Pith. sign in

REVIEW 2 major objections 6 minor 41 references

Locally Robust Kernel Specification Tests for Conditional Moment Restrictions

T0 review · 2 major / 6 minor · reviewed 2026-07-31 · grok-4.5

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

Pith's one-line read Kernel specification tests stay valid with machine-learned nuisances by orthogonalizing the whole RKHS process.

desk verdict Solid process-level orthogonal RKHS specification tests for ML nuisances; the growing-dictionary bridge in Example 1 is thinner than advertised, but the core package is real and referee-worthy. read the letter →

arxiv 2607.24382 v1 pith:YGZNGKIR submitted 2026-07-27 stat.ME

Juan Carlos Escanciano This is my paper · ORCID
classification stat.ME MSC 62G1062G2062J07
keywords modelcheckshigh-dimensionalmodelskernelmethodsorthogonalmomentsmultiplierbootstrapconditionalmomentrestrictionscross-fittingNeymanorthogonality
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

Specification tests for conditional moment models usually break when the nuisance pieces are fit with modern machine learning, because those estimators are slow, nonlinear, and inject first-order error into plug-in statistics. This paper builds locally robust kernel tests that combine Neyman-orthogonal moments, cross-fitting, and an RKHS of test functions so the feasible test process is first-order insensitive to that error. The main theoretical guarantee is process-level oracle equivalence under local alternatives and only weak product-rate conditions on the nuisances: the feasible statistic behaves like the infeasible oracle that knows the true nuisances. A multiplier bootstrap then delivers critical values without re-estimating anything. The same template covers high-dimensional regression checks, significance testing with flexible regressions, and tests that conditional treatment effects are constant.

What carries the argument

The locally robust kernel (LRK) statistic: the squared RKHS norm of a cross-fitted orthogonal empirical process whose adjustment is the first-step influence process (FSIP). Orthogonality plus cross-fitting make the feasible process oracle-equivalent; the RKHS structure collapses the infinite-dimensional supremum to a quadratic form in a kernel matrix.

What would settle it

In a high-dimensional regression design with Lasso shrinkage on nonzero coefficients, check whether the cross-fitted LRK rejection rate under the null stays near 5% while a non-orthogonal plug-in test over-rejects, and whether power vanishes exactly when the alternative direction is absorbed into the nuisance dictionary.

Watch

Extended reading notes

Core claim

Under local alternatives and mild product-rate conditions, the feasible cross-fitted Neyman-orthogonal RKHS process is uniformly equivalent to its oracle counterpart, so the resulting kernel specification test has the same local limiting distribution as the oracle and a multiplier bootstrap that never re-estimates nuisances consistently recovers the null law.

Load-bearing premise

Nuisance estimators must be accurate enough that residual-and-representer product errors shrink faster than one over square-root n (a common sufficient rule is faster than n to the minus one-fourth), or the oracle equivalence that protects size can fail.

Editorial extensions

If this is right

  • Specification tests for high-dimensional linear and logistic models can use Lasso or other ML nuisances without requiring asymptotic linearity of those estimators.
  • Significance of a covariate can be tested after arbitrary ML regression on the remaining covariates, under standard cross-fitting and product rates.
  • Constancy of conditional average treatment effects (CATE/CATT) can be tested with doubly robust scores and ML propensity/outcome fits, with a bootstrap that skips nuisance re-estimation.
  • Local power is confined to directions not absorbed by the model's nuisance components; alternatives inside the nuisance space are undetectable by design.

