REVIEW 3 major objections 3 minor 1 cited by
Ties, Tails and Spectra: On Rank-Based Dependency Measures in High Dimensions
T0 review · 3 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read Rank-based dependency matrices have universal limiting spectra under mild non-degeneracy.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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
What would settle it
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.
Extended reading notes
Core claim
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
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (3)
- [Section 4.6] 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.
- [Lemma 4.2] 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.
- [Lemma 3.4] 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.
minor comments (3)
- [Throughout] 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.
- [Remark 2.6] 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 3.2.1, footnote] 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.
Circularity Check
No significant circularity: Spearman and Kendall LSDs are derived from Theorem 2.8, which is proven via Stieltjes transforms; self-citations are ancillary, and earlier continuous-case results are recovered as consistency checks.
full rationale
The derivation chain is first-principles. Theorem 2.3 for Spearman's rho is obtained by verifying the moment conditions of Theorem 2.8 for the normalized rank matrix Z; the difficult condition (45) is proven directly in Section 3.2.1 (modulo an omitted proof of Lemma 3.4, a completeness gap, not a circular one). Theorem 2.5 for Kendall's tau is reduced in Section 4.3 to three claims: D can be replaced by tilde D (Lemma 4.1), the residual off-diagonal/U-statistic terms vanish (Section 4.5), and the leading term tilde D^{-1/2}M^{(1,1)}tilde D^{-1/2} equals (2/3)YY'-(2/3)I with Y = sqrt(3/n) tilde D^{-1/2}(u_1,...,u_n). The LSD of YY' is then supplied by Theorem 2.8, which is proved independently via Stieltjes transform/martingale arguments in Section 3.1. No parameter is fitted to the spectral data, and no limiting law is imported from the papers being generalized: the continuous-case results of [3] and [5] are recovered as special cases (Remarks 2.4, 2.6), not used as inputs. The only self-citations ([14], [25]) appear in remarks/examples and are not load-bearing for Theorems 2.3, 2.5, 2.8, or 2.9. The assertion in Section 4.6 that Y 'satisfies the conditions of Theorem 2.8' is abbreviated ('one can check'), and the proof of Lemma 3.4 is omitted; these are unverified proof obligations that bear on correctness, not circularity, because neither claim defines the target LSD in terms of itself or forces the conclusion by construction. Accordingly the circularity score is 0.
Assumptions & free parameters
assumptions (4)
- domain assumption Assumption 2.2: the triangular scheme is asymptotically uniformly non-degenerated
- domain assumption 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)
- domain assumption Exchangeability-type moment equalities in Theorem 2.8 (E[Y_ki^2] independent of i, etc.)
- standard math Standard Stieltjes transform and martingale concentration tools (Azuma, Burkholder, DKW)
Cite this review
Pith. "Pith review of Ties, Tails and Spectra: On Rank-Based Dependency Measures in High Dimensions." pith.science (2026). https://pith.science/paper/WN7J2Q43
@misc{pith2026250814992,
author = {Pith},
title = {Pith review of: Ties, Tails and Spectra: On Rank-Based Dependency Measures in High Dimensions},
year = {2026},
howpublished = {\url{https://pith.science/paper/WN7J2Q43}},
note = {Machine review of arXiv:2508.14992}
}
abstract
This work is concerned with the limiting spectral distribution of rank-based dependency measures in high dimensions. We provide distribution-free results for multivariate empirical versions of Kendall's $\tau$ and Spearman's $\rho$ in a setting where the dimension $p$ grows at most proportionally to the sample size $n$. Although rank-based measures are known to be well suited for discrete and heavy-tailed data, previous works in the field focused mostly on the continuous and light-tailed case. We close this gap by imposing mild assumptions and allowing for general types of distributions. Interestingly, our analysis reveals that a non-trivial adjustment of classical Kendall's $\tau$ is needed to obtain a pivotal limiting distribution in the presence of tied data. The proof for Spearman's $\rho$ is facilitated by a result regarding the limiting eigenvalue distribution of a general class of random matrices with rows on the Euclidean unit sphere, which is of independent interest. For instance, this finding can be used to derive the limiting spectral distribution of sample correlation matrices, which, in contrast to most existing works, accommodates heavy-tailed data.
Figures
Figures from the paper (1 more)
Forward citations
Cited by 1 Pith paper
-
Spectra of high-dimensional Spearman correlation matrices under scale-mixture dependence
Under a shared latent scale, the high-dimensional Spearman matrix has a generalized Marchenko–Pastur limit determined by the law of the conditional rank-score variance, not the classical MP law.
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.