{"id":"277c4e43-b371-403c-9a93-a9bbdcc2fe6d","arxiv_id":"2506.06550","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A hybrid two-sample covariance test combines a Frobenius-norm U-statistic with leading-eigenvalue statistics via Fisher's method, justified by a new joint central limit theorem.","lead":"This paper proposes a new statistical test that checks whether two high-dimensional datasets have the same overall spread and structure. It combines two existing detection strategies so that it can spot both small dense changes and isolated large changes in the covariance matrix.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The manuscript does not itself prove the asymptotic-independence step: Lemma 4.5 is omitted and the vanishing of the cross terms C3–C5 in Lemma 4.6, exactly the terms that make the covariance in (4.17) zero, is only asserted as analogous.","rationale":"The reader's weakest_assumption is A4/A5, the supercritical-spike condition. I agree that this is a genuine scope limitation: if the population has no supercritical spike, the leading-eigenvalue statistic is not Gaussian and the Fisher combination is not justified; the paper is explicit about this. But it is a stated assumption, not a hidden flaw. The more load-bearing problem for the central theorem is verification: Theorem 2.1/2.3 rests on Lemma 4.2, which rests on Lemma 4.3, which rests on Lemma 4.4. In the proof of Lemma 4.4, after reducing to a martingale CLT, the cross-covariance (4.17) is decomposed into C1-C5. Only C1 and C2 are proved; C3-C5, including the term C5 involving the cross-sample matrix G^(1,2), are dismissed as analogous. Since T_{n,1} contains the cross-sample U-statistic C_n, the covariance between T_{n,1} and the second-sample quadratic forms is not ruled out by sample independence; the omitted terms are exactly what must vanish for asymptotic independence. Lemma 4.5, the high-probability event B_n, is also omitted. These are not cosmetic: without them the manuscript text does not establish the key independence claim. The reader's verdict of CONDITIONAL is therefore appropriate, with no change needed.","tokens_in":35560,"tokens_out":10447,"duration_ms":112518,"concrete_test":"Independently derive the omitted bounds for C3,n, C4,n and C5,n in Lemma 4.6 from the martingale representation (4.15)-(4.16) and Lemma A.1, following the pattern used for C1,n and C2,n. If any of these three terms is not o_P(1), then the conditional covariance in (4.17) is nonzero and the claimed asymptotic independence of T_{n,1} and the leading eigenvalues, and hence the chi-square calibration of T_{n,FC}, is unsupported. This derivation is the decisive check; a finite-sample simulation of the covariance of T_{n,1} and T_{n,2} can only be supplementary.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorems 2.1 and 2.3) is that the Frobenius-norm U-statistic and the leading sample eigenvalues are asymptotically independent, which makes the Fisher combination chi-square. The proof chain is Lemma 4.1 -> Lemma 4.2 -> Lemma 4.3 -> Lemma 4.4 -> martingale CLT. In the proof of Lemma 4.4, all cross terms in the conditional covariance (4.17) must vanish. Equation (4.18) writes this covariance as a sum of five terms C1-C5. Only C1,n and C2,n are actually bounded in Lemma 4.6; C3,n, C4,n and C5,n, the corresponding terms for the second sample, where C5,n contains the cross-sample matrix G^(1,2)_(n,n), are declared analogous and their proofs omitted. These terms are not automatically zero by the independence of the two samples, because T_{n,1} itself contains the cross-sample statistic C_n. The paper also omits the proof of Lemma 4.5, the high-probability event B_n used before truncation and before the martingale representation. Thus the manuscript text does not itself close the argument for the asymptotic independence that is the foundation of the new test.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-sample test for equality of p-dimensional covariance matrices in the high-dimensional regime p/n -> y in (0,infinity). The test statistic T_{n,FC} combines, via Fisher's method, a Frobenius-norm U-statistic of Li and Chen (2012) with a leading-eigenvalue statistic of Zhang et al. (2022). The central theoretical result (Theorem 2.1) states that under the null hypothesis and the spiked-eigenvalue assumptions (A1)-(A4), the two components are asymptotically standard normal and independent, so T_{n,FC} is asymptotically chi-squared with 4 degrees of freedom. A multi-spike extension (Theorems 2.3-2.4) is also provided, and power consistency is claimed for dense alternatives (large Frobenius difference) and sparse alternatives (large gap in leading eigenvalues). The proofs are based on martingale central limit theorems. However, several load-bearing steps are omitted: Lemma 4.5 (the high-probability event) is stated without proof, the cross-term bounds C3-C5 in Lemma 4.6 are declared analogous and not proved, and the consistency of the kurtosis estimator in the p/n in (0,infinity) regime is asserted without proof. The simulation study compares the new test with existing methods under normal, t7, and Laplace data.","tokens_in":35816,"tokens_out":9923,"duration_ms":100590,"significance":"If the missing proofs can be supplied, the result is a useful contribution: it addresses a question left open in Zhang et al. (2022) by establishing asymptotic independence between the Frobenius-norm U-statistic and the leading eigenvalues, and it yields a simple Fisher-combined test that is sensitive to both dense and sparse alternatives. The multi-eigenvalue extension is natural, and the paper is transparent about which results are imported from Li and Chen (2012) and Zhang et al. (2022) and about the spiked-eigenvalue scope. The main reservations are completeness: the central independence claim is not closed within the manuscript, and the kurtosis consistency extension is asserted without proof. These issues are fixable in revision.","major_comments":[{"comment":"Lemma 4.5, the high-probability event B_n, is stated but its proof is omitted with the explanation that it is 'similar to proof of Lemma 4.2 in Zhang et al. (2022), which can be found in the supplement to their paper.' This event is used before truncation and in the martingale representation: the indicator functions in (4.10) and the bounds in (4.23) depend on it. Since the central theorem rests on the martingale CLT for the truncated variables, the proof of Lemma 4.5 must be included in the manuscript or in a supplement, not merely referenced.","section":"Section 4.5 (Lemma 4.5)"},{"comment":"The off-diagonal covariance between the Frobenius statistic and the eigenvalue statistic is written in (4.18) as a sum of five cross terms C1,n through C5,n, and asymptotic independence requires each term to be o_P(1). The proof establishes only C1,n and C2,n for j=1; C3,n, C4,n, and C5,n are declared 'analogous' and omitted. In particular, C5,n contains the cross-sample matrix G^(1,2)_{n,n} built from the first sample, so it is not an immediate consequence of the independence of the two samples; it is precisely the term coupling the second-sample eigenvalue statistic with the first-sample U-statistic. The proof of Lemma 4.6 must be completed for all five terms and for all j=1,...,K.","section":"Section 4.5.1 (Lemma 4.6, Eq. (4.18))"},{"comment":"The consistency of the kurtosis estimator \\hat\\gamma_{4,n} is stated for p/n in (0,infinity), while the cited result of Dörnemann and Dette (2024) is only for p/n in (0,1). The sentence 'A closer examination of the proof reveals that the consistency of this estimator remains valid in the case p/n in (0,infinity)' is an unproved extension. Since Lemma 4.1 uses \\hat\\gamma_{4,n} to estimate the normalization of both components of the test statistic, this is load-bearing. Please provide the proof or a precise reference to a theorem covering the full p/n in (0,infinity) regime.","section":"Appendix A.2 (Lemma A.4(d))"},{"comment":"The reduction from the leading sample eigenvalues to the quadratic forms Q^(i)_{j,n} is made by citing 'identity (S.3.9)' and 'similar arguments to (S.4.4)' in the supplement to Zhang et al. (2022), without stating those identities in the paper. This reduction is an essential step in the proof of Theorem 2.3. Please state the identities explicitly and either prove them or provide a self-contained derivation in the appendix.","section":"Section 4.3 (proof of Lemma 4.2)"}],"minor_comments":[{"comment":"The acronym 'LDD' is used in Figures 1 and 2 without being defined; specify that it denotes the new test (2.2), and similarly define LDD(1) and LDD(3) in Figures 3-5.","section":"Section 3 (figure captions)"},{"comment":"There are several typos that impede reading, including 'Due the the CLTs' at the start of the proof of Lemma 4.4 and 'there exists there exists' in the proof of Lemma 4.6; please correct them throughout.","section":"Section 4.5 and 4.5.1"},{"comment":"The estimator \\hat\\sigma^2_{n,1} is defined as 4/n^2 (B^(1)_n+B^(2)_n)^2, whereas in Section 2.1 it is described as an estimator of Var(B^(1)_n+B^(2)_n-2C_n); state explicitly that this is the estimator under H0 and how the cross term C_n enters under H0.","section":"Appendix A.2"},{"comment":"The simulations with t7-distributed data do not satisfy Assumption (A2), which requires a finite eighth moment; please state clearly that these results are robustness checks outside the theorem's assumptions, rather than presenting them on the same footing as the normal and Laplace cases.","section":"Section 3 and Remark 2.1"}],"recommendation":"major_revision","confidential_remarks":"The