Characterization based Goodness-of-Fit for Generalized Pareto Distribution: A Blend of Stein's Identity and Dynamic Survival Extropy

Gaurav Kandpal; Nitin Gupta

REVIEW 4 major objections 5 minor 2 references

Characterization based Goodness-of-Fit for Generalized Pareto Distribution: A Blend of Stein's Identity and Dynamic Survival Extropy

T0 review · 4 major / 5 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 Two new goodness-of-fit tests for the generalized Pareto distribution are built from Stein's identity and dynamic survival extropy, with U-statistic form, asymptotic normality, and a censored-data extension.

desk verdict A promising but currently unsound GOF paper: the new test statistics are real, but the asymptotic proof has a gap that undermines the claimed normality, and Theorem 1 is false for beta <= -1. read the letter →

arxiv 2506.01473 v1 pith:QFMIUVWW submitted 2025-06-02 stat.ME math.STstat.APstat.TH

Gaurav Kandpal , Nitin Gupta This is my paper

classification stat.MEmath.STstat.APstat.TH MSC 62G1062G2062B1094A17

keywords generalizedParetodistributiongoodness-of-fittestingStein'sidentitydynamicsurvivalextropyU-statisticsright-censoreddatabootstrapcriticalvaluesasymptoticnormality

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 proposes two goodness-of-fit tests for the generalized Pareto distribution (GPD), one for positive shape parameter and one for negative. The tests are built from two new characterizations: a Stein-identity fixed-point characterization that works across both support types, and a dynamic-survival-extropy characterization based on the constancy of the product of dynamic survival extropy and the hazard rate. Both test statistics are U-statistics, so they are simple to compute and have asymptotic normal limits. The paper argues that the tests have high power against many alternatives and provides a right-censored extension.

What carries the argument

The load-bearing machinery is the pair of characterizations. The Stein-type identity $F(t)=E\left[\frac{\beta+1}{\theta+\beta X}\,\min(X,t)\right]$ is derived from fixed-point characterizations for semi-bounded and bounded supports and handles both $\beta>0$ and $\beta<0$. The extropy characterization $J_s(X;t)h_F(t)=k$ reduces the GPD to a constant product, with different $k$ recovering exponential, uniform, and modified Pareto hazard rates. Squared deviations from these identities define $\Delta_P$ and $\Delta_N$; symmetrized U-statistic kernels $h_1,h_2,g$ estimate them, and asymptotic normality follows from the central limit theorem for U-statistics after the parameter-estimation term is argued to vanish.

What would settle it

Simulate samples from a known GPD with parameters $(\theta,\beta)$, estimate the parameters with a standard $\sqrt{n}$-consistent estimator such as maximum likelihood, and compare the empirical distribution of $\sqrt{n}\,\widehat{\Delta}_P$ (or $\sqrt{n}\,\widehat{\Delta}_N^*$) with the claimed normal null distribution with variance $9\sigma_0^2$ (or $\sigma_0^2$). If the empirical size under the asymptotic critical values deviates materially from the nominal level while the bootstrap-calibrated test holds its level, the stated null variance is not the right one.

Watch

Extended reading notes

Core claim

The central claim is that a random variable $X$ has the generalized Pareto distribution $\mathrm{GPD}(\theta,\beta)$ if and only if its CDF satisfies $F(t)=E\left[\frac{\beta+1}{\theta+\beta X}\,\min(X,t)\right]$ for $t>0$, and if and only if the product $J_s(X;t)h_F(t)$ is constant, where $J_s$ is the dynamic survival extropy and $h_F$ is the hazard rate. These characterizations are converted into Cramér–von Mises-type departure measures that vanish only under the GPD null hypothesis; replacing expectations by symmetrized U-statistic kernels and parameters by consistent estimators yields test statistics $\widehat{\Delta}_P$ and $\widehat{\Delta}_N$. The paper establishes consistency and asymptotic normality of these statistics (Theorems 4–7), supplies Monte Carlo critical values and power tables, and gives a censored-data version via inverse probability of censoring weighting.

