Pith. sign in

REVIEW 2 major objections 5 minor 296 references

This paper proves that aggregating block means through a radial Huber center—HOMER—matches median-of-means deviation bounds under a finite second moment while adding a threshold-controlled path to the sample mean and asymptotic inference.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 06:10 UTC pith:3ERJ72NC

load-bearing objection Sound, honest Hilbert-space Huber-of-means theory; the practical tuning rule is outside all the theorems, and the guarantee is block-summary contamination only — send to peer review. the 2 major comments →

arxiv 2607.27532 v1 pith:3ERJ72NC submitted 2026-07-29 stat.ML cs.LG

HOMER: Huber-of-Means for Efficient and Robust Estimation in Hilbert Spaces

Kisung You , Boram Cho This is my paper
classification stat.ML cs.LG MSC 62G3562G2062R10
keywords HOMERrobust mean estimationmedian-of-meansHilbert spaceHuber lossheavy tailssandwich covarianceasymptotic linearity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Heavy-tailed data make the ordinary sample mean unreliable. Median-of-means fixes that by averaging within blocks and then taking a robust median of the block averages, but its outer median has no dial for recovering the sample mean's efficiency on clean data. This paper proposes HOMER, which replaces that outer median with a radial Huber center: block averages that are close to the center are treated quadratically, like the mean, while distant ones are compressed linearly, like a median. The paper proves a 5/8-majority theorem—if that fraction of block averages are within radius r of the truth, every HOMER solution lies within 2r—and, under only a finite second moment, an exponential deviation bound of order sigma sqrt(log(1/delta)/n). It further shows that the smooth pseudo-Huber version approaches the sample mean as its threshold grows and, with a finite third moment and balanced block growth, gives asymptotically normal inference for finite-dimensional projections. The payoff is one estimator family that interpolates between median-like robustness and mean efficiency while supplying a consistent sandwich covariance.

Core claim

The central claim: aggregating block averages through a radial Huber center keeps the median-of-means high-confidence guarantee while gaining a threshold-controlled route to the sample mean and to inference. HOMER minimizes the average radial loss rho_tau(||Z_j - theta||); the canonical loss is quadratic up to tau and linear beyond, and the smooth pseudo-Huber loss converges to the sample mean as tau grows. If at least 5/8 of the block means lie within r of the truth and tau <= r, every solution is within 2r; under only a finite second moment this yields an exponential deviation bound of order sigma sqrt(log(1/delta)/n). With a finite third moment and k=o(m), fixed finite-dimensional project

What carries the argument

The load-bearing object is the radial Huber score map: the canonical map Psi_tau(y)=y when ||y||<=tau and tau y/||y|| beyond, and the pseudo-Huber map Psi_tau(y)=y/sqrt(1+||y||^2/tau^2). Both are odd and bounded by tau, which makes the HOMER objective median-like for distant block means and mean-like for central ones. The proof of the majority theorem uses a Hilbert angle lemma: if a candidate estimate is more than 2r from the true mean, each good block mean within r contributes a projected score exceeding tau sqrt(3/8), while each contaminated block can contribute no less than -tau. With at least 5/8 good blocks, the projected average score is strictly positive, contradicting the first-orde

Load-bearing premise

The robustness guarantee rests on the assumption that at least five-eighths of the block means are within r of the true mean and that the threshold tau is no larger than r using an oracle block-scale sigma; if contamination spreads across most blocks or tau is chosen by the median-distance rule, the stated bounds do not apply.

