{"id":"ecce8171-5b55-405f-bbfe-4002e4c673ee","arxiv_id":"2412.20019","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A Gaussian-replacement universal bootstrap is shown to be consistent for operator-norm spectral statistics when p/n is bounded or diverges to infinity, with no eigenvalue-decay assumptions.","lead":"This paper proposes replacing observed data with Gaussian draws that share the population covariance, then bootstrapping the operator norm of the sample covariance minus the hypothesized matrix, to get valid tests even when dimension p is much larger than sample size n. If the underlying random matrix universality holds, this gives a practical way to do covariance inference without assuming fast eigenvalue decay.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2 is not derivable from the marginal edge universality stated in Theorem 3.1; the operator norm is a joint extreme event, and the equal-endpoint case is left unresolved.","rationale":"The reader's weakest_assumption is Assumption 3, but I do not think that is the most load-bearing issue. Assumption 3 restricts the pair (Sigma, Sigma0), yet the main covariance-testing application sets Sigma = Sigma0 under H0, so commutativity holds automatically; the advertised 'no structural assumptions' is about eigenvalue decay, not pair commutativity. The more serious concern is the inference from Theorem 3.1 to Theorem 3.2: the operator norm ball event is the intersection of events on the two extreme eigenvalues, and the stated universality theorem gives only marginal CDFs. Since the central theorem's proof is entirely in the supplement, this gap may be closed there, but it is not visible in the main text. The concrete error in Corollary 1, centering at p instead of at (sqrt(p) + sqrt(n))^2, makes the gamma > 1 normalization suspect and independently supports the conditional verdict. I therefore keep the reader's CONDITIONAL verdict: the correct check is a careful reading of the supplement's proof of Theorem 3.2 and a re-derivation of the gamma > 1 edge scaling. I do not see grounds to accept unconditionally or to reject outright; the burden is on the deferred proof.","tokens_in":18834,"tokens_out":20580,"duration_ms":227076,"concrete_test":"Locate the proof of Theorem 3.2 in the supplement and identify the step that bounds rho_n(Sigma0) = sup_t |P(lambda1 <= t, -lambda_p <= t) - P_Gau(lambda1 <= t, -lambda_p <= t)|. Verify that this step either establishes joint universality for the pair of edge eigenvalues or proves that one edge dominates on the relevant t-range, including the |E+| = |E-| case. If the step only applies the marginal bounds of Theorem 3.1, the central claim is not supported. As a secondary check, recompute Corollary 1's centering for gamma > 1: the Wishart edge is (sqrt(p) + sqrt(n))^2, so mu = p needs reconciliation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"T = ||hatSigma - Sigma0|| = max(lambda1(M_n), -lambda_p(M_n)) with M_n = hatSigma - Sigma0, so P(T <= t) = P(lambda1(M_n) <= t, -lambda_p(M_n) <= t). This is a joint event involving both extreme eigenvalues. Theorem 3.1 establishes universality only for the marginal CDF of lambda1 and, when gamma = 1, separately for lambda_p; no statement is made about their joint law or about lambda_p when gamma > 1. Marginal universality does not imply joint universality: correlation between the two edges can differ between X and the Gaussian surrogate and change the CDF of their maximum. In the gamma > 1 regime, Theorem 3.1 does not even provide lambda_p universality, so the reduction of rho_n(Sigma0) to a single edge must be justified, e.g., by showing the opposite edge is deterministic on the relevant scale. None of this appears in the main text, and Theorem 3.2's proof is deferred to the supplement. Corollary 1 compounds the concern: for gamma > 1, Sigma = I, R = 0, the Wishart edge is (sqrt(p) + sqrt(n))^2, so the centering mu = p in Corollary 1 is off by 2 sqrt(pn), which diverges against the claimed fluctuation scale p^{1/2} n^{-1/6}. This is a concrete sign that the gamma > 1 edge normalization needs checking.