pith. sign in

arxiv: 2605.20681 · v1 · pith:H5WAIAX2new · submitted 2026-05-20 · 📊 stat.ME · cs.LG

Scale-Calibrated Median-of-Means for Robust Distributed Principal Component Analysis

Kisung You This is my paper

Pith reviewed 2026-05-21 03:00 UTC · model grok-4.3

classification 📊 stat.ME cs.LG
keywords distributed PCAmedian-of-meansrobust estimationproduct manifoldGrassmann manifoldscale calibrationasymptotic equivalence
0
0 comments X

The pith

The scale-calibrated median-of-means estimator on the product manifold is asymptotically equivalent to a scaled spatial median of node influence errors in distributed PCA.

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

The paper shows how to combine PCA results computed separately at many distributed nodes into a single robust summary by treating the mean vector and the principal subspace as a single object on their product space. It calibrates the relative scale between mean errors and subspace errors so that a median-of-means procedure on this combined space behaves like taking a scaled spatial median of the influence functions from each node. A sympathetic reader would care because the resulting limits are explicit: non-Gaussian when the number of nodes stays fixed, Gaussian with a bias term when nodes increase, and with a covariance that depends on the chosen scale. This gives concrete rules for choosing the calibration, high-probability bounds, and valid bootstrap inference even when some nodes are unreliable.

Core claim

We prove a local reduction showing that the proposed product-manifold median-of-means estimator is asymptotically equivalent to a scaled spatial median of node influence errors. This yields fixed-node non-Gaussian limits, growing-node Gaussian limits with finite-block bias, and an explicit scale-dependent covariance formula. The reduction rests on a node-level PCA expansion in which the mean component has the usual linear influence while the subspace component is an eigengap-weighted covariance perturbation.

What carries the argument

The product-manifold median-of-means estimator on the combined space of Euclidean means and Grassmann subspaces, with explicit calibration of the relative scale between the two error types.

Load-bearing premise

Node-level PCA estimates admit a first-order expansion separating a linear influence term for the mean from an eigengap-weighted perturbation term for the subspace.

What would settle it

A controlled simulation in which node influence errors are known exactly, the scale is deliberately set away from the eigengap-driven value, and the empirical distribution of the aggregated estimator is checked against the claimed non-Gaussian or Gaussian limit and covariance formula.

Figures

Figures reproduced from arXiv: 2605.20681 by Kisung You.

Figure 1
Figure 1. Figure 1: Spiked PCA simulation with p = 200, r = 5, n = 40000, and K = 80 nodes. As the eigengap increases, subspace estimation becomes easier and the calibrated scale adjusts to the changing mean–subspace uncertainty. Panels show (a) Grassmann subspace error, (b) Euclidean mean error, (c) selected scale αbrPCA, and (d) subspace error relative to scale-calibrated MoM. and 0.28 for projector averaging. Full-sample P… view at source ↗
Figure 2
Figure 2. Figure 2: Scale diagnostics. Panels show (a) the calibrated scale [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Effect of increasing the fraction of corrupted nodes in the spiked PCA model with eigengap [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of increasing bad-node perturbation magnitude at a fixed contamination fraction. [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: 10x mouse brain aligned PCA embeddings. Two-dimensional embeddings of a common [PITH_FULL_IMAGE:figures/full_fig_p036_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: 10x mouse brain clean splitting. Distributed PCA comparison using [PITH_FULL_IMAGE:figures/full_fig_p037_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 10x mouse brain bad-node stress test. Node-level corruption experiment with [PITH_FULL_IMAGE:figures/full_fig_p038_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Oracle covariance check. Gaussian node errors are generated directly in tangent coordi [PITH_FULL_IMAGE:figures/full_fig_p039_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Oracle replication distributions. Distributional summaries from the Gaussian node-error [PITH_FULL_IMAGE:figures/full_fig_p040_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Fixed-K, growing-K, and aggressive-K regimes. Simulation illustrating the finite-block bias term in the growing-node CLT. The fixed-K regime moves toward centered behavior as b increases, while the aggressive regime has a visibly shifted first root-n coordinate. Panels show (a) mean of the first root-n coordinate, (b) standard deviation of the first root-n coordinate, (c) mean root-n norm, and (d) boxplot… view at source ↗
Figure 11
Figure 11. Figure 11: Tabula Sapiens blood-cell PCA. Distributed PCA comparison on the Tabula Sapiens [PITH_FULL_IMAGE:figures/full_fig_p042_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: NOAA OISST EOF/PCA analysis. Distributed EOF/PCA comparison on a regional [PITH_FULL_IMAGE:figures/full_fig_p043_12.png] view at source ↗
read the original abstract

