Pith. sign in

REVIEW 4 major objections 5 minor 27 references

Inverse Norm Weighted Maxsum Test for High Dimensional Location Parameters

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper establishes that weighted spatial-sign max and sum statistics are asymptotically independent, so a Cauchy-combined inverse-norm test can stay powerful for both sparse and dense alternatives in high dimensions.

desk verdict A useful but incomplete extension of the maxsum framework to weighted spatial signs: the inverse-norm optimality result is genuinely new, but the key asymptotic-independence theorem is sketched by analogy rather than proved. read the letter →

arxiv 2501.14168 v1 pith:NNPAWS2Y submitted 2025-01-24 stat.ME

Guowei Yan , Ping Zhao , Long Feng This is my paper · ORCID
classification stat.ME MSC 62H1562G1062G20
keywords asymptoticindependenceCauchycombinationtesthigh-dimensionallocationtestingspatialsigninversenormweightingmax-typesum-typeheavy-taileddistributions
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

The paper studies the one-sample location test in high dimensions where p can far exceed n and the data may be heavy-tailed. It proposes a max-type statistic built from the weighted spatial sign with weight function $w(x)=x^m$, and argues that the inverse-norm weight $m=-1$, i.e. $w(x)=1/\|x\|$, is the best choice in this class. The second claim is that this max-type statistic is asymptotically independent of the weighted spatial-sign sum-type statistic introduced earlier, so the two p-values can be combined by the Cauchy method into one adaptive test. If these claims are correct, the resulting IN-CC test should control size under the null and match the better of the sum and max procedures across all sparsity regimes, which is what the simulations show under normal, $t(4)$, and mixture-normal settings. The central proof target is Theorem 7, which states the asymptotic independence under the null.

What carries the argument

The machinery is the weighted spatial sign: for radius $R=\|D^{-1/2}(X-\theta)\|$, each observation contributes $w(R)\,U(D^{-1/2}(X-\theta))$ with $U(x)=x/\|x\|$, so magnitude information enters the direction-only sign. The location and diagonal scale estimates $(\hat\theta, \hat D)$ solve the pair of estimating equations in (5), and Theorem 1 gives the Bahadur representation $n^{1/2}\hat D^{-1/2}(\hat\theta-\theta)=n^{-1/2}\zeta_{m-1}^{-1}\sum w(R_i)U_i + C_n$. The independence proof then compares the sum-type U-statistic and the squared sup-norm of the weighted sign average through a Gaussian approximation over rectangles, delivering Theorem 7. The final piece is the Cauchy combination rule $p_{CC} = 1 - G(0.5\tan\{(0.5-p_{MAX})\pi\}+0.5\tan\{(0.5-p_{SUM})\pi\})$, whose validity under arbitrary dependence makes it the natural way to merge the two asymptotically independent p-values.

What would settle it

Simulate the null distribution of the IN-CC p-value under $n=80$, $p=400$ or $600$, with a multivariate $t(4)$ distribution and the paper's Toeplitz covariance $\Sigma_{ij}=0.5^{|i-j|}$: the p-value must be approximately uniform, and the empirical size at 5% should stay within simulation error of 0.05. A systematic inflation or deflation across many seeds, or a nonzero correlation between $T_{SUM}/\hat\sigma_n$ and $T_{MAX}$ in the same data, would contradict the asymptotic independence in Theorem 7. As an additional check, run the iterative estimator (5) from different starting values; if its output changes materially, Assumption 1 fails and the test statistic is not well defined.

Watch

Extended reading notes

Core claim

Under the null hypothesis and a set of regularity conditions (sub-exponential coordinates, bounded radius moments, and a correlation structure that gives the maximum a Gumbel limit), the paper's weighted max statistic $T_{MAX}^{(m)}$ and the weighted sum statistic $T_{SUM}^{(m)}$ are asymptotically independent, in the sense that $P(T_{SUM}/\sigma_n \le x, T_{MAX} \le y) \to \Phi(x)F(y)$. This lets the paper reject $H_0$ when the Cauchy-combined p-value $p_{CC}$ is small; Theorem 8 extends the independence to a local alternative with sparse nonzero coordinates. Along the way the paper derives a Bahadur representation and Gaussian approximation for the weighted spatial median, and shows by asymptotic relative efficiency comparisons that the inverse-norm weight $m=-1$ dominates other power weights in the class. The real-data example on S&P 500 excess returns agrees with the simulations: the combined inverse-norm test has the highest rejection rate among the procedures compared.

