Isotropic randomization for one-sample testing in metric spaces

Helle S{\o}rensen; Matthieu Bult\'e

REVIEW 4 major objections 4 minor 39 references

Isotropic randomization for one-sample testing in metric spaces

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Randomizing with fixed-point isometries tests Fréchet means.

desk verdict A promising new randomization framework for Fréchet mean testing, but the theoretical guarantees promised in the abstract are not actually delivered—the paper itself admits this. read the letter →

arxiv 2501.15945 v1 pith:MNMBO567 submitted 2025-01-27 stat.ME

Matthieu Bulté , Helle S{o}rensen This is my paper · ORCID

classification stat.ME MSC 62G1062R20

keywords FréchetmeanmetricspacesrandomizationtestisotropygroupadmissibleempiricalvariancedirectionalstatisticsBures-Wassersteindistance

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

This paper aims to extend the classical one-sample mean test to random objects that live in a metric space, where only distances are available and there is no vector structure. The proposed procedure randomizes each observation by an isometry that fixes the hypothesized mean, and compares the empirical Fréchet variance of the original sample with the distribution of variances under this randomization. The paper proves that such randomization preserves the Fréchet mean and all distance moments, and that an 'admissible' randomization makes the randomized Fréchet variance strictly larger unless the null hypothesis is true. From this it argues that the test can detect alternatives while keeping the nominal level, supporting the claim with simulations on a circle, a booklet space, and covariance matrices under the Bures–Wasserstein distance, plus a wind-direction example. The paper itself notes in the conclusion that proving the test has the correct size under the null remains an open problem.

What carries the argument

The load-bearing object is the isotropy group $G_\mu$, the subgroup of the metric space's isometry group consisting of all isometries that fix the hypothesized mean $\mu$. An admissible randomization is a random isometry whose support is a subgroup of $G_\mu$ and that does not almost surely fix any other point; this admissibility condition is what keeps the test from confusing the null with alternatives such as antipodal points on the circle. The argument is carried by Proposition 3.1 (mean and distance-moment preservation), Proposition 3.2 (the variance inequality that yields power), and Theorem 3.1 (asymptotic normality of the empirical Fréchet variance), with the randomization test itself following the standard randomization-hypothesis construction.

What would settle it

Simulate $n$ observations on the circle from a distribution whose Fréchet mean is exactly $\mu$ but whose law is not invariant under reflection about $\mu$—for example an unequal mixture of two von Mises components arranged so the mean is $\mu$—and run the isotropic test with the reflection randomization at level 0.05. If the rejection rate over many replications stays above 0.05 for large $n$, the test does not control size without full distributional invariance.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a one-sample Fréchet-mean test can be built in any metric space whose hypothesized point admits enough isometries. The construction starts from the isotropy group $G_\mu$ of all isometries fixing the hypothesized mean $\mu$; a random isometry supported on a subgroup of $G_\mu$ is called admissible if no other point is almost surely fixed by every map in its support. Proposition 3.1 shows that randomizing $X$ by such an isometry preserves both the Fréchet mean and all distance moments, and Proposition 3.2 shows that the Fréchet variance of $g\cdot X$ is at least that of $X$, with equality only when $\mathbb{E}[X]=\mu$. Together with the asymptotic normality of the empirical Fréchet variance (Theorem 3.1), this justifies rejecting the null when the observed empirical variance falls below the lower tail of the randomization distribution. The intended conclusion is that the test has correct size and power against whatever alternatives the isotropy group can separate, although the paper explicitly leaves a proof of size control under the null for future work.

Load-bearing premise

For the test to have the promised false-rejection rate, the data's whole probability distribution—not just its Fréchet mean and distance moments—must be unchanged by the randomizing isometries; the paper proves the mean and moments are preserved but never proves this distributional invariance, and its conclusion states that a proof of correct size under the null remains open.

