⋆(s)∥ ≥ 1 2 ·c 1BκD2 · 3 4 ≥ 3c1 8 BκD2. Stochastic regime.Construct s0, s1 ∈ S B differing only in mean by δsto :=B/ √n, e.g., take s1(x) =s 0(x) + (δsto/B)P x0→xw for a fixed unit vector w∈E x0 chosen orthogonal to v. Th...

  44. [52]

    Combining.Since the two pairs (s0, s1) are drawn from SB independently of each other, both lower bounds hold simultaneously for the worst-case risk over SB

    Le Cam’s inequality gives inf ˆmn sup s Es∥ˆmn −m ⋆(s)∥ ≥ 1 2 · B√n · 1 2 = B 4√n . Combining.Since the two pairs (s0, s1) are drawn from SB independently of each other, both lower bounds hold simultaneously for the worst-case risk over SB. Adding and absorbing all numerical c...

  45. [53]

    Then ∥x∥ ≤2 max u∈N ⟨u, x⟩,(40) and a1/2-net can be constructed with cardinality|N | ≤5 k. Proof. Let v∗ =x/∥x∥ be the unit vector in the direction of x. By definition of the net, there exists u∈ N with ∥u−v ∗∥ ≤ 1

  46. [54]

    holonomy ambiguity

    By the polarization identity for unit vectors in a Hilbert space, ⟨u, v∗⟩= 1− 1 2 ∥u−v ∗∥2 (which follows from ∥u∥2 =∥v ∗∥2 = 1 and expanding ∥u−v ∗∥2 = 2−2⟨u, v ∗⟩). Therefore, ⟨u, v∗⟩= 1− 1 2 ∥u−v ∗∥2 ≥1− 1 2 · 1 4 = 7 8 . Hence ⟨u, x⟩=∥x∥⟨u, v ∗⟩ ≥ 7 8 ∥x∥. Thus maxu∈N ⟨u, ...

    2024

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.