{"id":"398bc198-6c98-42e6-a4d7-20d7ef8ff256","arxiv_id":"2508.14992","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Spearman's rho and a rescaled Kendall's tau matrices have universal limiting spectral distributions (semicircle and Marchenko-Pastur) for high-dimensional data with ties or heavy tails.","lead":"This paper proves that the eigenvalue distributions of rank-based dependency matrices like Spearman's rho and an adjusted Kendall's tau converge to universal random matrix laws even when data have ties and heavy tails. The result opens the door to nonparametric high-dimensional inference on discrete or heavy-tailed data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Kendall proof's pivotal step — §4.6 assertion that Y satisfies Theorem 2.8's moment conditions (8) — is unverified; if wrong, Theorem 2.5's LSD fails, though explicit bounds suggest the gap is fixable.","rationale":"The reader's weakest assumption is exactly the unverified application of Theorem 2.8 in Section 4.6, and I agree this is the single load-bearing gap: every Kendall LSD conclusion funnels through it. My own audit of the construction indicates the moment conditions are probably satisfiable: Y's rows are exchangeable, E[Y_{i1}Y_{i2}]=0 by antisymmetry, and Assumption 2.2 gives a uniform lower bound on E[U^2], which should make E[Y_{i1}^4]=O(n^{-2}). So the central result is likely correct, but the written proof omits a nontrivial verification and the paper itself flags the step with 'one can check.' Secondary gaps (Lemma 3.4 and Lemma 4.2 proof sketches) also deserve completion, but they do not change the picture. The appropriate status remains CONDITIONAL: the proof is not complete as written, but no evidence indicates the theorem itself is false.","tokens_in":25076,"tokens_out":32905,"duration_ms":334042,"concrete_test":"Write out the missing verification for §4.6. For a generic row, with S=Σ_{s=1}^n U_s^2, prove (a) E[U_1U_2/S]=0 by swapping U_1 and U_2; (b) under Assumption 2.2, sup_i E[U_{i1}^4]/E[U_{i1}^2]^2 < ∞ and P(S_i/n < E[U_{i1}^2]/2) ≤ e^{-c n}, so E[Y_{i1}^4] ≤ C n^{-2} with C independent of i. Substitute into (8): n^2/p^2 Σ_i E[Y_{i1}^4] ≤ C/p → 0, and the second sum is 0. If these bounds cannot be established with C independent of i, Theorem 2.5's claim (iii) is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.6 reduces Theorem 2.5 to Theorem 2.8 by defining Y = √(3/n) \\tilde D^{-1/2}(u_1,...,u_n) and saying 'one can check that it satisfies the conditions of Theorem 2.8.' This is the load-bearing point: the Kendall LSD passes entirely through the LSD of YY'. Row-sphere and exchangeability conditions are immediate, but the quantitative moment conditions (8) are not demonstrated. In particular, (8) requires n^2/p^2 Σ E[Y_{i1}^4] = o(1) and n^2/p^{3/2} Σ |E[Y_{i1}Y_{i2}]| = o(1). The second is likely 0 by antisymmetry of U_1U_2/S since the U's are i.i.d. and S is symmetric. The first is plausible: |U|≤1 and Assumption 2.2 should force E[U^2]≥c>0, giving E[Y_{i1}^4]=E[U_1^4/S^2]≤C n^{-2}, so n^2/p^2 Σ E[Y_{i1}^4]≤C/p→0. But the paper supplies none of this, and Lemma 4.2's lower bound on \\tilde D is only sketched. As written, the pivotal reduction is unsupported; if the uniform bound fails in some tied-data regime, the universal LSD in Theorem 2.5 would not follow.