{"id":"57108c19-cb9a-44a5-a472-e689eeca9501","arxiv_id":"2507.02640","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The Li-Chen U-statistic test for equality of high-dimensional covariance matrices is shown to be asymptotically valid and powerful under generalized elliptical models, with a new central limit theorem and variance formula.","lead":"This paper proposes a two-sample test for whether two high-dimensional datasets have the same covariance matrix, under generalized elliptical distributions rather than only normal or independent-component data. The key result is a new central limit theorem for the Li-Chen U-statistic, giving the first asymptotic guarantees for this test in elliptical models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma A.3's proof of the Lyapunov condition relies on an unjustified bound S_{n,2,2}=o(σ^4) when p/n→∞, leaving the martingale CLT step in Theorem 2.1 unverified.","rationale":"The reader's weakest assumption was about Assumptions (A1) and (A3) being unverified in applications. My stress-test identified a different, more internal concern: a specific step in the proof of Lemma A.3 that is not justified by the cited moment lemma when p/n_1 is large, which is precisely the flexible regime the paper advertises. The theorem may be salvageable, because the true fourth moment of an elliptical quadratic form is much smaller than the crude (tr C)^4 bound used in the paper; however, the proof as written does not supply the necessary argument, so the central claim is not fully verified. This warrants conditional acceptance: the authors should either replace the crude bound with a sharp fourth-moment expansion or state an additional rate condition such as p/n_1 = o(1/ε_n) tied to the o(1/p) in Lemma B.2. The concern is not an ad hominem or a disagreement with consensus; it is a concrete gap in the technical derivation of the CLT.","tokens_in":28111,"tokens_out":46811,"duration_ms":464457,"concrete_test":"Set n_1 = n_2 = n, p = n^2 / log n (so p/n → ∞), Σ_1 = I_p, Σ_2 = (1-η)I_p with η fixed, and choose ξ^2 with (ξ^2-p)/√p having variance τ>0 and finite 8th moment. Compute (or Monte-Carlo estimate with 10^6 replications) the exact ratio Σ_{k=1}^{n_1} E[T_{n,k}^4] / σ_n^4 using the elliptical moment identities of Lemma B.1 rather than the crude Lemma B.2 bound. If the ratio tends to 0, Lemma A.3 holds and the paper only needs a sharper proof; if it does not, the claimed CLT (and hence the level and power guarantees) fails in this regime. As a second check, re-derive Lemma B.2 for q=4 with C = ηI to determine whether the correct second-order term is o(1/p)(tr C)^4 or o(1/p)tr(C^4); if the latter, the displayed proof of S_{n,2,2}=o(σ^4) is invalid and must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 2.1 is established via the martingale CLT, which requires Lemma A.3, the claim that Σ E[T_{n,k}^4] = o(σ_n^4). In the proof of Lemma A.3, the deterministic part is bounded by S_{n,2,2} = o(1/p)n_1^{-3}(tr Σ_1(Σ_1-Σ_2))^4, and this is asserted to be o(σ_n^4) 'similarly to S_{n,1,2}'. The analogy fails: S_{n,1,2} = n_1^{-3}[tr(Σ_1Δ)^2]^2 is o(σ^4) because σ^4 ≥ c n_1^{-2}[tr(Σ_1Δ)^2]^2. For S_{n,2,2}, the best available inequality tr(Σ_1Δ)^4 ≤ p^2[tr(Σ_1Δ)^2]^2 only yields S_{n,2,2} ≤ o(p/n_1) times the same lower bound, which is not o(1) when p ≫ n_1. The configuration Σ_1 = I_p, Σ_2 = (1-η)I_p with p/n_1 → ∞ satisfies (A1)-(A3) and makes the displayed bound o(p/n_1)σ^4, so the written proof does not establish Lemma A.3. The root cause is Lemma B.2: for C = ηI, (tr C)^4 = η^4 p^4, whereas the true fourth moment of x^T C x − tr C is O(η^4 p^2), so the lemma's o(1/p)tr^4(CΣ) term is far too crude to justify the claimed o(σ^4). The theorem may still be true with a sharper fourth-moment expansion, but as written the proof has a genuine gap exactly in the regime where the paper claims no growth condition on p/n is needed.