What would settle it

Test the 5/8-majority theorem directly: in R^2, place k=16 block means with 10 within the radius-r ball around a fixed mu and 6 arbitrary far points, solve the HOMER minimization exactly for both losses, and check whether every solution stays within 2r. A single configuration violating 2r disproves Theorem 3; conversely, a dispersed-contamination case where one observation in every block shifts each block mean beyond r—so the 5/8 premise fails—will show the theorem's scope is block summaries, not raw contamination.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • With k=32 log(1/delta) blocks and tau on the block-error scale, the deviation bound ||hat_mu - mu|| <= 4 sigma sqrt(k/n) holds with probability at least 1-delta under only E||X||^2 < infinity.
  • Canonical HOMER is exactly the sample mean whenever all block means lie inside its quadratic basin, while pseudo-HOMER approaches the sample mean as tau -> infinity, giving a formal interpolation path from geometric median to mean.
  • The pseudo-Huber version provides a plug-in sandwich covariance that is consistent in trace norm; with a finite third moment and k=o(m), fixed finite-dimensional projections are asymptotically normal around the true mean at the parametric rate.
  • The same block-mean machinery applies directly to functional means, covariance-operator means with eigenspace perturbation bounds, kernel mean embeddings with RKHS bounds, and aggregation of distributed gradient summaries.
  • Pseudo-Huber is strictly convex, so the majorization-minimization weighted-average update converges monotonically to the unique minimizer using only the k block means and their Gram matrix.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The data-driven median-distance threshold used in experiments is outside all three theorems; a Lepski-type grid, mentioned in the paper, is the natural candidate for an adaptive threshold that would bring practice under the theory.
  • Because the theorems hold for block summaries, replacing ordinary block means with robust within-block estimators should extend the guarantees to dispersed raw contamination; the paper identifies this as an open direction.
  • The 5/8 majority and e^{-k/32} exponent are convenient constants, not optimal; the appendix's sharper per-loss fractions suggest the same two-radius argument could be tightened for specific contamination fractions.
  • The reported small-block undercoverage suggests a finite-sample calibration of the sandwich intervals—for example, bootstrap or bias-corrected thresholds—could be tested against the asymptotic theory before routine use.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper introduces HOMER, a robust aggregator for block means in Hilbert spaces that replaces the geometric median in median-of-means (MOM) with a radial Huber center. For canonical and pseudo-Huber losses, the paper proves existence and a score equation (Prop. 1), endpoint behavior connecting HOMER to MOM and the sample mean (Prop. 2), a two-radius majority theorem (Thm. 3), and an exponential high-confidence deviation bound under only a finite second moment (Thm. 4). For pseudo-Huber aggregation, it proves fixed-block asymptotic linearity around a population block-Huber target, consistent sandwich covariance estimation (Thm. 5, Prop. 14), and, under a finite third moment with k=o(m), a projected CLT centered at the true mean (Prop. 6, Cor. 7). Experiments cover finite-dimensional, functional, covariance, RKHS, distributed-gradient, and real-data settings, and the paper explicitly reports undercoverage of finite-block sandwich intervals and failure when contamination spreads across most blocks.

Significance. If the theorems hold, HOMER provides a useful threshold interpolation between geometric MOM and the sample mean, with a smooth Hessian that supports plug-in sandwich inference in Hilbert spaces. The main proof chain appears internally sound: the angle lemma, the Chebyshev/Hoeffding constants, and the k=o(m) centering argument are coherent, and the paper is unusually explicit about the limits of its guarantees. In particular, Appendix B disclaims the data-driven threshold rule, and Section 9 disclaims raw-point contamination. These disclosures are a real strength. The central caveat is that the oracle-threshold theory does not cover the median-distance tuning rule used in the experiments; thus the practical robustness/inference claims are supported by simulation rather than by the theorems. This is a scope gap rather than an algebraic inconsistency, but it is load-bearing for the paper's overall message.

major comments (2)
  1. [Appendix B and §8] Theorem 4 requires a user-chosen threshold 0<τ≤2σ/√m, with σ an oracle block-scale. The estimator used in the displayed robustness and efficiency experiments instead sets τ̂=c·median_j ||Z_j−μ̃|| for c∈{0.5,2,8}, and Appendix B states this rule is not covered by Theorem 4, Theorem 5, or Corollary 7. The gap is quantitative, not merely formal: on the event used by Theorem 4, with at least 5/8 block means within r=2σ/√m of μ, the median distance can exceed r by a factor that grows with pilot error; for c≥1/3, τ̂ can then exceed r, so Theorem 3's τ≤r condition fails and its 2r conclusion is unavailable. Thus the headline high-confidence guarantee does not transfer to the algorithm used in the main experiments. The theorems are internally consistent, but the paper needs either an adaptive-threshold theorem or a clear repositioning of the oracle-threshold results as the theoretical contributi
  2. [§6.2, Cor. 7 and §8] The inference theory covers only fixed deterministic λ with τ=λ/√m and, for centering at μ, the triangular-array limit m,k→∞ with k=o(m). The functional real-data and bike-demand studies use c·median thresholds with k=8 or 16 and are explicitly labeled descriptive; finite-block sandwich intervals undercover, as the paper acknowledges. This is acceptable as a stated limitation, but the abstract's wording that HOMER 'supports mean inference at the usual parametric rate' should be read strictly for the oracle-threshold pseudo-Huber estimator in the asymptotic regime. To make the practical inference claim match the theory, the authors should either extend the results to data-dependent thresholds or narrow the abstract/introduction claims accordingly.
minor comments (5)
  1. [§8 and Figs. 2–3] The caption or text should state clearly at the figure level that the reported coverages are finite-k diagnostics rather than validations of Theorem 5 or Corollary 7. The prose already says this, but a figure-level note would prevent misinterpretation.
  2. [Appendix C.1] The distinction between the 'configured but unreported paper profile', the 'executed reduced grid', and the 'full grid' is confusing. Please specify unambiguously which configuration produced each displayed result and treat the other configurations as reproducible options.
  3. [References] Several references are incomplete or non-standard, e.g., 'Jorge Reyes-Ortiz, D. A. (2013)' and 'Fanaee-T, H. and Gama, J. (2014)'. Full titles, venues, and identifiers would improve the bibliography.
  4. [Algorithm 1] The algorithm returns the final iterate even if the maximum number of iterations is reached without satisfying the tolerance. It would be helpful to state explicitly that the tolerance is a stopping criterion and to report the number of iterations used in the experiments.
  5. [§5, KME applications] The sentence about replacing empirical kernel means in robust MMD calculations could cite the MONK comparison more precisely, since MONK already provides robust kernel mean estimation; the current wording is slightly vague about what HOMER adds beyond the smooth threshold path.

