Pith. sign in

REVIEW 2 major objections 2 minor 27 references

On the Bernstein-smoothed lower-tail Spearman's rho estimator

T0 review · 2 major / 2 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proves that Bernstein-smoothed lower-tail Spearman's rho keeps the classical estimator's Gaussian limit while strictly reducing mean squared error, with simulated gains of up to 70% in deep tails.

desk verdict Useful niche estimator with a solid first-order theory, but the headline MSE improvement is asserted without a covariance analysis. read the letter →

arxiv 2506.08857 v1 pith:ZB4EQPPR submitted 2025-06-10 math.ST stat.MEstat.TH

Frédéric Ouimet , Selim Orhun Susam This is my paper
classification math.STstat.MEstat.TH MSC 62H1262E2062G0562H05
keywords BernsteinestimatorSpearman'srhocopulalower-taildependencenonparametricestimationsmoothingempiricalmeansquarederror
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

This paper proposes a Bernstein-smoothed version of lower-tail Spearman's rho, the concordance measure obtained by integrating a copula over the square [0,p]^2 rather than over the whole unit square. The estimator replaces the empirical copula by its Bernstein-polynomial smooth, and the paper proves it is strongly consistent and asymptotically normal with the same limiting variance as the classical estimator. The sharper result is a second-order comparison: with Bernstein degree $m \asymp n^{2/3}$, the variance drops by an extra term of order $n^{-1}m^{-1/2}$ while the bias costs only $O(m^{-1})$, so the mean squared error is strictly smaller than the classical estimator's. This matters because tail concordance is exactly where sample sizes are effectively small, so a variance reduction that leaves the limit distribution untouched makes tail-dependence inference more precise, and simulations with the FGM copula show MSE reductions up to roughly 70% at the deep-tail threshold $p=0.1$.

What carries the argument

The object doing the work is the Bernstein copula estimator $C_{m,n}(u,v)=\sum_{k,\ell=0}^m C_n(k/m,\ell/m)P_{k,m}(u)P_{\ell,m}(v)$, where the $P_{k,m}(w)=\binom{m}{k}w^k(1-w)^{m-k}$ are binomial kernels used as smoothing weights on the empirical copula $C_n$. The paper feeds this estimator through the integral operator $T_p(f)=D(p)^{-1}\int_{[0,p]^2}f(u,v)\,du\,dv$ so that the pointwise bias and variance expansions of $C_{m,n}$ become expansions of the integrated statistic. The key mechanism is that the pointwise variance-reduction term $-V(u,v)/(n\sqrt{m})$ survives integration as $-T_p(V)/(n m^{1/2})$, while the pointwise bias contributes $O(m^{-1})$; equating the two orders gives the degree rule $m\asymp n^{2/3}$ and a net loss of $O(n^{-4/3})$ in MSE relative to the empirical estimator.

What would settle it

Compute the exact covariance structure of the integrated Bernstein copula process for a fixed copula and finite or moderate $n$, and check whether $\operatorname{Var}[\hat\rho_{m,n}(p)]$ equals $\sigma_p^2/n - T_p(V)/(n m^{1/2}) + o(n^{-1}m^{-1/2})$ with the coefficient exactly $-T_p(V)$; a Monte Carlo evaluation of the variance at several $m$ for a copula with known $T_p(V)$ could also settle it.

