Pith. sign in

REVIEW 2 major objections 4 minor 34 references

Testing the equality of estimable parameters

T0 review · 2 major / 4 minor · reviewed 2026-07-10 · grok-4.5

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

Pith's one-line read A single U-statistic framework tests equality of many parameters across populations, with valid asymptotics even as dimension grows slower than sample size.

desk verdict Solid unification of multi-sample U-statistic equality tests with usable fixed-d and moderate high-d procedures; theory is standard but carefully done, simulations informative. read the letter →

arxiv 2607.07588 v2 pith:2HKRHNOI submitted 2026-07-08 stat.ME

M. Romero-Madroñal , M. Remedios Sillero-Denamiel , M. Dolores Jiménez-Gamero This is my paper
classification stat.ME MSC 62G1062H1562E20
keywords non-parametrictestingmultivariateinferenceU-statisticsincreasingdimensionjackknifecovarianceANOVA-typestatisticWald-typeweightedbootstrap
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

Researchers often need to decide whether several populations share the same variance, correlation, Gini index, or other functional of the data. Existing tests are usually hand-crafted for one parameter and two groups. This paper shows that any parameter that can be written as a smooth function of expectations of symmetric kernels can be handled by the same two quadratic-form statistics, once the jackknife supplies consistent covariance estimates. For fixed dimension the Wald statistic is asymptotically chi-squared and the ANOVA-type statistic is a weighted sum of chi-squares; a weighted bootstrap also approximates the latter. When the parameter dimension d grows with total sample size n but d/n tends to zero, the ANOVA-type statistic, after centering and scaling by traces, becomes standard normal under the null. The result therefore supplies a single, distribution-free procedure that covers classical comparisons and still works when many coordinates are tested at once.

What carries the argument

The ANOVA-type quadratic form Q_n = n θ̂ᵀ H θ̂, where H is the orthogonal projection that encodes the equality contrast C = P_k ⊗ I_d. Its null behaviour is controlled by the eigenvalues (or traces) of the projected asymptotic covariance of the stacked U-statistic estimators.

What would settle it

Generate independent samples from k populations that truly share the same parameter vector of growing dimension d, with d/n o 0, compute the centered and scaled ANOVA-type statistic with jackknife covariances, and check whether its empirical distribution is standard normal and its type-I error stays near the nominal level; systematic departure would falsify the increasing-dimension claim.

Watch

Extended reading notes

Core claim

Under independence of the samples, finite second moments of the kernels, positive first-order variance components, comparable sample sizes, and continuous differentiability of the map f, the ANOVA-type statistic Q_n = n θ̂ᵀ H θ̂ converges under the null to a weighted sum of chi-squares when d is fixed, and, after centering by tr(Σ_n H) and scaling by the square root of 2 tr((Σ_n H)^{2}), converges to N(0,1) when d o ∞ with d/n o 0. The same limiting normal law remains valid when the unknown covariance is replaced by its jackknife estimator.

Load-bearing premise

The normal approximation for growing dimension requires that the parameter dimension stays much smaller than the total sample size and that the projected influence functions have uniformly bounded fourth moments and non-degenerate eigenvalues; if dimension grows as fast as sample size or the covariance becomes extremely sparse, the normal limit can fail.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper develops a unified nonparametric framework for testing equality of parameters θ_i = f(η_i) across k independent populations, where each η_i is a vector of expectations of symmetric kernels and is estimated by U-statistics. Two quadratic-form statistics are studied: a Wald-type statistic T_n whose null limit is χ^{2}_{(k-1)d} under fixed d (Proposition 1), and an ANOVA-type statistic Q_n = n θ̂ᵀ H θ̂ whose null limit is a weighted sum of χ^{2}_1 variables for fixed d (Proposition 3). For the latter, a weighted bootstrap calibration is proved consistent (Theorem 1), and under d o ∞ with d/n o 0, uniform fourth-moment bounds on projected influence functions, and eigenvalue sandwich conditions on HΣH, the centered and scaled ATS converges to N(0,1) (Theorem 2 / Proposition 5). Covariance matrices are estimated by the jackknife (Lemma 1, extending Arvesen). Local-power and consistency results are given, finite-sample behavior is examined in extensive simulations (univariate variance/Gini and multivariate mean-covariance settings), and the methods are illustrated on CPS1988 wage data.