Editorial extensions

If this is right

For any metric space where the hypothesized point has a nontrivial isotropy group, the same algorithm applies: sample isometries from a subgroup of $G_\mu$, recompute the Fréchet variance, and reject when the observed variance is too small.
The test is fully nonparametric in the sense that it requires no parametric family and no tangent-space normality assumption, only a choice of isometry subgroup and the ability to compute distances.
Admissibility becomes a concrete design criterion: a randomization subgroup is usable exactly when its common fixed points reduce to the hypothesized mean, preventing the antipodal ambiguity seen on the circle.
For radially symmetric distributions, Lemma 3.1 supplies conditions under which the symmetry point is the Fréchet mean, giving a natural class of null models where the isotropic test applies.
The wind-direction case study shows that inverting the test over a grid yields a confidence interval for a mean direction, with the caveat that antipodal points can produce a second interval component.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension the paper leaves implicit: if the null distribution is not invariant under the chosen isotropy subgroup, the rejection rate should exceed the nominal level; a simulation with a non-symmetric distribution whose Fréchet mean is exactly $\mu$ would isolate whether full distributional invariance is required for size control.
Because the paper proves only invariance of the mean and distance moments, not of the full law, the method's validity for non-symmetric nulls may hold only asymptotically or only for the specific statistic used; this distinction could be probed by comparing rejection rates across different test statistics under the same non-symmetric null.
In spaces with a rich isotropy group, a larger admissible subgroup should give more power without breaking admissibility, so comparing subgroups of $G_\mu$ is a practical way to tune the test; the paper does not explore this trade-off.

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.

Desk Editor's Note

read the letter

Read this one if you work on Fréchet means or randomization tests. The genuine new piece is the isotropic randomization framework: using isotropy groups of the hypothesized mean to randomize data in metric spaces, with an admissibility condition (Definition 3.1) that guarantees power in the sense of a variance inequality (Proposition 3.2). That inequality is clean, the examples—circle, SPD matrices, booklets—are instructive, and the wind data application is a reasonable illustration. The simulations look careful, with code and a proper null-size check.

The soft spot is real and acknowledged: the paper promises 'theoretical guarantees' in the abstract and says admissible randomizations ensure 'correct size' in the conclusion, but Section 3.2 and Section 6 both state that no proof of consistency under the null exists. Proposition 3.1 only preserves distance moments and the Fréchet mean, not the full distribution. The randomization test in Algorithm 1 only has exact level if the null distribution is invariant under the isotropy subgroup. That holds for the radially symmetric distributions defined in Section 3.3, but not for the general null hypothesis 'mean equals mu0'. So the method has a proven power heuristic and unproven level, plus simulations suggesting the level is okay in practice. The reader's rejection is too harsh—the authors are transparent about the gap, they just don't reconcile it with the abstract. The right fix is to prove level under the invariance condition and explicitly frame the general case as heuristic, or tone down the claims.

The math otherwise is sound as far as I can see; citations are appropriate, and the admissibility notion is a real contribution even if the power result is definition-driven rather than a distributional theorem.

Send it to peer review. It deserves a serious referee who will push for a proof of level under the stated conditions. I wouldn't cite it as establishing theoretical guarantees, but the framework is worth engaging with.

Referee Report

4 major / 4 minor

Summary. The paper proposes a randomization test for the null hypothesis that the Fréchet mean of a distribution on a metric space equals a prespecified value. It extends the classical Euclidean randomization idea by using random isometries that fix the hypothesized mean, first on Riemannian manifolds via exponential/logarithm maps and then on general metric spaces via what the authors call admissible randomizations. The paper defines admissible randomizations, proves a population-level variance inequality (Proposition 3.2), introduces a group-action-based notion of radial symmetry, and reports simulations on the circle, on an SPD matrix space with Bures-Wasserstein distance, and on a booklet space, together with a wind-direction data application. The abstract and conclusion claim theoretical guarantees, including correct size and power against alternatives.