Watch

Extended reading notes

Core claim

The paper's central claim is that the Bernstein-smoothed lower-tail Spearman's rho, defined by $\hat\rho_{m,n}(p)=D(p)^{-1}(\int_{[0,p]^2} C_{m,n}(u,v)\,du\,dv - p^4/4)$ with $C_{m,n}$ the Bernstein copula estimator and $D(p)=p^3/3-p^4/4$, inherits the strong consistency and the $\sqrt{n}$-Gaussian limit of the empirical-copula estimator while improving the variance expansion. Specifically, the paper obtains $\operatorname{Bias}[\hat\rho_{m,n}(p)] = T_p(b)/m + o(m^{-1})$ and $\operatorname{Var}[\hat\rho_{m,n}(p)] = \sigma_p^2/n - T_p(V)/(n m^{1/2}) + o(n^{-1}m^{-1/2})$, where $T_p(f)$ is the normalized integral of $f$ over $[0,p]^2$. Balancing bias and variance at $m_{\mathrm{opt}}\asymp n^{2/3}$ yields $\operatorname{MSE}[\hat\rho_{m_{\mathrm{opt}},n}(p)] = \sigma_p^2/n - 3T_p(V)^{4/3}/(4^{4/3}T_p(b)^{2/3}n^{4/3}) + o(n^{-4/3})$, which is strictly smaller than the classical estimator's $\sigma_p^2/n + o(n^{-1})$. In plain terms, smoothing the empirical copula with Bernstein polynomials removes part of the sampling variance in the tail without paying for it at first order in the limit distribution.

Load-bearing premise

The load-bearing premise is that the pointwise variance expansion of the Bernstein copula estimator can be integrated term by term over the tail square, so that the negative $-T_p(V)/(n\sqrt{m})$ term survives and any cross-covariance between different points is of lower order; if that fails, the strictly-smaller-MSE claim is not established.

Editorial extensions

If this is right

  • With $m\asymp n^{2/3}$, $\hat\rho_{m,n}(p)$ has the same $\sqrt{n}$-normal limit as $\hat\rho_n(p)$, so standard normal-based confidence intervals and tests for lower-tail Spearman's rho remain valid while the estimator's MSE is smaller to second order.
  • The estimator is strongly consistent at rate $O(n^{-1/2}(\log\log n)^{1/2})$, so it can be used reliably when interest is confined to a deep lower-tail window where few observations effectively matter.
  • For the FGM copula at $p=0.1$, simulated MSE reductions reach about 70% for weak to moderate dependence and $n=50$, and remain substantial at $n=200$; at $p=0.5$ reductions are smaller but typically positive, while at $p=1$ smoothing can mildly increase MSE under strong dependence.
  • The rule-of-thumb degree $m=\lfloor n^{2/3}\rfloor$ follows from the bias-variance balance, and the MSE-minimizing degree shifts to smaller values when the tail window narrows, giving practical guidance for choosing smoothing strength.

Reading between the lines

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

  • Editorial inference: the same integrate-then-smooth argument should apply to upper-tail Spearman's rho and to other integral functionals of the copula, such as weighted Cramér-von Mises statistics, suggesting a general second-order variance-reduction phenomenon for Bernstein smoothing that the paper does not state.
  • Editorial inference: the asymptotic degree rule $m\asymp n^{2/3}$ may be suboptimal in very deep tails, where the effective sample within $[0,p]^2$ is small; a plug-in or data-driven degree selection could outperform the fixed rule, a direction the paper itself flags.
  • Editorial inference: because the bias is $O(m^{-1})$ and grows as $p$ shrinks, smoothing is not uniformly beneficial; the simulation's negative MSE reductions at $p=1$ under strong dependence suggest boundary cases where the variance gain is outweighed by bias, and this trade-off deserves explicit quantification.
  • Editorial inference: if the variance-reduction mechanism extends to higher-dimensional copulas, Bernstein smoothing could soften the curse of dimensionality for tail-dependence estimation, but the present proof is bivariate and does not establish that.
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 / 2 minor

Summary. The paper proposes a Bernstein-smoothed estimator of the lower-tail Spearman's rho, bρm,n(p), defined by integrating the Bernstein copula estimator over [0,p]^2 instead of the empirical copula. The main theoretical results (Theorems 1 and 2) establish strong consistency and asymptotic normality with the same asymptotic variance as the classical empirical-copula estimator. Section 2.3 then claims a second-order variance reduction of order n^{-1}m^{-1/2} and, by balancing this against an O(m^{-1}) bias, concludes that the smoothed estimator attains a strictly smaller MSE than the unsmoothed estimator when m ≍ n^{2/3}. A Monte Carlo study on the FGM copula reports substantial finite-sample MSE reductions, up to about 70% for deep tails and small sample sizes, with accompanying R code made available online.

