{"id":"6db1b088-73ba-4d1d-83a9-314225dde281","arxiv_id":"2502.07699","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For maxima of identically distributed random variables, the paper gives sharp anti-concentration bounds under arbitrary dependence, and sharper bounds under a new convexity condition on the copula's diagonal.","lead":"This paper derives exact worst-case bounds on the probability that the maximum of a set of random variables lands in a short interval. It shows that unless the dependence structure is restricted, this probability must grow linearly with the dimension, and introduces a new 'diagonally convex copula' condition that restores slow growth for many common distributions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Theorem 2 is correct as stated, and the diagonal-convexity assumption is explicit, scoped, and verified for the main copula families.","rationale":"I agree with the reader's ACCEPT verdict. The strongest claim, Theorem 2, is mathematically sound: the proof is short and correct, the extremal construction is valid, and the examples follow. The reader's identified weakest assumption—convexity of the copula diagonal—is a genuine restriction on applicability, and the paper's own counterexample shows it is not automatic. However, this is not a flaw in the argument: the theorem is conditional on that assumption, the assumption is verified for Gaussian, Archimedean, and mixture copulas, and the paper transparently discusses the limitation of identical marginals. The statistical application relies on a coupling from a forthcoming paper, but the anti-concentration bound itself is derived unconditionally from Theorem 2 and does not depend on that coupling for its validity. The only concrete issue found is a missing factor d in an auxiliary equality in Lemma 3, which is typographical and does not change the results. Therefore, no verdict adjustment is warranted.","tokens_in":24025,"tokens_out":32787,"duration_ms":251793,"concrete_test":"As a verification step worth running despite the non-finding: independently re-derive the formula for Ψ in Lemma 3 from Δ'(x) = d·(ψ^{-1})'(d ψ(x))·ψ'(x) and then numerically evaluate the diagonal Δ(x) = 1/log(d e^{1/x} - (d-1)e) on a fine grid for d=2 to confirm it is non-convex, checking that the monotonicity condition on Ψ indeed fails. Additionally, verify the extremal diagonal Δ_u satisfies the three conditions of Lemma 1 for representative (d,u) pairs with d=2,3,5 and u in both regimes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After checking the proof of Theorem 2 line by line, I find no load-bearing error in the central claim. The upper bound follows from Jensen's inequality applied to the convex diagonal at the endpoints u and 1, combined with the Lipschitz bound on the diagonal from Lemma 1; both inequalities are valid simultaneously, so taking the minimum is legitimate. The extremal diagonal Δ_u defined in Section 5.2 is convex (slope 0 then c), satisfies the three conditions of Lemma 1 for every u ∈ [0,1] (I verified c ≤ d, Δ_u(1)=1, and Δ_u(t) ≤ t in both regimes u ≤ (d-1)/d and u > (d-1)/d), and achieves Δ_u(u+δ)-Δ_u(u) = δ min(1/(1-u), d). The quantile transform handles arbitrary marginal CDFs, including atoms. The only issue I noticed is a typographical error in Lemma 3: the displayed equality for Ψ should read Ψ(x) = d·(ψ^{-1})'(d x)/(ψ^{-1})'(x), not (ψ^{-1})'(d x)/(ψ^{-1})'(x). The factor d does not affect any subsequent monotonicity argument or the counterexample in Section 3.1. The downstream examples (8, 10-13) correctly translate the pointwise bound into the claimed dimension dependence, and the paper states its main limitation (identical marginals) honestly in the conclusions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops sharp pointwise anti-concentration bounds for the maximum of d identically distributed real-valued random variables. Using the fact that the law of the maximum depends on the joint distribution only through the copula diagonal, the authors first give exact worst-case upper and lower bounds over all copulas (Theorem 1), showing that a strictly sublinear dependence on d is impossible without