Significance. If the asymptotics hold as stated, the paper supplies a single, assumption-light procedure that recovers many classical multi-sample tests (variances, correlations, Gini indices, coefficients of variation, etc.) and extends them to arbitrary k ≥ 2 and to moderately growing dimension. The jackknife covariance estimator, the weighted-bootstrap validity proof, and the martingale-CLT argument for the increasing-dimension normal limit are concrete technical contributions that go beyond ad-hoc special cases. The simulation design and the practical guidelines in §4.3 (kd/n thresholds, effective-dimension requirements for ATS-ID, relative computational cost) make the methods usable. The work is therefore a useful consolidation and moderate extension of the U-statistic multi-sample literature.

major comments (2)
  1. The increasing-dimension theory (Theorem 2, Proposition 5, Remark 2) is restricted to fixed k and d/n o 0. The paper correctly flags that joint growth of k and d would require additional tracking of n/n_i factors and moment constants, yet the simulations vary k up to 10 while reporting only d/n. Because the effective dimension of the quadratic form is of order kd, the finite-sample guidance in §4.3 (kd/n ≲ 0.03, kd ≳ 600) should be stated as the primary regime indicator already in the statement of Theorem 2 / Remark 2, not only in the discussion of simulations; otherwise readers may misapply ATS-ID when k is large and d moderate.
  2. Condition (16) requires a uniform fourth-moment bound on the projected influence functions e_rᵀ D_i h_1(X_i1) that is independent of d. For parameters whose dimension grows by stacking many distinct functionals (e.g., all pairwise correlations or a full covariance matrix), this bound is not automatic and can fail under heavy tails. The manuscript should either supply a verifiable sufficient condition on the original kernels that implies (16) for the leading examples (means + covariances, correlation matrices), or explicitly list the examples for which (16) is known to hold.
minor comments (4)
  1. In §2.4 the claim “with this particular choice of C we have H = C” is true for the chosen projection, but the surrounding text first introduces a general C and then specialises; a one-sentence clarification that H coincides with C only for this Kronecker choice would avoid confusion.
  2. Table 1 (CPU times) reports averages over the three weight distributions for WBS; a short note on the Monte-Carlo size used for the bootstrap p-values would help readers reproduce the timings.
  3. Several typographical inconsistencies appear (e.g., “ANOV A-type”, “bpImhof”, “estimates as follows”). A careful copy-edit pass is needed.
  4. The real-data section (§4.2) refers to Tables S9 etc. in the SM; a one-sentence summary of the log-wage versus raw-wage contrast already in the main text would make the application self-contained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; null limits follow from classical U-statistic CLTs, Delta method, jackknife consistency and quadratic-form/martingale theory under explicitly stated conditions.

full rationale

The paper derives asymptotic null distributions for a Wald-type statistic (fixed d) and an ANOVA-type statistic (fixed d and d o∞ with d/n o0) for multi-sample equality of parameters that are smooth functions of U-statistic expectations. All load-bearing steps invoke standard external results (Hoeffding CLT and SLLN, multivariate Delta method, Arvesen jackknife consistency, Graybill quadratic-form theorems, Helland martingale CLT, Weyl perturbation) under verifiable moment, smoothness and eigenvalue conditions that do not embed the target conclusions. Jackknife estimators of Σ are used only for feasible critical values and are proved consistent separately; they do not define the parameters or the null. Weighted-bootstrap validity is likewise proved from the same influence-function representation. Self-citations to the authors’ related k o∞ univariate work appear only as background motivation and are not used to justify any uniqueness claim or any step of the present fixed-k proofs. No free parameters are fitted to data and then re-presented as predictions, and no ansatz is smuggled via citation. The framework unifies existing special-case procedures but does so by embedding them in a larger, independently derived asymptotic theory rather than by renaming a known empirical pattern. The derivation chain is therefore self-contained against external classical benchmarks.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The theory rests on classical U-statistic asymptotics, independence of samples, moment and non-degeneracy conditions, smooth f, sample-size comparability, and for high-d results eigenvalue and fourth-moment bounds plus d/n→0. No free parameters are fitted into the asymptotic critical values; jackknife and bootstrap are estimators, not tuning constants of the claim.