Distributed principal component analysis (PCA) produces node-level estimates of both a mean vector and a principal subspace. Robustly aggregating these heterogeneous objects requires a relative scale between mean error and subspace error. We study a scale-calibrated median-of-means estimator for this problem using the product geometry of Euclidean space and the Grassmann manifold. A node-level PCA expansion shows that the mean component has the usual linear influence, whereas the subspace component is an eigengap-weighted covariance perturbation. We prove a local reduction showing that the proposed product-manifold median-of-means estimator is asymptotically equivalent to a scaled spatial median of node influence errors. This yields fixed-node non-Gaussian limits, growing-node Gaussian limits with finite-block bias, and an explicit scale-dependent covariance formula. We propose robust block-scale and inference-optimal calibration rules, establish high-probability median-of-means bounds, characterize factorwise bad-node influence, and prove node-bootstrap validity. Simulations and large-scale single-cell RNA-seq data show that scale calibration adapts to eigengap-driven subspace uncertainty and provides a robust distributed PCA summary.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 3 minor

Summary. The manuscript proposes a scale-calibrated median-of-means estimator for robust distributed PCA that aggregates node-level mean vectors and principal subspaces on the product manifold of Euclidean space and the Grassmann manifold. A node-level PCA expansion separates the linear influence of the mean component from an eigengap-weighted covariance perturbation on the subspace component. The central theoretical result is a local reduction establishing asymptotic equivalence of the product-manifold estimator to a scaled spatial median of node influence errors, which yields fixed-node non-Gaussian limits, growing-node Gaussian limits with finite-block bias, and an explicit scale-dependent covariance formula. The paper also develops robust block-scale and inference-optimal calibration rules, high-probability median-of-means bounds, a characterization of factorwise bad-node influence, and node-bootstrap validity, with supporting simulations and an application to large-scale single-cell RNA-seq data.

Significance. If the local reduction and limit theorems hold under the stated assumptions, the work supplies a theoretically justified approach to robust aggregation in distributed PCA that explicitly handles the relative scale between mean and subspace errors—an issue that arises whenever eigengaps vary across nodes. The explicit scale-dependent covariance, the high-probability bounds, and the bootstrap validity result are concrete strengths that could be useful for downstream inference in high-dimensional settings such as genomics. The simulations and real-data example provide empirical support for the claim that scale calibration adapts to subspace uncertainty.

major comments (2)
  1. [Section 3 (proof of local reduction)] The local reduction to the scaled spatial median (central to all limit statements) rests on the first-order node-level PCA expansion that separates Euclidean mean influence from the eigengap-weighted Grassmann perturbation. The manuscript should state the precise lower bound on the eigengap and the perturbation order required for the equivalence to hold uniformly in the number of nodes; without these rates it is difficult to verify that no additional bias terms enter the limiting distribution.
  2. [Section 4.2 (growing-node limit and scale calibration)] The growing-node Gaussian limit is stated to contain a finite-block bias term whose magnitude depends on the calibrated scale. It is not immediately clear whether this bias vanishes or remains when the scale is estimated from the same data; an explicit statement of the joint asymptotics for the estimator and the scale would strengthen the claim.
minor comments (3)
  1. [Abstract] The abstract refers to 'finite-block bias' without a parenthetical definition or forward reference; adding a short clause or pointer to the relevant theorem would improve readability for readers outside the immediate subfield.
  2. [Section 2 (model and notation)] Notation for the product-manifold distance and the relative scale parameter is introduced without an explicit reminder of its dimension; a one-line display equation collecting the definitions would reduce cross-referencing.
  3. [Section 6 (real-data example)] The single-cell RNA-seq experiment reports qualitative improvement but does not include a quantitative table of subspace estimation error or downstream clustering metrics across calibration choices; adding such a table would make the empirical evidence more persuasive.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading, positive assessment, and constructive suggestions. The comments help clarify the conditions for our local reduction and the joint asymptotics under scale estimation. We address each major comment below and will incorporate the requested clarifications in the revised manuscript.

read point-by-point responses

  1. Referee: [Section 3 (proof of local reduction)] The local reduction to the scaled spatial median (central to all limit statements) rests on the first-order node-level PCA expansion that separates Euclidean mean influence from the eigengap-weighted Grassmann perturbation. The manuscript should state the precise lower bound on the eigengap and the perturbation order required for the equivalence to hold uniformly in the number of nodes; without these rates it is difficult to verify that no additional bias terms enter the limiting distribution.