Load-bearing premise

The asymptotic normality of the test statistics assumes that the plug-in estimators of $\beta$ and $\theta$ are so accurate that their estimation error vanishes after multiplying by $\sqrt{n}$; the proof only assumes consistency, which is weaker than what the stated variance formula needs.

Editorial extensions

If this is right

If the paper is correct, practitioners get two easy-to-evaluate GOF statistics that avoid the computational burden of some classical procedures and work for both positive and negative shape parameters.
For $\beta>0$, $\widehat{\Delta}_P$ has an asymptotically normal null distribution and rejects for large values; for $\beta<0$, the scale-invariant $\widehat{\Delta}_N^*$ gives a two-sided test.
The power tables indicate detection of alternatives such as Weibull, Gamma, and absolute Gumbel distributions often exceeds 0.9 even at sample size 50.
The censored-data test $\widehat{\Delta}_c^*$ extends the method to right-censored lifetimes with a variance estimator, so GPD checks can be run on survival data.
The parametric bootstrap algorithm provides critical values without requiring the difficult null variance.

Reading between the lines

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

The asymptotic-normality proof implicitly requires the plug-in estimators to satisfy $\sqrt{n}(\widehat{\beta}-\beta)\to 0$ and $\sqrt{n}(\widehat{\theta}-\theta)\to 0$ in probability; with standard $\sqrt{n}$-consistent estimators the estimation term is $O_p(1)$ and can change the null variance.
The same construction can be specialized to the exponential and uniform boundary cases by taking $\beta=0$ and $\beta=-1$, yielding GOF tests for those distributions as limiting cases.
Using the bootstrap critical values recommended in the paper sidesteps the difficult null variance, and a direct power comparison against the bootstrap test of Villaseñor-Alva and González-Estrada (2009) on identical seeds would quantify the improvement.

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.

Desk Editor's Note

read the letter

Colleague,

Two things to know before you spend time on arXiv:2506.01473. First, the core idea is genuinely new: two U-statistic GOF tests for the GPD, one for positive shape based on a Stein-type characterization and one for negative shape based on dynamic survival extropy. The test statistics are not in the prior literature, and the simulation work is extensive. Second, the paper as written has two load-bearing problems you should not miss.

The asymptotic normality proof for Delta-hat_P (Theorem 5) assumes only consistency of beta-hat and theta-hat, but the proof needs sqrt(n)(beta-hat - beta) → 0 in probability for the plug-in term to vanish. Standard root-n consistent estimators — including the AML and CMM estimators used in the simulations — give sqrt(n)(beta-hat - beta) = O_p(1), so the estimation term contributes to the limit distribution and the stated variance 9 sigma^2 is not justified. The same gap appears in Theorem 6 and in the censored case (Theorem 7). The bootstrap algorithm A1 also resamples the original observations with replacement rather than generating parametric samples from the fitted GPD, so the resulting critical values are not null critical values.

Second, Theorem 1, the Stein characterization for GPD, is outright false for beta = -1 and beta < -1. The proof claims the GPD density vanishes at the upper endpoint, but for beta = -1 the density is constant (uniform case) and for beta < -1 the density diverges. The theorem is not used for the beta < 0 test — that test relies on the extropy characterization — so the negative-side test survives, but the overreach should be corrected.

What earns credit: the test statistics are new; the U-statistic manipulations are competently done; the power study is reasonably thorough, though 1000 replications is low and Table 6 contains an impossible power of 2.000; and the censored-data extension via IPCW is a sensible template. The real-data examples are a plus. This is not a sloppy idea; it is an under-polished paper with localized but serious errors.

Bottom line: the central claim — a simple, asymptotically normal test with valid rejection regions — is not supported as written. The errors are correctable: fix Theorem 1, replace Algorithm A1 with a genuine parametric bootstrap, and either derive the plug-in variance or soften the asymptotics to a consistency result. If the authors repair these, the paper could be a useful contribution to extreme-value GOF testing. For now I would not cite it, but I would not desk-reject it either. It deserves a serious referee and a major revision.