Reading between the lines

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

  • The same orthogonal RKHS template should extend, with work, to time-series or cluster dependence once a suitable multiplier or block bootstrap replaces the iid multipliers.
  • When practitioners only have very slow ML rates, the theory flags that size control may require explicit undersmoothing or stronger double robustness rather than default off-the-shelf fits.
  • Characterizing undetectable directions gives a diagnostic: if a suspected alternative lies in the nuisance span, one should change the nuisance class rather than blame the test.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper constructs specification tests for conditional moment restrictions E[ε(W,η₀)|X]=0 with high-/infinite-dimensional nuisances estimated by machine learning. The proposal combines Neyman-orthogonal moments indexed by the unit ball of an RKHS with cross-fitting: a first-step influence process (FSIP) corrects the whole testing-moment collection for nuisance estimation (Prop. 1), and the resulting LRK statistic is a supremum that reduces to a quadratic form in an adjusted kernel matrix. Main results: uniform oracle equivalence of the feasible and oracle processes under n^{-1/2} local alternatives (Thm 1); limiting Gaussian distribution and local power with an explicit adjoint characterization of undetectable directions (Thm 2, Prop. 2, Table 1); validity of a multiplier bootstrap that avoids nuisance re-estimation (Thm 3); consistency (Thm 4). Three examples (high-dimensional GLM, significance testing, constant CATE/CATT) are worked out, with assumption verification in Supp. B, Monte Carlos, and an NSW application that does not reject constant CATT.

Significance. If correct, this is a useful extension of orthogonal/double-ML ideas from finite-dimensional inference to global, omnibus specification testing, a setting where plug-in tests genuinely fail under regularization bias (Table 2: plug-in reaches 17.7% at 5% nominal while LRK-CF stays near nominal). Named strengths: an explicit, example-by-example FSIP construction rather than an abstract existence claim; a sharp, interpretable local-power theory (undetectable directions are exactly the nuisance-space directions, and the tensor-product design confirms this falsifiable prediction in Figure 1/Table 7); a computationally cheap multiplier bootstrap with no nuisance re-estimation; and honest comparisons against oracle and non-orthogonal benchmarks. The advertised novelty over Escanciano (2024) is precisely the growing-dictionary regime, which is where my main concern lies.

major comments (2)
  1. [Supp. B.3, Example 1 verification] Supp. B.3 (Assumption 2 paragraph, p. S-23/24): the growing-dictionary verification requires the Gram eigenvalues bounded below AND sup_x b_J(x)'G_J^{-1}b_J(x) uniformly bounded in J, yielding sup_J r̄_J < ∞ for the sieve representer r_{γ,0,J}. For the Fourier dictionaries of Design 1 each basis function is uniformly O(1), so with λ_min(G_J) bounded below, b_J(x)'G_J^{-1}b_J(x) ≍ b_J(x)'b_J(x) = Σ_j(sin²+cos²) ≍ J for every x: the stated hypothesis fails as J_n→∞. Since the representer bound is derived from this leverage, the advertised claim (§2.4: 'we allow the nuisance dimension to increase with the sample size'; B.3 summary: s log J = o(√n)) is not covered by the sufficient conditions as written. Two resolutions seem available: (i) bound ∥r_{γ,0,J}(x,·)∥_K directly rather than through leverage — the representer norm involves the kernel-smoothed Gram E[λ₀b_J K(X,X')b_J'], whose entrie
  2. [§6.1, Design 1 simulations] §6.1 (Design 1, Tables 2, 5–7): the simulations hold J ∈ {50,100,150} fixed while n ∈ {200,400}, so they do not probe the growing-J regime the theory advertises as its main advance over Escanciano (2024). The size results are therefore consistent with at least two explanations — the stated sufficient conditions, or a genuinely milder requirement masked at these (J,n) — and the paper provides no evidence distinguishing them (for J=150, n=200 the crude bound J·s log J/n already exceeds 1, yet size is accurate, suggesting orthogonality helps beyond what Supp. B.3 captures). I recommend adding a design with J_n growing along a polynomial path in n (and ideally a rougher kernel than the Gaussian) to document where size control begins to degrade; this would also discipline the resolution of the previous comment.
minor comments (6)
  1. [Supp. A, proofs of Theorems 1 and 5] Supp. A, proof of Theorem 1 (and proof of Theorem 5, bias part): the text cites 'eq. (10)' for the local-alternative conditional mean E_{F_n}[ε_i|X_i]=n^{-1/2}a(X_i); in the main text that display is eq. (13) — eq. (10) is the weighted projection Π_{Γ,λ}. Please correct both cross-references.
  2. [§3.2 / Supp. A] Notation for the adjusted test function is inconsistent: the main text (§3.2) writes ê_k(x)=k(x)−α_{0θ}(x,k), while the proof of Proposition 1 (Supp. A, p. S-2) uses k̃(x). Please unify.
  3. [§5.2, Table 1] Table 1: the CATE/CATT rows are hard to parse in the current layout (adjoint and undetectable-direction columns run together). The formulas themselves check out — for CATT, (Π⊥)'a = a − e_0 E[a]/E[e_0] and undetectable directions a = c e_0 — but the table would benefit from restructuring (e.g., separate rows for CATE and CATT).
  4. [Supp. B.4, Example 2 verification] Supp. B.4: the claim q_{n,ℓ}=O_p(p_{n,ℓ}) — that estimating the conditional kernel-mean embedding μ₀(z,·)=E[K(X,·)|Z=z] inherits the scalar regression rate because the response is bounded in H_K — is argued by analogy. A citation to kernel mean embedding/nonparametric regression rates in Hilbert space would strengthen this step, particularly for the random-forest case.
  5. [§6.1, Design 1] §6.1: the Lasso penalty λ_n = 0.5√(log J/n) is held fixed across all designs. A brief sensitivity check on the leading constant (or a data-dependent choice) would help assess how much the size results depend on it, given that regularization bias is the object of study.
  6. [§1; §6.4] Typo, §1: 'semiparametricC(α)-type tests' (missing space). Also, in §6.4 and Table 10 the phrase 'both quantities are multiplied by 10^{-6}' is ambiguous about the direction of rescaling; please clarify whether the reported values are scaled up or down.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: oracle equivalence, local power, and bootstrap validity are derived from Neyman orthogonality and high-level rate assumptions, not from fitted targets or load-bearing self-citation.