main issue is completeness of the proof of the asymptotic-independence claim, which is the foundation of the Fisher combination. I recommend asking the authors to provide a supplementary appendix with full proofs of Lemma 4.5, Lemma 4.6 (including C3-C5), and the extension of kurtosis consistency to p/n in (0,infinity). The paper also relies heavily on the supplement of Zhang et al. (2022) for several identities; if the journal requires self-containedness, the authors should reproduce those identities. The scope is restricted to supercritical spiked alternatives, and this should be prominently stated in the abstract or introduction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, useful paper. The joint CLT between a Frobenius-norm U-statistic and leading sample eigenvalues is genuinely new, and the Fisher-combined test is a smart, practical payoff. The simulations are thorough and include heavy-tailed settings.\n\nWhat's new: Theorem 2.1/2.3 establish asymptotic independence, i.e., the vector (T_{n,1}, T_{n,2}) is asymptotically N(0,I). This was open after Zhang et al. (2022). The proof strategy is standard martingale CLT, and the paper gives a detailed outline. Estimators for unknown variances are provided and shown consistent, including a kurtosis estimator claimed valid for all p/n in (0,∞). The multi-eigenvalue generalization is a natural extension and handled cleanly.\n\nWhere it gets soft: the proof of the central independence claim is not fully contained. Lemma 4.5, which defines the high-probability event B_n used before truncation, is stated and then omitted with a reference to Zhang et al. (2022). More important, in Lemma 4.6 the cross terms C3, C4, and C5—the second-sample terms, one of which involves the cross-sample matrix G^(1,2)—are declared analogous to C1 and C2 and their proofs skipped. Since the whole point is to show the conditional covariance in (4.17) vanishes, this is the load-bearing step. The analogy is plausible but not automatic; C5 in particular involves a term that couples the two samples, and that coupling is exactly what the independence claim is about. A referee should ask for those proofs or a supplement.\n\nSecond soft spot: the theory only covers supercritical spiked eigenvalues (A4/A5). The test's null distribution is not justified when there are no spikes above the BBP threshold. That is acknowledged and framed as a limitation, but it means the test is for a specific alternative regime, not a universal two-sample covariance test. The paper is honest about this; the simulations do not cover the non-spiked regime.\n\nThird, the consistency of the kurtosis estimator for p/n in (0,∞) is asserted by \"closer examination of the proof\" with reference to Dörnemann and Dette (2024). That's acceptable, but a referee may want a few lines of justification.\n\nThe citation pattern is fine: they build directly on Li and Chen (2012) and Zhang et al. (2022) and credit them properly. Overall, the core result is new and the test is well motivated and well tested. The missing proofs are the main issue, but they appear to be the standard sort of omitted technical details, not a hint of a wrong claim. The paper deserves a serious referee; I would send it out and ask for a supplement with Lemma 4.5 and the C3–C5 proofs.","headline":"Solid new joint CLT and a practical hybrid test; the main proofs have omitted steps that a referee should push to have filled in.","tokens_in":36357,"tokens_out":2094,"would_cite":true,"duration_ms":20747,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A18","60F05","62H15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A new two-sample test for high-dimensional covariance matrices combines a Frobenius-norm U-statistic with leading sample eigenvalues; the paper proves the two components are asymptotically independent under the null hypothesis.","keywords":["two-sample test","covariance matrices","high-dimensional statistics","spiked covariance model","leading eigenvalues","U-statistics","p-value combination","central limit theorem"],"falsifier":"Simulate, say, p equal to 1000 and n equal to 100, with i.i.d. standard normal entries and two samples drawn from the same covariance matrix that has one leading eigenvalue well above the bulk, so the paper's assumptions hold. Repeat many times and estimate the joint distribution of (T_{n,1}, T_{n,2}); if the empirical covariance matrix is not close to the identity or the marginals deviate systematically from standard normal, Theorem 2.1 would be contradicted.","tokens_in":35344,"feed_emoji":"📊","tokens_out":6919,"duration_ms":67446,"temperature":0.7,"pith_summary":"This paper proposes a single two-sample test for covariance matrices that works when the dimension grows proportionally to the sample size. The