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Summary: The paper develops a two-sample test for equality of high-dimensional covariance matrices under generalized elliptical models (EM). The test statistic T_n is the U-statistic estimator of the squared Frobenius norm ||Σ_{n,1}−Σ_{n,2}||_F^2 introduced by Li and Chen (2012) for independent component models. The central theoretical result, Theorem 2.1, asserts that under assumptions (A1)–(A3), (T_n − ||Σ_{n,1}−Σ_{n,2}||_F^2)/σ_n converges in distribution to a standard normal, with σ_n^2 given in (2.4), and no explicit growth condition on p/n. Proposition 2.2 establishes ratio consistency of the null variance estimator, Corollary 2.3 gives asymptotic level control, and Theorem 2.4 gives power consistency. The proofs use a martingale decomposition of T_n, conditional variance computations (Lemmas A.1–A.2), and a fourth-moment Lindeberg condition (Lemma A.3). Simulations compare the method against LRT, DHW, SY, and ZZPZ across five elliptical distributions and three covariance structures, and an application to S&P 500 quarterly returns is reported.","tokens_in":28445,"tokens_out":43708,"duration_ms":400164,"significance":"If the main result holds, the paper delivers the first two-sample covariance test with level and power guarantees under generalized elliptical models, filling a clear gap between the ICM literature and elliptical models; the absence of growth conditions on p/n is a genuinely attractive feature. The manuscript is commendable for the transparency of its proof architecture (explicit martingale differences, separated treatment of the lower-order U-statistic terms, ratio-consistent variance estimation), for the variance formula (2.4) reducing correctly to the normal-case variance when τ_i = 2, and for a substantial simulation study covering five distributions and spiked/Toeplitz/bulk-eigenvalue structures. However, the verification of the Lindeberg-type condition in Lemma A.3 contains a genuine gap that is load-bearing for Theorem 2.1 exactly in the p ≫ n regime, so the significance claim is conditional on a successful repair of that proof step (see Major Comment 1).","major_comments":[{"comment":"This is the load-bearing issue. In the proof of Lemma A.3, the deterministic part of the fourth-moment bound is S_{n,2,2} = (1/n_1^4) o(1/p) Σ_{k=1}^{n1} tr^4(Σ_1(Σ_1−Σ_2)), and the text asserts 'similarly to S_{n,1,2}' that S_{n,2,2} = o(σ_n^4). The analogy fails: for S_{n,1,2} = n_1^{-3} tr^2(Σ_1Δ)^2, the lower bound σ_n^4 ≥ c n_1^{-2} tr^2(Σ_1Δ)^2 (from the term 8 n_1^{-1} tr{Σ_1Δ}^2 in (2.4)) absorbs the extra n_1^{-1}; for S_{n,2,2}, the Cauchy–Schwarz inequality |tr(Σ_1Δ)| ≤ √p·(tr(Σ_1Δ)^2)^{1/2} yields only S_{n,2,2} ≤ o(p/n_1)·σ_n^4, which is not o(σ_n^4) when p/n_1 → ∞. The configuration Σ_1 = I_p, Σ_2 = (1−η)I_p with n_1 = n_2 and p/n_1 → ∞ satisfies (A1)–(A3) and has tr^4(Σ_1Δ) = η^4 p^4, so the displayed bound cannot establish the lemma in the advertised no-growth-on-p/n regime. The root cause is Lemma B.2 (imported from Wang and Lopes, 2025): for C = ηI, its second term o(1/p)tr^4(CΣ) is of order o(η^4 p^3), which can dominate the first term tr^2((CΣ)^2) = η^4 p^2, whereas the actual fourth moment of x^T Cx − tr C is of order η^4 p^2 in the light-tailed cases (e.g., χ^2_p radii) permitted by (A1); the crudeness of the trace-power term is exactly what breaks the subsequent bound. Since Lemma A.3 provides the conditional Lindeberg condition in the martingale CLT (Billingsley, Theorem 18.1) used to conclude Theorem 2.1, the proof of the main theorem is incomplete exactly in the regime the paper emphasizes. The gap appears fixable by a sharper fourth-moment expansion that keeps the deterministic Δ = Σ_1 − Σ_2 component inside the tr^2((CΣ)^2)-type term instead of bounding its trace by a power of p; the authors should supply such an argument (or a strengthened version of Lemma B.2) and re-verify Lemma A.3, including the k > n_1 case and the question of