Circularity Check

0 steps flagged

No circularity: HOMER's theorems derive from explicit block-mean objective and standard probabilistic arguments; no fitted input is passed off as prediction.

full rationale

The paper's load-bearing chain is not circular. Section 4's Theorem 3 follows from Lemma 8's Hilbert angle identity and the score equation (6); Theorem 4 is a Chebyshev+Hoeffding application to the same block means. The efficiency/inference results (Theorem 5, Proposition 6, Corollary 7) are derived via the smooth-test Gaussian replacement and asymptotic expansions around the population block-Huber target, with the k=o(m) condition removing the bias to center at μ. The threshold τ is an oracle-constrained user parameter in the theoretical statements; Appendix B explicitly states that the data-driven median rule 'is not covered by Theorem 4, Theorem 5, or Corollary 7,' so no uncovered fitted value is relabeled as a theorem-based prediction. HOMER's target is not defined in terms of the estimator's output, and the cited geometric-median constants are external (Minsker 2015) rather than self-citations. The main limitation—that practical tuning lies outside the formal guarantee—is a scope gap, not circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 6 axioms · 1 invented entities

The central results are self-contained under the stated moment and contamination assumptions. No parameter is fitted to the target; the main exogenous inputs are the threshold tau and the good-block fraction. The invented block-Huber target is openly introduced and removed by the k=o(m) condition.

free parameters (1)
  • Huber threshold tau (or standardized lambda) = tau_hat = c * median_j ||Z_j - tilde_mu|| with c in {0.5, 2, 8} in experiments; tau = lambda / sqrt(m) with lambda fixed
    Controls the robustness-efficiency tradeoff and is chosen by the user, not fit to the target. Theorem 4 requires an oracle bound tau <= 2 sigma / sqrt(m), so the data-driven value is not covered by the theory.
axioms (6)
  • domain assumption The sample is an iid sequence of strongly measurable elements of a real separable Hilbert space with finite second moment.
    Used throughout Sections 3-4; needed for block-mean independence and the Chebyshev step in Theorem 4.
  • domain assumption The robustness problem is at the block-summary level: at least 5/8 of block means lie within r of the true mean.
    Theorem 3's deterministic majority condition; the paper's own discussion notes robustness does not extend to raw-point contamination that spreads into most blocks.
  • domain assumption Finite third moment E||X - mu||^3 < infinity for mean inference.
    Proposition 6 and Corollary 7 require it for the Gaussian-replacement bias bound and the k=o(m) centering.
  • domain assumption An oracle bound on the block-error scale sigma / sqrt(m) is available for the non-asymptotic theorem.
    Theorem 4 sets tau <= 2 sigma / sqrt(m); Appendix B states the practical median-distance rule is not covered by the theorems.
  • domain assumption Under central symmetry, u_{m,lambda}=0 and mean inference does not need k=o(m).
    Corollary 7 and the remark after it; symmetry removes the block-bias term.
  • domain assumption For covariance applications, a finite fourth moment of X and a Davis-Kahan eigengap condition are assumed.
    Section 5's covariance-operator corollary uses the Hilbert-Schmidt second moment and the standard Davis-Kahan bound.
invented entities (1)
  • Population block-Huber target theta_{m,lambda} = mu + u_{m,lambda} / sqrt(m) no independent evidence
    purpose: Defines the quantity around which fixed-m pseudo-Huber inference is centered; the bias to mu is O(m^{-1/2}) and vanishes for projections when k=o(m).
    It is constructed from the same population objective used for estimation and carries no external falsifiable prediction; the paper explicitly treats it as an intermediate target in Theorem 5 and Corollary 7.

pith-pipeline@v1.3.0-daily-deepseek · 24520 in / 18211 out tokens · 175692 ms · 2026-08-01T06:10:08.198715+00:00 · methodology

0 comments
read the original abstract

Heavy tails weaken high-confidence control for the empirical mean. Geometric median-of-means (MOM) also lacks a threshold that moves toward mean efficiency. We propose \emph{HOMER}, or Huber-of-Means for Efficient and Robust Estimation. HOMER aggregates block means through a radial Huber center. Its canonical and pseudo-Huber forms bound each block score and interpolate between median-like robustness and the empirical mean. We establish a Hilbert-space majority theorem and a MOM-order deviation bound under a finite second moment. Canonical HOMER recovers the sample mean inside its quadratic region. Pseudo-HOMER approaches the sample mean as the threshold grows. It also admits asymptotic linearity and consistent sandwich covariance estimation around the population block-Huber target. Under a finite third moment, fixed finite-dimensional projections support mean inference at the usual parametric rate. This result requires growing block sizes and counts, with block sizes increasing faster. Heavy-tailed simulations show that HOMER remains stable when a minority of block summaries is displaced. On clean Gaussian data, both versions closely approach the empirical mean's efficiency. Finite-block sandwich intervals undercovered, especially for skewed functional data. Further studies show failure when contamination affects most blocks or compromises ordinary within-block means.

