Convex combinations of random variables stochastically dominate the parent for a new class of heavy-tailed distributions

Idir Arab; Paulo Eduardo Oliveira; Tommaso Lando

REVIEW 3 major objections 6 minor 8 references

Convex combinations of random variables stochastically dominate the parent for a new class of heavy-tailed distributions

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A random variable whose inverted distribution is subadditive is stochastically dominated by any convex combination of its copies.

desk verdict InvSub gives a genuinely broader sufficient condition for convex-combination stochastic dominance, but the main proof has a repairable algebraic slip that must be fixed before the paper is citable. read the letter →

arxiv 2411.14926 v2 pith:EOFSMZCC submitted 2024-11-22 math.PR

Idir Arab , Tommaso Lando , Paulo Eduardo Oliveira This is my paper · ORCID

classification math.PR MSC 60E1591G7062P05

keywords stochasticdominanceconvexcombinationinverteddistributionsubadditivityheavy-taileddistributionsinfinitemeanoddsfunctiontransformorder

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 tries to establish that a random variable can be stochastically dominated by a convex combination of its independent copies under a new, relatively mild condition called InvSub: the cumulative distribution function of 1/X is subadditive. If true, this means that for such heavy-tailed variables, the sample mean is stochastically larger than a single observation, even though the mean is infinite. The authors show that the InvSub class contains the earlier super-heavy-tailed family and all nonnegative super-Pareto laws, and that it is the first of these classes to include discrete distributions. A consequence is that the surprising dominance phenomenon is not tied to absolute continuity or to concave odds functions.

What carries the argument

The key object is the Inverted-subadditive (InvSub) class. A nonnegative X with FX(0) = 0 is InvSub when F−1/X(x) = 1 − FX(1/x) is subadditive; Lemma 2.3 rewrites this as FX(x/θ) + FX(x/(1−θ)) ≤ FX(x) + 1 for all θ ∈ (0,1). That inequality is the engine of the proof: it gives the bound used in the n = 2 case and in the induction step of Theorem 3.1, and it makes the family closed under star-shaped transformations. The inverted-subadditive order (Definition 5.1) recasts the same property as a benchmark comparison with Pareto, FX ≤i−sb P, and with Fréchet for super-heavy-tailed laws.

What would settle it

A direct check of the n = 2 case for the discrete InvSub law of Example 2.7 would settle the claim: compute P(θX1 + (1−θ)X2 > x) at a jump point x and a weight θ such as θ = 1/3, and see whether it is at least F̄X(x); if it is not, Theorem 3.1 fails in the discrete setting.

Watch

Extended reading notes

Core claim

The central claim is Theorem 3.1: if X is InvSub, then for any n and any weights θ1 + ... + θn = 1, X ≤st θ1X1 + ... + θnXn. The proof conditions on one variable and bounds the survival integral of the convex combination; the InvSub inequality (2) ensures the bound reaches the parent survival function. A corollary via Proposition 3.2 is that any non-degenerate X satisfying the dominance must have infinite mean, and InvSub distributions are therefore heavy-tailed. The paper also proves that InvSub is closed under continuous star-shaped transformations (Theorem 2.8), admits a hazard-rate sufficient condition (Proposition 2.5), and contains discrete examples, so the class is broader than earlier sufficiency classes.

Load-bearing premise

The load-bearing premise is that the inverted-subadditivity inequality (2) is strong enough to make the integral lower bound in the induction step of Theorem 3.1 produce the desired survival bound; if that step fails, the theorem is not established by the proof as written.

Editorial extensions

If this is right

If X is InvSub, then every fixed convex combination of independent copies, including the ordinary sample mean, stochastically dominates a single observation.
InvSub implies infinite mean, so the dominance phenomenon cannot occur for finite-mean variables, consistent with the convex-order argument.
The super-heavy-tailed class of Chen and Shneer is contained in InvSub, so the new theorem subsumes that earlier dominance result.
Nonnegative super-Pareto laws are InvSub, and some distributions with non-concave, non-convex odds functions also satisfy the condition, widening the range of examples.
Unlike earlier classes, InvSub admits discrete distributions, so the stochastic dominance result is not an artifact of absolute continuity.

Reading between the lines

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

Editorial inference: the InvSub inequality (2) is checkable directly from tail quantiles, so the class is easy to test on empirical data without estimating densities.
Editorial inference: because InvSub is closed under continuous star-shaped transformations, new examples can be built by applying such transformations to known base laws such as the Pareto or the geometric-type law of Example 2.7.
Editorial inference: the fact that super-Cauchy distributions are not contained in InvSub suggests there may be a broader common condition containing both families; identifying it would unify the currently disjoint sufficiency classes.