Significance. If the claimed guarantees were established, the paper would offer a reasonably general nonparametric approach to one-sample Fréchet mean testing that avoids tangent-space coordinates and works directly with the metric. The paper has several genuine strengths: it formulates a clean notion of admissible randomization; Proposition 3.2 is a correct and useful population-level observation; the numerical experiments cover three structurally different metric spaces; the code is publicly available; and the real-data application illustrates a practical workflow. However, the central theoretical claim, that the proposed test has the advertised size, is not proved and is in fact explicitly disclaimed in the text. Because the method's validity as stated depends on an unstated distributional invariance assumption, the paper's main advertised contribution is currently unsupported. The framework may be salvageable by restricting the scope to radially symmetric distributions and proving exact level under that condition, but the present text does not do so.

major comments (4)

[Section 3.2 / Algorithm 1] The claim that the test has correct size is not proved and is contradicted by the text itself. After presenting the randomization procedure, the paper states 'While no proof of the consistency of the test is available here', and Section 6 repeats 'it remains to prove that the randomization scheme is consistent under the null hypothesis, which would provide a theoretical guarantee that the test has the correct size.' Yet the Abstract claims 'We establish theoretical guarantees for our testing procedure' and Section 6 calls admissible randomization the key contribution that 'ensures that the resulting test has correct size.' A formal level control argument is the core of any hypothesis test; its absence is a load-bearing gap, not a presentational issue.
[Section 3.2 / Proposition 3.1] The validity of a randomization test requires that, under the null, the distribution of the randomized sample matches the null distribution of the sample (or at least that the test statistic's null distribution is preserved). The paper does not state or prove such an invariance for the general metric-space test. Proposition 3.1 only shows that a random isotropy of the Fréchet mean preserves the Fréchet mean and the distance moments, which is a much weaker property than distributional invariance. The Euclidean construction in Section 2.2 explicitly assumes distributions symmetric about the tested mean, but Algorithm 1 and the surrounding theory in Section 3.2 drop that assumption. Without an invariance or exchangeability condition, the randomization distribution used to compute the p-value is not a valid null distribution. If the intended scope is radially symmetric distributions (Definition 3.2), that restriction must be stated in the theorem and the level property proved under it; the current text does neither.
[Section 3.2 / Proposition 3.2] Proposition 3.2 does not establish power for the finite-sample test. The inequality Var[X*] ≥ Var[X] is a statement about population variances, whereas the test compares the empirical Fréchet variance of the observed sample to the empirical distribution of variances obtained under random resampling. No argument shows that, under an alternative, the observed empirical variance is stochastically larger than the randomization quantile. The admissibility condition only rules out equality in the population inequality; it does not yield a stochastic ordering or a consistency statement for the rejection rule. The heuristic sentence 'This, together with the following theorem, suggests...' is not a proof, and Theorem 3.1 is a CLT for the original-sample variance, not a statement about the randomization distribution.
[Section 4.1 / Figure 4] The simulation claiming 'correct size' for the non-symmetric mixture of von Mises distributions is not justified by any theoretical result in the paper. For this distribution, the null hypothesis E[X]=μ holds at the true mean, but the distribution is not invariant under the isotropy group of that mean, so the randomization test's level is not covered by the framework even under the radial-symmetry reading. The paper should either restrict such empirical claims to the symmetric case or provide a different argument for why the level is controlled in this non-symmetric setting.

minor comments (4)

[Throughout] There are several typos and misspellings: 'progresively' in the Abstract, 'Riemmanian' in Section 3.1, 'Lebegues' in Section 3.3, and 'Assuption' in Theorem 3.1. The manuscript would benefit from a careful proofreading pass.
[Section 3.3 / Definition 3.2] The definition of radial symmetry via invariance under the full isotropy group Gμ is clear, but the paper does not connect it to the level of Algorithm 1 in a formal proposition. A short lemma showing that, when X is radially symmetric around μ, the randomized sample (g_i X_i) has the same joint distribution as (X_i) under the null would make the intended scope precise.
[Section 4 / General] Theorem 3.1 assumes a bounded metric space, but the SPD matrix example in Section 4.2 uses the full space S2+ with the Bures-Wasserstein distance. The authors do not verify that the boundedness condition holds or explain why the CLT still applies. This deserves at least a remark.
[Figure 7 / Section 5] The application is interesting, but the comparison with the naive t-test on raw angles is not particularly informative because the t-test is applied to a circular variable in an inappropriate way. The comparison with the score test is more relevant and could be given more prominence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation rests on external CLT results and elementary isometry arguments; the unproved size guarantee is a support gap, not a circular reduction.

