{"id":"3ea6bf6a-4102-49a1-9484-48269a1e5c84","arxiv_id":"2411.15401","paper_version":4,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For fourth-order and higher stochastic dominance, whether one random variable dominates another depends on the chosen reference interval; for third-order and below it does not.","lead":"This paper studies whether the answer to 'does one risky gamble dominate another' depends on the range of possible outcomes you assume. It proves that the dominance ranking can flip when you widen or narrow the range, but only for dominance orders of four or higher. A smart generalist might read it to see why supposedly preference-free comparisons of risk can secretly depend on a modeler's choice of interval.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1 is false as stated (the factor n makes the limit diverge for constant Z), and Proposition 1's proof relies on it; the main theorems appear correct after the standard correction, so acceptance should be conditional on fixing this step.","rationale":"The reader's weakest-assumption note about Jean's boundary conditions is a legitimate scope caveat, but it is not the most pressing defect. The paper contains a concrete false mathematical statement in Lemma 1, and that lemma is used in the proof of Proposition 1, which in turn supports the n ≤ 3 half of Theorem 1. I independently checked the central constructions: Example 1, Lemma 2, the scaling argument in Theorem 1, and the cone argument in Theorem 2 all appear sound, and the corrected Lemma 1 is a standard asymptotic that restores the intended proof. So the mathematical headline of the paper is very likely correct, but the submitted proof is not formally valid as written. A conditional acceptance with a mandatory correction of Lemma 1 and the associated M-boundary argument is the appropriate outcome rather than an unconditional accept.","tokens_in":12876,"tokens_out":50216,"duration_ms":429163,"concrete_test":"Evaluate the expression in Lemma 1 for Z ≡ 1 and n = 2: the limit is 2, not E[Z] = 1, confirming the lemma is false as printed. Then, after replacing Lemma 1 with the correct identity lim_{η→∞} η − (E[(η−Z)^n_+])^{1/n} = E[Z], re-check Proposition 1 Steps (a) and (b): in particular, prove from the expansion F^{[k]}_Z(η) = η^{k−1}/(k−1)! − E[Z] η^{k−2}/(k−2)! + O(η^{k−3}) that E[X] > E[Y] forces F^[k]_X(M) ≤ F^[k]_Y(M) for all k ≤ n−1 at sufficiently large M; if this asymptotic step goes through, the central theorems are unaffected.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Lemma 1 states lim_{η→∞} n(η − (E[(η−Z)^n_+])^{1/n}) = E[Z]. For Z ≡ 1 the expression is n(η − (η−1)) = n, so the displayed limit is n, not 1; for n > 1 the limit does not even exist. The standard correct identity is lim_{η→∞} η − (E[(η−Z)^n_+])^{1/n} = E[Z]. Proposition 1 Step (a) invokes Lemma 1 to infer E[X] ≥ E[Y] from X ≥_3 Y, and Step (b) invokes it again to assert the existence of M ≥ b with F^[k]_X(M) ≤ F^[k]_Y(M) for all k ∈ [n−1] when E[X] > E[Y]. Both conclusions are true and follow from the polynomial asymptotics of F^{[k]}_Z(η), but they do not follow from the lemma as printed. Since the n ≤ 3 consistency is the 'if' half of Theorem 1, the submitted proof has a genuine gap at a load-bearing point. The gap is easily repaired, but the manuscript should not be accepted without correcting Lemma 1 and supplying the missing asymptotic argument.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares two formulations of higher-order stochastic dominance for bounded random variables: the whole-line relation nSDR and Jean's interval-based relation nSD[a,b], which adds boundary conditions at b. Proposition 1 provides utility-class characterizations and asserts that nSDR implies nSD[a,b] for all pairs supported in [a,b] exactly when n ≤ 3. Theorem 1 shows that enlarging the right endpoint of the reference interval preserves rankings for all such pairs exactly when n ≤ 3, while for n ≥ 4 there exist pairs supported inside the smaller interval that dominate on the larger interval but fail