    Authors: We agree that the uniformity conditions should be stated explicitly. In the revision we will add to Section 3 the standing assumption that the eigengap is bounded below by a positive constant δ > 0 independent of the number of nodes n, together with the requirement that local perturbations are of order O_p(m^{-1/2}) where m denotes the per-node sample size. Under these rates the first-order node-level expansion holds uniformly in n and no extra bias terms enter the limiting distribution of the product-manifold median-of-means estimator. A new remark will summarize the resulting uniformity statement. revision: yes

  2. Referee: [Section 4.2 (growing-node limit and scale calibration)] The growing-node Gaussian limit is stated to contain a finite-block bias term whose magnitude depends on the calibrated scale. It is not immediately clear whether this bias vanishes or remains when the scale is estimated from the same data; an explicit statement of the joint asymptotics for the estimator and the scale would strengthen the claim.

    Authors: We thank the referee for pointing out the need for joint asymptotics. The current statement of the growing-node limit treats the scale as fixed. In the revision we will add to Section 4.2 an explicit joint expansion showing that, under the robust block-scale calibration, the estimated scale converges at rate o_p(n^{-1/2}) and therefore the finite-block bias term remains asymptotically negligible. The resulting covariance formula will be stated jointly for the estimator and the data-driven scale. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained

full rationale

The paper derives its central local reduction from a node-level PCA expansion that separates the Euclidean mean influence (standard linear term) from an eigengap-weighted covariance perturbation on the Grassmann component. This expansion, combined with product-manifold geometry, directly yields the asymptotic equivalence to a scaled spatial median of node influence errors, producing the stated fixed-node non-Gaussian limits, growing-node Gaussian limits, and explicit scale-dependent covariance. The proposed block-scale and inference-optimal calibration rules are data-dependent tuning steps that feed into the estimator but do not redefine or tautologically reproduce the asymptotic limits or covariance formula by construction. No self-citation chain, ansatz smuggling, or renaming of known results is required for the load-bearing steps; the argument remains internally consistent with standard manifold perturbation theory and is not forced to equal its inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claims rest on the validity of the node-level PCA expansion, the appropriateness of the product manifold, and the existence of robust block-scale and inference-optimal calibration rules derived from the local reduction.

free parameters (1)
  • relative scale between mean and subspace errors
    Calibrated via robust block-scale and inference-optimal rules to adapt to eigengap-driven subspace uncertainty.
axioms (1)
  • domain assumption Node-level PCA expansion accurately decomposes mean influence as linear and subspace influence as eigengap-weighted covariance perturbation.
    Invoked to justify the product geometry and the reduction to scaled spatial median.

pith-pipeline@v0.9.0 · 5707 in / 1266 out tokens · 49557 ms · 2026-05-21T03:00:35.098362+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

300 extracted references · 300 canonical work pages · 4 internal anchors

  1. [1]

    Science , author =

    The. Science , author =. 2022 , pages =. doi:10.1126/science.abl4896 , abstract =

    work page doi:10.1126/science.abl4896 2022

  2. [2]

    Perturbation

    Katō, Tosio , year =. Perturbation

    work page

  3. [3]

    Pearson,Liii

    The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science , author =. 1901 , pages =. doi:10.1080/14786440109462720 , language =

    work page doi:10.1080/14786440109462720 1901

  4. [4]

    Stewart, G. W. and Sun, Ji-guang , year =. Matrix

    work page

  5. [5]

    Journal of Climate 34, 2923–2939

    Improvements of the. Journal of Climate , author =. 2021 , pages =. doi:10.1175/JCLI-D-20-0166.1 , abstract =

    work page doi:10.1175/jcli-d-20-0166.1 2021

  6. [6]

    Journal of Climate , author =

    An. Journal of Climate , author =. 2002 , pages =. doi:10.1175/1520-0442(2002)015<1609:AIISAS>2.0.CO;2 , language =

    work page doi:10.1175/1520-0442(2002)015 2002

  7. [7]