Load-bearing premise

Assumption 1 asserts that the estimating equations defining the weighted location and scale estimators have a unique solution; the paper gives no convergence proof for the iterative algorithm and no proof of existence or uniqueness, so the entire procedure silently presupposes that these estimators can be computed deterministically.

Editorial extensions

If this is right

  • The inverse-norm weighted max test $T_{MAX}^{(-1)}$ has asymptotic relative efficiency at least 1 against the plain spatial-sign max test ($m=0$) and the Gaussian max test ($m=1$), with equality only when the standardized radius is degenerate.
  • Under $H_0$, the Cauchy combination $p_{CC}$ inherits asymptotic size control from the independence in Theorem 7, so practitioners can use one test without knowing whether the alternative is sparse or dense.
  • Under the local alternative in (8), the power of IN-CC is at least as large as the power of the max or sum component alone at half the level, and can exceed both in intermediate sparsity.
  • Because the statistics are scalar-invariant after diagonal standardization, the procedure remains usable when coordinate variances are unequal, as in the S&P 500 application with heterogeneous asset volatilities.
  • Across the four simulation settings (normal, $t(4)$, and two mixture-normal distributions), IN-CC matched IN-MAX in very sparse cases and IN-SUM in dense cases while outperforming all competitors in between.

Reading between the lines

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

  • Editorial inference: the same independence mechanism should extend to the two-sample location problem and to alpha tests in linear factor pricing models, the two extensions the conclusion names as open.
  • Editorial inference: because the Cauchy combination is valid under arbitrary dependence among inputs, IN-CC would remain a usable test even if the independence in Theorem 7 holds only approximately at finite sample sizes; the practical cost would be a slight size distortion near the boundary.
  • Editorial inference: the paper's ARE identity implies that the inverse-norm weighting advantage grows with heavier tails; for near-Gaussian data all three max tests are asymptotically equivalent, so practitioners should expect little gain from IN-CC over simpler max tests in that regime.
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

4 major / 5 minor

Summary. The paper proposes a weighted spatial-sign max-type test for the high-dimensional one-sample location problem, claims that the inverse-norm weight is asymptotically optimal within the weighted class, and constructs an adaptive max-sum test by combining the weighted max and sum statistics through a Cauchy combination. The theoretical core is Theorem 7, which asserts asymptotic independence between the weighted max and weighted sum statistics under the null and thereby justifies the Cauchy p-value combination. The paper reports simulation comparisons and a financial data application supporting the power and robustness of the proposed IN-CC test.

Significance. If the central claims are correct, the paper makes a useful contribution: it extends spatial-sign max tests with a weight function that provably improves asymptotic relative efficiency for heavy-tailed data, and it provides an adaptive combination that performs well under both sparse and dense alternatives. The ARE derivation in Section 2.2 is a genuine parameter-free comparison, and the simulation evidence in Figures 1-4 and Table 3 is extensive and favorable. The main value depends on the asymptotic independence in Theorem 7, because the Cauchy combination test has no exact size guarantee without it. The paper also gives credit for machine-checkable derivations only in the sense that the proof structure follows established strategies; no code is supplied.