full rationale

The paper is a methodological statistics contribution. The null is the population conditional moment restriction; the orthogonal moment ψ is constructed from pathwise derivatives (FSIP) under Assumption 1 and shown Neyman-orthogonal in Proposition 1 by direct differentiation along paths F_τ—not by fitting to data or importing an unverified uniqueness claim. Theorem 1 (oracle equivalence) is proved from high-level Assumptions 2–6 via an explicit residual/representer decomposition in the supplement; Theorems 2–3 and Proposition 2 then follow from a Hilbert-space CLT and the adjoint of Π⊥_0. Locally undetectable directions (Table 1) are correctly identified as the range of the nuisance projection—an intentional, stated consequence of orthogonalization, not a hidden fit presented as prediction. Self-citations (Escanciano 2006/2024; Chernozhukov–Escanciano et al. 2022) supply standard tools (ICM/RKHS tests, locally robust moments) but the main theorems are proved self-contained. Simulations and the NSW application illustrate finite-sample behavior; they do not calibrate the limiting law or the testable-direction characterization. Separately noted concerns about whether sufficient conditions in Supp. B.3 hold for growing Fourier dictionaries are satisfiability/correctness issues, not circularity. No step reduces a claimed prediction to its own inputs by construction.

Assumptions & free parameters 3 free parameters · 5 assumptions · 2 invented entities

The paper is a semiparametric testing methodology piece. Load-bearing content is standard i.i.d. empirical-process and Neyman-orthogonality machinery plus high-level rate conditions on generic ML nuisances; no new physical entities. Free choices are implementation tuning (kernel, folds, multipliers, penalties) that affect finite-sample behavior but not the population null.