Recommendation: send to peer review with a referee who checks the asymptotics carefully. Conditional acceptance after revision is the right target.

Referee Report

4 major / 5 minor

Summary. The paper proposes goodness-of-fit tests for the generalized Pareto distribution (GPD) based on two characterizations: one derived from Stein-type identities for the positive shape parameter case and one based on dynamic survival extropy for the negative shape parameter case. Test statistics are built as U-statistics, with parameters estimated by the AML and CMM estimators of Villaseñor-Alva and González-Estrada (2009). Monte Carlo critical values and power comparisons are provided, along with an extension to right-censored data and two real-data applications. The paper claims asymptotic normality of the test statistics under the null and alternatives, simple computation, and good power.

Significance. If the claims were fully established, the proposed tests would be a useful addition to the GPD goodness-of-fit toolbox: the characterizations are explicit, the test statistics have closed-form kernels, the simulation study covers many alternatives and sample sizes, and the censored-data extension addresses a practical gap. The authors also make a laudable effort to connect Stein-type identities and dynamic survival extropy. However, the central asymptotic normality proof has a load-bearing plug-in gap, the bootstrap algorithm in the paper is not a parametric bootstrap, and Theorem 1 is false as stated for a range of shape parameters. These issues must be resolved before the paper's main claims can be accepted.

major comments (4)

[Section 2.1, Theorem 1] Theorem 1 is false as stated for beta = -1 and for beta < -1. For beta = -1 the GPD is the Uniform(0,theta) distribution, and the right-hand side of (2.5) is identically zero because beta + 1 = 0, while F(t) = t/theta. The proof's assertion that lim_{x->-theta/beta} f(x) = 0 is also false here: the limit equals 1/theta. For beta < -1 the density diverges at the upper endpoint, so the limit in Lemma 2 does not even exist and condition (iii) of Lemma 2 fails. The characterization should be restricted to beta > -1 (or explicitly to beta > 0 if it is used only for the positive test), and the statement and proof need to be corrected accordingly.
[Section 4.1, proof of Theorem 5 and Theorem 6] The plug-in argument in the proof of Theorem 5 is incomplete. The decomposition splits sqrt(n)(DeltaHat_P - Delta_P) into sqrt(n)(DeltaHat_P - DeltaTilde_P) and sqrt(n)(DeltaTilde_P - Delta_P). For the first term, the proof shows only that ((betaHat+1)^2 - (beta+1)^2)U1 and ((betaHat+1) - (beta+1))U2 converge in probability to zero, and then invokes Chebyshev to claim sqrt(n) times these products is o_p(1). Consistency alone does not give this. If betaHat is root-n consistent, as is usual for the AML and CMM estimators used in Section 5.1, then sqrt(n)(betaHat - beta) is O_p(1), and the plug-in term is O_p(1), not negligible. In addition, the proof treats only the multiplicative factors (betaHat+1)^2 and (betaHat+1); it does not address the fact that U1 and U2 themselves are computed using estimated thetaHat and betaHat inside the kernels (3.8)-(3.9). A correct proof needs a full plug-in/delta-method expansion that adds the estimation contribution to the asymptotic variance, or an explicit assumption of sqrt(n)-consistent estimators together with the expanded variance. Without this, the null variance sigma_0^2 in Corollary 1, the critical region (4.3), and the analogous statements in Theorem 6, Corollary 3, and Theorem 7 are not justified. This is load-bearing for the paper's central claim that the tests have asymptotic normality and controlled size.
[Section 3.1, Algorithm A1] Algorithm A1 is not a parametric bootstrap. It estimates beta, k, and theta from the original data and then resamples the original observations with replacement (y <- x[i]), i.e., it draws from the empirical distribution of the observed sample. Under the alternative hypothesis the bootstrap samples are drawn from the non-GPD empirical distribution, not from a fitted GPD null model, so the resulting critical values C1 and C2 are not null critical values. The text before the algorithm correctly describes parametric bootstrap, but the algorithm implements something different. The bootstrap rejection rule described in Section 3.1 is therefore not valid as stated. The authors should implement a true parametric bootstrap (generate samples from the fitted GPD) or rely solely on the Monte Carlo tables, which are generated under H0.
[Section 4.2, Theorem 7 and Corollary 4] The censored-data theorem inherits the same plug-in problem. The kernel g in (4.6) contains the estimated quantity kHat_c (via betaHat_c), while Theorem 7 is imported from Datta et al. (2010) for a fixed kernel. The variance sigma_{1c}^2 is the projection variance for a fixed kernel; when the kernel depends on estimated parameters, the asymptotic variance changes and the martingale representation in (4.10)-(4.11) does not directly apply. Corollary 4 adds only the consistency of thetaHat_c through Slutsky's theorem, which does not account for the estimation of k and beta inside the kernel. A separate delta-method calculation is needed, or the theoretical result should be stated only for fixed parameters with the censored test calibrated by simulation.