major comments (4)
  1. [§7.3.6, Theorem 7] The asymptotic independence proof is not carried out. After reducing the problem to Equation (25), the text says the normal case follows from Theorem 2 of Feng et al. (2024) and the non-normal case is 'almost the same as the proof in Theorem 6 in Liu et al. (2024),' with only Sigma_u replaced by zeta_{2m}^{-1}Sigma_w. This substitution changes the U-statistic kernel and the linear statistic in a substantive way, and the paper never verifies the Gaussian-coupling, anti-concentration, and covariance-comparison conditions needed for the joint convergence in Equation (25). Since Theorem 7 is the load-bearing justification for the Cauchy combination p_CC, the proof must be supplied in full or replaced by a precise theorem from prior work whose conditions are shown to follow from Assumptions 1-7 and 9-10.
  2. [§2.1, Assumption 1] Assumption 1 simply postulates that Equations (5) have a unique solution, and the text explicitly admits that there is no convergence proof for the iterative algorithm and no proof of existence or uniqueness, even in low-dimensional settings. Because both T_MAX and T_SUM are defined through the estimators (theta_hat, D_hat) solving (5), an uncomputable or non-unique estimator leaves the entire testing procedure not well-defined. This is a structural gap, not just a technical nicety; at minimum, existence, uniqueness, and algorithm convergence should be established for the recommended weight w(x)=x^{-1} used by IN-MAX and IN-CC.
  3. [§3.1, displayed T_SUM statistic] The displayed formula for T_SUM^(m) is degenerate: the second factor is U(D_ij^{-1/2} X_i), the same observation as the first factor, so the kernel is U(X_i)^T U(X_i) rather than U(X_i)^T U(X_j). The subsequent proofs in §7.3.5 and §7.3.6 silently use the corrected version with X_j. As printed, the statistic is identically a sum of squared unit vectors and cannot be the statistic studied in the theorems. This needs to be corrected and the corrected definition used uniformly.
  4. [Theorem 1, Lemma 1, §7.3.1-§7.3.2] There are inconsistent statements of the Bahadur representation that matter for the later proofs. Lemma 1 states G ~ N(0, zeta_{m-1}^{-2} Sigma_w), but the proof of Lemma 1 writes G ~ N(0, zeta_1^{-2} Sigma_u) with Sigma_u not defined there. Similarly, the final representation in the proof of Theorem 1 in §7.3.1 reads n^{1/2} D_hat^{-1/2}(theta_hat - theta) = n^{-1/2} zeta_{m-1}^{-1} sum U_i + C_n, omitting the weight factor R_i^m that appears in the theorem statement. These are not purely typographical: the weighted representation is what justifies the covariance matrix zeta_{2m}^{-1} Sigma_w used in Theorem 7, and the printed proof must be consistent with it.
minor comments (5)
  1. [Notation, p. 16 and Appendix §7] The definition of zeta_k is inconsistent: Section 2.1 defines zeta_k = E(R_i^k), while the Appendix defines zeta_k = E(R_i^{-k}). The proofs and theorems use zeta_k = E(R_i^k), so the Appendix definition should be corrected.
  2. [Assumption 10] The condition 'zeta_{-k}/zeta_k^{-1} -> gamma_k' is garbled and does not parse as written. Since this assumption is used in the null distribution of T_SUM, it needs to be stated cleanly, presumably as a ratio of moments of the radius variable.
  3. [Theorem 8] Theorem 8 says it holds under 'the alternative hypothesis 7', but Equation (7) is the sum-test local alternative; the local alternative used for the max-sum independence discussion is Equation (8). The theorem statement should refer to (8) and the conditions should be checked against (8).
  4. [§7.3.6, after Equation (25)] The sentence 'When Equation holds, combined with O_p(L_{n,p}) = o_p(1), Equation holds obviously' omits the equation numbers; the reader cannot tell which display is being referenced.
  5. [§2.1, line following Equation (5)] The algorithm update in step (ii) is written as a vector equation with no denominator or normalization shown for the weighted sum; adding the exact update would help reproducibility, especially since no convergence proof is provided.

Circularity Check

1 steps flagged · score 4.0 of 10

Theorem 7's asymptotic-independence proof is delegated to overlapping-author prior work after only a covariance substitution; the weighted kernel's needed conditions are never verified, making the central claim partially self-citation-dependent.