further structure. They then introduce a new class of copulas with convex diagonal section (Definition 3) and prove the sharp bound P(x < max_i X_i <= x+epsilon) <= (F(x+epsilon)-F(x)) min(1/(1-F(x)), d) (Theorem 2), together with an extremal copula diagonal attaining it. The paper verifies diagonal convexity for Gaussian, Archimedean, and mixture copulas, derives consequences for Gaussian, Weibull, reverse Gumbel, Pareto, and gamma marginals, and gives high-dimensional inference applications via coupling and quantile-based tests, including a Gaussian-mixture factor-model example.","tokens_in":24273,"tokens_out":15110,"duration_ms":141334,"significance":"The results are a genuine contribution to anti-concentration theory for extremal statistics. Theorem 1 settles the worst-case copula behavior exactly, and the diagonal-convexity condition in Theorem 2 is a novel, explicitly scoped structural assumption that yields dimension-free or poly-logarithmic bounds for a wide class of non-Gaussian dependence structures. The paper is notable for its sharpness: the extremal copula diagonals are explicitly constructed and shown to satisfy the copula-diagonal characterization, so the bounds are not merely one-sided. The proofs are transparent and mostly self-contained modulo standard external facts (Sklar's theorem, the diagonal characterization, and cited Gaussian maximum-density lemmas). I found no load-bearing technical error. The identical-marginals restriction and the worst-case nature of the bounds are stated honestly in the conclusion. The remaining issues are local presentation and edge-case matters.","major_comments":[],"minor_comments":[{"comment":"The displayed definition of Ψ is internally inconsistent: the first expression contains an extra factor d, while the second expression d·(ψ^{-1})'(d x)/(ψ^{-1})'(x) is not equal to it. The subsequent proof and all worked examples use Ψ(x) = (ψ^{-1})'(d x)/(ψ^{-1})'(x), so the displayed equality should be corrected accordingly; the monotonicity argument is unaffected because the discrepancy is a constant positive factor, but the statement as written is false.","section":"Section 5.4, Lemma 3"},{"comment":"Equation (7) defines Δ_up using the denominator d-1 and is therefore undefined for d=1, although Theorem 1 is stated for all d in N. The d=1 case is trivial and can be handled separately or by a limiting interpretation, but as written the proof of the extremal upper diagonal does not cover d=1.","section":"Section 5.1, proof of Lemma 4 / equation (7)"},{"comment":"The bound is stated for every x in R and epsilon >= 0, but when F(x)=1 the expression 1/(1-F(x)) is undefined and the convexity inequality in the proof divides by 1-u with u=1. Since in that case both the left-hand side and the increment F(x+epsilon)-F(x) are zero, the statement should either exclude F(x)=1 explicitly or add a one-line convention that the bound is interpreted by continuity; the same remark applies to the extremal diagonal Δ_u at u=1.","section":"Theorem 2 and its proof in Section 5.2"},{"comment":"There is a typographical error in the first sentence of Example 2: 'Xi ∼ Uare uniformly distributed' should read 'Xi ∼ U are uniformly distributed' or, preferably, 'Xi ∼ U[0,1]'. This is a presentation issue only.","section":"Section 2.2, Example 2 and Notation"},{"comment":"The claim that Example 14 offers a 'superior' anti-concentration inequality compared with Nazarov's inequality should be qualified. The displayed first bound contains the factor 1/p1 and the comparison with the Nazarov-type bound is asymptotic in the regime where d is large and the minimal variance σ is small; for fixed small p1 and finite d the two bounds need not be ordered. The authors' subsequent symmetric case clarifies the point, but the surrounding text should state the regime explicitly.","section":"Section 4.1, Example 14"}],"recommendation":"minor_revision","confidential_remarks":"No substantive confidentiality concerns. The paper is well within the scope of the journal, and the local presentation and edge-case issues listed in the minor comments are sufficient to warrant a minor revision rather than any deeper reworking."