full rationale

The paper's derivation chain is largely self-contained and external. Proposition 3.1 is proved directly from the isometry/isotropy definitions; Theorem 3.1 is an external CLT (Dubey and Müller); Proposition 3.2 is a direct calculation from the Fréchet variance, Assumption 2.1, and the admissibility condition; Lemma 3.1 uses an external equivariance result from McCormack and Hoff. There is no fitted parameter later presented as a prediction, no load-bearing self-citation, and no uniqueness theorem imported from the authors' own prior work. The definition of admissible randomization is tailored so that the equality case in Proposition 3.2 is clean, but Proposition 3.2 is still a proved consequence of stated assumptions rather than an input renamed as an output. The main advertised guarantee that the test has correct size is not actually proved: the paper itself states in Section 6 that 'it remains to prove that the randomization scheme is consistent under the null hypothesis', and Section 3.2 says 'no proof of the consistency of the test is available here'. That is an honest limitation and a correctness/support gap, not a circular step. The finite-sample power evidence is simulation-based and independent of the definitions. Therefore no specific circular reduction is exhibited, and the circularity score is 0.

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

The method has no fitted parameters. The main load-bearing assumptions are the standard Fréchet mean and entropy conditions, plus an unstated distributional invariance that is necessary for a valid randomization test.

assumptions (4)

domain assumption Unique and well-separated Fréchet mean (Assumption 2.1)
Used to ensure consistency of the empirical Fréchet mean and to make the equality condition in Proposition 3.2 well defined.
domain assumption Metric space complexity control via covering numbers (Assumption 2.2)
Required for the CLT for the empirical Fréchet variance cited from Dubey and Müller (Theorem 3.1).
ad hoc to paper Existence of an admissible randomization for the space of interest
The method only guarantees power against alternatives if such a group action exists; the paper constructs examples but gives no general existence theorem.
domain assumption Distributional invariance under the randomization subgroup (unstated)
For the randomization test to have exact level, the null distribution must be invariant under the chosen isotropy subgroup; the paper only proves moment preservation, not distributional invariance.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Isotropic randomization for one-sample testing in metric spaces." pith.science (2026). https://pith.science/paper/MNMBO567

@misc{pith2026250115945,
  author       = {Pith},
  title        = {Pith review of: Isotropic randomization for one-sample testing in metric spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MNMBO567}},
  note         = {Machine review of arXiv:2501.15945}
}

read the original abstract

We address the problem of testing hypotheses about a specific value of the Fr\'echet mean in metric spaces, extending classical mean testing from Euclidean spaces to more general settings. We extend an Euclidean testing procedure progresively, starting with test construction in Riemannian manifolds, leveraging their natural geometric structure through exponential and logarithm maps, and then extend to general metric spaces through the introduction of admissible randomization techniques. This approach preserves essential geometric properties required for valid statistical inference while maintaining broad applicability. We establish theoretical guarantees for our testing procedure and demonstrate its effectiveness through numerical experiments across different metric spaces and distributional settings. The practical utility of our method is further illustrated through an application to wind data in western Denmark, showcasing its relevance for real-world statistical analysis.

Figures

Figures reproduced from arXiv: 2501.15945 by the authors.

**Figure 1.** For a fixed sample X1, . . . , X50 ∼ N(µ0, 1) with µ0 = 1, the panels display the distribution (grey histogram) of the randomized test statistic Vˆ n(X⋆ ) with 1000 randomizations for µ = µ0 (left) and µ = 0 ̸= µ0 (right) against the value of the variance on the original data Vˆ n(X) (dashed horizontal line). B and a significance level α, one can sample Z1, . . . , Zn iid∼ Bernoulli(1/2) and compute the test stati… view at source ↗

**Figure 2.** Illustration of the behavior of a random variable and its variance [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

**Figure 3.** Panel (a) shows a sample (light gray rays) from the mixture de [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Panel (a) shows a sample (light gray rays) from the VM( [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]

**Figure 5.** Figure 5: Panel (a) displays a visualization of 20 covariance matrices samples [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]