  1. self citation load bearing [Section 7.3.6, Proof of Theorem 7 (normal-case reduction cites Feng et al. 2024, Theorem 2; non-normal case cites Liu et al. 2024, Theorem 6).]
    "From the Theorem 2 in Feng et al. (2024), the Equation 25 holds if Ui follows the normal distribution. We then investigate the non-normal case. ... The following proof is almost the same as the proof in Theorem 6 in Liu et al. (2024), except that Σ u is replaced by ζ−1 2mΣ w"

    The paper's central claim—the asymptotic independence of T_SUM^(m) and T_MAX^(m) in Theorem 7—is not derived in the manuscript. The proof stops at an assertion that Equation (25) follows from Theorem 2 of Feng et al. (2024) in the Gaussian case and that the general case is 'almost the same' as Theorem 6 of Liu et al. (2024), with only Σ_u replaced by ζ_{2m}^{-1}Σ_w. Those cited results are by overlapping authors (Feng and Zhao appear in the present paper and in Liu et al. 2024; Feng is also an author of Feng et al. 2024), and they concern the unweighted spatial-sign kernel U_i^T U_j and statistic n^{-1/2}Σ U_i, not the weighted kernel R_i^m R_j^m U_i^T U_j or n^{-1/2}Σ R_i^m U_i introduced here.

full rationale

There is no fitted-input-called-prediction circularity: the inverse-norm weight is derived from the paper's own ARE calculation via Holder's inequality, not estimated from data. Theorems 1-6 (Bahadur representation, Gaussian approximation, null distribution of the weighted sum statistic) are proved in the appendix with substantive detail, and the simulation studies provide external, falsifiable evidence for the proposed IN-CC procedure. The main circularity concern is confined to Theorem 7: the asymptotic independence that licenses the Cauchy combination is not actually proven in the paper but is delegated to prior theorems by the same research group after a covariance substitution. The paper never checks that the weighted kernel satisfies the coupling and anti-concentration conditions required by those cited theorems. This is a load-bearing self-citation, but not a by-construction equivalence or a renamed data fit, so a moderate score is appropriate. Two additional limitations, noted in the manuscript itself, are acknowledged but are not circularity: Assumption 1 assumes existence and uniqueness of the estimator without proof of convergence even in low dimensions, and the displayed T_SUM formula in Section 3.1 contains a typo (the second U is evaluated at X_i instead of X_j) that the proof silently corrects.

Assumptions & free parameters 0 free parameters · 10 assumptions · 0 invented entities

The central claims are conditional on eleven assumptions, most of which are distributional and covariance constraints inherited from prior work. Assumption 1 is the most fragile because it postulates, without proof, the existence and uniqueness of the estimator used throughout. No parameters are fitted to data and no new physical or model entities are introduced.

assumptions (10)
  • ad hoc to paper Assumption 1: Equations (5) have a unique solution (theta_hat, D_hat).
    The paper postulates existence and uniqueness of the estimator and explicitly states there is no proof of convergence for the algorithm that computes it.
  • domain assumption Assumption 2: The components W_i,j are i.i.d. symmetric with mean 0, variance 1, and a sub-exponential Orlicz norm bound.
    Used to control the Gaussian approximation and moment behavior of the spatial sign vectors.
  • ad hoc to paper Assumption 3: The weight function is w(x) = x^m with m <= 1.
    Restricts the class of weights to a single-parameter family; the paper notes that the theoretical analysis would be more complex for general weights.
  • domain assumption Assumption 4: Moments of the radius R_i, including negative moments, are bounded uniformly in p.
    Required for the Bahadur representation and for the asymptotic variance expressions involving zeta_k.
  • domain assumption Assumption 5: The shape matrix R has bounded row sums and the diagonal entries of D are bounded away from zero.
    Needed to control the operator norm of the covariance shape and to ensure the scaled observations are well behaved.
  • domain assumption Assumption 6: tr(R^2) - p = o(n^{-1} p^2), n^{-2} p^2 / tr(R^2) = O(1), and log p = o(n).
    A trace condition relating dimension, sample size, and covariance structure, standard in high-dimensional testing.
  • domain assumption Assumption 7: The correlation matrix has mostly small off-diagonal entries, with a sparse set of larger correlations, to ensure Gumbel convergence of the maximum.
    Needed for the limiting distribution of the max-type statistic.
  • domain assumption Assumption 8: The data follow an elliptically symmetric density for the sum-type test.
    Imported from Feng et al. (2021) to establish the null distribution of the weighted sum statistic.
  • domain assumption Assumptions 9-10: Trace and moment growth conditions, including tr(R^4) = o(tr^2(R^2)) and zeta_{4m} = O(zeta_{2m}^2).
    These are technical conditions used in the asymptotic normality of the sum-type test and in the independence proof.
  • domain assumption Assumption 11: Strong eigenvalue and correlation bounds on R, needed for the asymptotic independence between sum and max statistics.
    Imported from Feng et al. (2024) and used in the proof of Theorem 7.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inverse Norm Weighted Maxsum Test for High Dimensional Location Parameters." pith.science (2026). https://pith.science/paper/NNPAWS2Y