assumptions (5)
  • domain assumption Independent samples from k populations with ni/n → κi ∈ (0,1).
    Assumption (1) and (8); independence is used for joint normality and block-diagonal Σ.
  • domain assumption Kernels have finite second moments and positive first-order variance components ζi1^(r) > 0.
    Conditions (3)–(4) for non-degenerate U-statistic CLTs.
  • domain assumption f has continuous first (and for bootstrap/high-d, second) partial derivatives near ηi.
    Delta method and jackknife/bootstrap expansions in Lemma 1, Theorem 1, Theorem 2.
  • domain assumption For increasing-d ATS: max_r E|e_rᵀ Di h1(Xi1)|⁴ ≤ M and eigenvalue sandwich on HΣH; d/n→0.
    Conditions (16)–(17) and regime of Theorem 2; load-bearing for normal approximation.
  • standard math Standard results on U-statistics, Delta method, continuous mapping, Weyl perturbation, martingale CLT, quadratic forms of Gaussians.
    Cited via Hoeffding, Serfling, Graybill, Helland, Bhatia, etc., throughout §6.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Testing the equality of estimable parameters." pith.science (2026). https://pith.science/paper/2HKRHNOI

@misc{pith2026260707588,
  author       = {Pith},
  title        = {Pith review of: Testing the equality of estimable parameters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2HKRHNOI}},
  note         = {Machine review of arXiv:2607.07588}
}
abstract

This paper proposes a general and unified framework for testing the equality of a broad class of parameters, defined via $U$-statistics, across multiple independent populations. This approach encompasses various common statistical problems, such as comparing variances, correlation coefficients, or Gini indices, among many others. We consider two test statistics, a Wald-type statistic and an ANOVA-type statistic. The asymptotic distribution of the first one is derived under a fixed-dimension regime, whereas the second one is studied under both fixed and increasing-dimension regimes, where the parameter dimension diverges with the sample size. Based on these limiting distributions, we construct test procedures enabling asymptotically exact inference without parametric assumptions. Additionally, an alternative null distribution estimator based on a weighted bootstrap approximation is studied, which is applicable to the ANOVA-type statistic under a fixed-dimension regime. The finite-sample performance and computational efficiency of the proposed procedures are evaluated through an extensive simulation study. Finally, an application to a real dataset illustrates the usefulness of the proposed methodology.

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

  1. [1]

    Aleksi´ c, D. G. and Miloˇ sevi´ c, B. (2025). Two-sample testing with missing data via energy distance: Weighting and imputation approaches.arXiv preprint arXiv:2508.11421

    work page Pith review arXiv 2025

  2. [2]

    Arvesen, J. N. (1969). Jackknifing U-statistics.The Annals of Mathematical Statistics, 40(6):2076–2100

    work page 1969

  3. [3]

    and Saranadasa, H

    Bai, Z. and Saranadasa, H. (1996). Effect of high dimension: By an example of a two sample problem. Statistica Sinica, 6(2):311–329

    work page 1996

  4. [4]

    Bartlett, M. S. (1937). Properties of sufficiency and statistical tests.Proceedings of the Royal Society of London. Series A, 160:268–282

    work page 1937

  5. [5]

    (2013).Matrix Analysis

    Bhatia, R. (2013).Matrix Analysis. Graduate Texts in Mathematics. Springer New York

    work page 2013

  6. [6]

    Bhoj, D. S. and Ahsanullah, M. (1993). Testing equality of coefficients of variation of two populations. Biometrical Journal, 35(3):355–359

    work page 1993

  7. [7]

    Chen, S. X. and Qin, Y. (2010). A two sample test for high dimensional data with applications to gene-set testing. MPRA Paper 59642, University Library of Munich, Germany

    work page 2010

  8. [8]

    Davidson, R. (2009). Reliable inference for the gini index.Journal of Econometrics, 150(1):30–40

    work page 2009

Show all 34 references
  1. [9]

    Ditzhaus, M., Fried, R., and Pauly, M. (2021). QANOVA: quantile-based permutation methods for general factorial designs.TEST, 30:960–979

    2021

  2. [10]

    and Smaga, L

    Ditzhaus, M. and Smaga, L. (2025). Inference for all variants of the multivariate coefficient of variation in factorial designs.Scandinavian Journal of Statistics, 52(1):270–294

    2025

  3. [11]

    and de Micheaux, P

    Duchesne, P. and de Micheaux, P. L. (2010). Computing the distribution of quadratic forms: Further comparisons between the liu-tang-zhang approximation and exact methods.Computational Statistics and Data Analysis, 54:858–862. 24

    2010

  4. [12]