minor comments (5)

[Table 6] The entry for Gen-Gamma(2,1/2) at n=30, beta=0.2 is reported as 2.000, which is outside the valid range [0,1] for a power estimate; this appears to be a typographical error.
[Section 5.3] The sentence 'we reject the null hypothesis H0, when both the alternative hypothesis for positive and negative beta, HP1 and HP1, cannot be rejected' is unclear and contains an apparent typo: the second hypothesis should presumably be HN1, and the stated decision rule should be reworded.
[Equation (4.4)] The expression for sigma^2 has an unmatched parenthesis: it reads Var[E(g(X1,X2,X3)|X1] and should be Var[E(g(X1,X2,X3)|X1)].
[Corollary 3 and Corollary 1] The notation sigma_0 is used for different quantities in Corollary 1 and Corollary 3; renaming one of them would avoid confusion.
[Algorithm A1] The algorithm's notation 'beta <- - Xbar/(Xbar - max(x))' is the CMM estimator; if a parametric bootstrap is intended, the algorithm should state explicitly that beta, theta, and k are re-estimated from each parametric bootstrap sample rather than fixed at the original estimates.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: characterizations come from external lemmas; tests are standard U-statistics with Monte Carlo calibration.

full rationale

The paper's derivation chain is self-contained relative to external mathematical results rather than circular. Theorem 1 instantiates the fixed-point characterizations of Betsch and Ebner (2021) to the GPD density and is independently checked by integration by parts. Theorem 3's forward direction is a direct calculation of dynamic survival extropy for the GPD, and its converse is explicitly imported from Sathar and Nair (2021), an external prior work, not from the authors' own prior results. The test statistics are U-statistics based on these characterizations; plugging consistent estimators into such statistics is standard practice and does not make the statistic equal to its own input. The Monte Carlo critical values are calibrated under simulated GPD samples, and power is evaluated against external alternative distributions, so no fitted parameter is renamed as a prediction. The skeptically noted asymptotic gap—consistency alone does not imply that the parameter-estimation term is o_p(1) after multiplying by sqrt(n)—and the bootstrap issue in Algorithm A1 (resampling the empirical distribution rather than parametric null samples) are genuine correctness or validity concerns, but they are not circularity: they do not amount to defining a result in terms of itself or to a fitted input being called a prediction. There is no self-citation used as load-bearing evidence, and no equation in the paper reduces by construction to the quantity being tested. Therefore the circularity score is 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard characterization theorems (Betsch-Ebner, Sathar-Nair) and on the consistency of the AML/CMM estimators. The only user-chosen free parameter is the AML tuning k, which is left unspecified. No new entities are introduced. The paper's own flaw is the misuse of the endpoint limit term in Theorem 1, which is an unflagged assumption rather than a free parameter.

free parameters (1)

AML tuning parameter k (number of upper order statistics) = unspecified
The AML estimator in equations (5.5)-(5.6) depends on k, which the paper leaves free ('1<=k<=n'); the choice affects beta-hat, theta-hat, and hence the test statistic and critical values, but no selection rule is given.