whether the stated eighth-moment condition in (A1) suffices once the expansion is sharpened.","section":"Appendix A.1.1 (Proof of Lemma A.3)"},{"comment":"Section 3.2 applies the test to S&P 500 quarterly log-returns with p = 434, n_1 = 30, n_2 = 22, i.e., p/n_1 ≈ 14.5 and p/n_2 ≈ 19.7, which is precisely the p ≫ n regime in which the proof of Lemma A.3 (and hence of Theorem 2.1) is currently incomplete. In addition, the paper offers no diagnostic supporting (A1) (notably E|(ξ^2 − p)/√p|^8 = o(p^2)) or (A3) for these returns, despite citing the Wang–Lopes goodness-of-fit test for ellipticity in the introduction. The headline empirical claim that the proposed test 'is the only one to detect a difference' therefore rests on the unproven regime. The authors should either repair the proof so that it covers this regime, or qualify the application claim and add a discussion (or check) of the model assumptions on the data.","section":"Section 3.2 (S&P 500 application)"}],"minor_comments":[{"comment":"The caption contains the typo 'repititions' (should be 'repetitions'), and the figure would be clearer if it stated the sample sizes and the values of τ_i for the two displayed elliptical distributions.","section":"Figure 1 caption"},{"comment":"The indices in (2.3) are inconsistent: the first summand writes x^{(1)T}_i x^{(2)}_j without defining i, and the starred sums are not accompanied by the 'pairwise distinct' convention in that display; this appears to be a typesetting artifact but should be corrected.","section":"Equation (2.3)"},{"comment":"The definition of (EM) in the Introduction states E[ξ_p] = p, but consistency with (A1) and with the normalization Cov(x) = Σ requires E[ξ_{i,1}^2] = p; as written the normalization is incorrect.","section":"Introduction (definition of the elliptical model)"},{"comment":"The quantity σ̃_n is defined with factor 2/n_i in Proposition 2.2 but with factor 1/n_i in the discussion before Theorem 2.4; the factor is asymptotically irrelevant but the two displays should be aligned.","section":"Before Theorem 2.4"},{"comment":"The sentence 'where used Lemma B.4 for the last step' is missing a subject and should read 'where we used Lemma B.4'.","section":"Proof of Lemma A.1"},{"comment":"Remark 2 uses ν_{i,3} for the fourth moment of the independent component; the subscript is confusing and should be renamed (for instance κ_i).","section":"Remark 2"},{"comment":"The Introduction contains several typos, e.g., 'they different fundamentally in their modeling capacities' should read 'they differ fundamentally'.","section":"Introduction (typos)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is genuine: I checked the bound S_{n,2,2} in the proof of Lemma A.3 and the displayed inequality only gives o(p/n_1)σ_n^4, which is not sufficient when p ≫ n_1; the diagonal configuration Σ_1 = I, Σ_2 = (1−η)I confirms that the written argument cannot go through. This is the single load-bearing defect I found. The authors should also be asked to verify that, after sharpening the fourth-moment expansion, the moment condition in (A1) still suffices, since E[(ξ^2 − p)^4] = p^2 E[((ξ^2−p)/√p)^4] can grow slowly under E|(ξ^2−p)/√p|^8 = o(p^2). The rest of the paper — the variance formula, the ratio-consistency argument, the power analysis, and the simulations — is sound as far as I checked. No concerns about attribution or novelty; the citations to Li–Chen (2012) and Wang–Lopes are appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something genuinely useful: it takes Li and Chen's U-statistic and derives the correct variance and normal limit under generalized elliptical models, where the variance has an extra tau_i term that the independent-component formula misses. That is a real technical contribution, and the paper is honest about what is new and what is borrowed. The proof is detailed, the martingale decomposition is standard and mostly careful, and the imported moment lemmas are cited properly. The numerical study is broad and the comparison against ICM-based tests is informative. I agree with the reader that the central claim is worth taking seriously.