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

Short version: the paper gets the direction right. The InvSub class (subadditivity of the inverted CDF) is a real, reasonably mild sufficient condition for X ≤st Σθ_i X_i, and the inclusions the authors prove are useful: nonnegative super-Pareto and super-heavy-tailed both fall inside InvSub, while the discrete InvSub example is a genuine extension over the earlier continuous-only classes. I agree with your reader that this is a broadening of Chen et al. 2024, Chen-Shneer 2024, and Müller 2024 rather than a brand-new phenomenon, but the broadening looks substantial, not cosmetic.

What needs fixing: the induction step in Theorem 3.1. As printed, the line

P ≥ b + a(1−c) + c(c−b) ≥ b + a + c(c−b−a) ≥ b + a ≥ c

uses c(c−b−a) as if it were nonnegative; InvSub actually gives c ≤ a+b, so that term is ≤ 0. Dropping it is invalid. The stress test is right that a one-line rearrangement fixes it: the expression equals c + (a+b−c)(1−c), which is ≥ c under InvSub because a+b ≥ c and c ≤ 1. So the theorem survives, but only after that correction. The same caution applies to the n=2 base case display, which claims an upper bound by a similar route; that part is recoverable but should be rewritten.

Second soft spot: the examples proving strict inclusion (4.13 and 4.16, also 2.6) are asserted with "it can be verified" rather than proved. For a paper whose main contribution is a class definition and its boundaries, these checks should be written out. They are probably true, but "probably" is not a proof.

Citation pattern looks fine; the self-cited Arab et al. lemma is standard and used only as a tool. No fitted parameters or circularity.

Bottom line: this deserves a serious referee. With the induction step corrected and the examples completed, it should be acceptable. As is, I wouldn't cite it yet. If you have a reading group that likes spotting repairable gaps, it's a good candidate; otherwise wait for the revision.

Referee Report

3 major / 6 minor

Summary. The paper introduces a class of distributions called InvSub, defined by subadditivity of the cumulative distribution function of the reciprocal (i.e. of the inverted distribution), and claims that if a random variable X is InvSub then any convex combination of independent copies of X stochastically dominates X. The authors show that this class contains the previously studied super-heavy-tailed class and, under suitable transformations, the super-Pareto class, while being incomparable with the super-Cauchy class, and they provide a characterization via a new 'inverted-subadditive order'. They also prove that the stochastic dominance property forces infinite mean, consistent with earlier results.

Significance. If the main result is correct, the paper gives a substantially broader and conceptually simple sufficient condition for the counterintuitive stochastic dominance of a convex combination over its parent, unifying several recent results and including discrete distributions for the first time. The InvSub condition is mild, comes with a useful closure property under star-shaped transformations, and yields a clean characterization. The paper does not ship machine-checked proofs or code, and several supporting examples are asserted rather than proved, but the central theorem is likely correct and repairable; the present proof, however, is not complete as written.

major comments (3)

[Theorem 3.1, induction step (§3, displayed inequalities after Figure 2)] The displayed inference that the lower bound is at least Fbar(x/theta_n)+Fbar(x/(1-theta_n)) is invalid: after rewriting the bound as b + a(1-c) + c(c-b) with a=Fbar(x/(1-theta_n)), b=Fbar(x/theta_n), c=Fbar(x), the term c(c-b-a) is nonpositive under condition (2), so dropping it does not give a lower bound. The gap is repairable: the same expression equals c^2 + (a+b)(1-c), which is at least c because (2) gives a+b >= c and c <= 1. The proof should be corrected accordingly.
[Theorem 3.1, base case (§3, first display)] The expression Fbar(x/(1-theta))Fbar(x) + Fbar(x)(Fbar(x/theta)-Fbar(x)) is a lower bound for the integral, not an upper bound: on [0,x] the integrand is bounded below by Fbar(x/(1-theta)) and on [x,x/theta] it is bounded below by Fbar(x). Therefore the chain of inequalities claiming an upper bound for the integral and concluding P >= Fbar(x) for n=2 is not justified as written. The case n=2 requires a separate correct argument.
[Examples 4.13 and 4.16 (§4.2, §4.3)] The strict inclusions are asserted by phrases such as 'it can be verified' and 'it is easily verified' without giving the verification. Since these examples are used to support the paper's claim that the InvSub class strictly enlarges the super-heavy-tailed family and is not contained in the super-Cauchy family, the relevant inequalities (2) and (3), and the claimed failure of convexity/concavity of C^{-1} o V, should be proved explicitly or accompanied by a reproducible computation.

minor comments (6)