test combines two detectors: a U-statistic estimator of the squared Frobenius norm of the difference between the population covariance matrices, which sees dense changes, and the difference of leading sample eigenvalues, which sees sparse changes. The paper's main theoretical result is a joint central limit theorem showing these two components are asymptotically independent and standard normal under the null hypothesis, provided the leading population eigenvalues are supercritical. That independence justifies combining the two p-values through a classical p-value combination method, yielding a test statistic with an asymptotic chi-squared distribution with 4 degrees of freedom, so the level can be controlled asymptotically. A sympathetic reader would care because the test is consistent against both dense and sparse alternatives while existing methods tend to be specialized to one type.","feed_headline":"Frobenius and top-eigenvalue tests are asymptotically independent","feed_subtitle":"A hybrid covariance test merges both detectors, keeping level control and catching dense or sparse alternatives.","key_machinery":"The central object is the pair (T_{n,1}, T_{n,2}). T_{n,1} is a U-statistic estimator of the trace of the squared difference of the two population covariance matrices, built from sums over distinct indices, standardized by a consistent variance estimator. T_{n,2} is the scaled difference of the two leading sample eigenvalues. The argument rests on identifying supercritical eigenvalues, meaning population eigenvalues strong enough to separate from the bulk spectrum, and on a martingale representation of both statistics; the joint central limit theorem follows once the cross-covariance terms are shown to vanish. For the multi-eigenvalue extension, the covariance matrix of the normalized leading eigenvalue differences is estimated from sample eigenvector components and kurtosis estimators.","core_discovery":"The paper establishes that, under the null hypothesis Sigma_n^(1) equals Sigma_n^(2) and under assumptions (A1)-(A4), the vector (T_{n,1}, T_{n,2}) consisting of the centered and scaled Frobenius-norm U-statistic and the scaled leading-eigenvalue difference converges in distribution to N_2(0, I_2); in particular, T_{n,1} and T_{n,2} are asymptotically independent. The proof works by replacing the leading sample eigenvalues with associated random quadratic forms, reducing to the centered-data case, and then applying a martingale central limit theorem whose key step is showing that the cross-covariance terms vanish. Consequently, the combined statistic T_{n,FC} converges to a chi-squared distribution with 4 degrees of freedom under the null hypothesis, and the test is consistent whenever either the squared Frobenius difference of the population covariance matrices diverges or the scaled leading-eigenvalue drift diverges. The same structure extends to K leading eigenvalues through a generalized statistic and a multivariate normal limit for the eigenvalue differences.","pith_inferences":["A natural stress test the authors did not run: set both populations to have no supercritical eigenvalue and compare the empirical rejection rates of the combined test to the nominal level; the theory here predicts breakdown because the leading-eigenvalue component has no proven central limit theorem, but quantifying the distortion would map the method's applicability boundary.","The independence mechanism may extend to other pairs of statistics whose first component is a smooth spectral U-statistic and whose second is a spiked eigenvalue; if the vanishing cross-covariance argument generalizes, the same p-value combination strategy could equip other high-dimensional detectors.","In practice, estimating kurtosis and eigenvector alignments from the sample introduces finite-sample sensitivity, so the chi-squared approximation likely degrades when the estimated spike variances are near the boundary of the supercritical regime, a region the paper's appendix probes only for eigenvalue multiplicity violations."],"forward_implications":["If the joint central limit theorem is correct, the combined test statistic T_{n,FC} has an asymptotic chi-squared null distribution with 4 degrees of freedom, so critical values are available without simulation under the null hypothesis.","The same reasoning yields a multi-eigenvalue version whose combined statistic again has an asymptotic chi-squared distribution with 4 degrees of freedom, provided the leading population eigenvalues are supercritical and separated.","Under alternatives with a diverging Frobenius difference or a diverging scaled leading-eigenvalue drift, the rejection probability tends to 1, so the test is consistent for both dense and sparse signal classes.","The asymptotic independence means the two detectors