**Figure 6.** Figure 6: Panel (a) displays a visualization of the booklet space [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]

**Figure 7.** Figure 7: Panel (a) displays the dataset of wind directions collected from the [PITH_FULL_IMAGE:figures/full_fig_p024_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages

[1]

Computing Medians and Means in Hadamard Spaces

Miroslav Baˇ c´ ak. Computing Medians and Means in Hadamard Spaces. SIAM Journal on Optimization , 24(3):1542–1566, January 2014. Pub- lisher: Society for Industrial and Applied Mathematics

work page 2014
[2]

Riemannian geometry and matrix geometric means

Rajendra Bhatia and John Holbrook. Riemannian geometry and matrix geometric means. Linear Algebra and its Applications , 413(2):594–618, March 2006

work page 2006
[3]

On the Bu- res–Wasserstein distance between positive definite matrices

Rajendra Bhatia, Tanvi Jain, and Yongdo Lim. On the Bu- res–Wasserstein distance between positive definite matrices. Exposi- tiones Mathematicae, 37(2):165–191, June 2019

work page 2019
[4]

Omnibus Clts for Fr´ echet Means and Nonparametric Inference on Non-Euclidean Spaces

Rabi Bhattacharya and Lizhen Lin. Omnibus Clts for Fr´ echet Means and Nonparametric Inference on Non-Euclidean Spaces. Proceedings of the American Mathematical Society , 145(1):413–428, 2017. Publisher: American Mathematical Society

work page 2017
[5]

Concentration of empirical barycenters in metric spaces

Victor-Emmanuel Brunel and Jordan Serres. Concentration of empirical barycenters in metric spaces, March 2023. arXiv:2303.01144 [math, stat]

work page Pith review arXiv 2023
[6]

An Autoregressive Model for Time Series of Random Objects, September 2024

Matthieu Bult´ e and Helle Sørensen. An Autoregressive Model for Time Series of Random Objects, September 2024. arXiv:2405.03778

arXiv 2024
[7]

Medoid splits for efficient random forests in metric spaces

Matthieu Bult´ e and Helle Sørensen. Medoid splits for efficient random forests in metric spaces. Computational Statistics & Data Analysis , 198:107995, October 2024

work page 2024
[8]

Cambridge studies in advanced mathematics: Riemannian geometry: A modern introduction series number 98

Isaac Chavel. Cambridge studies in advanced mathematics: Riemannian geometry: A modern introduction series number 98 . Cambridge stud- ies in advanced mathematics. Cambridge University Press, Cambridge, England, 2 edition, April 2006. 27

work page 2006

Show all 39 references

[9]

Sinho Chewi, Tyler Maunu, Philippe Rigollet, and Austin J. Stromme. Gradient descent algorithms for Bures-Wasserstein barycenters. In Pro- ceedings of Thirty Third Conference on Learning Theory , pages 1276–
[10]

Fr´ echet analysis of variance for random objects

Paromita Dubey and Hans-Georg M¨ uller. Fr´ echet analysis of variance for random objects. Biometrika, 106(4):803–821, December 2019

2019
[11]

Fr´ echet change-point detec- tion

Paromita Dubey and Hans-Georg M¨ uller. Fr´ echet change-point detec- tion. The Annals of Statistics , 48(6):3312–3335, December 2020. Pub- lisher: Institute of Mathematical Statistics

2020
[12]

Huckemann

Benjamin Eltzner and Stephan F. Huckemann. A Smeary Central Limit Theorem for Manifolds with Application to High Dimensional Spheres, January 2018. arXiv:1801.06581

2018 arXiv
[13]

Les ´ el´ ements al´ eatoires de nature quelconque dans un espace distanci´ e

Maurice Fr´ echet. Les ´ el´ ements al´ eatoires de nature quelconque dans un espace distanci´ e. In Annales de l’institut Henri Poincar´ e, volume 10, pages 215–310, 1948. Issue: 4

1948
[14]

Uni- versal Bayes consistency in metric spaces

Steve Hanneke, Aryeh Kontorovich, Sivan Sabato, and Roi Weiss. Uni- versal Bayes consistency in metric spaces. The Annals of Statistics , 49(4):2129–2150, August 2021. Publisher: Institute of Mathematical Statistics