Figures

Figures reproduced from arXiv: 2607.27532 by Boram Cho, Kisung You.

Figure 1
Figure 1. Figure 1: HOMER at a glance. (a) The sample is split into blocks whose means are combined by a radial Huber center. The center balances block scores, and distant contributions are bounded. (b) Normalized radial losses for u = t/τ . (c) Normalized score magnitudes. Rescaling the MOM objective by τ preserves its minimizer and sets its normalized score to one. orem. Chebyshev and Hoeffding inequalities then give a MOM-… view at source ↗
Figure 2
Figure 2. Figure 2: Finite-dimensional Hilbert mean: robustness, efficiency, and inference. (a) Student-t3 data have n = 256, d = 40, k = 16, and c = 2. Error is the 95th percentile over 500 replications. (b) Clean Gaussian MSE is relative to the empirical mean for c ∈ {0.5, 2, 8}. (c) Empirical coverage uses a fixed projection with n = 256, d = 20, and k = 16. The horizontal line is 0.95. 0.0 0.1 0.2 Corrupted-block fraction… view at source ↗
Figure 3
Figure 3. Figure 3: Functional and subject-level applications. Panels (a) and (b) show reduced-grid functional robustness and evening-minus-morning coverage. Each panel uses 500 replications. Panel (c) shows distance from the unstressed empirical WALKING trajectory under subject spikes at c = 2. Panel (d) gives held-out nearest-template accuracy under sensor failure. gets θm,λ rather than the mean unless symmetry re￾moves the… view at source ↗
Figure 4
Figure 4. Figure 4: Supplementary Hilbert-mean diagnostics. (a) Student-t3 complete-block contamination. (b) Clean tail regimes. (c) Spectral-decay sensitivity on the reduced grid. (d) Monte Carlo estimate of finite-block target displacement. Under symmetric laws, the curves reflect simulation error around the theoretical value zero. Error summaries use 500 replications. space. Apply empirical averaging, geometric MOM, and HO… view at source ↗
Figure 5
Figure 5. Figure 5: Functional-mean simulation. (a) Representative Student-t3 curve with two of eight complete blocks contaminated. (b) Replicated functional robustness over 500 Monte Carlo replications. (c) Fixed-k coverage for the evening-minus-morning contrast. Literal point evaluation is not used. 0 1 2 Time (s) −0.005 0.000 0.005 0.010 0.015 0.020 Body acceleration x (a) WALKING subject-spike fit 0.0 0.1 0.2 0.3 Corrupte… view at source ↗
Figure 6
Figure 6. Figure 6: Subject-level wearable trajectories. (a) Representative WALKING fit under subject-spike stress. (b) Distance to the unstressed empirical trajectory as corrupted subjects increase at c = 2. (c) Held-out nearest￾template accuracy under sensor failure. Panels (a) and (b) use a different stress mechanism from panel (c). All comparisons are descriptive. Gram matrix among block embeddings, ⟨Zj , Zℓ⟩K = 1 m2 X i∈… view at source ↗
Figure 7
Figure 7. Figure 7: Covariance operators and PCA. (a) Student-t5 Frobenius robustness with n = 256, d = 20, k = 8, known center, and population eigengap 0.0458. (b) Rank-three projector error under two contaminated summaries. (c) Clean unknown-center decomposition by tail regime. In this diagnostic panel, “mean covariance error” averages only the four displayed methods. All Monte Carlo summaries use 500 replications. error ro… view at source ↗
Figure 8
Figure 8. Figure 8: Robust kernel means and MMD. (a) RBF embedding error under complete-block outlier clusters. (b) Mean absolute MMD-squared error across the RBF and linear kernels with two of eight blocks contaminated. (c) Clean outer-aggregation time at n = 128. Sample and block Gram construction is excluded. The reduced grid uses 500 replications. 0.0 0.1 0.2 Corrupted-worker fraction 0.0 0.5 1.0 1.5 2.0 2.5 3.0 95th perc… view at source ↗
Figure 9
Figure 9. Figure 9: Distributed gradient summaries. (a) Student-t3 pointwise aggregation error. (b) Clean one-step progress. (c) Illustrative quadratic-objective paths with two of eight workers corrupted at each iteration. The reduced grid uses 500 replications. Panel (c) averages independently redrawn paths. lecting four of 16 training blocks raised empirical mean holdout error to 0.2185 under the crash shift. MOM, canonical… view at source ↗
Figure 10
Figure 10. Figure 10: Industry-return stability under artificial crash shifts. Mean holdout error, covariance holdout error, and rank-three projector error averaged over 31 overlapping rolling windows. The horizontal axis is the realized fraction of the 16 original training blocks selected for stress. These are descriptive dependent-data comparisons [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Bike-demand curves. (a) Representative overall-demand fit with one of eight event-spike blocks. (b) Holdout curve error over 15 overlapping windows. The horizontal axis is the realized block fraction. (c) Descriptive unstressed next-window coverage for the 7 to 9 a.m. average and the 5 to 7 p.m. minus 7 to 9 a.m. contrast. 0.0 0.1 0.2 0.3 Label-noise fraction 0.815 0.820 0.825 0.830 0.835 0.840 0.845 Near… view at source ↗
Figure 12
Figure 12. Figure 12: Class-conditional HAR kernel embeddings. (a) Held-out nearest-embedding accuracy under label noise. (b) Mean clean-to-stressed RKHS displacement under outlier contamination. (c) Absolute shift in squared MMD between WALKING and WALKING UPSTAIRS. Panels use k = 16 and c = 2. The legend distinguishes the equal-block mean from the Weiszfeld geometric-MOM approximation [PITH_FULL_IMAGE:figures/full_fig_p023_… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

