{"id":"d1c95153-e5e7-469a-8fa2-0d160e2abc0b","arxiv_id":"2411.14926","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Subadditivity of the distribution of 1/X is sufficient for X to be stochastically dominated by any convex combination of its independent copies.","lead":"Researchers identify a broad new sufficient condition, subadditivity of the inverted distribution, under which a convex combination of independent copies of a heavy-tailed random variable stochastically dominates the original variable. The condition extends earlier 'super-heavy-tailed' and 'super-Pareto' classes and, for the first time, covers some discrete distributions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's induction step drops a nonpositive term from a lower bound as printed, but the displayed bound can be rearranged to give the result; the gap is real yet repairable.","rationale":"The reader's weakest_assumption identifies exactly the induction-step inference in Theorem 3.1: a nonpositive term is dropped from a lower bound, so the printed conclusion does not follow. My analysis confirms that this is the most load-bearing concern in the paper, since Theorem 3.1 is the central claim and the final inference in its proof is invalid as written. However, the concern does not sink the theorem. The displayed lower bound can be rearranged to show P ≥ c directly: the expression b + a(1−c)+c(c−b) simplifies to c + (a+b−c)(1−c), and InvSub gives a+b ≥ c, so the extra term is nonnegative. Thus the proof is repairable by a short algebraic correction. The base case has a similar display issue but also appears repairable; the strict-inclusion examples are stated without proof, which is a secondary weakness. Since the reader's CONDITIONAL verdict already requires correcting the proof before acceptance, my read does not change that verdict. I therefore recommend UNCHANGED, meaning the paper should remain conditional on completing the proof and examples.","tokens_in":9560,"tokens_out":15231,"duration_ms":133653,"concrete_test":"Re-derive the induction step algebraically: verify that the lower bound b + a(1−c) + c(c−b) equals c + (a+b−c)(1−c). If this identity holds, then InvSub (a+b ≥ c) immediately yields P ≥ c, confirming the theorem despite the misstep in the printed final inference.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the induction step of Theorem 3.1 the authors obtain P ≥ b + a(1−c) + c(c−b), with a = Fbar(x/(1−θn)), b = Fbar(x/θn), c = Fbar(x). They rewrite this as b + a + c(c−b−a) and, because InvSub (2) gives b+a ≥ c, conclude P ≥ b+a ≥ c. But c(c−b−a) ≤ 0, so b+a+c(c−b−a) ≤ b+a; the conclusion does not follow as written. This is a genuine algebraic slip in a load-bearing step. However, the same lower bound equals c + (a+b−c)(1−c), which is ≥ c since InvSub gives a+b ≥ c and c ≤ 1. Thus the theorem's inductive claim survives a one-line algebraic correction; the gap is not fatal to the central result. The n=2 base case contains a similar display where a lower bound on a subinterval is used where an upper bound is needed, but direct integral evaluation for the InvSub examples still supports the base case. A secondary weakness is that the strict-inclusion examples in §4.2 and §4.3 are asserted rather than proved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":9720,"tokens_out":19979,"duration_ms":163325,"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":[{"comment":"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.","section":"Theorem 3.1, induction step (§3, displayed inequalities after Figure 2)"},{"comment":"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.","section":"Theorem 3.1, base case (§3, first display)"},{"comment":"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.","section":"Examples 4.13 and 4.16 (§4.2, §4.3)"}],"minor_comments":[{"comment":"The title contains a typo ('do minate' instead of 'dominate') and the abstract uses 'heavy tailed' without a hyphen; these should be corrected.","section":"Title and Abstract"},{"comment":"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.","section":"Example 2.6 (§2)"},{"comment":"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.","section":"Example 2.4 (§2)"},{"comment":"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":"Theorem 2.8 (§2)"},{"comment":"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.","section":"Section 5 (§5)"},{"comment":"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.","section":"Proposition 3.2 (§3)"}],"recommendation":"major_revision","confidential_remarks":"The central result is likely correct and the new class is interesting, but the proof of Theorem 3.1 has two genuine gaps: an invalid inference in the induction step and a sign error in the base case. Both appear repairable with the algebra indicated in the report. The paper would be acceptable after those proofs are corrected and the inclusion examples are substantiated. The scope is appropriate for a probability journal; the novelty is moderate and builds on very recent work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nWhat needs fixing: the induction step in Theorem 3.1. As printed, the line\n\nP ≥ b + a(1−c) + c(c−b) ≥ b + a + c(c−b−a) ≥ b + a ≥ c\n\nuses 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.\n\nSecond 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.\n\nCitation pattern looks fine; the self-cited Arab et al. lemma is standard and used only as a tool. No fitted parameters or circularity.\n\nBottom 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.","headline":"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.","tokens_in":10330,"tokens_out":4435,"would_cite":false,"duration_ms":39920,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60E15","91G70","62P05"],"pacs":[],"model":"deepseek-v4-flash","headline":"A random variable whose inverted distribution is subadditive is stochastically dominated by any convex combination of its copies.","keywords":["stochastic dominance","convex combination","inverted distribution","subadditivity","heavy-tailed distributions","infinite mean","odds function","convex transform order"],"falsifier":"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.","tokens_in":9283,"feed_emoji":"📈","tokens_out":6755,"duration_ms":63528,"temperature":0.7,"pith_summary":"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.","feed_headline":"New heavy-tailed class: sample mean beats a single draw","feed_subtitle":"Subadditivity of the inverted distribution makes every convex combination stochastically larger than the parent.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Introduces the super-Pareto class and proves the original stochastic-dominance result for it, the statement this paper generalises.","marker":"Chen et al. (2024)"},{"why":"Defines super-heavy-tailed distributions and proves the dominance result for them; Theorem 4.12 shows that class is contained in InvSub.","marker":"Chen and Shneer (2024)"},{"why":"Establishes the super-Fréchet and super-Cauchy dominance results and the inclusion relations that serve as comparison benchmarks in Section 4.3.","marker":"Müller (2024)"},{"why":"Supplies the standard definitions of stochastic dominance, NWU/NBU classes, and convex order used throughout the paper.","marker":"Shaked and Shanthikumar (2007)"},{"why":"Introduces the decreasing-odds-rate family, used to identify super-Pareto laws as DOR in the discussion around Proposition 4.3.","marker":"Lando et al. (2023)"},{"why":"Provides the anti-star-shaped inverse property used in the proof of the closure Theorem 2.8.","marker":"Arab et al. (2024)"}],"fun_headline_variants":["Sample mean stochastically beats a single draw for new heavy-tailed class","Convex combinations of iid copies dominate the parent for InvSub class","Subadditivity of inverted CDF gives stochastic dominance by convex sums","New heavy-tailed distributions: sample mean stochastically larger than X"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Sample mean stochastically beats a single draw for new heavy-tailed class","Convex combinations of iid copies dominate the parent for InvSub class","Subadditivity of inverted CDF gives stochastic dominance by convex sums","New heavy-tailed distributions: sample mean stochastically larger than X"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000636,"raw_usage":{"total_tokens":2873,"prompt_tokens":824,"completion_tokens":2049,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":1981}},"tokens_in":440,"tokens_out":2049,"duration_ms":13984,"temperature":1.0,"reasoning_tokens":1981,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:43:36.148055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}