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 'universal bootstrap' for spectral statistics of sample covariance matrices: replace the original data by Gaussian observations with the same covariance and use the resulting quantiles of statistics such as T = ||hat Sigma - Sigma0||_op. Theorems 3.1–3.3 claim edge universality and uniform Gaussian-approximation bounds for the operator-norm ball and its generalizations, in both proportional p/n and ultra-high-dimensional p/n -> infinity regimes. The paper also gives power analysis for covariance tests and simultaneous confidence intervals, supported by simulations and a climate data application.","tokens_in":19086,"tokens_out":9038,"duration_ms":92858,"significance":"If the deferred proofs are correct, the universal bootstrap is an appealing and potentially important idea: it would provide a way to obtain distributional approximations for extreme spectral statistics in regimes where empirical and multiplier bootstrap are known to fail. The paper is explicit about algorithms (Algorithm 1 and 2), reports extensive simulation evidence, and includes a real-data demonstration. The claim that the asymptotic distribution depends only on the first two moments for the operator-norm-based statistic, in contrast to Frobenius/supremum norms, is a noteworthy conceptual contribution. The main theorems are, however, not fully checkable from the preprint body because all proofs are in the supplementary material, and the issues described below affect the central claims.","major_comments":[{"comment":"The centering mu = p is not the correct deterministic location of the largest eigenvalue of Z^T Z when p/n -> infinity and Z is an n x p standard normal matrix. The Wishart edge is approximately (sqrt(p)+sqrt(n))^2 = p + 2 sqrt(pn) + n, so centering at p leaves a deterministic bias of order sqrt(pn). Compared with the claimed fluctuation scale sigma = p^{1/2} n^{-1/6}, this bias tends to infinity like n^{2/3}. Thus (10) cannot follow from Theorem 3.1 as stated, and the ultra-high-dimensional Tracy-Widom assertion needs either a corrected centering (e.g., (sqrt(p)+sqrt(n))^2) or a different normalization.","section":"Section 3.1, Corollary 1 (Eq. (10))"},{"comment":"The statistic T = ||hat Sigma - Sigma0||_op is the maximum of lambda_1(M_n) and -lambda_p(M_n), a joint event involving both extreme eigenvalues. Theorem 3.1 provides universality only for the marginal CDF of lambda_1, and for gamma > 1 it gives no statement about lambda_p. Uniform Gaussian approximation for the CDF of the maximum requires either a joint universality result for the pair of extreme eigenvalues, or a deterministic argument showing that the opposite edge is negligible on the n^{-2/3} scale. No such argument appears in the main text, and since Theorem 3.2 is the central bootstrap-consistency claim, this gap is load-bearing. The proof in the supplement must be made available and its key step stated in the main text.","section":"Section 3.2, Theorem 3.2"},{"comment":"Assumption 3 requires Sigma R = R Sigma with R = -Sigma0, i.e., Sigma and Sigma0 share all eigenvectors. This is a structural condition on the pair of matrices, not merely an eigenvalue-decay or low-rank condition. The abstract's statement that the method works 'without requiring structural assumptions on the population covariance matrix' is therefore too strong. The assumption is automatic for the null hypothesis H0: Sigma = Sigma0, but for general pairs (Sigma, Sigma0) and for the power analysis in Section 4.1 it restricts the model to common-eigenvector structures. The paper should explicitly delineate the class of covariance pairs covered by the theory.","section":"Assumption 3 and abstract"},{"comment":"Even if the joint-edge issue is resolved, the rate rho_n(Sigma0) <= n^{-delta} is left as an unspecified positive delta. Since the bootstrap quantile error in (13) is stated as O(rho_n(Sigma0) + B^{-1/2}), an explicit delta (or at least an explicit modulus) would be needed to assess whether the result is non-vacuous for practical sample sizes. This is not a fatal objection, but it should be addressed in a revision.","section":"Section 3.2, Eq. (11)-(13)"}],"minor_comments":[{"comment":"The reference 'Alex, B., Erdős, L., ... ' appears to be a garbled version of a standard isotropic local law paper; please correct the author name and bibliographic details.","section":"References"},{"comment":"The sentence 'we examine the following distributions: Gaussian distribution, Uniform and t-distributions, Gaussian and uniform distributions , and Gaussian and t-distributions .' is incomplete and unclear; please specify exactly which of the first n/2 and last n/2 rows receive which entry distribution in each scenario.","section":"Section 5.1"},{"comment":"The displayed formula for kappa contains a double fraction that is hard to read; please rewrite it with explicit parentheses and define all quantities (E_+, tilde m) in the statement of Theorem 4.1, not only by reference to Section 3.","section":"Theorem 4.1, Eq. (19)"},{"comment":"All proofs are deferred to the supplementary material. For a journal submission, please ensure the supplement is part of the review package and provide a short proof sketch or a precise statement of the key lemmas in the main text, especially for Theorem 3.2 and Theorem 3.3, which are the main contributions.