@misc{pith2026250114168,
  author       = {Pith},
  title        = {Pith review of: Inverse Norm Weighted Maxsum Test for High Dimensional Location Parameters},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NNPAWS2Y}},
  note         = {Machine review of arXiv:2501.14168}
}
read the original abstract

In the context of high-dimensional data, we investigate the one-sample location testing problem. We introduce a max-type test based on the weighted spatial sign, which exhibits exceptional performance, particularly in the presence of sparse alternatives. Notably, we find that the inverse norm test significantly enhances the power of the test compared to several existing max-type tests. Next, we prove the asymptotic independence between the newly proposed max-type test statistic and the sum-type test statistic based on the weighted spatial sign. Then, we propose an innovative max-sum type testing procedure that integrates both test statistics. This novel procedure demonstrates remarkable robustness and effectiveness across a wide range of signal sparsity levels and heavy-tailed distributions. Through extensive simulation studies, we highlight the superior performance of the proposed method, showcasing its robustness and efficiency compared to traditional alternatives in various high-dimensional settings.

Figures

Figures reproduced from arXiv: 2501.14168 by the authors.

Figure 1
Figure 1. Power of tests with different sparsity levels over ( [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Power of tests with different distribution for [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Power of tests with different distribution for [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Power of tests with different distribution for [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Histogram of standard deviation of US securities. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Q-Q plots of the weekly return rates of some stocks with h [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: t test statistics of the weekly excess return rates of ea [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 26 canonical work pages

  1. [1]

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

    work page 1996

  2. [2]

    Liu, and Y

    Cai, T., W. Liu, and Y. Xia (2014). Two-sample test of high dimensional means under dependence. Journal of the Royal Statistical Society Series B: Statistical Meth odology 76 (2), 349–372

    work page 2014

  3. [3]

    Chen, S. X. and Y.-L. Qin (2010). A two-sample test for high-dimens ional data with applications to gene-set testing. The Annals of Statistics 38 (2), 808–835

    work page 2010

  4. [4]

    Peng, and C

    Cheng, G., L. Peng, and C. Zou (2023). Statistical inference for u ltrahigh dimensional location parameter based on spatial median. arXiv preprint , arXiv:2301.03126

    arXiv 2023

  5. [5]

    Chetverikov, and K

    Chernozhukov, V., D. Chetverikov, and K. Kato (2015). Comparis on and anti-concentration bounds for maxima of gaussian random vectors. Probability Theory and Related Fields 162 , 47–70. 33

    work page 2015

  6. [6]

    Chetverikov, and K

    Chernozhukov, V., D. Chetverikov, and K. Kato (2017). Central limit theorems and bootstrap in high dimensions. Annals of Probability 45 (4), 2309–2352

    work page 2017

  7. [7]

    Jiang, X

    Feng, L., T. Jiang, X. Li, and B. Liu (2024). Asymptotic independenc e of the sum and maximum of dependent random variables with applications to high-dimensional tests. Statistica Sinica 34 , 1745–1763

    work page 2024

  8. [8]