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a clean, correct paper. The core results are real. Theorem 1 pins down the exact worst-case pointwise concentration for maxima over all copulas, showing linear-in-d is unavoidable in general. Theorem 2 introduces the convex-diagonal condition and shows it is enough to get dimension-free or poly-log bounds for many common marginals. I checked the extremal diagonal constructions and the proof of Theorem 2 line by line; both hold up. The stress-test note found only a typo in Lemma 3: the displayed formula for Ψ is missing a factor d in the second expression. It does not affect anything downstream.\n\nWhat's new: the diagonal-convexity condition (Definition 3) is not in the literature I know, and the paper verifies it for Gaussian, Clayton, Frank, Gumbel–Hougaard, and mixture copulas. The examples giving explicit dimension dependence for Gaussian, Weibull, reverse Gumbel, Pareto, and gamma marginals are useful. The improvements over Nazarov-type bounds for non-Gaussian joint laws are genuine. The paper is also honest about its limitations: it only handles identical marginals, and the counterexample in Section 3.1 shows diagonal convexity is not automatic.\n\nSoft spots, in proportion: the Section 4 application to Gaussian mixture factor models leans on a coupling from Cattaneo et al. (2025), which is forthcoming and self-cited. That part is illustrative rather than a self-contained theorem. The conclusion's caveat that the bounds may not be 'dimension-free' in the sense of depending only on E[max|Xi|] is fair. The typo in Lemma 3 should be fixed, but it is minor.\n\nOverall, the math is sound, the new condition is interesting, and the paper deserves serious refereeing. I would send it out, and I would cite the diagonal-convexity result in my own work.","headline":"A correct and genuinely useful paper: sharp worst-case anti-concentration for maxima, plus a new convex-diagonal copula condition that yields dimension-free bounds; deserves a serious referee.","tokens_in":24841,"tokens_out":1845,"would_cite":true,"duration_ms":16677,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60E15","62H05","62G32"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves sharp two-sided bounds on the probability that a maximum of d identically distributed variables falls in a short interval, and shows that convexity of the copula's diagonal section converts a linear-in-d worst case into…","keywords":["anti-concentration","concentration function","maximum statistic","copula","diagonal section","convex diagonal","high-dimensional probability","extreme value theory"],"falsifier":"Optimize $\\Delta(u+\\delta)-\\Delta(u)$ over all convex functions $\\Delta:[0,1]\\to[0,1]$ with $\\Delta(1)=1$ and $0\\le \\Delta'\\le d$, the exact constraints from Lemma 1; the paper's proof shows the maximizing diagonal is piecewise linear with at most two breakpoints, so this is a finite linear program, and any convex $\\Delta$ whose maximal increment exceeds $\\delta\\min(1/(1-u),d)$ would refute Theorem 2.","tokens_in":23791,"feed_emoji":"📈","tokens_out":11415,"duration_ms":93552,"temperature":0.7,"pith_summary":"The paper asks how likely it is that the maximum of $d$ identically distributed real-valued random variables falls in a small window $(x,x+\\varepsilon]$, without assuming independence or a known joint law. Its first theorem gives sharp worst-case bounds over all copulas: the probability can be as large as $d(F(x+\\varepsilon)-F(x)) \\wedge F(x+\\varepsilon)$, so linear-in-dimension concentration is unavoidable in full generality. The second theorem introduces a new structural condition, convexity of the copula's diagonal section, under which the worst case drops to $(F(x+\\varepsilon)-F(x))\\min(1/(1-F(x)),d)$, a bound that is attained and therefore improvable only by stronger assumptions. From this the paper derives dimension-independent or poly-logarithmic anti-concentration inequalities for Gaussian, Weibull, reverse Gumbel, Pareto, and gamma margins, and applies them to high-dimensional testing and Gaussian mixture approximations. A sympathetic reader would care because it replaces joint Gaussianity, a