\n\nBut I cannot join the Accept verdict. The stress-test note is right. Lemma A.3 is load-bearing: it supplies the Lyapunov-type condition that lets the martingale CLT go through. In the proof, the term S_{n,2,2} is dismissed as o(sigma_n^4) by analogy with S_{n,1,2}, and the analogy fails. The best available bound gives only tr^4(C) ≤ p^2 [tr(C^2)]^2, which turns S_{n,2,2} into o(p/n_1) times the same lower bound that makes S_{n,1,2} negligible. When p/n_1 → ∞, e.g. with Sigma_1 = I, Sigma_2 = (1-eta)I, that is not o(1). That configuration satisfies (A1)–(A3), so the gap sits exactly in the regime the abstract advertises: no explicit growth condition on p/n. The theorem may be true, and a sharper fourth-moment expansion might fix it, but as written the proof does not establish it.\n\nMinor issues are minor: the E[xi^p]=p typo, omitted symmetric cases, no code, and the S&P application with n_1=30, n_2=22, p=434 is far below any asymptotic regime. None of those matter next to the Lemma A.3 gap.\n\nWho should read this: anyone working on high-dimensional covariance testing or elliptical models. It deserves a serious referee, because the claim is important and the gap looks fixable, but it should not be accepted until Lemma A.3 is repaired. I would not cite the main theorem in its current form, though I would happily bring the paper to a reading group to see whether the proof can be patched.","headline":"The new CLT under elliptical models is a real contribution, but Lemma A.3 has a genuine gap in the p≫n regime, so the main theorem is not yet proven as written.","tokens_in":29009,"tokens_out":3139,"would_cite":false,"duration_ms":34142,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A18","60F17","62H15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that a U-statistic estimator of the squared Frobenius norm of the covariance difference is asymptotically normal under generalized elliptical models, yielding the first two-sample covariance test with level and power…","keywords":["Central limit theorem","covariance matrix","elliptical models","high-dimensional statistics","hypothesis testing","two-sample test","U-statistic","Frobenius norm"],"falsifier":"Simulate two elliptical samples with $p$ large and $n_1=n_2$ moderately large, using covariance matrices with a few dominating spikes, for instance $\\Sigma = \\operatorname{diag}(p, 1, \\ldots, 1)$, so that $\\operatorname{tr}(\\Sigma^4)$ is comparable to $\\operatorname{tr}^2(\\Sigma^2)$ and assumption (A3) fails, and inspect whether $(T_n - \\|\\Sigma_{n,1}-\\Sigma_{n,2}\\|_F^2)/\\sigma_n$ is approximately normal. Alternatively, use an elliptical radius whose squared law is heavier-tailed than the eighth-moment condition permits, such as a scaled $\\chi^2_p$ with few degrees of freedom, and record the empirical rejection rate of the proposed test under the null.","tokens_in":27846,"feed_emoji":"📊","tokens_out":8282,"duration_ms":78518,"temperature":0.7,"pith_summary":"This paper proposes a two-sample test for whether two high-dimensional covariance matrices are equal, under generalized elliptical distributions in which each observation is a common random scale times a uniformly distributed direction. The test statistic is the same U-statistic estimator of the squared Frobenius norm of the covariance difference that Li and Chen introduced for independent component models. The paper's central contribution is a new central limit theorem for this statistic under elliptical data, valid under both the null and the alternative hypotheses, together with a variance estimator that is ratio-consistent under the null. If the theorem is right, this is the first two-sample covariance test with asymptotic level control and power guarantees for elliptical models, and