296 extracted references · 10 canonical work pages · 1 internal anchor

  1. [1]

    Jorge Reyes-Ortiz, Davide Anguita , year =. Human. doi:10.24432/C54S4K , urldate =

    doi:10.24432/c54s4k

  2. [2]

    Dimension-free PAC-Bayesian bounds for the estimation of the mean of a random vector

    Catoni, Olivier and Giulini, Ilaria , year =. Dimension-free. doi:10.48550/ARXIV.1802.04308 , abstract =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1802.04308

  3. [3]

    Electronic Journal of Statistics , author =

    Distributed statistical estimation and rates of convergence in normal approximation , volume =. Electronic Journal of Statistics , author =. doi:10.1214/19-EJS1647 , number =

    doi:10.1214/19-ejs1647

  4. [4]

    SIAM Journal on Mathematics of Data Science , author =

    The. SIAM Journal on Mathematics of Data Science , author =. 2024 , pages =. doi:10.1137/23M1592420 , language =

    work page doi:10.1137/23m1592420 2024

  5. [5]

    and Judin, David B

    Nemirovskij, Arkadij S. and Judin, David B. and Dawson, E. R. and Nemirovskij, Arkadij S. , year =. Problem

  6. [6]

    M-estimation and

    Mathieu, Timothée , month = jan, year =. M-estimation and. doi:10.70675/93361937z7fe3z44c0z8742zcb273d1a3dcc , abstract =

    work page doi:10.70675/93361937z7fe3z44c0z8742zcb273d1a3dcc

  7. [7]

    Proceedings of the 36th

    Lerasle, Matthieu and Szabo, Zoltan and Mathieu, Timothée and Lecue, Guillaume , editor =. Proceedings of the 36th. 2019 , pages =

    2019

  8. [8]

    Electronic Journal of Statistics , author =

    On the estimation of the mean of a random vector , volume =. Electronic Journal of Statistics , author =. doi:10.1214/17-EJS1228 , number =

    work page doi:10.1214/17-ejs1228

  9. [9]

    TEST , author =

    Fast estimation of the median covariation matrix with application to online robust principal components analysis , volume =. TEST , author =. 2017 , pages =. doi:10.1007/s11749-016-0519-x , language =

    work page doi:10.1007/s11749-016-0519-x 2017

  10. [10]

    Bernoulli , author =

    Efficient and fast estimation of the geometric median in. Bernoulli , author =. doi:10.3150/11-BEJ390 , number =

    work page doi:10.3150/11-bej390

  11. [11]

    The Annals of Statistics , author =

    Online estimation of the geometric median in. The Annals of Statistics , author =. doi:10.1214/16-AOS1460 , number =

    work page doi:10.1214/16-aos1460

  12. [12]

    Blanchard, Peva and El Mhamdi, El Mahdi and Guerraoui, Rachid and Stainer, Julien , editor =. Machine. Advances in

  13. [13]

    Improving the

    Perronnin, Florent and Sánchez, Jorge and Mensink, Thomas , editor =. Improving the. Computer. 2010 , pages =

    2010

  14. [14]

    Perronnin, Florent and Dance, Christopher , month = jun, year =. Fisher. 2007. doi:10.1109/CVPR.2007.383266 , urldate =

    arXiv 2007

  15. [15]

    Journal of Computational and Graphical Statistics , author =

    Search. Journal of Computational and Graphical Statistics , author =. 2022 , pages =. doi:10.1080/10618600.2022.2069779 , language =

    arXiv 2022

  16. [16]

    Cuturi, Marco and Avis, David , month = jan, year =. Ground. Journal of Machine Learning Research , publisher =

  17. [17]

    Biometrika , author =

    Bayesian cluster analysis , volume =. Biometrika , author =. 1978 , pages =. doi:10.1093/biomet/65.1.31 , language =

    doi:10.1093/biomet/65.1.31 1978

  18. [18]