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper contains an interesting idea and a substantial amount of work, but the erroneous centering in Corollary 1 and the missing joint-edge justification for Theorem 3.2 are serious enough that the central claims need to be re-examined before publication. Please ask the authors to make the supplementary proofs available and to clarify whether the wrong centering is a typo or reflects a deeper scaling issue in the ultra-high-dimensional edge normalization."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper deserves attention: the central idea is clean and genuinely new. A Gaussian-replacement bootstrap, justified by edge universality, is claimed to work for operator-norm spectral statistics even when p/n diverges and no eigen-decay is imposed. That goes beyond Han et al., Lopes et al., and Giessing, all of which need low effective rank or p much smaller than n. The generalized T_ExS class is useful, the power comparison between T, T_Roy and T_Com is a nice touch, and the simulations are honest and fairly extensive, showing size control across several heavy-tailed and non-identical designs.\n\nThe soft spots, in proportion. First, Assumption 3 (Sigma R = R Sigma) restricts the pair to share eigenvectors. That is a real structural condition, not just a technical footnote, and the abstract's 'no structural assumptions' overstates the actual scope. The non-commuting case is simply not addressed.\n\nSecond, the proof of Theorem 3.2 is entirely in the supplement, and the main text alone does not bridge the gap. T is the maximum of the largest and minus-smallest eigenvalue. Theorem 3.1 states marginal universality for lambda_1, and for lambda_p only in the proportional regime. Marginal universality does not automatically give joint universality of the two edges, and for gamma > 1 the negative edge is not even covered by the stated theorem. Unless the supplement shows that the opposite edge is deterministic on the relevant scale or establishes the joint law, the operator-norm bootstrap consistency does not follow from the displayed results. The equal-endpoint case in Assumption 4' is exactly where this bites hardest.\n\nThird, Corollary 1 looks wrong as written. For gamma > 1, the Wishart edge for lambda_1(Z^T Z) is (sqrt(p)+sqrt(n))^2, not p. The missing 2 sqrt(pn) term is much larger than the claimed fluctuation scale p^{1/2} n^{-1/6}, so Tracy-Widom convergence with that centering cannot hold. Unless a different normalization is hiding somewhere, this is a concrete error in the main text.\n\nThe simulations and the generalized universality idea suggest this is not a throwaway preprint. But the central theorem is not checkable from the body, the commutativity assumption narrows the advertised scope, and Corollary 1 should be fixed. A serious referee should look at it; if the supplementary proofs hold up, this is a meaningful contribution. If not, the edge cases need substantial rework.","headline":"Serious, novel attempt at Gaussian-replacement bootstrap consistency for operator-norm spectral statistics in the ultra-high-dimensional regime, but the main text has a likely normalization error and a gap between the stated edge-universality theorem and the bootstrap consistency claim.","tokens_in":19631,"tokens_out":4534,"would_cite":true,"duration_ms":50066,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H15","60B20","62G09","62E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a Gaussian-proxy bootstrap is uniformly consistent for spectral statistics such as the operator norm of the sample covariance minus a hypothesized matrix, even when p/n converges to a positive constant or diverges…","keywords":["universal bootstrap","spectral statistics","random matrix universality","operator norm","covariance testing","ultra-high-dimensional inference","deformed Marchenko-Pastur law","Tracy-Widom law"],"falsifier":"Run the universal bootstrap at level 0.05 for a non-commuting pair, e.g., $\\Sigma$ diagonal with eigenvalues evenly spaced on $[1,3]$ and $\\Sigma_0$ obtained by rotating a diagonal spectrum through a fixed angle, with $n=300$, $p=300$, and many replications. If the empirical rejection rate under $H_0$ does not approach 0.05 as $n$ and $p$ grow, the advertised scope without structural assumptions on $\\Sigma$ and $\\Sigma_0$ is false; a vanishing $\\rho_n(\\Sigma_0)$ in (12) is the precise quantity to check.","tokens_in":18593,"feed_emoji":"📊","tokens_out":8428,"duration_ms":77650,"temperature":0.7,"pith_summary":"High-dimensional covariance tests based on spectral statistics have lacked a usable bootstrap: standard Gaussian and multiplier bootstraps fail when p/n is not small. The paper claims that universality from random matrix theory fixes this: drawing Gaussian data with the same covariance matrix and recomputing the statistic gives a valid reference distribution for the operator norm of the sample covariance minus a hypothesized matrix. Its main theorem bounds the uniform Gaussian approximation error by $O(n^{-\\delta})$ and the bootstrap size error by $O(n^{-\\delta}+B^{-1/2})$, for $p/n$ converging to a nonzero constant or to infinity and without eigenvalue-decay assumptions. The same principle covers a broad class of extreme-singular-value statistics and yields sharp simultaneous confidence intervals for all linear functionals of the covariance. A sympathetic reader should care because genuinely high-dimensional covariance inference becomes computationally routine: one algorithm, no analytic distributional derivation.","feed_headline":"Bootstrap for spectral statistics valid as p/n grows without bound","feed_subtitle":"Replacing data with Gaussian draws of the same covariance gives valid spectral tests with no eigenvalue decay assumptions.","key_machinery":"The engine is a weighted anisotropic local law for matrices of the form $\\hat\\Sigma + R$, with the perturbation $R = -\\Sigma_0$. When $\\Sigma$ and $R$ commute (Assumption 3), the resolvent is controlled through the $z$-dependent covariance $\\Sigma(z) = z\\Sigma(zI_p - \\phi^{-1/2}R)^{-1}$ and the fixed point $\\tilde m(z)$ of the deformed Marchenko-Pastur equation (7). This local law pins the extreme eigenvalues of $\\phi^{-1/2}(\\hat\\Sigma+R)$ at scale $n^{-1/6}$ near deterministic edges $E_\\pm$, which yields edge universality for the largest and smallest eigenvalues (Theorem 3.1). The bootstrap in Algorithms 1 and 2 is the practical mechanism; the local law is what makes Gaussian substitution without loss of distributional accuracy possible.","core_discovery":"The central claim is that the distribution of the spectral statistic $T = \\|\\hat\\Sigma - \\Sigma_0\\|_{\\mathrm{op}}$ is asymptotically unchanged when the data's entry distribution is replaced by a Gaussian with the same covariance matrix, so the bootstrap can be computed by drawing $Y_1,\\dots,Y_n \\sim N(0,\\Sigma)$ and using the empirical quantiles of $T^{\\mathrm{ub}} = \\|\\hat\\Sigma^{\\mathrm{ub}} - \\Sigma_0\\|_{\\mathrm{op}}$. Theorem 3.2 states that for any admissible pair $(\\Sigma,\\Sigma_0)$ — nonnegative matrices satisfying boundedness, a bulk non-degeneracy condition, and the commutation $\\Sigma(-\\Sigma_0) = (-\\Sigma_0)\\Sigma$ — the uniform Gaussian approximation error $\\rho_n(\\Sigma_0) = \\sup_{t\\ge 0}|P(\\hat\\Sigma \\in B_{\\mathrm{op}}(\\Sigma_0,t)) - P(\\hat\\Sigma^{\\mathrm{ub}} \\in B_{\\mathrm{op}}(\\Sigma_0,t))|$ is $O(n^{-\\delta})$, and $\\sup_\\alpha |P(T \\ge \\hat q^{\\mathrm{ub}}) - \\alpha| = O(n^{-\\delta} + B^{-1/2})$. Theorem 3.3 extends the same statement to statistics built from several extreme singular values, and Theorem 4.2 converts it into simultaneous confidence intervals for $\\langle A,\\Sigma\\rangle$ over all $A$. As a byproduct, the paper derives the Tracy-Widom law for the largest sample-covariance eigenvalue in the ultra-high-dimensional regime for general entry distributions, where only the Gaussian case was previously known.","pith_inferences":["Editorial inference: if the commutation assumption can be relaxed to approximate commutation with a small commutator norm, the same bootstrap should extend to arbitrary covariance-testing pairs; the obstacle is a non-commuting analog of the deformed Marchenko-Pastur local law, which the paper does not address.","Editorial inference: because the operator norm is insensitive to fourth moments, tests calibrated by this bootstrap may remain valid for heavier-tailed data than the $t(12)$ used in simulations, as long as Assumption 1's moment bounds hold; this is testable by simulation with $t(5)$ or lognormal entries.","Editorial inference: the simultaneous confidence intervals of Theorem 4.2 give a direct route to portfolio- or factor-model diagnostics: any null hypothesis expressible as all linear functionals $\\langle A,\\Sigma\\rangle$ lying in prescribed ranges can be checked with one bootstrap computation, at