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies limiting spectral distributions (LSDs) of multivariate Spearman's rho and a modified Kendall's tau for p×n data matrices with independent rows, under p/n → γ ∈ [0,∞). Assumption 2.2 (asymptotic uniform non-degeneracy) allows discrete and heavy-tailed entries. Theorem 2.3 states that Spearman's rho matrix R has ESD converging a.s. to the Marchenko–Pastur law for γ>0 and, after scaling, to the semicircle law for γ=0. Theorem 2.5 states that for T, a version of Kendall's tau normalized by the diagonal matrix D of (6), the ESD converges in probability to (2/3)(η−1) or (2/3)ζ. The proofs rest on a general unit-sphere random-matrix theorem (Theorem 2.8) with moment conditions (8), which is also applied to uncentered and centered sample correlation matrices (Theorem 2.9). Remark 2.7 explains, with simulations, why the D-normalization is necessary for a pivotal limit under ties: the unnormalized off-diagonal Kendall matrix has a distribution-dependent LSD.","tokens_in":25447,"tokens_out":14228,"duration_ms":158022,"significance":"If the main results are correct, the paper provides the first LSD results for rank-based dependency matrices in the tied/heavy-tailed high-dimensional regime. The nontrivial finding that classical Kendall's tau must be rescaled by D to obtain a universal limit is interesting and well illustrated. The unit-sphere theorem (Theorem 2.8) is of independent interest and extends sample-correlation LSD results beyond the finite fourth moment condition. The paper is largely self-contained and the proofs are from first principles via Stieltjes transforms. However, two load-bearing verifications are explicitly deferred ('one can check' in §4.6 and 'for brevity, we omit details' in Lemma 4.2), which prevents immediate acceptance.","major_comments":[{"comment":"The reduction of claim (iii) to Theorem 2.8 is the pivotal step for Theorem 2.5, but the sentence 'one can check that it satisfies the conditions of Theorem 2.8' is not accompanied by a verification of (8). For Y_ij = U_ij / (Σ_t U_it^2)^{1/2}, the cross-moment condition requires estimating E[Y_i1Y_i2] = E[U_i1U_i2 / Σ_t U_it^2]; since the denominator depends on all entries, this is not an immediate consequence of E[U_i1]=0. The fourth-moment condition requires a uniform lower bound on E[U_i1^2] under Assumption 2.2 and a bound E[Y_i1^4]=O(n^{-2}). If either condition fails, the universal LSD in Theorem 2.5 would not follow. Please supply the calculation.","section":"Section 4.6"},{"comment":"The uniform high-probability lower bounds min_i D_ii > C and min_i D̃_ii > C are stated with proof omitted ('follows along the lines of (45)', 'for brevity, we omit details'). These bounds are used in §4.4–4.5 to justify the Frobenius-norm approximations (62) and (63), so they are load-bearing. Since Assumption 2.2 is only asymptotic and uniform in i, the lower bound requires a genuine simultaneous-in-i argument; please include it.","section":"Lemma 4.2"},{"comment":"The proof of Lemma 3.4 is omitted ('simple application of Hölder's inequality'). This lemma is used in §3.2.1 to reduce the tied-data analysis of (45) to the two-point case, and the proof of (45) is the difficult part of Theorem 2.3. Either give the proof explicitly or provide a precise reference for the reduction.","section":"Lemma 3.4"}],"minor_comments":[{"comment":"There are several LaTeX/formulation errors: 'WriteQi j' (missing space), 'the largest of of the Xi j's', 'we may that assume' in §4.1, and broken expressions 'ofp n/p' in Theorem 2.3, Theorem 2.8(1), and Theorem 2.9(1). These should be fixed.","section":"Throughout"},{"comment":"The displayed equality after '= 2 n(n−1)' is malformed; the final relation to D^{1/2}TD^{1/2}+I should be typeset cleanly.","section":"Remark 2.6"},{"comment":"The footnote in the proof of (45) is terse: passing from limsup in Assumption 2.2 to a fixed η>0 for all i and t deserves more explanation, since the argument is used to make Hoeffding bounds uniform in i.","section":"Section 3.2.1, footnote"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central claims are plausible and original, and the missing verifications in §4.6 and Lemma 4.2 are likely routine to fill in given Assumption 2.2. I do not see a fundamental flaw, but the proof of the main Kendall theorem currently contains unsupported load-bearing assertions. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real step forward. They give universal LSDs for rank-based dependency matrices under tied/heavy-tailed data: Spearman's rho under asymptotic non-degeneracy, and a newly defined D-scaled Kendall's tau whose LSD is (2/3)(MP-1) or (2/3)semicircle. Theorem 2.8 (unit-sphere rows) is clean and independently useful; the application to sample correlation matrices under infinite fourth moments is a genuine extension. The paper recovers earlier continuous-case results as checks, not as inputs, and there are no fitted parameters. The proofs are standard Stieltjes transform/martingale machinery and, as far as I can tell, the skeleton is sound.\n\nThe soft spots are real but concentrated. Lemma 3.4 is asserted without proof (\"simple application... omit the proof\") — minor, likely fine. Lemma 4.2's lower bound on D_ii is also only sketched — notable but probably follows from the same argument used for Spearman's rho. The load-bearing gap is in Section 4.6: reducing Kendall's tau to Theorem 2.8 via Y = sqrt(3/n) tilde D^{-1/2}(u_1,...,u_n) and saying \"one can check that it satisfies the conditions of Theorem 2.8.\" The second moment condition is immediate because E[U_i1 U_i2]=0; the first requires a uniform lower bound on tilde D_ii and a fourth-moment estimate. The stress-test sketch (|U|<=1, Assumption 2.2 forcing E[U^2]>=c, and p/n bounded) suggests it is fixable, but the paper does not supply the check, and Theorem 2.5 collapses without it. This should be requested in review, not assumed.\n\nOverall: worth a serious referee. The novelty is clear, the statements are precise, and the issue is a missing verification rather than a suspicious conclusion. I'd send to a good math.ST referee and ask specifically for the Section 4.6 details and the omitted proof of Lemma 3.4. If the authors fill that in, this is a solid paper I'd cite.","headline":"Genuinely new universal LSDs for Spearman's rho and a cleverly D-scaled Kendall's tau under ties and heavy tails; the main theorems are plausible, but the Kendall proof's pivotal 'one can check' in Section 4.6 needs to be written out before I'd call it complete.","tokens_in":25906,"tokens_out":3425,"would_cite":true,"duration_ms":40209,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","60F05","60F10","60G10","60G55","60G70"],"pacs":[],"model":"deepseek-v4-flash","headline":"Rank-based dependency matrices have universal limiting spectra under mild non-degeneracy.","keywords":["limiting spectral distribution","Kendall's tau","Spearman's rho","sample correlation matrix","Marchenko-Pastur law","semicircle law","ties","heavy-tailed data"],"falsifier":"Take rows with a fixed atom-heavy distribution such as Bernoulli(1/2) or a t-distribution with 3 degrees of freedom, let p/n tend to a fixed gamma > 0, and compute the two quantities in condition (8) for the Kendall-based matrix Y = sqrt(3/n) tilde D^{-1/2}(u_1,...,u_n): n^2/p^2 sum_i E[Y_{i1}^4] and n^2/p^{3/2} sum_i |E[Y_{i1}Y_{i2}]|. If either fails to converge to 0, or if the simulated empirical spectral distribution of T does not match the distribution of (2/3)(eta - 1) at that gamma, the Kendall theorem is refuted.","tokens_in":25002,"feed_emoji":"📊","tokens_out":10539,"duration_ms":111506,"temperature":0.7,"pith_summary":"This paper asks what the eigenvalue distributions of rank-based dependency matrices look like when the number of variables p grows with sample size n. It claims that, under a mild non-degeneracy condition, Spearman's rho matrix has the same limiting spectral distribution as a sample covariance matrix: Marchenko-Pastur when p/n approaches a positive constant, semicircle when p/n approaches 0. For Kendall's tau, the raw sign-based matrix does not have a distribution-free limit when ties are present, so the paper rescales it by a data-dependent diagonal matrix and proves the rescaled matrix converges to a universal shifted or scaled Marchenko-Pastur or semicircle law for