free parameters (3)
  • Gaussian kernel bandwidth σ (median pairwise distance) = sample median of nonzero pairwise distances
    Chosen by a data-dependent rule in all simulations and the NSW application; affects finite-sample power/size though not the asymptotic null under stated conditions.
  • Lasso penalty level λ_n and dictionary size J = λ_n=0.5√(log J/n); J up to 150 / 80-term dict.
    Design 1 uses λ_n=0.5√(log J/n) and J∈{50,100,150}; Design 3 uses an 80-term dictionary. These are tuning choices for the nuisance estimators.
  • Number of cross-fitting folds L and bootstrap draws = L=5; 999 bootstrap replications
    Fixed at L=5 and 999 Mammen multipliers in reported experiments; theory only needs L fixed and finite.
assumptions (5)
  • domain assumption i.i.d. observations and standard pathwise differentiability / nonsingularity conditions for the conditional moment and nuisance identifying equations (Assumption 1).
    Foundation for the FSIP representation and Neyman orthogonality in §3; standard in semiparametric inference.
  • domain assumption Mean-square consistency, product rates o_p(n^{−1/2}), and orthogonality bias √n||b_Fn(η̂)−b_Fn(η_0)||=o_p(1) for cross-fitted ML nuisances (Assumptions 4–6), often via ||η̂−η_0||=o_p(n^{−1/4}).
    Load-bearing for Theorem 1 oracle equivalence; verified under sparsity/smoothness primitives in examples but not free for arbitrary black-box learners.
  • domain assumption Bounded continuous kernel K and RKHS representation of the first-step influence process (Assumptions 1(b), 2), plus moment and feature-map conditions (Assumptions 7–8).
    Needed to reduce the supremum statistic to a quadratic form and run Hilbert-valued CLTs.
  • standard math Hilbert-space triangular-array CLT / conditional multiplier CLT (e.g. Kundu et al. 2000; van der Vaart–Wellner framework).
    Used in proofs of Theorems 2–3 for weak convergence of R_n^0 and bootstrap processes in H_K.
  • domain assumption Unconfoundedness and overlap when specializing to CATE/CATT (Example 3 / NSW application).
    Identification of doubly robust scores; standard causal assumption, not tested by LRK itself.
invented entities (2)
  • Locally robust kernel (LRK) statistic / orthogonal RKHS empirical process ν̂_n independent evidence
    purpose: Feasible test process and quadratic-form statistic that is first-order insensitive to nuisance estimation error.
    Constructed object of the paper (Neyman-orthogonal moments + cross-fitting + RKHS indexing), not a new physical entity; independent evidence is the derived asymptotics and simulations.
  • First-step influence process (FSIP) φ independent evidence
    purpose: RKHS-indexed adjustment that orthogonalizes the whole collection of testing moments with respect to γ and θ.
    Extension of classical influence functions to a process; defined via pathwise derivatives in §3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Locally Robust Kernel Specification Tests for Conditional Moment Restrictions." pith.science (2026). https://pith.science/paper/YGZNGKIR

@misc{pith2026260724382,
  author       = {Pith},
  title        = {Pith review of: Locally Robust Kernel Specification Tests for Conditional Moment Restrictions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YGZNGKIR}},
  note         = {Machine review of arXiv:2607.24382}
}
read the original abstract

We develop kernel-based specification tests for semiparametric conditional moment models with high-dimensional nuisance parameters, extending existing conditional moment tests---which typically require asymptotically linear nuisance estimators---to accommodate modern machine-learning methods. The proposed locally robust kernel tests combine Neyman-orthogonal moments, cross-fitting, and reproducing kernel Hilbert space methods, yielding inference that is first-order insensitive to nuisance estimation error. We establish oracle equivalence between the feasible and infeasible test processes under local alternatives and weak nuisance-rate conditions, and characterize the resulting local power. A fast multiplier bootstrap avoids nuisance re-estimation. Applications include specification testing in high-dimensional linear and logistic regression, significance testing with machine-learning regressions, and tests of constant conditional treatment effects. Monte Carlo simulations and an application to the National Supported Work program illustrate the finite-sample performance of the proposed tests.

Figures

Figures reproduced from arXiv: 2607.24382 by the authors.

Figure 1
Figure 1. Design 1: Rejection probabilities for J = 100. The perturbation lies outside the additive nuisance space but inside the tensor-product nuisance space. LRK-CF and LRK￾NoCF use cross-fitted and full-sample residuals, respectively; the dashed line is the nominal 5% level. 6.2 Design 2: Significance Testing with Machine Learning The second experiment corresponds to Example 2. The DGP is Yi = g0(Zi) + δ √ n Di + εi , g0(… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 1 canonical work pages

  1. [1]

    Following these studies, we combine NSW treated participants with a nonexperimental comparison group from the Panel Study of Income Dynamics (PSID)

    do not appear Empirical Application: Constant Treatment Effects in the NSW Data We apply the LRK test to the National Supported Work (NSW) job-training data studied by LaLonde (1986) and Dehejia and Wahba (1999). Following these studies, we combine NSW treated participants with a nonexperimental comparison group from the Panel Study of Income Dynamics (PS...