Significance. If the MSE-improvement claim were rigorously established, the paper would be a useful contribution: it would show that a simple smoothing procedure can reduce the tail-focused Spearman's rho estimator's variance without altering the first-order limit distribution. The first-order results are soundly derived from known weak convergence and strong consistency results for Bernstein copula estimators, and the paper is clearly written. The availability of reproducible simulation code is a strength. However, the central second-order claim is currently unsupported because the integrated variance expansion is asserted rather than derived from the covariance structure of the smoothed process.

major comments (2)
  1. [Section 2.3, Eq. (5) and the variance expansion] The expansion Var[bρm,n(p)] = σ_p^2/n - T_p(V)/(n m^{1/2}) + o(n^{-1}m^{-1/2}) is stated without proof. This quantity is a double integral of Cov(Cm,n(u,v), Cm,n(s,t)) over [0,p]^4, and the pointwise variance expansion (5) only controls the diagonal s=u, t=v. Off-diagonal covariances can contribute to the integrated second-order term, and no argument is given that they are negligible at order n^{-1}m^{-1/2}. The subsequent strict-MSE claim and the optimal-degree formula rest on this unproved covariance integration, so the central theoretical conclusion is not established.
  2. [Section 2.3, m_opt formula] The balancing formula m_opt = (4 T_p(b)^2/(T_p(V) n))^{2/3} and the accompanying MSE expansion involve T_p(b) in the denominator. For the FGM copula with θ=0, which is explicitly included in Table 1, the copula is the independence copula Π(u,v)=uv, so b(u,v)=0 identically and T_p(b)=0. The optimal-degree formula is then undefined, and the claimed n^{-4/3} MSE reduction term has a zero denominator. The paper should either exclude such cases from the theoretical claim or provide a separate treatment for the case T_p(b)=0, where the bias is of smaller order and the balance between bias and variance changes.
minor comments (2)
  1. [Section 3, Table 1] Table 1 reports MSE reduction percentages without Monte Carlo standard errors or confidence intervals. Since K=10,000 replications are used, standard errors for the estimated variances and MSEs could be easily computed and would help the reader assess whether the reported reductions are statistically distinguishable from zero.
  2. [Section 3.4] The statement that the empirical MSE-minimizing degree 'shifts to smaller degrees when the integration window narrows' is made without an explicit quantitative summary. A small table or text listing the argmin m for each setting would make the comparison with the rule-of-thumb m=⌊n^{2/3}⌋ more transparent.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: central results rely on external Janssen-Swanepoel-Veraverbeke theorems and the functional delta method; self-citations are motivational only.

full rationale

The derivation chain is self-contained with respect to external inputs. The estimator in (6) is defined by plugging the Bernstein copula C_{m,n} into Schmid-Schmidt's lower-tail rho functional, and its consistency (Theorem 1) and asymptotic normality (Theorem 2) follow directly from external uniform consistency and weak convergence results of Janssen-Swanepoel-Veraverbeke [6] applied through the functional delta method. The claimed second-order MSE improvement in Section 2.3 uses the pointwise bias/variance expansions (5) from [6]; the balancing m approximately n^{2/3} is the same rule as in [6, Remark 4]. The only self-citations ([8], [9], [21]) are motivational or code-repository links and are not load-bearing. One could question Section 2.3 because the integrated variance expansion Var[bHat_{m,n}(p)] = sigma_p^2/n - T_p(V)/(n m^{1/2}) + o(n^{-1} m^{-1/2}) is asserted without deriving it from the double integral of the covariance of C_{m,n}; in particular, the assertion that off-diagonal covariance terms are negligible is not proved. That concern is a correctness risk, not circularity, since the expansion does not assume the conclusion it is used to prove. No fitted parameter is renamed a prediction, and no self-citation chain forces the main conclusion.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The central MSE claim relies on external pointwise expansions from [6] and on an unproved exchange of integration with the variance expansion. No free parameters are fitted and no new entities are postulated.