cost independent of the number of $A$'s."],"forward_implications":["The universal bootstrap controls the type I error of operator-norm covariance tests uniformly in the level $\\alpha$, in both the proportional regime $p/n \\to c>0$ and the ultra-high-dimensional regime $p/n \\to \\infty$, without eigenvalue decay or low effective rank.","The combined statistic $T^{\\mathrm{Com}} = T^2/\\mathrm{tr}(\\Sigma_0) + (T^{\\mathrm{Roy}})^2$ inherits bootstrap consistency and is shown numerically to keep high power in both spike and white-noise alternatives.","Power analysis under the generalized spike model yields a sharp BBP-type phase transition: spikes above $\\kappa$ are detected with power tending to 1, while spikes below $\\kappa$ are asymptotically invisible to the test.","Sharp simultaneous confidence intervals for $\\langle A,\\Sigma\\rangle$ over all $A \\in \\mathbb{R}^{p\\times p}$ hold with error $O(n^{-\\delta} + d_{\\mathrm{Sp}}(\\widehat{\\mathrm{Sp}}(\\Sigma), \\mathrm{Sp}(\\Sigma)) + B^{-1/2})$, covering quadratic forms $c_1^\\top \\Sigma c_2$ as a special case.","The distribution of operator-norm spectral statistics depends only on the first two moments of the entries, in contrast to Frobenius- and supremum-norm statistics whose limiting laws require fourth-moment information."],"supporting_citations":[{"why":"Establishes the Tracy-Widom limit for the largest Wishart eigenvalue when $p/n \\to \\infty$ in the Gaussian case, the result whose general-entry version Corollary 1 provides.","marker":"Karoui (2003)"},{"why":"Supplies the global and local CLTs and local law for general sample covariance matrices when the dimension diverges, the ultra-high-dimensional techniques refined here.","marker":"Ding & Wang (2023)"},{"why":"Provides two-sample covariance tests in ultra-high dimension and local-law machinery that the paper adapts to the perturbed matrix $\\hat\\Sigma + R$.","marker":"Ding et al. (2024)"},{"why":"Develops anisotropic local laws for sample covariance matrices with bounded $p/n$, the framework extended to include the perturbation $R$.","marker":"Knowles & Yin (2017)"},{"why":"Demonstrates inconsistency of nonparametric and Gaussian-approximation bootstraps for spectral statistics in high dimension, the failure mode the universal bootstrap overcomes.","marker":"El Karoui & Purdom (2019)"},{"why":"Proves universality for the largest eigenvalue of sample covariance matrices with general population, a precedent for the edge universality used here.","marker":"Bao et al. (2015)"},{"why":"Introduces the generalized spike model and four-moment theorem that underpin the power analysis in Theorem 4.1.","marker":"Jiang & Bai (2021)"}],"fun_headline_variants":["Universal bootstrap for spectral statistics beyond Gaussian","Bootstrap for spectral stats valid when p/n diverges","Spectral bootstrap works without eigenvalue decay assumptions","Universal bootstrap for high-dimensional spectral statistics","Spectral bootstrap: universal across dimensions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the commuting-pair condition: the hypothesized covariance $\\Sigma_0$ and the true covariance $\\Sigma$ must share all eigenvectors; if a testing problem violates this, the paper gives no guarantee that the bootstrap size converges.","fun_headline_variants_meta":{"raw":{"variants":["Universal bootstrap for spectral statistics beyond Gaussian","Bootstrap for spectral stats valid when p/n diverges","Spectral bootstrap works without eigenvalue decay assumptions","Universal bootstrap for high-dimensional spectral statistics","Spectral bootstrap: universal across dimensions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001042,"raw_usage":{"total_tokens":4448,"prompt_tokens":1078,"completion_tokens":3370,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":3304}},"tokens_in":694,"tokens_out":3370,"duration_ms":24744,"temperature":1.0,"reasoning_tokens":3304,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:42:09.323620+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the universal bootstrap at level 0.05 for a non-commuting pair, e.g., $\\Sigma$ diagonal with eigenvalues evenly spaced on $[1,3]$ and $\\Sigma_0$ obtained by rotating a diagonal spectrum through a fixed angle, with $n=300$, $p=300$, and many replications. If the empirical rejection rate under $H_0$ does not approach 0.05 as $n$ and $p$ grow, the advertised scope without structural assumptions on $\\Sigma$ and $\\Sigma_0$ is false; a vanishing $\\rho_n(\\Sigma_0)$ in (12) is the precise quantity to check.","supporting_citations":[],"review_version":1}