    Pisco, Angela , year =. Tabula. doi:10.6084/M9.FIGSHARE.27921984 , abstract =

    work page doi:10.6084/m9.figshare.27921984

  8. [8]

    SIAM Journal on Matrix Analysis and Applications , author =

    The. SIAM Journal on Matrix Analysis and Applications , author =. 1998 , pages =. doi:10.1137/S0895479895290954 , language =

    work page doi:10.1137/s0895479895290954 1998

  9. [9]

    and Judin, David B

    Nemirovskij, Arkadij S. and Judin, David B. and Dawson, E. R. and Nemirovskij, Arkadij S. , year =. Problem complexity and method efficiency in optimization , isbn =

    work page

  10. [10]

    Foundations of Computational Mathematics , author =

    Mean. Foundations of Computational Mathematics , author =. 2019 , note =. doi:10.1007/s10208-019-09427-x , language =

    work page doi:10.1007/s10208-019-09427-x 2019

  11. [11]

    The Annals of Statistics , author =

    Sub-. The Annals of Statistics , author =. doi:10.1214/17-AOS1639 , number =

    work page doi:10.1214/17-aos1639

  12. [12]

    Bernoulli , author =

    Geometric median and robust estimation in. Bernoulli , author =. doi:10.3150/14-BEJ645 , number =

    work page doi:10.3150/14-bej645

  13. [13]

    Hsu, Daniel and Sabato, Sivan , month = jan, year =. Loss. Journal of Machine Learning Research , publisher =

    work page

  14. [14]

    and Oliveira, R

    Lerasle, M. and Oliveira, R. I. , year =. Robust empirical mean. doi:10.48550/ARXIV.1112.3914 , abstract =

    work page doi:10.48550/arxiv.1112.3914

  15. [15]

    Journal of Computer and System Sciences , author =

    The. Journal of Computer and System Sciences , author =. 1999 , pages =. doi:10.1006/jcss.1997.1545 , language =

    work page doi:10.1006/jcss.1997.1545 1999

  16. [16]

    Theoretical Computer Science , author =

    Random generation of combinatorial structures from a uniform distribution , volume =. Theoretical Computer Science , author =. 1986 , pages =. doi:10.1016/0304-3975(86)90174-X , language =

    work page doi:10.1016/0304-3975(86)90174-x 1986

  17. [17]

    The Annals of Statistics , author =

    Distributed estimation of principal eigenspaces , volume =. The Annals of Statistics , author =. doi:10.1214/18-AOS1713 , number =

    work page doi:10.1214/18-aos1713

  18. [18]

    SIAM Journal on Numerical Analysis , author =

    The. SIAM Journal on Numerical Analysis , author =. 1970 , pages =. doi:10.1137/0707001 , language =

    work page doi:10.1137/0707001 1970

  19. [19]

    Probability content of regions under spherical normal distributions, IV: The distribution of homogeneous and non-homogeneous quadratic functions of normal variables,

    Asymptotic. The Annals of Mathematical Statistics , author =. 1963 , pages =. doi:10.1214/aoms/1177704248 , language =

    work page doi:10.1214/aoms/1177704248 1963

  20. [20]

    , volume =

    Analysis of a complex of statistical variables into principal components. , volume =. Journal of Educational Psychology , author =. 1933 , pages =. doi:10.1037/h0071325 , language =

    work page doi:10.1037/h0071325 1933

  21. [21]

    Brain Connectivity , author =

    Geodesic. Brain Connectivity , author =. 2021 , pages =. doi:10.1089/brain.2020.0881 , language =

    work page doi:10.1089/brain.2020.0881 2021

  22. [22]

    SIAM Journal on Matrix Analysis and Applications , author =

    Schubert. SIAM Journal on Matrix Analysis and Applications , author =. 2016 , pages =. doi:10.1137/15M1054201 , language =

    work page doi:10.1137/15m1054201 2016

  23. [23]

    , year =

    Schieferdecker, Dennis and Huber, Marco F. , year =. Gaussian mixture reduction via clustering , booktitle =

    work page

  24. [24]

    Huber means on

    Lee, Jongmin and Jung, Sungkyu , month = jul, year =. Huber means on. doi:10.48550/arXiv.2407.15764 , abstract =

    work page doi:10.48550/arxiv.2407.15764

  25. [25]