assumptions (3)
  • domain assumption The copula C has continuous first-order partial derivatives on [0,1]^2.
    Required for weak convergence of the empirical copula process [13, Theorem 3] and invoked in Theorem 2.
  • ad hoc to paper The pointwise bias and variance expansions (5) from Janssen et al. hold uniformly enough to be integrated over [0,p]^2, with cross-covariance terms of lower order.
    Section 2.3 asserts the integrated variance expansion without analyzing the covariance kernel of the Bernstein copula process.
  • domain assumption The regularity conditions for [6, Lemma 3 (iii)] are satisfied, including existence and continuity of second-order derivatives appearing in b(u,v) and V(u,v).
    The bias and variance computations in Section 2.3 rest on these conditions, which the paper does not restate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Bernstein-smoothed lower-tail Spearman's rho estimator." pith.science (2026). https://pith.science/paper/ZB4EQPPR

@misc{pith2026250608857,
  author       = {Pith},
  title        = {Pith review of: On the Bernstein-smoothed lower-tail Spearman's rho estimator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZB4EQPPR}},
  note         = {Machine review of arXiv:2506.08857}
}
abstract

This note develops a Bernstein estimator for lower-tail Spearman's rho and establishes its strong consistency and asymptotic normality under mild regularity conditions. Smoothing the empirical copula yields a strictly smaller mean squared error (MSE) in tail regions by lowering sampling variance relative to the classical Spearman's rho estimator. A Monte Carlo simulation experiment with the Farlie--Gumbel--Morgenstern copula demonstrates variance reductions that translate into lower MSE estimates (up to $\sim 70\%$ lower) at deep-tail thresholds under weak to moderate dependence and small sample sizes. To facilitate reproducibility of the findings, the R code that generated all simulation results is readily accessible online.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 18 canonical work pages

  1. [1]

    On the Bernstein-smoothed lower-tail Spearman's rho estimator

    Introduction Copulas provide a flexible way to separate marginal behavior from joint dependence. By Sklar’s theorem [1], any continuous bivariate distribution function H with marginals F and G admits a unique copula C : [0, 1]2 → [0, 1] satisfying H(x, y) =C(F(x), G(y)). In many fields, including hydrology, risk theory, and financial econometrics, practit...

    work page Pith review arXiv 2025

  2. [2]

    Estimation of the lower-tail Spearman’s rho Spearman’s rho is a widely used concordance measure for two continuous random variables. Expressed in terms of the underlying copula C, it is given by ρS = 12 Z [0,1]2 {C(u, v) − Π(u, v)}dudv = 12 Z [0,1]2 C(u, v)dudv − 3, where Π(u, v) is the independence copula; see, e.g., Nelsen [12, Section 5.1.2 ]. It offer...

    work page

  3. [3]

    Simulation study The FGM copula is defined, for every u, v ∈ [0, 1] and θ ∈ [−1, 1], by Cθ(u, v) =uv{1 + θ(1 − u)(1 − v)}; see, e.g., Nelsen [12, p. 77]. A simulation study is conducted for parameter values θ ∈ {−1, −0.5, 0, 0.5, 1}, ranging from moderate discordance (θ = −1) to moderate concordance ( θ = 1), sample sizes n ∈ {50, 200}, and lower-tail thr...

    work page

  4. [4]

    Summary and future research The Bernstein estimator bρm,n(p) for lower-tail Spearman’s rho has been proposed. Under mild regularity conditions, it retains the strong consistency and the same √n-limit distribution as the empirical copula-based counterpart, bρn(p), while reducing its variance by a term of order ≍ n−1m−1/2 at the cost of O(m−1) bias. Consequ...