    Understanding uncertainty in

    Balocchi, Cecilia and Wade, Sara , year =. Understanding uncertainty in. doi:10.48550/ARXIV.2506.16295 , abstract =

    doi:10.48550/arxiv.2506.16295

  19. [19]

    Statistical Science , author =

    Revisiting the. Statistical Science , author =. doi:10.1214/20-STS812 , number =

    doi:10.1214/20-sts812

  20. [20]

    The R Journal , author =

    rotations:. The R Journal , author =. 2014 , pages =. doi:10.32614/RJ-2014-007 , language =

    work page doi:10.32614/rj-2014-007 2014

  21. [21]

    Statistical

    Cazelles, Elsa and Pauwels, Edouard and Portales, Léo , year =. Statistical. doi:10.48550/ARXIV.2604.22366 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.22366

  22. [22]

    Rates of estimation of optimal transport maps using plug-in estimators via barycentric projections , isbn =

    Deb, Nabarun and Ghosal, Promit and Sen, Bodhisattva , year =. Rates of estimation of optimal transport maps using plug-in estimators via barycentric projections , isbn =. Proceedings of the 35th

  23. [23]

    SIAM Journal on Imaging Sciences , author =

    Regularized. SIAM Journal on Imaging Sciences , author =. 2014 , pages =. doi:10.1137/130929886 , language =

    work page doi:10.1137/130929886 2014

  24. [24]

    Outlier-

    Mukherjee, Debarghya and Guha, Aritra and Solomon, Justin M and Sun, Yuekai and Yurochkin, Mikhail , editor =. Outlier-. Proceedings of the 38th. 2021 , pages =

    2021

  25. [25]

    Near-optimal estimation of smooth transport maps with kernel sums-of-squares , copyright =

    Muzellec, Boris and Vacher, Adrien and Bach, Francis and Vialard, François-Xavier and Rudi, Alessandro , year =. Near-optimal estimation of smooth transport maps with kernel sums-of-squares , copyright =. doi:10.48550/ARXIV.2112.01907 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2112.01907

  26. [26]

    Minimax estimation of discontinuous optimal transport maps: the semi-discrete case , abstract =

    Pooladian, Aram-Alexandre and Divol, Vincent and Niles-Weed, Jonathan , year =. Minimax estimation of discontinuous optimal transport maps: the semi-discrete case , abstract =. Proceedings of the 40th

  27. [27]

    Entropic estimation of optimal transport maps , copyright =

    Pooladian, Aram-Alexandre and Niles-Weed, Jonathan , year =. Entropic estimation of optimal transport maps , copyright =. doi:10.48550/ARXIV.2109.12004 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2109.12004

  28. [28]

    International Journal of Computer Vision , author =

    The. International Journal of Computer Vision , author =. 2000 , pages =. doi:10.1023/A:1026543900054 , language =

    doi:10.1023/a:1026543900054 2000

  29. [29]

    Journal of Mathematical Imaging and Vision , author =

    A. Journal of Mathematical Imaging and Vision , author =. 2016 , pages =. doi:10.1007/s10851-016-0653-9 , language =

    work page doi:10.1007/s10851-016-0653-9 2016

  30. [30]

    Staerman, Guillaume and Laforgue, Pierre and Mozharovskyi, Pavlo and d'Alché-Buc, Florence , editor =. When. Proceedings of. 2021 , pages =

    2021

  31. [31]

    PeerJ , author =

    scikit-image: image processing in. PeerJ , author =. 2014 , pages =. doi:10.7717/peerj.453 , language =

    doi:10.7717/peerj.453 2014

  32. [32]

    Wang, Lili and Zheng, Chao and Zhou, Wen and Zhou, Wen-Xin , month = oct, year =. A. Statistica Sinica , publisher =. doi:10.5705/ss.202019.0045 , abstract =

    arXiv

  33. [33]

    Journal of the American Statistical Association , author =

    Adaptive. Journal of the American Statistical Association , author =. 2020 , pages =. doi:10.1080/01621459.2018.1543124 , language =

    arXiv 2020

  34. [34]

    The Annals of Statistics , author =

    Mean estimation with sub-. The Annals of Statistics , author =. doi:10.1214/19-AOS1843 , number =

    doi:10.1214/19-aos1843

  35. [35]

    The Annals of Statistics , author =

    Sub-. The Annals of Statistics , author =. doi:10.1214/16-AOS1440 , number =

    doi:10.1214/16-aos1440

  36. [36]

    Annales de l'Institut Henri Poincaré, Probabilités et Statistiques , author =

    Challenging the empirical mean and empirical variance:. Annales de l'Institut Henri Poincaré, Probabilités et Statistiques , author =. doi:10.1214/11-AIHP454 , number =

    doi:10.1214/11-aihp454

  37. [37]

    Progress in Artificial Intelligence , author =

    Event labeling combining ensemble detectors and background knowledge , volume =. Progress in Artificial Intelligence , author =. 2014 , pages =. doi:10.1007/s13748-013-0040-3 , language =

    doi:10.1007/s13748-013-0040-3 2014

  38. [38]