continuous, discrete, and heavy-tailed data alike. If this is right, rank-based spectral methods become pivotal across a much wider class of distributions than earlier continuous-light-tail results allowed.","feed_headline":"Rescale Kendall's tau and its spectrum becomes universal","feed_subtitle":"Spearman's rho matches the same universal laws; tied and heavy-tailed data included.","key_machinery":"The workhorse is Theorem 2.8, a limiting spectral distribution result for a general class of p x n random matrices with independent rows on the Euclidean unit sphere, under moment conditions (8): n^2/p^2 sum_i E[Y_{i1}^4] = o(1) and n^2/p^{3/2} sum_i |E[Y_{i1}Y_{i2}]| = o(1). It states that the empirical spectral distribution of sqrt(n/p)(YY' - I) converges almost surely to the semicircle law when p/n -> 0, and that of YY' converges almost surely to the Marchenko-Pastur law F_gamma when p/n -> gamma > 0. Spearman's rho matrix is put in this form by centering fractional ranks and normalizing rows to unit length; Kendall's matrix is rewritten as T = (2/3)YY' - (2/3)I with Y = sqrt(3/n) tilde D","core_discovery":"The central claim is that, under Assumption 2.2 (asymptotic non-degeneracy: no single atom captures almost all probability), the empirical spectral distribution of Spearman's rho matrix R converges almost surely to the Marchenko-Pastur law F_gamma when p/n -> gamma > 0, and after scaling by sqrt(n/p) on R - I it converges to the semicircle law when p/n -> 0. For Kendall's tau, the paper introduces T = D^{-1/2} offdiag(tau) D^{-1/2}, where D measures the row-wise variance of the effective rank scores, and proves that sqrt(n/p) T converges in probability to (2/3) times a semicircle variable in the gamma = 0 regime, while T converges in probability to (2/3)(eta - 1) with eta Marchenko-Pastur wh","pith_inferences":["The row-wise normalization idea should transfer to other U-statistic-based matrices, such as distance covariance or spatial sign matrices: normalize rows by a variance proxy before studying spectra, then check whether the resulting limit is pivotal.","A direct verification of the Section 4.6 moment conditions would close the only visible gap; if that check fails for some distribution near the boundary of Assumption 2.2, the Kendall theorem likely holds only under a stronger non-degeneracy condition.","Since the bulk of the adjusted Kendall spectrum already matches a covariance-type model, edge statistics of the adjusted Kendall matrix may follow Tracy-Widom laws under the same assumptions, paralleling known continuous-case results."],"forward_implications":["Spearman's rho matrices can be used for spectral inference when data are discrete or heavy-tailed, not only in the continuous light-tailed setting covered by earlier work.","Any high-dimensional procedure based on Kendall's tau should use the adjusted matrix T = D^{-1/2} offdiag(tau) D^{-1/2} (or its population analogue) when ties are possible; the unadjusted matrix has a distribution-dependent limit.","In the continuous case, the new theorems recover the known Marchenko-Pastur limits for Kendall's tau and Spearman's rho, so the results extend rather than contradict previous findings.","Theorem 2.8 gives a route to sample correlation matrix limiting spectral distributions under moment conditions weaker than finite fourth moments, covering distributions in the normal domain of attraction, including regularly varying tails when p grows as a power of n.","Both asymptotic regimes, gamma > 0 and gamma = 0, are covered by the same assumptions, so rank-based matrices have known bulk spectra across the full range of moderate-to-proportional high dimensionality."],"supporting_citations":[{"why":"Supplies the Stieltjes transform toolkit, martingale difference bounds, and the spectral comparison lemma used throughout the proofs of Theorems 2.8 and 2.5.","marker":"[2]"},{"why":"Prior limiting