on the smaller one. Theorem 2 extends the threshold to nth-degree m-mean preserving stochastic dominance: consistency holds exactly when n − m ≤ 3. Example 1 gives an explicit 4SD counterexample with distributions supported in [0,1] and reference intervals [0,1] and [0,2].","tokens_in":13157,"tokens_out":18819,"duration_ms":177286,"significance":"If correct, the results give a clean quantitative answer to a question that has circulated in the literature: the reference interval only matters for stochastic dominance of order at least four, and for mean-preserving variants only when n − m ≥ 4. This is economically relevant because fourth-order dominance (temperance) is now routinely used, and the paper shows concretely that rankings then depend on a subjective modeling choice. The paper also contains explicit, checkable counterexamples and a self-contained proof strategy built on known utility characterizations. The main weakness is a misstated asymptotic lemma in the proof of Proposition 1; the gap is real but appears routine to repair, so the central claims remain plausible and worth publishing after revision.","major_comments":[{"comment":"Lemma 1 is false as stated. For Z ≡ 1 and any η > 1, E[(η−Z)^n_+]^{1/n} = η−1, so the displayed expression equals n, not E[Z] = 1; the correct identity is either lim_{η→∞} (η − (E[(η−Z)^n_+])^{1/n}) = E[Z], or the displayed statement with right-hand side nE[Z]. Because Proposition 1, Step (a) and Step (b), both invoke Lemma 1 to infer E[X] ≥ E[Y] from X ≥_n Y, the proof of Proposition 1 currently rests on an incorrect statement. The inference itself is true, but the manuscript must replace Lemma 1 with the correct identity and adjust the two uses accordingly.","section":"Section 3.1, Lemma 1"},{"comment":"The assertion that 'Using Lemma 1 again and noting that E[X] > E[Y], there exists M ≥ b such that F^{[k]}_X(M) ≤ F^{[k]}_Y(M) for all k ∈ [n−1]' does not follow from the first-order limit in Lemma 1. What is needed is the polynomial expansion F^{[k]}_Z(η) = η^{k−1}/(k−1)! − E[Z]η^{k−2}/(k−2)! + O(η^{k−3}) for η beyond the supports; with E[X] > E[Y] this makes each difference negative for sufficiently large η. This assertion is true but must be proved, otherwise the integration-by-parts proof of (iv) ⇒ (v) is incomplete.","section":"Section 3.1, Proof of Proposition 1, Step (b)"}],"minor_comments":[{"comment":"The sentence 'we study whether one dominates the other in higher-order stochastic dominance depends on the reference interval' is ungrammatical and should be rewritten.","section":"Abstract"},{"comment":"The text 'the equivalence between (i) and (iii) in Proposition 1 holds for n ≤ 3' appears to be a misreference; it should refer to the equivalence (iv) ⇔ (i).","section":"Section 3.2, Proof of Theorem 1"},{"comment":"The displayed computation of the second-moment ratio is compressed: the denominator '45mn' appears after an implicit division by ε_n, and the intermediate algebra is hard to follow. Please add the missing steps or parentheses.","section":"Section 3.2, Lemma 2"},{"comment":"The long integration-by-parts display uses nonstandard notation such as '(−1)^{-1}u(η)(F_Y − F_X)|...' and omits some differentials; rewriting this display with explicit integration measures would improve clarity.","section":"Section 3.1, Step (b)"},{"comment":"The reduction from Lemma 2's interval [0,9] to a general b should mention the affine rescaling argument; as written, 'the existence is due to Lemma 2' skips a step, even though the rescaling is straightforward.","section":"Section 3.2, Proof of Theorem 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely correct after a routine repair of Lemma 1 and the Step (b) asymptotic argument; the counterexample arithmetic in Example 1 and the structure of Lemma 2 and Theorem 2 are consistent. The main editorial caution is that Lemma 1 is explicitly attributed to Proposition 6 of Ogryczak and Ruszczynski (2001); the authors should verify the original statement and correct the factor n, since a false citation in a load-bearing lemma is exactly the kind of issue that future readers will trip over. The paper's scope and contribution are appropriate for math.PR."