    Journal of Financial Economics , author =

    Common risk factors in the returns on stocks and bonds , volume =. Journal of Financial Economics , author =. 1993 , pages =. doi:10.1016/0304-405X(93)90023-5 , language =

    doi:10.1016/0304-405x(93)90023-5 1993

  39. [39]

    Cruciani, Federico and Sun, Chen and Zhang, Shuai and Nugent, Chris and Li, Chunping and Song, Shaoxu and Cheng, Cheng and Cleland, Ian and Mccullagh, Paul , month = aug, year =. A. 2019. doi:10.1109/SmartWorld-UIC-ATC-SCALCOM-IOP-SCI.2019.00071 , urldate =

    arXiv 2019

  40. [40]

    Pattern Recognition , author =

    Comparative study on classifying human activities with miniature inertial and magnetic sensors , volume =. Pattern Recognition , author =. 2010 , pages =. doi:10.1016/j.patcog.2010.04.019 , language =

    doi:10.1016/j.patcog.2010.04.019 2010

  41. [41]

    The Annals of Statistics , author =

    Invariant. The Annals of Statistics , author =. doi:10.1214/aos/1176343283 , number =

    arXiv

  42. [42]

    and Gretton, Arthur and Fukumizu, Kenji and Schölkopf, Bernhard and Lanckriet, Gert R.G

    Sriperumbudur, Bharath K. and Gretton, Arthur and Fukumizu, Kenji and Schölkopf, Bernhard and Lanckriet, Gert R.G. , month = aug, year =. Hilbert. Journal of Machine Learning Research , publisher =

  43. [43]

    R. On. Proceedings of the. 1961 , pages =

    1961

  44. [44]

    Journal of the American Statistical Association , author =

    Sequential. Journal of the American Statistical Association , author =. 1998 , pages =. doi:10.1080/01621459.1998.10473765 , language =

    arXiv 1998

  45. [46]

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

    Sequential. Journal of the Royal Statistical Society Series B: Statistical Methodology , author =. 2006 , pages =. doi:10.1111/j.1467-9868.2006.00553.x , language =

    arXiv 2006

  46. [47]

    Ecology , author =

    Diversity and. Ecology , author =. 1973 , pages =. doi:10.2307/1934352 , abstract =

    doi:10.2307/1934352 1973

  47. [48]

    Measuring

    Gorham, Jackson and Mackey, Lester , year =. Measuring. Proceedings of the 34th

  48. [49]

    Measuring

    Gorham, Jackson and Mackey, Lester , year =. Measuring. Proceedings of the 29th

  49. [50]

    and Giné, Evarist , editor =

    De La Peña, Víctor H. and Giné, Evarist , editor =. Decoupling , isbn =. doi:10.1007/978-1-4612-0537-1 , urldate =

    doi:10.1007/978-1-4612-0537-1

  50. [51]

    Journal of Multivariate Analysis , author =

    Modifications of the. Journal of Multivariate Analysis , author =. 2001 , pages =. doi:10.1006/jmva.2000.1922 , language =

    arXiv 2001

  51. [52]

    Causality:

    Pearl, Judea , year =. Causality:

  52. [53]

    and Ivanov, Sergej and Ivanov, Sergei V

    Burago, Dmitrij Jurʹevič and Burago, Jurij D. and Ivanov, Sergej and Ivanov, Sergei V. , year =. A

  53. [54]

    Methods of

    Amari, Shun'ichi and Nagaoka, Hiroshi and Amari, Shun'ichi and Amari, Shun'ichi , translator =. Methods of

  54. [55]

    Lange, Richard D and Kwok, Devin and Matelsky, Jordan Kyle and Wang, Xinyue and Rolnick, David and Kording, Konrad , editor =. Deep. Proceedings of 2nd. 2023 , pages =

    2023

  55. [56]

    Nature Communications , author =

    Single-cell analysis reveals inflammatory interactions driving macular degeneration , volume =. Nature Communications , author =. 2023 , pages =. doi:10.1038/s41467-023-37025-7 , language =

    doi:10.1038/s41467-023-37025-7 2023

  56. [57]

    The Annals of Statistics , author =

    Data-driven. The Annals of Statistics , author =. doi:10.1214/009053607000000541 , number =

    doi:10.1214/009053607000000541

  57. [58]

    , month = jan, year =

    Brown, Lawrence D. , month = jan, year =. Fundamentals of. doi:10.1214/lnms/1215466757 , urldate =

    arXiv

  58. [59]

    Transactions of the American Mathematical Society , author =

    Strong laws for \. Transactions of the American Mathematical Society , author =. 1996 , pages =. doi:10.1090/S0002-9947-96-01681-9 , language =

    doi:10.1090/s0002-9947-96-01681-9 1996

  59. [60]