    1986

  2. [2]

    and Wager, S

    Athey, S., Tibshirani, J. and Wager, S. (2019). Generalized random forests. Ann. Statist. 47 1148--1178

    2019

  3. [3]

    J., Ritov, Y

    Bickel, P. J., Ritov, Y. and Stoker, T. M. (2006). Tailor-made tests for goodness of fit to semiparametric hypotheses. Ann. Statist. 34 721--741

    2006

  4. [4]

    Bierens, H. J. (1982). Consistent model specification tests. J. Econometrics 20 105--134

    1982

  5. [5]

    Bravo, F., Escanciano, J. C. and van Keilegom, I. (2020). Two-step semiparametric empirical likelihood inference. The Annals of Statistics 48 1--26

    2020

  6. [6]

    and van de Geer, S

    B\"uhlmann, P. and van de Geer, S. (2011). Statistics for High-Dimensional Data: Methods, Theory and Applications. Springer, Berlin

    2011

  7. [7]

    and Fan, Y

    Chen, X. and Fan, Y. (1999). Consistent hypothesis testing in semiparametric and nonparametric models for econometric time series. J. Econometrics 91 373--401

    1999

  8. [8]

    and Spindler, M

    Chernozhukov, V., Hansen, C. and Spindler, M. (2015). Valid post-selection and post-regularization inference: An elementary, general approach. Annual Review of Economics 7 649--688

    2015

Show all 41 references
  1. [9]

    Chernozhukov, V., Chetverikov, D., Demirer, M., Duflo, E., Hansen, C., Newey, W. K. and Robins, J. M. (2018). Double/debiased machine learning for treatment and structural parameters. Econometrics Journal 21 C1--C68

    2018

  2. [10]

    C., Ichimura, H., Newey, W

    Chernozhukov, V., Escanciano, J. C., Ichimura, H., Newey, W. K. and Robins, J. M. (2022). Locally robust semiparametric estimation. Econometrica 90 1501--1535

    2022

  3. [11]

    and Gretton, A

    Chwialkowski, K., Strathmann, H. and Gretton, A. (2016). A kernel test of goodness of fit. In Proceedings of the 33rd International Conference on Machine Learning 48 2606--2615

    2016

  4. [12]

    K., Hotz, V

    Crump, R. K., Hotz, V. J., Imbens, G. W. and Mitnik, O. A. (2008). Nonparametric tests for treatment effect heterogeneity. Rev. Econ. Statist. 90 389--405

    2008

  5. [13]

    Dehejia, R. H. and Wahba, S. (1999). Causal effects in nonexperimental studies: Reevaluating the evaluation of training programs. J. Amer. Statist. Assoc. 94 1053--1062

    1999

  6. [14]

    Delgado, M. A. and Gonz\'alez-Manteiga, W. (2001). Significance testing in nonparametric regression based on the bootstrap. Ann. Statist. 29 1469--1507

    2001

  7. [15]

    Escanciano, J. C. (2006). A consistent diagnostic test for regression models using projections. Econometric Theory 22 1030--1051

    2006

  8. [16]

    Escanciano, J. C. (2024). A Gaussian process approach to model checks. Ann. Statist. 52 2456--2481

    2024

  9. [17]

    Escanciano, J. C. and de U\ na-\'Alvarez, J. (2025). Goodness-of-fit tests for censored and truncated data: Maximum mean discrepancy over regular functionals. Working paper

    2025

  10. [18]

    Escanciano, J. C. and Goh, C. (2014). Specification analysis of linear quantile models. J. Econometrics 178 495--507

    2014

  11. [19]

    and Gonz\'alez-Manteiga, W

    Gaio, R., Costa-Miranda, R. and Gonz\'alez-Manteiga, W. (2026). Estimation-robust model checking for generalized partially linear models. ISNPS 2026 contributed talk

    2026

  12. [20]

    M., Rasch, M

    Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch\"olkopf, B. and Smola, A. J. (2012). A kernel two-sample test. J. Mach. Learn. Res. 13 723--773