    , year =

    Huber, Peter J. , year =. Robust

    work page

  26. [26]

    and Hanebeck, Uwe D

    Huber, Marco F. and Hanebeck, Uwe D. , year =. Progressive. doi:10.5445/IR/1000034859 , abstract =

    work page doi:10.5445/ir/1000034859

  27. [27]

    A first-order primal-dual algorithm for convex prob- lems with applications to imaging.J

    A. Journal of Mathematical Imaging and Vision , author =. 2011 , pages =. doi:10.1007/s10851-010-0251-1 , language =

    work page doi:10.1007/s10851-010-0251-1 2011

  28. [28]

    Proceedings of the National Academy of Sciences , author =

    The multivariate. Proceedings of the National Academy of Sciences , author =. 2000 , pages =. doi:10.1073/pnas.97.4.1423 , language =

    work page doi:10.1073/pnas.97.4.1423 2000

  29. [29]

    van der Vaart, A. W. and Wellner, Jon A. , year =. Weak convergence and empirical processes , isbn =

    work page

  30. [30]

    van der Vaart, A. W. , year =. Asymptotic

    work page

  31. [31]

    Journal of Machine Learning Research , author =

    Visualizing data using t-. Journal of Machine Learning Research , author =. 2008 , pages =

    work page 2008

  32. [32]

    , month = nov, year =

    Serfling, Robert J. , month = nov, year =. Approximation. doi:10.1002/9780470316481 , language =

    work page doi:10.1002/9780470316481

  33. [33]

    The Annals of Mathematical Statistics , author =

    A. The Annals of Mathematical Statistics , author =. 1966 , pages =. doi:10.1214/aoms/1177699450 , language =

    work page doi:10.1214/aoms/1177699450 1966

  34. [34]

    , year =

    Kosorok, Michael R. , year =. Introduction to

    work page

  35. [35]

    and Ronchetti, Elvezio M

    Hampel, Frank R. and Ronchetti, Elvezio M. and Rousseeuw, Peter J. and Stahel, Werner A. , month = mar, year =. Robust. doi:10.1002/9781118186435 , language =

    work page doi:10.1002/9781118186435

  36. [36]

    Constant

    You, Kisung , year =. Constant. doi:10.48550/ARXIV.2601.10992 , abstract =

    work page doi:10.48550/arxiv.2601.10992

  37. [37]

    Bernoulli , author =

    Fréchet means and. Bernoulli , author =. doi:10.3150/17-BEJ1009 , number =

    work page doi:10.3150/17-bej1009

  38. [38]

    and Zemel, Yoav , year =

    Panaretos, Victor M. and Zemel, Yoav , year =. An. doi:10.1007/978-3-030-38438-8 , language =

    work page doi:10.1007/978-3-030-38438-8

  39. [39]

    Projection

    Huang, Minhui and Ma, Shiqian and Lai, Lifeng , editor =. Projection. Proceedings of the 38th international conference on machine learning , publisher =. 2021 , pages =

    work page 2021

  40. [40]

    Robust Estimation of a Location Parameter

    Robust. The Annals of Mathematical Statistics , author =. 1964 , pages =. doi:10.1214/aoms/1177703732 , language =

    work page doi:10.1214/aoms/1177703732 1964

  41. [41]

    The Annals of Applied Probability , author =

    Statistical inference for. The Annals of Applied Probability , author =. doi:10.1214/20-AAP1618 , number =

    work page doi:10.1214/20-aap1618

  42. [42]

    Journal of the European Mathematical Society , author =

    Fast convergence of empirical barycenters in. Journal of the European Mathematical Society , author =. 2022 , pages =. doi:10.4171/JEMS/1234 , language =

    work page doi:10.4171/jems/1234 2022

  43. [43]

    Probability Theory and Related Fields , author =

    Existence and consistency of. Probability Theory and Related Fields , author =. 2017 , pages =. doi:10.1007/s00440-016-0727-z , language =

    work page doi:10.1007/s00440-016-0727-z 2017

  44. [44]

    Journal of Mathematical Analysis and Applications , author =

    A fixed-point approach to barycenters in. Journal of Mathematical Analysis and Applications , author =. 2016 , pages =. doi:10.1016/j.jmaa.2016.04.045 , language =

    work page doi:10.1016/j.jmaa.2016.04.045 2016

  45. [45]