    Journal of the American Statistical Association , author =

    Asymptotic. Journal of the American Statistical Association , author =. 1987 , pages =. doi:10.1080/01621459.1987.10478472 , language =

    arXiv 1987

  60. [61]

    Journal of Machine Learning Research , author =

    Scikit-learn:. Journal of Machine Learning Research , author =. 2011 , pages =

    2011

  61. [62]

    Biometrika , author =

    Saddlepoint approximations for the. Biometrika , author =. 2005 , pages =. doi:10.1093/biomet/92.2.465 , language =

    doi:10.1093/biomet/92.2.465 2005

  62. [63]

    Computational Statistics , author =

    On maximum likelihood estimation of the concentration parameter of von. Computational Statistics , author =. 2014 , pages =. doi:10.1007/s00180-013-0471-0 , language =

    doi:10.1007/s00180-013-0471-0 2014

  63. [64]

    IEEE Transactions on Information Theory , author =

    Information projections revisited , volume =. IEEE Transactions on Information Theory , author =. 2003 , pages =. doi:10.1109/TIT.2003.810633 , language =

    arXiv 2003

  64. [65]

    The Annals of Statistics , author =

    Exponential. The Annals of Statistics , author =. doi:10.1214/aos/1176344838 , number =

    arXiv

  65. [66]

    IEEE Transactions on Information Theory , author =

    Information geometry on hierarchy of probability distributions , volume =. IEEE Transactions on Information Theory , author =. 2001 , pages =. doi:10.1109/18.930911 , number =

    doi:10.1109/18.930911 2001

  66. [67]

    Nature Reviews Neuroscience , author =

    Neural tuning and representational geometry , volume =. Nature Reviews Neuroscience , author =. 2021 , pages =. doi:10.1038/s41583-021-00502-3 , language =

    doi:10.1038/s41583-021-00502-3 2021

  67. [68]

    New insights and perspectives on the natural gradient method , url =

    Martens, James , month = sep, year =. New insights and perspectives on the natural gradient method , url =

  68. [69]

    International Statistical Review , author =

    Directions. International Statistical Review , author =. 2022 , pages =. doi:10.1111/insr.12481 , abstract =

    doi:10.1111/insr.12481 2022

  69. [70]

    doi:10.1007/978-981-19-1044-9 , language =

    Directional. doi:10.1007/978-981-19-1044-9 , language =

    doi:10.1007/978-981-19-1044-9

  70. [71]

    On the spherical

    You, Kisung and Shung, Dennis , year =. On the spherical. doi:10.48550/ARXIV.2208.11929 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.2208.11929

  71. [72]

    Statistics and Computing , author =

    Learning over von. Statistics and Computing , author =. 2026 , pages =. doi:10.1007/s11222-026-10844-6 , language =

    doi:10.1007/s11222-026-10844-6 2026

  72. [73]

    Diethe, Tom , month = feb, year =. A. doi:10.48550/arXiv.1502.07104 , abstract =

    Pith/arXiv arXiv doi:10.48550/arxiv.1502.07104

  73. [74]

    doi:10.18637/jss.v058.i10 , language =

    Journal of Statistical Software , author =. doi:10.18637/jss.v058.i10 , language =

    doi:10.18637/jss.v058.i10

  74. [75]

    Communications in Statistics - Theory and Methods , author =

    A solution to the multivariate behrens-fisher problem , volume =. Communications in Statistics - Theory and Methods , author =. 1986 , pages =. doi:10.1080/03610928608829342 , language =

    doi:10.1080/03610928608829342 1986

  75. [76]

    Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , author =

    Dispersion on a. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , author =. 1953 , pages =. doi:10.1098/rspa.1953.0064 , language =

    arXiv 1953

  76. [77]

    Annals of Eugenics , author =

    The. Annals of Eugenics , author =. 1936 , pages =. doi:10.1111/j.1469-1809.1936.tb02137.x , language =

    arXiv 1936

  77. [78]

    Translational Psychiatry , author =

    Brain network dysfunctions in addiction: a meta-analysis of resting-state functional connectivity , volume =. Translational Psychiatry , author =. 2022 , pages =. doi:10.1038/s41398-022-01792-6 , number =

    doi:10.1038/s41398-022-01792-6 2022

  78. [79]

    SIAM Journal on Matrix Analysis and Applications , author =

    Permutation-. SIAM Journal on Matrix Analysis and Applications , author =. 2024 , pages =. doi:10.1137/22M1538144 , language =

    doi:10.1137/22m1538144 2024

  79. [80]

    NeuroImage , author =

    Resting state functional connectivity in addiction:. NeuroImage , author =. 2012 , pages =. doi:10.1016/j.neuroimage.2012.01.117 , language =

    doi:10.1016/j.neuroimage.2012.01.117 2012

  80. [81]

    Frontiers in Human Neuroscience , author =

    Sex. Frontiers in Human Neuroscience , author =. 2019 , pages =. doi:10.3389/fnhum.2019.00104 , urldate =

    arXiv 2019

Showing first 80 references.