    work page 2024

  5. [5]

    The Bernstein copula and its applications to modeling and approximations of multivariate distributions

    Sancetta, A.; Satchell, S. The Bernstein copula and its applications to modeling and approximations of multivariate distributions. Econometric Theory 2004, 20, 535–562. https://doi.org/10.1017/S0266466604203 05X

    work page doi:10.1017/s0266466604203 2004

  6. [6]

    Also, they showed that when the polynomial degree is set to m ≍ n2/3, the second-order term of its (pointwise) variance is strictly smaller than for Cn; see (5) below for details

    proved that it is uniformly strongly consistent, just as the classical empirical copulaCn is. Also, they showed that when the polynomial degree is set to m ≍ n2/3, the second-order term of its (pointwise) variance is strictly smaller than for Cn; see (5) below for details. Building on this variance reduction, the present study proposes a Bernstein version...

    work page

  7. [7]

    Hudaverdi and Susam [8] adapted the same Bernstein smoothing idea to weighted Cramér– von Mises statistics and documented substantial power gains across a broad spectrum of copulas

    proposed three nonparametric tests of independence based on the Bernstein copula estimator and its density, yielding higher power than a closely related Cramér–von Mises test built using the empirical copula. Hudaverdi and Susam [8] adapted the same Bernstein smoothing idea to weighted Cramér– von Mises statistics and documented substantial power gains ac...

    work page

  8. [8]

    Fonctions de répartition à n dimensions et leurs marges

    Sklar, M. Fonctions de répartition à n dimensions et leurs marges. Publ. Inst. Statist. Univ. Paris 1959, 8, 229–231

    work page 1959

Show all 27 references
  1. [9]

    Multivariate conditional versions of Spearman’s rho and related measures of tail dependence

    Schmid, F.; Schmidt, R. Multivariate conditional versions of Spearman’s rho and related measures of tail dependence. J. Multivariate Anal. 2007, 98, 1123–1140. https://doi.org/10.1016/j.jmva.2006.05.005

    2007 doi

  2. [10]

    La fonction de dépendance empirique et ses propriétés

    Deheuvels, P . La fonction de dépendance empirique et ses propriétés. Un test non paramétrique d’indépendance. Acad. Roy. Belg. Bull. Cl. Sci. (5) 1979, 65, 274–292

    1979

  3. [11]

    Bernstein Polynomials, second ed.; Chelsea Publishing Co., New York, 1986; pp

    Lorentz, G.G. Bernstein Polynomials, second ed.; Chelsea Publishing Co., New York, 1986; pp. x+134

    1986

  4. [12]

    Large sample behavior of the Bernstein copula estimator.J

    Janssen, P .; Swanepoel, J.; Veraverbeke, N. Large sample behavior of the Bernstein copula estimator.J. Statist. Plann. Inference 2012, 142, 1189–1197. https://doi.org/10.1016/j.jspi.2011.11.020

    2012 doi

  5. [13]

    Testing independence based on Bernstein empirical copula and copula density

    Belalia, M.; Bouezmarni, T.; Lemyre, F.C.; Taamouti, A. Testing independence based on Bernstein empirical copula and copula density. J. Nonparametr. Stat. 2017, 29, 346–380. https://doi.org/10.1080/10485252.2017. 1303063

    2017

  6. [14]

    On the weighted tests of independence based on Bernstein empirical copula

    Hudaverdi, B.; Susam, S.O. On the weighted tests of independence based on Bernstein empirical copula. Comm. Statist. Simulation Comput. 2023, 52, 404–424. https://doi.org/10.1080/03610918.2020.1859535

    2023

  7. [15]

    A flexible parameter estimation method for the Farlie–Gumbel–Morgenstern copula: a simula- tion study

    Susam, S.O. A flexible parameter estimation method for the Farlie–Gumbel–Morgenstern copula: a simula- tion study. J. Statist. Comput. Simul. 2025, pp. 1–19. https://doi.org/10.1080/00949655.2025.2494140