    Fisher, R. A. (1925).Statistical Methods for Research Workers. Oliver and Boyd, Edinburgh

    1925

  5. [13]

    Graybill, F. A. (1976).Theory and Application of the Linear Model. Duxbury Press, North Scituate, Massachusetts

    1976

  6. [14]

    Helland, I. S. (1982). Central limit theorems for martingales with discrete or continuous time.Scandi- navian Journal of Statistics, 9(2):79–94

    1982

  7. [15]

    Hoeffding, W. (1948). A class of statistics with asymptotically normal distribution.The Annals of Mathematical Statistics, 19(3):293–325

    1948

  8. [16]

    Horn, R. A. and Johnson, C. R. (1985).Matrix Analysis. Cambridge University Press

    1985

  9. [17]

    Hu, J., Bai, Z., Wang, C., and Wang, W. (2017). On testing the equality of high dimensional mean vectors with unequal covariance matrices.Annals of the Institute of Statistical Mathematics, 69(2):365–387

    2017

  10. [18]

    Huang, Y., Li, C., Li, R., and Yang, S. (2022). An overview of tests on high-dimensional means.Journal of Multivariate Analysis, 188:104813. 50th Anniversary Jubilee Edition

    2022

  11. [19]

    Imhof, J. P. (1961). Computing the distribution of quadratic forms in normal variables.Biometrika, 48(3/4):419–426

    1961

  12. [20]

    Jennrich, R. I. (1970). An asymptotic chi 2 test for the equality of two correlation matrices.Journal of the American Statistical Association, 65(330):904–912. Jim´ enez-Gamero, M. and Sillero-Denamiel, M. (2025). The k-sample problem using gini covariance for large k.Journal ...

    1970

  13. [21]

    and Zeileis, A

    Kleiber, C. and Zeileis, A. (2008).Applied Econometrics with R. Springer-Verlag, New York

    2008

  14. [22]

    Lee, A. J. (1990).U-Statistics: Theory and Practice. Routledge, 1st edition

    1990

  15. [23]

    and Chen, S

    Li, J. and Chen, S. X. (2012). Two sample tests for high-dimensional covariance matrices.The Annals of Statistics, 40(2):908–940

    2012

  16. [24]

    Mammen, E. (1993). Bootstrap and wild bootstrap for high-dimensional linear models.Ann. Statist., 21(1):255–285

    1993

  17. [25]

    Mincer, J. A. (1974).Schooling, Experience, and Earnings. Number minc74-1 in NBER Books. National Bureau of Economic Research, Inc, none edition

    1974

  18. [26]

    and Finn, J

    Olkin, I. and Finn, J. D. (1995). Correlations redux.Psychological Bulletin, 118(1):155–164

    1995

  19. [27]

    Owen, A. B. (2001).Empirical Likelihood. Chapman and Hall/CRC

    2001

  20. [28]

    Paul, S. R. (1989). Test for the equality of several correlation coefficients.The Canadian Journal of Statistics / La Revue Canadienne de Statistique, 17(2):217–227

    1989

  21. [29]

    and Schick, A

    Peng, H. and Schick, A. (2018). Asymptotic normality of quadratic forms with random vectors of increasing dimension.Journal of Multivariate Analysis, 164:22–39. Romero-Madro˜ nal, M., Sillero-Denamiel, M.R., and Jim´ enez-Gamero, M.D. (2025). Testing the equality of estimable ...

    2018

  22. [30]

    Schott, J. R. (2007). Some high-dimensional tests for a one-way manova.Journal of Multivariate Analysis, 98(9):1825–1839

    2007

  23. [31]

    (2008).A Matrix Handbook for Statisticians, volume 746 ofWiley Series in Probability and Statistics

    Seber, G. (2008).A Matrix Handbook for Statisticians, volume 746 ofWiley Series in Probability and Statistics. Wiley, New York

    2008

  24. [32]

    Serfling, R. J. (1980).Approximation Theorems of Mathematical Statistics. Wiley, New York

    1980

  25. [33]

    S., Katayama, S., and Kano, Y

    Srivastava, M. S., Katayama, S., and Kano, Y. (2013). A two sample test in high dimensional data. Journal of Multivariate Analysis, 114(C):349–358

    2013

  26. [34]

    Zheng, S., Lin, R., Guo, J., and Yin, G. (2020). Testing homogeneity of high-dimensional covariance matrices.Statistica Sinica, 30:35–53. 25

    2020

Pith tools

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