spectral distribution result for Spearman's rho in the continuous case, which Theorem 2.3(2) recovers.","marker":"[3]"},{"why":"Marchenko-Pastur law for Kendall's tau in the continuous case, recovered by Theorem 2.5(2) when ties are absent.","marker":"[5]"},{"why":"Spectral statistics for the large Spearman rank correlation matrix; supplies the fractional-rank representation used to align with continuous-case results.","marker":"[8]"},{"why":"Connects the fourth-moment condition to the domain of attraction of the normal law for correlation matrices, used in Remark 2.10.","marker":"[14]"},{"why":"Martingale concentration method for random matrices with rows on the unit sphere, used in the proof of Lemma 3.1.","marker":"[16]"},{"why":"Asymptotic ratio results for sums of squares, used to check the moment conditions for regularly varying entries in Example 2.11.","marker":"[22]"},{"why":"Provides the bound on Stieltjes transform differences used in the martingale difference argument in Lemma 3.1.","marker":"[39]"}],"fun_headline_variants":["Ties and tails tamed: rank spectra go universal","Kendall's tau fixed for ties: spectrum goes universal","Spearman's rho hits Marchenko-Pastur in high dimensions","Rank measures: universal spectra even with ties and heavy tails","Rescaled rank-based spectra: universal laws for tied data"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is a calculation left unstated: the paper asserts in Section 4.6 that the unit-sphere matrix built from the Kendall scores satisfies the same weak moment conditions as the general theorem, and the whole Kendall result collapses if that check turns out to fail for some distribution allowed by Assumption 2.2.","fun_headline_variants_meta":{"raw":{"variants":["Ties and tails tamed: rank spectra go universal","Kendall's tau fixed for ties: spectrum goes universal","Spearman's rho hits Marchenko-Pastur in high dimensions","Rank measures: universal spectra even with ties and heavy tails","Rescaled rank-based spectra: universal laws for tied data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1133,"prompt_tokens":758,"completion_tokens":375,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":289}},"tokens_in":502,"tokens_out":375,"duration_ms":5098,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:11:22.239165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take rows with a fixed atom-heavy distribution such as Bernoulli(1/2) or a t-distribution with 3 degrees of freedom, let p/n tend to a fixed gamma > 0, and compute the two quantities in condition (8) for the Kendall-based matrix Y = sqrt(3/n) tilde D^{-1/2}(u_1,...,u_n): n^2/p^2 sum_i E[Y_{i1}^4] and n^2/p^{3/2} sum_i |E[Y_{i1}Y_{i2}]|. If either fails to converge to 0, or if the simulated empirical spectral distribution of T does not match the distribution of (2/3)(eta - 1) at that gamma, the Kendall theorem is refuted.","supporting_citations":[{"cited_title":"Large sample covariance matrices without independence structures in columns","cited_arxiv_id":null,"evidence_quote":"Prior limiting spectral distribution result for Spearman's rho in the continuous case, which Theorem 2.3(2) recovers."},{"cited_title":"Mar ˇcenko-pastur law for Kendall’s tau","cited_arxiv_id":null,"evidence_quote":"Marchenko-Pastur law for Kendall's tau in the continuous case, recovered by Theorem 2.5(2) when ties are absent."},{"cited_title":"Concentration of measure and spectra of random matrices: applications to correlation matrices, elliptical distributions and beyond","cited_arxiv_id":null,"evidence_quote":"Martingale concentration method for random matrices with rows on the unit sphere, used in the proof of Lemma 3.1."},{"cited_title":"Fuchs, A","cited_arxiv_id":null,"evidence_quote":"Asymptotic ratio results for sums of squares, used to check the moment conditions for regularly varying entries in Example 2.11."}],"review_version":1}