    2025

  8. [16]

    Bernstein copula characteristic function

    Bahraoui, T. Bernstein copula characteristic function. Comm. Statist. Theory Methods 2024, 53, 6513–6526. https://doi.org/10.1080/03610926.2023.2247107

    2024

  9. [17]

    Bernstein-based estimation of the cross ratio function

    Abrams, S.; Sercik, O.; Veraverbeke, N. Bernstein-based estimation of the cross ratio function. Statistics 2024, 58, 230–246. https://doi.org/10.1080/02331888.2024.2320924

    2024

  10. [18]

    An Introduction to Copulas, second ed.; Springer Series in Statistics, Springer, New York, 2006; pp

    Nelsen, R.B. An Introduction to Copulas, second ed.; Springer Series in Statistics, Springer, New York, 2006; pp. xiv+269. https://doi.org/10.1007/0-387-28678-0

    2006 doi

  11. [19]

    Weak convergence of empirical copula processes.Bernoulli 2004, 10, 847–860

    Fermanian, J.D.; Radulovi´ c, D.; Wegkamp, M. Weak convergence of empirical copula processes.Bernoulli 2004, 10, 847–860. https://doi.org/10.3150/bj/1099579158

    2004

  12. [20]

    zipfR: Statistical Models for Word Frequency Distributions, 2020

    Evert, S.; Baroni, M. zipfR: Statistical Models for Word Frequency Distributions, 2020. R package version 0.6-70. CRAN link

    2020

  13. [21]

    The oscillation behavior of empirical processes: the multivariate case

    Stute, W. The oscillation behavior of empirical processes: the multivariate case. Ann. Probab. 1984, 12, 361–379. https://doi.org/10.1214/aop/1176993295

    1984

  14. [22]

    Asymptotics of empirical copula processes under non-restrictive smoothness assumptions.Bernoulli 2012, 18, 764–782

    Segers, J. Asymptotics of empirical copula processes under non-restrictive smoothness assumptions.Bernoulli 2012, 18, 764–782. https://doi.org/10.3150/11-BEJ387

    2012 doi

  15. [23]

    The estimation of copulas: theory and practice

    Charpentier, A.; Fermanian, J.D.; Scaillet, O. The estimation of copulas: theory and practice. In Copulas: from theory to application in finance; Rank, J., Ed.; London: Risk Books, 2007; pp. 35–64. https://archive-ouverte. unige.ch/unige:41917

    2007

  16. [24]

    kdecopula: An R package for the kernel estimation of bivariate copula densities

    Nagler, T. kdecopula: An R package for the kernel estimation of bivariate copula densities. J. Stat. Softw. 2018, 84, 1–22. https://doi.org/10.18637/jss.v084.i07

    2018 doi

  17. [25]

    GeD spline estimation of multivariate Archimedean copulas

    Dimitrova, D.S.; Kaishev, V .K.; Penev, S.I. GeD spline estimation of multivariate Archimedean copulas. Comput. Statist. Data Anal. 2008, 52, 3570–3582. https://doi.org/10.1016/j.csda.2007.11.010

    2008 doi

  18. [26]

    Linear B-spline copulas with applications to nonparametric estimation of copulas

    Shen, X.; Zhu, Y.; Song, L. Linear B-spline copulas with applications to nonparametric estimation of copulas. Comput. Statist. Data Anal. 2008, 52, 3806–3819. https://doi.org/10.1016/j.csda.2008.01.002

    2008 doi

  19. [27]

    LowerTailSpearman, 2025

    Ouimet, F.; Susam, S.O. LowerTailSpearman, 2025. GitHub repository: https://github.com/ FredericOuimetMcGill/LowerTailSpearman. 10 of 14 Appendix A. Figure A1. Estimated absolute bias, variance, and MSE for bρn(p) and bρm,n(p) in the simulation study, with sample sizes n ∈ {50...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.