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Start with the punchline: the main results of this paper are correct, but there is a real error in the proof of Proposition 1 that has to be fixed before publication. The stress-test note is right. Lemma 1 as stated has a factor n that makes the limit diverge for a constant Z. For Z=1, the expression is n(eta - (eta-1)) = n, not E[Z]=1. The standard identity drops the factor n. That matters because Step (a) and Step (b) of Proposition 1 both rely on this lemma. The conclusions drawn there - that X >=_3 Y implies E[X] >= E[Y], and that when E[X] > E[Y] one can find M with all iterated integrals ordered - are true, and they follow from polynomial asymptotics of the iterated integrals as eta -> infinity. But that argument is not supplied. So the 'if' half of Theorem 1 has a genuine gap, even though it is an easily repairable one.\n\nWhat is genuinely new: the sharp thresholds. Theorem 1 gives the if-and-only-if n <= 3 for consistency between nSD[a,b] and nSDR, and Theorem 2 extends this to nth-degree m-mean preserving orders with the n-m <= 3 threshold. Example 1 is a clean explicit counterexample showing a 4SD ranking that flips when the reference interval changes from [0,1] to [0,2]; the boundary-violation arithmetic (341/72900) checks out. The attribution to Fang and Post is accurate, and the paper fills a gap they left open.\n\nThe proofs are mostly self-contained and careful. Lemma 2's construction and the cone-contradiction argument in Theorem 2 are legitimate. One minor caveat worth noting, not a flaw: the boundary conditions in Definition 2 are Jean's, so the result is about that formulation rather than about reference-interval dependence in the abstract; the paper is transparent about this.\n\nBottom line: the central argument holds up after a standard correction. This deserves a serious referee. Send it out, but the authors should be required to fix Lemma 1 and supply the missing asymptotic argument in Proposition 1 before acceptance.","headline":"Main results are right, but Lemma 1 as printed is false and Proposition 1 leans on it; the paper deserves review with a mandatory fix.","tokens_in":608,"tokens_out":884,"would_cite":true,"duration_ms":29992,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60E15","91B16"],"pacs":[],"model":"deepseek-v4-flash","headline":"Stochastic dominance rankings depend on the reference interval precisely when the order is at least four.","keywords":["higher-order stochastic dominance","reference interval","mean-preserving stochastic dominance","prudence","temperance","expected utility","boundary conditions","risk attitudes"],"falsifier":"Compute for Example 1 whether $X \\ge^{[0,c]}_4 Y$ holds as $c$ varies between 1 and 2; the theorem predicts it holds at $c=2$ and fails at $c=1$. If it held for every $c \\in [1,2]$, Theorem 1 would be false. A pair of variables supported in $[0,1]$ with $X \\ge_4 Y$ and $X \\ge^{[0,c]}_4 Y$ for all $c \\in (0,1]$ would also refute the theorem.","tokens_in":12679,"feed_emoji":"📉","tokens_out":9752,"duration_ms":85616,"temperature":0.7,"pith_summary":"The paper asks whether the statement \"X dominates Y in nth-order stochastic dominance\" is well-defined once the random variables are known to live in a bounded interval, or whether it secretly depends on which reference interval the modeler draws around that support. The answer is a sharp cutoff: for n ≤ 3 the two standard formulations — whole-line $n$SDR and interval-based $n$SD$_{[a,b]}$ — always rank bounded variables identically, while for n ≥ 4 shrinking the interval makes the dominance relation strictly harder to satisfy. The same phenomenon extends to mean-preserving versions, where interval-dependence appears exactly when $n-m \\ge 4$. If correct, rankings under fourth-order dominance (the \"temperance\" order) and higher are not robust to a subjective modeling choice, which matters for