it needs neither sparsity nor any explicit condition on the dimension-to-sample-size ratio. The paper also reports simulations and an S&P 500 application where the proposed test rejects a covariance change that two competing independent-component-based tests miss.","feed_headline":"First two-sample covariance test proven valid for elliptical data","feed_subtitle":"Same U-statistic as Li–Chen controls the level and detects S&P 500 covariance shifts, no sparsity or dimension limits.","key_machinery":"The carrying object is the U-statistic $T_n = U_{n,1} + U_{n,2} - 2V_n$, where each $U_{n,i}$ estimates $\\operatorname{tr}(\\Sigma_{n,i}^2)$ from one sample and $V_n$ estimates $\\operatorname{tr}(\\Sigma_{n,1}\\Sigma_{n,2})$ from cross-sample inner products, so that $T_n$ unbiasedly estimates $\\|\\Sigma_{n,1}-\\Sigma_{n,2}\\|_F^2$. The proof decomposes $T_n$ into martingale differences and computes their conditional variances using mixed moments of elliptical quadratic forms, whose leading term $E[\\xi^4]/(p(p+2))$ with the shared radius $\\xi$ produces the variance terms with $\\tau_i$ in (2.4). A moment bound for centered quadratic forms under the eighth-moment condition on $\\xi^2$ verifies the Lindeberg condition, and the trace-product decay in assumption (A3) kills the cross terms.","core_discovery":"At the center of the paper is a central limit theorem: if assumptions (A1) through (A3) hold, then $\\frac{1}{\\sigma_n}(T_n - \\|\\Sigma_{n,1}-\\Sigma_{n,2}\\|_F^2)$ converges in distribution to a standard normal random variable, where $\\sigma_n^2$ is the explicit variance formula in (2.4). The formula contains terms with $\\tau_i$, the limiting variance of $\\xi^2/\\sqrt{p}$ for the elliptical radius $\\xi$ of sample $i$; in the Gaussian case $\\tau_i=2$ and the variance collapses to the independent-component formula of Li and Chen, while for other elliptical radii it differs. From this CLT the paper derives two practical guarantees: the test that rejects when $\\hat T_n > z_{1-\\alpha}$ has asymptotic level $\\alpha$ under the null, and it is power-consistent whenever $\\|\\Sigma_{n,1}-\\Sigma_{n,2}\\|_F^2/\\sigma_n \\to \\infty$. The test needs no sparsity, no bootstrap, and no explicit condition on the ratio $p/n$.","pith_inferences":["Editorial inference: Because the test is valid across both model classes without pretesting, its availability weakens the argument for running a goodness-of-fit test for ellipticity first in high-dimensional pipelines, since such pretesting consumes sample information.","Editorial inference: The same variance-correction mechanism—replacing the independent-component variance with a $\\tau_i$-dependent expression built from $E[\\xi^4]/(p(p+2))$—could plausibly convert other quadratic-form-based tests from independent component models to elliptical models, not only the Li-Chen statistic.","Editorial inference: The S&P 500 finding, if it replicates on other assets and time windows, suggests the test could serve as a change-point diagnostic for portfolio covariance structure without assuming normality or sparsity.","Editorial inference: Since the variance formula involves only second- and fourth-order traces, one could attempt a version based on rank-based or robust scatter estimates, provided the needed quadratic-form moments can be controlled under ellipticity."],"forward_implications":["The same test is valid under both independent component models and generalized elliptical models, so a practitioner does not need to decide between the two model classes before testing.","Asymptotic level control and power consistency hold without sparsity assumptions and without any explicit growth condition linking dimension $p$ to the sample sizes, covering regimes with $p$ much larger than $n$.","Under the null the plug-in estimator $\\hat\\sigma_{n,0}$ is ratio-consistent, so the decision rule uses only a standard normal quantile and needs no bootstrap or tuning