2021
[15]

Differential geometry, lie groups and symmetric spaces

Sigurdur Helgason. Differential geometry, lie groups and symmetric spaces. Graduate studies in mathematics. American Mathematical So- ciety, Providence, RI, June 2001

2001
[16]

Huckemann

Shayan Hundrieser, Benjamin Eltzner, and Stephan F. Huckemann. A Lower Bound for Estimating Fr´ echet Means, February 2024. arXiv:2402.12290 [math, stat]

2024 arXiv
[17]

Testing serial indepen- dence of object-valued time series

Feiyu Jiang, Hanjia Gao, and Xiaofeng Shao. Testing serial indepen- dence of object-valued time series. Biometrika, page asad069, November 2023

2023
[18]

Wilfrid S. Kendall. Probability, Convexity, and Harmonic Maps with Small Image I: Uniqueness and Fine Existence. Proceedings of the London Mathematical Society , s3-61(2):371–406, 1990. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1112/plms/s3-61.2.371. 28

1990 doi
[19]

Robust Signal Recovery in Hadamard Spaces, July 2023

Georg K¨ ostenberger and Thomas Stark. Robust Signal Recovery in Hadamard Spaces, July 2023. arXiv:2307.06057 [math, stat]

2023 arXiv
[20]

On the Consistency of Procrustean Mean Shapes

Huiling Le. On the Consistency of Procrustean Mean Shapes. Advances in Applied Probability, 30(1):53–63, 1998. Publisher: Applied Probabil- ity Trust

1998
[21]

Testing statistical hypotheses

E L Lehmann and Joseph P Romano. Testing statistical hypotheses . Springer texts in statistics. Springer Nature, Cham, Switzerland, 4 edi- tion, June 2022

2022
[22]

Wasserstein Riemannian geometry of Gaussian densities

Luigi Malag` o, Luigi Montrucchio, and Giovanni Pistone. Wasserstein Riemannian geometry of Gaussian densities. Information Geometry , 1(2):137–179, December 2018

2018
[23]

Directional statistics

Kanti V Mardia and Peter E Jupp. Directional statistics. John Wiley & Sons, 2009

2009
[24]

Equivariant estimation of Fr´ echet means

A McCormack and P D Hoff. Equivariant estimation of Fr´ echet means. Biometrika, page asad014, February 2023

2023
[25]

The Stein effect for Fr´ echet means

Andrew McCormack and Peter Hoff. The Stein effect for Fr´ echet means. The Annals of Statistics , 50(6):3647–3676, December 2022. Publisher: Institute of Mathematical Statistics

2022
[26]

Panaretos and Yoav Zemel

Victor M. Panaretos and Yoav Zemel. An Invitation to Statistics in Wasserstein Space . SpringerBriefs in Probability and Mathematical Statistics. Springer International Publishing, Cham, 2020

2020
[27]

Frechet regression for ran- dom objects with Euclidean predictors

Alexander Petersen and Hans-Georg Muller. Frechet regression for ran- dom objects with Euclidean predictors. page 29, 2019

2019
[28]

Functional data analysis for density functions by transformation to a Hilbert space

Alexander Petersen and Hans-Georg M¨ uller. Functional data analysis for density functions by transformation to a Hilbert space. The Annals of Statistics , 44(1), February 2016. arXiv: 1601.02869

2016 arXiv
[29]

J. O. Ramsay and B. W. Silverman. Functional data analysis . Springer series in statistics. Springer, New York, 2nd ed edition, 2005. 29

2005
[30]

Convergence rates for the generalized Fr´ echet mean via the quadruple inequality

Christof Sch¨ otz. Convergence rates for the generalized Fr´ echet mean via the quadruple inequality. Electronic Journal of Statistics , 13(2):4280– 4345, January 2019. Publisher: Institute of Mathematical Statistics and Bernoulli Society

2019
[31]

Anuj Srivastava and Eric P. Klassen. Functional and Shape Data Anal- ysis. Springer Series in Statistics. Springer New York, New York, NY, 2016