standard but restrictive assumption, with a mild and checkable dependence condition that works for arbitrary common marginals.","feed_headline":"A single copula curve decides how sharply maxima concentrate","feed_subtitle":"Exact worst-case bounds for maxima hold for any marginal law—and drop to log-dimension under a convex copula diagonal.","key_machinery":"The carrying object is the diagonal section of the copula, $\\Delta(u)=C(u,\\dots,u)$, together with the class of diagonally convex copulas (Definition 3), meaning $\\Delta$ is convex on $[0,1]$. Because the maximum's distribution function is $\\Delta\\circ F$, the concentration probability is exactly an increment of $\\Delta$, so extremizing it reduces to a one-dimensional optimization over copula diagonals. Lemma 1 characterizes diagonals by $\\Delta(1)=1$, $\\Delta(u)\\le u$, and a Lipschitz slope bound of $d$; Theorem 1 solves the unconstrained extremization subject to those conditions, and Theorem 2 adds convexity, which forces $\\Delta$ to lie below the chord from $(u,\\Delta(u))$ to $(1,1)$ and yields the factor $1/(1-F(x))$.","core_discovery":"The paper's central claim is that pointwise anti-concentration of the maximum statistic is governed entirely, through the copula decomposition, by the diagonal section $\\Delta(u)=C(u,\\dots,u)$ of the dependence structure. Theorem 1 states exactly, for any common marginal law $F$ and any copula, that the worst-case probability is $d(F(x+\\varepsilon)-F(x)) \\wedge F(x+\\varepsilon)$ and the best-case probability is $0\\vee(1-F(x)-d(1-F(x+\\varepsilon)))$, with copulas attaining both, so the bounds cannot be improved without extra structure. Theorem 2 states that under a new condition, convexity of the diagonal section, the worst case improves to $(F(x+\\varepsilon)-F(x))\\min(1/(1-F(x)),d)$, and the paper constructs a copula attaining it. Because the proof only uses convexity of this one-dimensional curve, the result applies to any common marginal law and any copula whose diagonal is convex; the paper verifies the condition for the independence, Fréchet–Hoeffding, Gaussian, Clayton, Frank, Gumbel–Hougaard, and mixture copulas, and derives explicit dimension-dependent bounds for several marginals. The paper's application shows these bounds sharpen distributional approximations in high-dimensional inference, including a Gaussian mixture factor-model example where the dimension and small-variance terms separate additively.","pith_inferences":["The paper leaves open non-identical margins; a natural extension would be a multivariate analogue of the convex-diagonal condition controlling the probability of landing near the perimeter of a high-dimensional rectangle, which the authors note reduces to a generalized maximum problem.","Since only the diagonal enters the bound, practitioners could certify diagonal convexity rather than estimate the full copula; an empirical check of convexity of $\\widehat\\Delta(u)$ from data would indicate whether the favorable anti-concentration regime applies.","The mixing example suggests a general recipe: for mixture laws whose components have very different scales, conditioning on the component and applying the diagonal bound componentwise can yield additive rather than multiplicative dimension and scale factors."],"forward_implications":["For arbitrary dependence, no anti-concentration bound can be uniformly sublinear in the dimension: the worst-case probability is $d(F(x+\\varepsilon)-F(x))\\wedge F(x+\\varepsilon)$, so local concentration can be linear in $d$.","For diagonally convex copulas with Gaussian margins, the bound is $(\\varepsilon/\\sigma)(\\sqrt{2\\log d}+1)$, recovering the classical Gaussian anti-concentration inequality without requiring joint Gaussianity and with a slightly better constant.","For Weibull margins with shape $\\alpha\\ge 1$, the bound is $(\\varepsilon\\alpha/\\lambda)(\\log d+1)^{(\\alpha-1)/\\alpha}$; for Pareto and gamma margins with $\\alpha\\ge 1$, the bounds $\\alpha\\varepsilon/\\lambda$ and $\\varepsilon/\\lambda$ are dimension-independent.","In coupling-based high-dimensional inference, the pointwise bound at the quantile $q_\\alpha$ controls