assumptions (4)

standard math Betsch-Ebner fixed-point characterization for semi-bounded and bounded support (Lemmas 1 and 2)
Used directly in Theorem 1 to derive the Stein-type characterization; the bounded-support version requires the endpoint limit term to be handled correctly.
domain assumption Sathar-Nair characterization: constant J_s(X;t) h_F(t) implies hazard rate h_F(t) = 1/(c1 t + c2)
The converse direction of Theorem 3 is borrowed from the proof of Theorem 8 in Sathar and Nair (2021); no independent proof is given.
domain assumption Consistency (and in practice sqrt(n)-consistency) of the AML and CMM estimators
Theorems 4-6 assume consistent estimators of beta and theta; the paper cites Villasenor-Alva and Gonzalez-Estrada (2009) for these estimators but the sqrt(n)-consistency needed for the asymptotic argument is not stated or proven.
standard math IPCW U-statistic asymptotic theory of Datta, Bandyopadhyay and Satten (2010)
Used for the censored-data test statistic in Section 4.2, including the variance estimator and martingale representation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Characterization based Goodness-of-Fit for Generalized Pareto Distribution: A Blend of Stein's Identity and Dynamic Survival Extropy." pith.science (2026). https://pith.science/paper/QFMIUVWW

@misc{pith2026250601473,
  author       = {Pith},
  title        = {Pith review of: Characterization based Goodness-of-Fit for Generalized Pareto Distribution: A Blend of Stein's Identity and Dynamic Survival Extropy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QFMIUVWW}},
  note         = {Machine review of arXiv:2506.01473}
}

read the original abstract

This paper proposes a goodness of fit test for the generalized Pareto distribution (GPD). Firstly, we provide two characterizations of GPD based on Stein's identity and dynamic survival extropy. These characterizations are used to test GPD separately for the positive and negative shape parameter cases. A Monte Carlo simulation is conducted to provide the critical values and power of the proposed test against a good number of alternatives. Our test is simple to use and it has asymptotic normality and relatively high power, which strengthened the purpose of proposing it. Considering the case of right censored data, we provide the procedure to handle censored case too. A few real-life applications are also included.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

2 extracted references

[1]

Logistic or not Logistic?

Allison, James S., Bruno Ebner, and Marius Smuts (2023). “Logistic or not Logistic?” In:Statistica Neerlandica77.4, pp. 429–443.DOI:https://doi.org/10.1111/stan.12292. Anastasiou, Andreas et al. (2023). “Stein’s method meets computational statistics: a review of some recent developments”. In:Statist. Sci.38.1, pp. 120–139.DOI:https : / / doi . org / 10 . ...

arXiv 2023
[20]

Regression Analysis with Randomly Right-Censored Data

Koul, H., V . Susarla, and J. Van Ryzin (1981). “Regression Analysis with Randomly Right-Censored Data”. In:The Annals of Statistics9.6, pp. 1276–1288.DOI:https://doi.org/10.1214/ aos/1176345644. Koul, Hira L. and Vyaghreswarudu Susarla (1980). “Testing for new better than used in expectation with incomplete data”. In:Journal of the American Statistical A...

arXiv 1981

Pith tools

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

[1] [1]

Logistic or not Logistic?

Allison, James S., Bruno Ebner, and Marius Smuts (2023). “Logistic or not Logistic?” In:Statistica Neerlandica77.4, pp. 429–443.DOI:https://doi.org/10.1111/stan.12292. Anastasiou, Andreas et al. (2023). “Stein’s method meets computational statistics: a review of some recent developments”. In:Statist. Sci.38.1, pp. 120–139.DOI:https : / / doi . org / 10 . ...

arXiv 2023

[2] [20]

Regression Analysis with Randomly Right-Censored Data

Koul, H., V . Susarla, and J. Van Ryzin (1981). “Regression Analysis with Randomly Right-Censored Data”. In:The Annals of Statistics9.6, pp. 1276–1288.DOI:https://doi.org/10.1214/ aos/1176345644. Koul, Hira L. and Vyaghreswarudu Susarla (1980). “Testing for new better than used in expectation with incomplete data”. In:Journal of the American Statistical A...

arXiv 1981