    Cheng, Zixiong and Liu, Hang , year =. Robust. doi:10.48550/ARXIV.2603.07563 , abstract =

    work page doi:10.48550/arxiv.2603.07563

  46. [46]

    Journal of the Royal Statistical Society Series B: Statistical Methodology , author =

    Huber means on. Journal of the Royal Statistical Society Series B: Statistical Methodology , author =. 2026 , pages =. doi:10.1093/jrsssb/qkaf054 , abstract =

    work page doi:10.1093/jrsssb/qkaf054 2026

  47. [47]

    Journal of Multivariate Analysis , author =

    Geometric medians on product manifolds , volume =. Journal of Multivariate Analysis , author =. 2026 , pages =. doi:10.1016/j.jmva.2026.105653 , language =

    work page doi:10.1016/j.jmva.2026.105653 2026

  48. [48]

    Foundations and Trends® in Machine Learning , author =

    Kernel. Foundations and Trends® in Machine Learning , author =. 2017 , pages =. doi:10.1561/2200000060 , language =

    work page doi:10.1561/2200000060 2017

  49. [49]

    Learning from distributions via support measure machines , volume =

    Muandet, Krikamol and Fukumizu, Kenji and Dinuzzo, Francesco and Schölkopf, Bernhard , editor =. Learning from distributions via support measure machines , volume =. Advances in neural information processing systems , publisher =

    work page

  50. [50]

    and Salzmann, Mathieu and Jayasumana, Sadeep and Hartley, Richard and Li, Hongdong , editor =

    Harandi, Mehrtash T. and Salzmann, Mathieu and Jayasumana, Sadeep and Hartley, Richard and Li, Hongdong , editor =. Expanding the. Computer. 2014 , pages =. doi:10.1007/978-3-319-10584-0_27 , language =

    work page doi:10.1007/978-3-319-10584-0_27 2014

  51. [51]

    , year =

    Hamm, Jihun and Lee, Daniel D. , year =. Grassmann. Proceedings of the 25th international conference on. doi:10.1145/1390156.1390204 , language =

    work page doi:10.1145/1390156.1390204

  52. [52]

    Learning over

    Wolf, Lior and Shashua, Amnon , month = dec, year =. Learning over. Journal of Machine Learning Research , publisher =

    work page

  53. [53]

    Vehtari, A

    Rank-. Bayesian Analysis , author =. doi:10.1214/20-BA1221 , number =

    work page doi:10.1214/20-ba1221

  54. [54]

    doi:10.1214/ss/1177011136 , author =

    Inference from. Statistical Science , author =. doi:10.1214/ss/1177011136 , number =

    work page doi:10.1214/ss/1177011136

  55. [55]

    Probability Surveys , author =

    Basic. Probability Surveys , author =. doi:10.1214/154957805100000104 , number =

    work page doi:10.1214/154957805100000104

  56. [56]

    Journal of Machine Learning Research , author =

    Kernel. Journal of Machine Learning Research , author =. 2022 , pages =

    work page 2022

  57. [57]

    The Annals of Statistics , author =

    Batch means and spectral variance estimators in. The Annals of Statistics , author =. doi:10.1214/09-AOS735 , number =

    work page doi:10.1214/09-aos735

  58. [58]

    Journal of the American Statistical Association , author =

    Fixed-. Journal of the American Statistical Association , author =. 2006 , pages =. doi:10.1198/016214506000000492 , language =

    work page doi:10.1198/016214506000000492 2006

  59. [59]

    Geodesic exponential kernels:

    Feragen, Aasa and Lauze, Francois and Hauberg, Soren , month = jun, year =. Geodesic exponential kernels:. 2015. doi:10.1109/CVPR.2015.7298922 , urldate =

    work page doi:10.1109/cvpr.2015.7298922 2015

  60. [60]

    SIAM Journal on Matrix Analysis and Applications , author =

    Geometric. SIAM Journal on Matrix Analysis and Applications , author =. 2007 , pages =. doi:10.1137/050637996 , number =

    work page doi:10.1137/050637996 2007

  61. [61]

    and Salzmann, Mathieu and Hartley, Richard , editor =

    Harandi, Mehrtash T. and Salzmann, Mathieu and Hartley, Richard , editor =. From. Computer. 2014 , pages =. doi:10.1007/978-3-319-10605-2_2 , language =

    work page doi:10.1007/978-3-319-10605-2_2 2014

  62. [62]