[Title and Abstract] The title contains a typo ('do minate' instead of 'dominate') and the abstract uses 'heavy tailed' without a hyphen; these should be corrected.
[Example 2.6 (§2)] The statement that the distributions Y_b satisfy the hazard-rate condition for b <= 0.7 is not substantiated; the derivative calculation that verifies Proposition 2.5 should be included or sketched.
[Example 2.4 (§2)] The claim that the odds function of the Frechet distribution is convex is stated without proof; a one-line verification or a reference would be helpful.
[Theorem 2.8 (§2)] The proof uses the inverse h^{-1}, but a continuous star-shaped function is not necessarily strictly increasing; the hypothesis should be strengthened to strictly increasing or the argument should use a generalized inverse.
[Section 5 (§5)] The new inverted-subadditive order is stated and used transitively, but transitivity is not proved; a short proof or a reference would improve the presentation.
[Proposition 3.2 (§3)] The argument that X <=_{st} Y together with equal finite means implies equality in distribution is standard, but it would be useful to state this fact explicitly for readers.

Circularity Check

0 steps flagged · score 1.0 of 10

No circularity: InvSub is an independent sufficient condition; the only self-citation is minor and not load-bearing.

full rationale

The central claim is not circular. InvSub (Definition 2.2) is defined as subadditivity of the inverted CDF, a condition that does not mention convex combinations or stochastic dominance; the target relation (1) is a separate universal statement over weights. Theorem 3.1 derives it using Lemma 2.3 (an algebraic rewriting), direct CDF conditioning, and induction on n; no parameter is fitted to the conclusion and no normalization is imported from the target inequality. The one self-citation in the derivation chain is in the proof of Theorem 2.8, where the paper invokes 'Lemma 4.1 in Arab et al. (2024)' for the elementary fact that the inverse of a star-shaped function is anti-star-shaped. That lemma is parameter-free, does not involve stochastic dominance, and is not the paper's main theorem, so it is a minor self-citation rather than load-bearing circularity. The notable caveat in the proof is a correctness issue, not a circularity: in the induction step of Theorem 3.1 the displayed lower bound is rearranged into a form with a nonpositive correction term and then that term is dropped, so the printed inference does not follow as written; this is an algebraic gap, not an equivalence between the InvSub assumption and the stochastic-dominance conclusion. Accordingly, the circularity score is low.

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

The central claim rests on the definition of InvSub, standard subadditivity facts for anti-star-shaped functions, and the restriction to nonnegative variables. No numerical parameters are fitted. The proof gap in Theorem 3.1 is a derivation issue rather than an axiom; it is recorded in red flags.

assumptions (4)

domain assumption X is nonnegative with F_X(0)=0, that is, support in [0, infinity] with possible mass at infinity
Definition 2.2 restricts InvSub to such X; Theorem 3.1's proof uses nonnegativity in the integral decomposition and the survival function.
standard math For continuous star-shaped h, h^{-1} is anti-star-shaped; anti-star-shaped functions vanishing at 0 are subadditive
Used in Proposition 2.5 and Theorem 2.8; the inverse property is cited from Arab et al. (2024) rather than proved.
domain assumption Independence of X1,...,Xn and weights summing to one
Statement of Theorem 3.1 requires i.i.d. copies; the proof conditions on X_n and uses induction.
standard math For negative a and b, exp(a+b) is at least exp(a)+exp(b)-1
Used in the proof of Theorem 4.12 to pass from super-heavy-tailed to InvSub.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Convex combinations of random variables stochastically dominate the parent for a new class of heavy-tailed distributions." pith.science (2026). https://pith.science/paper/EOFSMZCC

@misc{pith2026241114926,
  author       = {Pith},
  title        = {Pith review of: Convex combinations of random variables stochastically dominate the parent for a new class of heavy-tailed distributions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EOFSMZCC}},
  note         = {Machine review of arXiv:2411.14926}
}

read the original abstract

Stochastic dominance of a random variable by a convex combination of its independent copies has recently been shown to hold within the relatively narrow class of distributions with concave odds function, and later extended to broader families of distributions. A simple consequence of this surprising result is that the sample mean can be stochastically larger than the underlying random variable. We show that a key property for this stochastic dominance result to hold is the subadditivity of the cumulative distribution function of the reciprocal of the random variable of interest, referred to as the inverted distribution. By studying relations and inclusions between the different classes for which the stochastic dominance was proved to hold, we show that our new class can significantly enlarge the applicability of the result, providing a relatively mild sufficient condition.

Figures

Figures reproduced from arXiv: 2411.14926 by the authors.