portfolio choice, precautionary-saving comparisons, and other applications built on higher-order risk preferences.","feed_headline":"Reference interval flips fourth-order dominance","feed_subtitle":"Whole-line and interval-based dominance agree only through order three; above that, rankings hinge on the chosen upper bound.","key_machinery":"The machinery is the iterated distribution function $F^{[n]}(\\eta)=\\int_{-\\infty}^{\\eta}F^{[n-1]}(\\xi)\\,d\\xi$, which turns dominance into pointwise inequalities of $E[(\\eta-X)_+^{n-1}]$ over all of $\\mathbb{R}$ or over a compact interval $[a,b]$. The interval version (Jean 1980) adds boundary conditions $F^{[k]}_X(b) \\le F^{[k]}_Y(b)$ for every $k \\in [n]$, equivalently $E[(b-X)^{k-1}] \\le E[(b-Y)^{k-1}]$. The load-bearing identity is $\\lim_{\\eta \\to \\infty}(\\eta - (E[(\\eta-Z)_+^n])^{1/n}) = E[Z]$, which converts whole-line dominance into a mean comparison; this is why at $n=3$ the only extra boundary condition (the mean) is automatic, and why at $n=4$ the second-moment boundary condition is not implied. Example 1 and the scaling construction behind Lemma 2 then produce four-point distributions that flip ranking between intervals.","core_discovery":"The central claim is Theorem 1: for random variables supported in $[a,b]$, $n$SD$_{[a,c]}$ implies $n$SD$_{[a,d]}$ whenever $c<d$, but the converse implication holds for all such variables if and only if $n \\le 3$. For $n \\ge 4$ there exist $X,Y$ supported inside $[a,b]$ such that $X$ dominates $Y$ over the larger interval $[a,d]$ but not over the smaller $[a,c]$, so the ranking depends on the right endpoint even when both variables stay inside the intersection of the two intervals. Theorem 2 generalizes the same cutoff to $n$th degree $m$-mean preserving stochastic dominance, where interval-independence holds exactly when $n-m \\le 3$. The paper identifies the mechanism at $n=4$: the interval criterion adds the boundary condition $E[(b-X)^2] \\le E[(b-Y)^2]$, which whole-line fourth-order dominance does not enforce.","pith_inferences":["The paper's comparison with Fishburn's criterion implies that interval-dependence is not a feature of stochastic dominance in the abstract but a feature of Jean's boundary-condition formulation; adopting a different interval definition would move the cutoff.","A direct diagnostic suggested by the results: when comparing risks by fourth- or higher-order dominance, rerun the ranking over nested intervals $[-R,R]$ for a range of $R$; any rank reversal indicates the conclusion is an artifact of the chosen bound.","The $n-m \\le 3$ threshold hints at a wider pattern — boundary information only matters once more than three moments beyond the preserved ones enter — and testing whether analogous cutoffs hold for other moment-based orderings would be a natural extension."],"forward_implications":["If Theorem 1 is correct, portfolio and saving decisions that invoke fourth-order dominance can be reversed by changing the assumed upper bound on possible wealth, even when all variables stay within both bounds.","Third-order (prudence) comparisons are interval-free, so precautionary-saving statements at order three do not inherit the ambiguity.","For mean-preserving comparisons, the interval choice is harmless when the gap between dominance degree and preserved moments is at most three; for larger gaps it starts to matter.","Enlarging a reference interval makes higher-order dominance easier to satisfy, so empirical studies should treat the interval length as a robustness parameter rather than a fixed modeling input."],"supporting_citations":[{"why":"Defines the interval-based $n$SD$_{[a,b]}$ criterion whose boundary conditions are the target of the consistency results.","marker":"Jean (1980, page 152)"},{"why":"Supplies the whole-line $n$SDR definition and Theorem 4.A.58 used in the mean-preserving direction of Theorem 2.","marker":"Shaked and Shanthikumar (2007, Section 4.A.7)"},{"why":"Section 2.2 contends that $n$SD$_{[a,b]}$ is more stringent than $n$SDR and may be inconsistent