    SIAM Journal on Matrix Analysis and Applications , author =

    Theoretically and. SIAM Journal on Matrix Analysis and Applications , author =. 2022 , pages =. doi:10.1137/22M1471729 , language =

    work page doi:10.1137/22m1471729 2022

  63. [63]

    Geodesics and

    Thanwerdas, Yann and Pennec, Xavier , editor =. Geodesics and. Geometric. 2021 , pages =. doi:10.1007/978-3-030-80209-7_11 , language =

    work page doi:10.1007/978-3-030-80209-7_11 2021

  64. [64]

    Journal of Machine Learning Research , author =

    Universality,. Journal of Machine Learning Research , author =. 2011 , pages =

    work page 2011

  65. [65]

    Bernoulli , author =

    Strictly and non-strictly positive definite functions on spheres , volume =. Bernoulli , author =. doi:10.3150/12-BEJSP06 , number =

    work page doi:10.3150/12-bejsp06

  66. [66]

    Duke Mathematical Journal , author =

    Positive definite functions on spheres , volume =. Duke Mathematical Journal , author =. doi:10.1215/S0012-7094-42-00908-6 , number =

    work page doi:10.1215/s0012-7094-42-00908-6

  67. [67]

    Reinhard Selten

    Equivalence of distance-based and. The Annals of Statistics , author =. doi:10.1214/13-AOS1140 , number =

    work page doi:10.1214/13-aos1140

  68. [68]

    and Rasch, Malte J

    Gretton, Arthur and Borgwardt, Karsten M. and Rasch, Malte J. and Schölkopf, Bernhard and Smola, Alexander , month = mar, year =. A. Journal of Machine Learning Research , publisher =

    work page

  69. [69]

    Positive

    Bhatia, Rajendra , month = dec, year =. Positive. doi:10.1515/9781400827787 , urldate =

    work page doi:10.1515/9781400827787

  70. [70]

    Medians and

    Arnaudon, Marc and Barbaresco, Frédéric and Yang, Le , editor =. Medians and. Matrix. 2013 , note =. doi:10.1007/978-3-642-30232-9_8 , language =

    work page doi:10.1007/978-3-642-30232-9_8 2013

  71. [71]

    Econometrics and Statistics , author =

    Modeling. Econometrics and Statistics , author =. 2022 , pages =. doi:10.1016/j.ecosta.2021.04.004 , language =

    work page doi:10.1016/j.ecosta.2021.04.004 2022

  72. [72]

    Springer, New York, 2 edition, 2006

    Petersen, Peter , year =. Riemannian. doi:10.1007/978-0-387-29403-2 , publisher =

    work page doi:10.1007/978-0-387-29403-2

  73. [73]

    You, Kisung , year =. Finite. doi:10.48550/ARXIV.2604.24895 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.24895

  74. [74]

    The Annals of Statistics , author =

    Fréchet regression for random objects with. The Annals of Statistics , author =. doi:10.1214/17-AOS1624 , number =

    work page doi:10.1214/17-aos1624

  75. [75]

    NeuroImage , author =

    The. NeuroImage , author =. 2012 , pages =. doi:10.1016/j.neuroimage.2012.02.018 , language =

    work page doi:10.1016/j.neuroimage.2012.02.018 2012

  76. [76]

    and Pedersen, M.S

    Petersen, K.B. and Pedersen, M.S. , month = nov, year =. The

    work page

  77. [77]

    Science , author =

    Prediction of. Science , author =. 2010 , pages =. doi:10.1126/science.1194144 , language =

    work page doi:10.1126/science.1194144 2010

  78. [78]

    A wrapped normal distribution on hyperbolic space for gradient-based learning , volume =

    Nagano, Yoshihiro and Yamaguchi, Shoichiro and Fujita, Yasuhiro and Koyama, Masanori , editor =. A wrapped normal distribution on hyperbolic space for gradient-based learning , volume =. Proceedings of the 36th international conference on machine learning , publisher =. 2019 , pages =

    work page 2019

  79. [79]

    , year =

    Lee, John M. , year =. Riemannian

    work page

  80. [80]

    Asymptotics in

    Le Cam, Lucien and Lo Yang, Grace , year =. Asymptotics in. doi:10.1007/978-1-4612-1166-2 , publisher =

    work page doi:10.1007/978-1-4612-1166-2

Showing first 80 references.