**Figure 2.** Lower bound for the integral in the induction step. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

8 extracted references · 4 canonical work pages

[2]

Technical note—an unexp ected stochastic domi- nance: Pareto distributions, dependence, and diversiﬁcation. Op er. Res. Articles in Advance. doi:10.1287/opre.2022.0505. Chen, Y., Shneer, S.,

work page arXiv 2022
[3]

doi:10.48550/arXiv.2408.15033, arXiv:2408.15033

Stochastic dominance for super heavy -tailed random variables. doi:10.48550/arXiv.2408.15033, arXiv:2408.15033. 11 Lando, T., Arab, I., Oliveira, P.E.,

arXiv doi:10.48550/arxiv.2408.15033
[6]

doi:10.48550/arXiv.2411.10139, arXiv:2411.10139

Some remarks on the eﬀect of risk sharing and dive rsiﬁcation for inﬁnite mean risks. doi:10.48550/arXiv.2411.10139, arXiv:2411.10139. Pomatto, L., Strack, P., Tamuz, O.,

work page doi:10.48550/arxiv.2411.10139
[8]

Springe r Series in Statistics, Springer, New York

Stochastic orders. Springe r Series in Statistics, Springer, New York. doi: 10.1007/978-0-387-34675-5 . 12

doi:10.1007/978-0-387-34675-5
[2007]

Springer Series in Sta tistics, Springer, New York

Life distributions. Springer Series in Sta tistics, Springer, New York. doi:10.1007/978-0-387-68477-2 . M¨ uller, A.,

doi:10.1007/978-0-387-68477-2
[2020]

Journal of Political Economy 128, 1877–1900

Stochastic dominance un der independent noise. Journal of Political Economy 128, 1877–1900. doi: 10.1086/705555. Shaked, M., Shanthikumar, J.G.,

work page doi:10.1086/705555 1900
[2023]

Nonparametric inference ab out increasing odds rate distribu- tions. J. Nonparametr. Stat. 36, 435–454. doi: 10.1080/10485252.2023.2220050. Marshall, A.W., Olkin, I.,

arXiv 2023
[2024]

Inequalities and bounds for expected order statistics from transform-ordered families

Inequalities and bounds for e xpected order statistics from transform-ordered families. arXiv 2403.03802 [stat.ME]. doi: 10.48550/arXiv.2403.03802. Chen, Y., Embrechts, P., Wang, R.,

work page Pith review arXiv doi:10.48550/arxiv.2403.03802

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [2]

Technical note—an unexp ected stochastic domi- nance: Pareto distributions, dependence, and diversiﬁcation. Op er. Res. Articles in Advance. doi:10.1287/opre.2022.0505. Chen, Y., Shneer, S.,

work page arXiv 2022

[2] [3]

doi:10.48550/arXiv.2408.15033, arXiv:2408.15033

Stochastic dominance for super heavy -tailed random variables. doi:10.48550/arXiv.2408.15033, arXiv:2408.15033. 11 Lando, T., Arab, I., Oliveira, P.E.,

arXiv doi:10.48550/arxiv.2408.15033

[3] [6]

doi:10.48550/arXiv.2411.10139, arXiv:2411.10139

Some remarks on the eﬀect of risk sharing and dive rsiﬁcation for inﬁnite mean risks. doi:10.48550/arXiv.2411.10139, arXiv:2411.10139. Pomatto, L., Strack, P., Tamuz, O.,

work page doi:10.48550/arxiv.2411.10139

[4] [8]

Springe r Series in Statistics, Springer, New York

Stochastic orders. Springe r Series in Statistics, Springer, New York. doi: 10.1007/978-0-387-34675-5 . 12

doi:10.1007/978-0-387-34675-5

[5] [2007]

Springer Series in Sta tistics, Springer, New York

Life distributions. Springer Series in Sta tistics, Springer, New York. doi:10.1007/978-0-387-68477-2 . M¨ uller, A.,

doi:10.1007/978-0-387-68477-2

[6] [2020]

Journal of Political Economy 128, 1877–1900

Stochastic dominance un der independent noise. Journal of Political Economy 128, 1877–1900. doi: 10.1086/705555. Shaked, M., Shanthikumar, J.G.,

work page doi:10.1086/705555 1900

[7] [2023]

Nonparametric inference ab out increasing odds rate distribu- tions. J. Nonparametr. Stat. 36, 435–454. doi: 10.1080/10485252.2023.2220050. Marshall, A.W., Olkin, I.,

arXiv 2023

[8] [2024]

Inequalities and bounds for expected order statistics from transform-ordered families

Inequalities and bounds for e xpected order statistics from transform-ordered families. arXiv 2403.03802 [stat.ME]. doi: 10.48550/arXiv.2403.03802. Chen, Y., Embrechts, P., Wang, R.,

work page Pith review arXiv doi:10.48550/arxiv.2403.03802