can be reported separately and combined without requiring a user to choose which type of alternative to optimize for.","The variance estimators are ratio-consistent under the null and under the alternative, so the level guarantee and the consistency result hold simultaneously."],"supporting_citations":[{"why":"Supplies the U-statistic estimator of the squared Frobenius norm and its marginal central limit theorem, which the new test uses as its dense detector.","marker":"Li and Chen (2012)"},{"why":"Provides the leading-eigenvalue central limit theorem and the quadratic-form machinery, and raises the independence question that this paper answers.","marker":"Zhang et al. (2022)"},{"why":"Establishes the phase transition that defines which population eigenvalues are supercritical and therefore visible in the leading sample eigenvalues.","marker":"Baik et al. (2005)"},{"why":"Gives the central limit theorem for spiked sample eigenvalues used to justify the marginal behavior of the leading-eigenvalue statistic.","marker":"Bai and Yao (2008)"},{"why":"Defines the p-value combination rule whose chi-squared null distribution is used for the combined statistic.","marker":"Fisher (1950)"},{"why":"Supports the combination of independent tests that underlies the asymptotic level control of the combined statistic.","marker":"Littell and Folks (1971)"},{"why":"Provides the martingale central limit theorem used in the proof of asymptotic independence.","marker":"Hall and Heyde (1980)"},{"why":"Supplies the consistency of estimated supercritical population eigenvalues used in the variance estimators.","marker":"Bai and Ding (2012)"}],"fun_headline_variants":["Frobenius and top-eigenvalue tests are asymptotically independent in high dimensions","Fisher's method merges two independent covariance tests for high-dimensional data","Joint CLT: Frobenius norm and top eigenvalue are independent under null","Hybrid test: U-statistic and leading eigenvalue independent in the limit","Covariance test catches both sparse and dense alternatives via Fisher"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that both populations have at least one leading eigenvalue strong enough to be supercritical, meaning separated from the bulk spectrum, so that the sample leading eigenvalue is asymptotically Gaussian; without such an outlier, the null distribution of the eigenvalue component is not established.","fun_headline_variants_meta":{"raw":{"variants":["Frobenius and top-eigenvalue tests are asymptotically independent in high dimensions","Fisher's method merges two independent covariance tests for high-dimensional data","Joint CLT: Frobenius norm and top eigenvalue are independent under null","Hybrid test: U-statistic and leading eigenvalue independent in the limit","Covariance test catches both sparse and dense alternatives via Fisher"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000422,"raw_usage":{"total_tokens":2154,"prompt_tokens":920,"completion_tokens":1234,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1137}},"tokens_in":536,"tokens_out":1234,"duration_ms":12273,"temperature":1.0,"reasoning_tokens":1137,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:54:02.432298+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate, say, p equal to 1000 and n equal to 100, with i.i.d. standard normal entries and two samples drawn from the same covariance matrix that has one leading eigenvalue well above the bulk, so the paper's assumptions hold. Repeat many times and estimate the joint distribution of (T_{n,1}, T_{n,2}); if the empirical covariance matrix is not close to the identity or the marginals deviate systematically from standard normal, Theorem 2.1 would be contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the leading-eigenvalue central limit theorem and the quadratic-form machinery, and raises the independence question that this paper answers."},{"cited_title":"B., and P \\'e ch \\'e , S","cited_arxiv_id":null,"evidence_quote":"Establishes the phase transition that defines which population eigenvalues are supercritical and therefore visible in the leading sample eigenvalues."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the p-value combination rule whose chi-squared null distribution is used for the combined statistic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the combination of independent tests that underlies the asymptotic level control of the combined statistic."},{"cited_title":"and Heyde, C","cited_arxiv_id":null,"evidence_quote":"Provides the martingale central limit theorem used in the proof of asymptotic independence."},{"cited_title":"and Ding, X","cited_arxiv_id":null,"evidence_quote":"Supplies the consistency of estimated supercritical population eigenvalues used in the variance estimators."}],"review_version":1}