    2012

  13. [21]

    and Mammen, E

    H\"ardle, W. and Mammen, E. (1993). Comparing nonparametric versus parametric regression fits. Ann. Statist. 21 1926--1947

    1993

  14. [22]

    and Zhu, L

    He, C., Chen, C. and Zhu, L. (2026). A goodness-of-fit assessment for general learning procedures in high dimensions. J. Amer. Statist. Assoc. 121 536--547. doi:10.1080/01621459.2025.2529602

    2026

  15. [23]

    and Newey, W

    Ichimura, H. and Newey, W. K. (2022). The influence function of semiparametric estimators. Quantitative Economics 13 29--61

    2022

  16. [24]

    D., B\"uhlmann, P

    Jankov\'a, J., Shah, R. D., B\"uhlmann, P. and Samworth, R. J. (2020). Goodness-of-fit testing in high-dimensional generalized linear models. J. Roy. Statist. Soc. Ser. B 82 773--795

    2020

  17. [25]

    Kennedy, E. H. (2023). Towards optimal doubly robust estimation of heterogeneous causal effects. Electron. J. Statist. 17 3008--3049

    2023

  18. [26]

    LaLonde, R. J. (1986). Evaluating the econometric evaluations of training programs with experimental data. American Economic Review 76 604--620

    1986

  19. [27]

    and Vergara Merino, P

    Lapenta, E., Strittmatter, A. and Vergara Merino, P. (2026). A machine-learning-compatible omnibus test for treatment effect heterogeneity. Working paper

    2026

  20. [28]

    and Song, X

    Lu, H. and Song, X. (2026). Orthogonal integrated conditional moment tests for treatment effect heterogeneity. arXiv preprint arXiv:2607.12622

    2026 arXiv

  21. [29]

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

    1993

  22. [30]

    and Sch\"olkopf, B

    Muandet, K., Fukumizu, K., Sriperumbudur, B. and Sch\"olkopf, B. (2017). Kernel mean embedding of distributions: A review and beyond. Found. Trends Mach. Learn. 10 1--141

    2017

  23. [31]

    M., Rotnitzky, A

    Robins, J. M., Rotnitzky, A. and Zhao, L. P. (1994). Estimation of regression coefficients when some regressors are not always observed. J. Amer. Statist. Assoc. 89 846--866

    1994

  24. [32]

    Rosenbaum, P. R. and Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika 70 41--55

    1983

  25. [33]

    Sancetta, A. (2022). Testing subspace restrictions in the presence of high dimensional nuisance parameters. Electron. J. Statist. 16 5277--5320

    2022

  26. [34]

    Shah, R. D. and B\"uhlmann, P. (2018). Goodness-of-fit tests for high dimensional linear models. J. Roy. Statist. Soc. Ser. B 80 113--135

    2018

  27. [35]

    Song, K. (2010). Testing semiparametric conditional moment restrictions using conditional martingale transforms. J. Econometrics 154 74--84

    2010

  28. [36]

    Stute, W. (1997). Nonparametric model checks for regression. Ann. Statist. 25 613--641

    1997

  29. [37]

    and Zhu, L

    Tan, F., Tang, S. and Zhu, L. (2026). Asymptotic distribution-free tests for ultra-high dimensional parametric regressions via projected empirical processes and p -value combination. arXiv preprint arXiv:2601.00541

    2026

  30. [38]

    Tibshirani, R. (1996). Regression shrinkage and selection via the Lasso. J. Roy. Statist. Soc. Ser. B 58 267--288

    1996

  31. [39]

    van der Vaart, A. W. and Wellner, J. A. (1996). Weak Convergence and Empirical Processes: With Applications to Statistics. Springer, New York

    1996

  32. [40]

    Zheng, J. X. (1996). A consistent test of functional form via nonparametric estimation techniques. J. Econometrics 75 263--289

    1996

  33. [41]

    and Mukherjee, K

    Kundu, S., Majumdar, S. and Mukherjee, K. (2000). Central limit theorems revisited. Statistics & Probability Letters 47 265--275. doi:10.1016/S0167-7152(99)00164-9

    2000 doi

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.