2016
[32]

Probability measures on metric spaces of non- positive curvature

Karl-Theodor Sturm. Probability measures on metric spaces of non- positive curvature. In Pascal Auscher, Thierry Coulhon, and Alexander Grigor’yan, editors, Contemporary Mathematics, volume 338, pages 357–
[33]

Theoretically and Computation- ally Convenient Geometries on Full-Rank Correlation Matrices

Yann Thanwerdas and Xavier Pennec. Theoretically and Computation- ally Convenient Geometries on Full-Rank Correlation Matrices. SIAM Journal on Matrix Analysis and Applications , 43(4):1851–1872, Decem- ber 2022. Publisher: Society for Industrial and Applied Mathematics

2022
[34]

J. R. Toggweiler. Shifting westerlies. Science, 323(5920):1434–1435, 2009

2009
[35]

A. W. van der Vaart. Asymptotic statistics . Cambridge series in statis- tical and probabilistic mathematics. Cambridge University Press, Cam- bridge, UK ; New York, NY, USA, 1998

1998
[36]

Bayes Hilbert Spaces

Karl Gerald van den Boogaart, Juan Jos´ e Egozcue, and Vera Pawlowsky- Glahn. Bayes Hilbert Spaces. Australian & New Zealand Journal of Statistics, 56(2):171–194, June 2014

2014
[37]

Law of large numbers in CAT(1)-spaces of small radii

Takumi Yokota. Law of large numbers in CAT(1)-spaces of small radii. Calculus of Variations and Partial Differential Equations , 57(2):35, February 2018. 30

2018
[390]

American Mathematical Society, Providence, Rhode Island, 2003

2003
[1304]

ISSN: 2640-3498

PMLR, July 2020. ISSN: 2640-3498

2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Computing Medians and Means in Hadamard Spaces

Miroslav Baˇ c´ ak. Computing Medians and Means in Hadamard Spaces. SIAM Journal on Optimization , 24(3):1542–1566, January 2014. Pub- lisher: Society for Industrial and Applied Mathematics

work page 2014

[2] [2]

Riemannian geometry and matrix geometric means

Rajendra Bhatia and John Holbrook. Riemannian geometry and matrix geometric means. Linear Algebra and its Applications , 413(2):594–618, March 2006

work page 2006

[3] [3]

On the Bu- res–Wasserstein distance between positive definite matrices

Rajendra Bhatia, Tanvi Jain, and Yongdo Lim. On the Bu- res–Wasserstein distance between positive definite matrices. Exposi- tiones Mathematicae, 37(2):165–191, June 2019

work page 2019

[4] [4]

Omnibus Clts for Fr´ echet Means and Nonparametric Inference on Non-Euclidean Spaces

Rabi Bhattacharya and Lizhen Lin. Omnibus Clts for Fr´ echet Means and Nonparametric Inference on Non-Euclidean Spaces. Proceedings of the American Mathematical Society , 145(1):413–428, 2017. Publisher: American Mathematical Society

work page 2017

[5] [5]

Concentration of empirical barycenters in metric spaces

Victor-Emmanuel Brunel and Jordan Serres. Concentration of empirical barycenters in metric spaces, March 2023. arXiv:2303.01144 [math, stat]

work page Pith review arXiv 2023

[6] [6]

An Autoregressive Model for Time Series of Random Objects, September 2024

Matthieu Bult´ e and Helle Sørensen. An Autoregressive Model for Time Series of Random Objects, September 2024. arXiv:2405.03778

arXiv 2024

[7] [7]

Medoid splits for efficient random forests in metric spaces

Matthieu Bult´ e and Helle Sørensen. Medoid splits for efficient random forests in metric spaces. Computational Statistics & Data Analysis , 198:107995, October 2024

work page 2024

[8] [8]

Cambridge studies in advanced mathematics: Riemannian geometry: A modern introduction series number 98

Isaac Chavel. Cambridge studies in advanced mathematics: Riemannian geometry: A modern introduction series number 98 . Cambridge stud- ies in advanced mathematics. Cambridge University Press, Cambridge, England, 2 edition, April 2006. 27

work page 2006