    Liu, and Y

    Feng, L., B. Liu, and Y. Ma (2021). An inverse norm sign test of locat ion parameter for high-dimensional data. Journal of Business & Economic Statistics 39 (3), 807–815

    work page 2021

Show all 27 references
  1. [9]

    Feng, L. and F. Sun (2016). Spatial-sign based high-dimensional loc ation test. Electronic Journal of Statis- tics 10 (2), 2420–2434

    2016

  2. [10]

    Zou, and Z

    Feng, L., C. Zou, and Z. Wang (2016). Multivariate-sign-based high -dimensional tests for the two-sample location problem. Journal of the American Statistical Association 111 (514), 721–735

    2016

  3. [11]

    Feng, L., C. Zou, Z. Wang, and L. Zhu (2015). Two-sample behrens -fisher problem for high-dimensional data. Statistica Sinica 25 (4), 1297–1312

    2015

  4. [12]

    He, Y., G. Xu, C. Wu, and W. Pan (2021). Asymptotically independent u-statistics in high-dimensional testing. Annals of statistics 49 (1), 154–181

    2021

  5. [13]

    Hettmansperger, T. P. and R. H. Randles (2002). A practical affin e equivariant multivariate median. Biometrika 89 (4), 851–860

    2002

  6. [14]

    Huang, X., B. Liu, Q. Zhou, and L. Feng (2023). A high-dimensional in verse norm sign test for two-sample location problems. Canadian Journal of Statistics 51 (4), 1004–1033

    2023

  7. [15]

    Koike, Y. (2021). Notes on the dimension dependence in high-dimens ional central limit theorems for hyper- rectangles. Japanese Journal of Statistics and Data Science 4 , 257–297

    2021

  8. [16]

    Feng, and Y

    Liu, B., L. Feng, and Y. Ma (2023). High-dimensional alpha test of line ar factor pricing models with heavy-tailed distributions. Statistica Sinica 33 , 1389–1410

    2023

  9. [17]

    Liu, J., L. Feng, P. Zhao, and Z. Wang (2024). Spatial-sign based ma xsum test for high dimensional location parameters. arXiv preprint , arXiv:2402.01381

    2024 arXiv

  10. [18]

    Liu, Y. and J. Xie (2020). Cauchy combination test: A powerful tes t with analytic p-value calculation under arbitrary dependency structures. Journal of the American Statistical Association 115 (529), 393–402

    2020

  11. [19]

    Ljung, G. M. and G. E. Box (1978). On a measure of lack of fit in time s eries models. Biometrika 65 (2), 297–303

    1978

  12. [20]

    Long, M., Z. Li, W. Zhang, and Q. Li (2023). The cauchy combination test under arbitrary dependence structures. The American Statistician 77 (2), 134–142

    2023

  13. [21]

    Ma, H., L. Feng, Z. Wang, and J. Bao (2024). Testing alpha in high dime nsional linear factor pricing models with dependent observations. arXiv preprint , arXiv:2401.14052

    2024 arXiv

  14. [22]

    Muirhead, R. J. (2009). Aspects of multivariate statistical theory . John Wiley & Sons

    2009

  15. [23]

    Oja, H. (2010). Multivariate nonparametric methods with R: an approach bas ed on spatial signs and ranks . Springer Science & Business Media

    2010

  16. [24]

    Park, J. and D. N. Ayyala (2013). A test for the mean vector in larg e dimension and small samples. Journal of Statistical Planning and Inference 143 (5), 929–943

    2013

  17. [25]

    Srivastava, M. S. (2009). A test for the mean vector with fewer o bservations than the dimension under non-normality. Journal of Multivariate Analysis 100 (3), 518–532. 34

    2009

  18. [26]

    Peng, and R

    Wang, L., B. Peng, and R. Li (2015). A high-dimensional nonparame tric multivariate test for mean vector. Journal of the American Statistical Association 110 (512), 1658–1669

    2015

  19. [27]

    Xu, G., L. Lin, P. Wei, and W. Pan (2016). An adaptive two-sample tes t for high dimensional means. Biometrika 103 (3), 609–624. 35

    2016

Pith tools

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