the error of maximum-statistic tests, and the Gaussian mixture example separates the $\\sqrt{\\log d}$ term from the small-variance term rather than multiplying them."],"supporting_citations":[{"why":"Supplies Sklar's theorem and the definition of copulas, which the paper uses to reduce the maximum's distribution to the copula diagonal.","marker":"Nelsen (2006)"},{"why":"Provides the characterization of d-dimensional copula diagonals used to construct the extremal copulas in Theorem 1 and the extremal diagonal in Theorem 2.","marker":"Fernández-Sánchez and Úbeda-Flores (2018)"},{"why":"Its lemmas on the density of Gaussian maxima are used in Lemma 2 to prove that every multivariate Gaussian copula has a convex diagonal section.","marker":"Chernozhukov et al. (2015)"},{"why":"Supplies the detailed classical Gaussian anti-concentration bound that the paper's Examples 8 and 9 improve upon and generalize.","marker":"Chernozhukov et al. (2017b)"},{"why":"The original Gaussian anti-concentration inequality for convex sets, which the paper recovers as a special case under weaker assumptions.","marker":"Nazarov (2003)"},{"why":"Provides the Mills-ratio inequality used in the Gaussian and Gaussian-mixture examples to control the hazard ratio $(1-\\Phi(x))/\\phi(x)$.","marker":"Birnbaum (1942)"},{"why":"Its refined version of the Gaussian anti-concentration inequality is the benchmark the paper compares against in the Gaussian mixture factor-model application.","marker":"Deng and Zhang (2020)"},{"why":"Supplies the martingale coupling for factor models that motivates the Gaussian mixture approximation analyzed in Example 14.","marker":"Cattaneo et al. (2025)"}],"fun_headline_variants":["Copula diagonal alone sets worst-case max bounds","Convex copula diagonal tightens max concentration","One curve controls anti-concentration of maxima","Sharp max bounds for any distribution via copula curve"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the curve describing the dependence among the variables, the copula's diagonal section, is convex; if it is not, only the weaker linear-in-dimension bound is guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Copula diagonal alone sets worst-case max bounds","Convex copula diagonal tightens max concentration","One curve controls anti-concentration of maxima","Sharp max bounds for any distribution via copula curve"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2895,"prompt_tokens":999,"completion_tokens":1896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":1837}},"tokens_in":615,"tokens_out":1896,"duration_ms":12511,"temperature":1.0,"reasoning_tokens":1837,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T11:52:06.148785+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Optimize $\\Delta(u+\\delta)-\\Delta(u)$ over all convex functions $\\Delta:[0,1]\\to[0,1]$ with $\\Delta(1)=1$ and $0\\le \\Delta'\\le d$, the exact constraints from Lemma 1; the paper's proof shows the maximizing diagonal is piecewise linear with at most two breakpoints, so this is a finite linear program, and any convex $\\Delta$ whose maximal increment exceeds $\\delta\\min(1/(1-u),d)$ would refute Theorem 2.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Sklar's theorem and the definition of copulas, which the paper uses to reduce the maximum's distribution to the copula diagonal."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Its lemmas on the density of Gaussian maxima are used in Lemma 2 to prove that every multivariate Gaussian copula has a convex diagonal section."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The original Gaussian anti-concentration inequality for convex sets, which the paper recovers as a special case under weaker assumptions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Mills-ratio inequality used in the Gaussian and Gaussian-mixture examples to control the hazard ratio $(1-\\Phi(x))/\\phi(x)$."},{"cited_title":"and Zhang, C.-H","cited_arxiv_id":null,"evidence_quote":"Its refined version of the Gaussian anti-concentration inequality is the benchmark the paper compares against in the Gaussian mixture factor-model application."}],"review_version":1}