for $n \\ge 4$; Proposition 1 formalizes this.","marker":"Fang and Post (2022)"},{"why":"Provides the limit identity that derives a mean comparison from whole-line dominance, the key step in the $n=3$ equivalence.","marker":"Ogryczak and Ruszczynski (2001) (Lemma 1)"},{"why":"Establishes the expected-utility equivalence for interval-based dominance, used in Proposition 1 and Theorem 1.","marker":"Eeckhoudt et al. (2009) (Theorem 1)"},{"why":"Defines $n$th degree $m$-mean preserving stochastic dominance, the generalization treated in Theorem 2.","marker":"Liu (2014)"},{"why":"Gives the cone-equality criterion used to prove strictness when $n-m \\ge 4$.","marker":"Muller (1997) (Corollary 3.8)"},{"why":"Represents $n$-increasing functions as positive combinations of singularity functions, supporting the expected-utility direction for whole-line dominance.","marker":"Williamson (1956)"}],"fun_headline_variants":["Interval bounds decide fourth-order stochastic dominance","For order >3, dominance rankings hinge on interval","Shrinking interval breaks stochastic dominance above order three","Fourth-order dominance flips with reference interval choice","Reference interval only matters for dominance order ≥4"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results rest on Jean's interval-based definition, which appends boundary conditions at the right endpoint for every order up to $n$; without those boundary conditions, as in Fishburn's criterion, the consistency cutoff shifts to $n \\in \\{1,2\\}$ and the paper's threshold is not definition-independent.","fun_headline_variants_meta":{"raw":{"variants":["Interval bounds decide fourth-order stochastic dominance","For order >3, dominance rankings hinge on interval","Shrinking interval breaks stochastic dominance above order three","Fourth-order dominance flips with reference interval choice","Reference interval only matters for dominance order ≥4"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000629,"raw_usage":{"total_tokens":2859,"prompt_tokens":846,"completion_tokens":2013,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":462,"completion_tokens_details":{"reasoning_tokens":1942}},"tokens_in":462,"tokens_out":2013,"duration_ms":15891,"temperature":1.0,"reasoning_tokens":1942,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:23:12.557130+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute for Example 1 whether $X \\ge^{[0,c]}_4 Y$ holds as $c$ varies between 1 and 2; the theorem predicts it holds at $c=2$ and fails at $c=1$. If it held for every $c \\in [1,2]$, Theorem 1 would be false. A pair of variables supported in $[0,1]$ with $X \\ge_4 Y$ and $X \\ge^{[0,c]}_4 Y$ for all $c \\in (0,1]$ would also refute the theorem.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the interval-based $n$SD$_{[a,b]}$ criterion whose boundary conditions are the target of the consistency results."},{"cited_title":"and Shanthikumar, J","cited_arxiv_id":null,"evidence_quote":"Supplies the whole-line $n$SDR definition and Theorem 4.A.58 used in the mean-preserving direction of Theorem 2."},{"cited_title":"and Post, T","cited_arxiv_id":null,"evidence_quote":"Section 2.2 contends that $n$SD$_{[a,b]}$ is more stringent than $n$SDR and may be inconsistent for $n \\ge 4$; Proposition 1 formalizes this."},{"cited_title":"and Ruszczy´ nski, A","cited_arxiv_id":null,"evidence_quote":"Provides the limit identity that derives a mean comparison from whole-line dominance, the key step in the $n=3$ equivalence."},{"cited_title":"and Tsetlin, I","cited_arxiv_id":null,"evidence_quote":"Establishes the expected-utility equivalence for interval-based dominance, used in Proposition 1 and Theorem 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines $n$th degree $m$-mean preserving stochastic dominance, the generalization treated in Theorem 2."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents $n$-increasing functions as positive combinations of singularity functions, supporting the expected-utility direction for whole-line dominance."}],"review_version":1}