parameters.","For Gaussian data the variance formula reduces to the Li-Chen variance, so the new result contains the classical independent-component case; for other elliptical radii the variances differ, which is why independent-component-based tests mis-calibrate under ellipticity."],"supporting_citations":[{"why":"Supplies the original two-sample covariance test statistic $T_n$ and the independent-component central limit theorem that this paper extends to elliptical models.","marker":"Li and Chen (2012)"},{"why":"Provides the elliptical moment formulas (their Lemma A.1) used for the variance kernel and illustrates assumption (A1).","marker":"Hu et al. (2019)"},{"why":"Gives moment identities for elliptical quadratic forms (their Lemma D.2) that the proof uses, together with the bootstrap framework for spectral statistics under ellipticity.","marker":"Wang and Lopes (2023)"},{"why":"Supplies Lemma D.4, the high-moment bound for centered quadratic forms that controls the Lindeberg condition under assumption (A1).","marker":"Wang and Lopes (2025)"},{"why":"Theorem 18.1 is the martingale central limit theorem applied to the martingale-difference decomposition of $T_n$.","marker":"Billingsley (1999)"}],"fun_headline_variants":["Covariance test for elliptical data: CLT proven","Two-sample covariance test: first with theory for ellipticals","Elliptical covariance testing: CLT enables level and power control","No-sparsity covariance test now valid for elliptical data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on assumption (A1), which requires $E|(\\xi^2-p)/\\sqrt{p}|^8 = o(p^2)$ for the elliptical radius $\\xi$, and on assumption (A3), which requires mixed fourth-order traces such as $\\operatorname{tr}(\\Sigma_i\\Sigma_j\\Sigma_k\\Sigma_l)$ to be of smaller order than the product of the corresponding squared traces; the paper states both conditions but does not verify them in its S&P 500 application.","fun_headline_variants_meta":{"raw":{"variants":["Covariance test for elliptical data: CLT proven","Two-sample covariance test: first with theory for ellipticals","Elliptical covariance testing: CLT enables level and power control","No-sparsity covariance test now valid for elliptical data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1248,"prompt_tokens":938,"completion_tokens":310,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":242}},"tokens_in":554,"tokens_out":310,"duration_ms":3764,"temperature":1.0,"reasoning_tokens":242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:27:49.956766+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate two elliptical samples with $p$ large and $n_1=n_2$ moderately large, using covariance matrices with a few dominating spikes, for instance $\\Sigma = \\operatorname{diag}(p, 1, \\ldots, 1)$, so that $\\operatorname{tr}(\\Sigma^4)$ is comparable to $\\operatorname{tr}^2(\\Sigma^2)$ and assumption (A3) fails, and inspect whether $(T_n - \\|\\Sigma_{n,1}-\\Sigma_{n,2}\\|_F^2)/\\sigma_n$ is approximately normal. Alternatively, use an elliptical radius whose squared law is heavier-tailed than the eighth-moment condition permits, such as a scaled $\\chi^2_p$ with few degrees of freedom, and record the empirical rejection rate of the proposed test under the null.","supporting_citations":[{"cited_title":"Lopes , Miles E M","cited_arxiv_id":null,"evidence_quote":"Gives moment identities for elliptical quadratic forms (their Lemma D.2) that the proof uses, together with the bootstrap framework for spectral statistics under ellipticity."},{"cited_title":"Lopes , Miles E M","cited_arxiv_id":null,"evidence_quote":"Supplies Lemma D.4, the high-moment bound for centered quadratic forms that controls the Lindeberg condition under assumption (A1)."},{"cited_title":"( 1999 )","cited_arxiv_id":null,"evidence_quote":"Theorem 18.1 is the martingale central limit theorem applied to the martingale-difference decomposition of $T_n$."}],"review_version":1}