{"id":"97d692e5-092a-4139-9ca1-2023fdfdcf34","arxiv_id":"2607.08388","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Separability of high-dimensional matrix covariances can be tested by Monte Carlo sphericity after separable MLE whitening, with an angular version robust to heavy-tailed elliptical laws and consistency under dense alternatives.","lead":"The paper gives a practical high-dimensional test for whether a matrix-valued covariance factors into row and column parts, by whitening with a separable MLE and testing sphericity. An angular (radial-normalized) variant keeps level control under heavy tails while retaining power, which matters for images, spectrograms, and other matrix data where full covariances are infeasible.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Consistency hinges on unproven pseudo-MLE rates under alternatives that the proofs treat as given.","rationale":"The reader correctly isolates the open pseudo-MLE rates as the weakest link. Exact level under equivariant whitening is rigorously established and does not depend on those rates; the angular construction is well-motivated and empirically robust. The consistency theorems, however, are the paper’s main theoretical advance beyond concurrent whitening work, and they rest squarely on (A4)/(B4). Because the authors themselves flag the rates as open and only support them numerically, the contribution remains conditional on either a proof of those rates or an explicit demotion of the consistency claims to “conditional on estimation rates.” No stronger internal inconsistency was found; the concern is precisely the one the reader identified.","tokens_in":36827,"tokens_out":716,"duration_ms":6584,"concrete_test":"Prove or disprove (A4) for the flip-flop MMLE under the rank-2 separable-component alternatives of Setup 2 (or the spiked family (5)): either obtain a high-probability Frobenius rate o_p(1) when pq~N and ρ(Σ) bounded away from zero, or exhibit a counter-example sequence of non-separable Σ_N where 1/√(pq)‖Σ̂₁⊗̃Σ̂₂−Σ₁⊗̃Σ₂‖_F stays bounded away from zero. If the latter occurs, Theorems 2–3 no longer guarantee consistency.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is finite-sample exact level (Corollaries 2 and Proposition 2) plus high-dimensional consistency of the MC tests against dense alternatives (Theorems 2 and 3). Level control is solid under matrix-affine equivariance of the MMLE. Consistency is not: both proofs reduce the problem to showing that the whitened (or angular) statistic concentrates around a positive separation that is inherited from Assumption (A5)/(B5), but that inheritance step requires the estimator to track the pseudo-true separable factors at rate (A4) 1/√(pq)‖Σ̂₁⊗̃Σ̂₂−Σ₁⊗̃Σ₂‖_F=o_p(1) (resp. operator-norm (B4)). The paper explicitly states that these rates under non-separable alternatives are open (Discussion §7; comparison to Franks et al. 2026 which only covers the separable case) and supplies only numerical box-plots in Appendix A.1. Without (A4)/(B4) the separation argument after whitening can collapse, so the consistency theorems are conditional on an unproved high-dimensional estimation claim that is load-bearing for the central theoretical contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes high-dimensional tests of covariance separability for matrix-variate data by whitening with the matrix-normal MMLE and testing sphericity of the whitened sample. After whitening, John’s, Nagao’s, and Ledoit–Wolf statistics coincide (Lemma 1); Monte Carlo calibration under a matrix-affine-equivariant estimator yields exact finite-sample level under the null (Theorem 1, Corollaries 1–2, Proposition 1). Consistency against dense Frobenius alternatives is claimed under Assumptions (A1)–(A5) for the elliptical statistic (Theorem 2) and under (B1)–(B5) for an angular (spatial-sign) version that projects whitened observations onto the unit sphere (Theorem 3, Proposition 3). Simulations and an acoustic phonetic application support power and improved robustness of the angular test under heavy-tailed elliptical laws with Gaussian calibration.","tokens_in":37215,"tokens_out":1136,"duration_ms":10157,"significance":"The problem is timely: high-dimensional matrix data make unstructured covariance estimation infeasible, so a computationally light, finite-sample valid test of separability is useful. The finite-sample level theory via matrix affine equivariance is clean and standard. Unifying classical sphericity statistics after MMLE whitening, the efficient inner-product form of the statistic, and the angular refinement for distributional robustness are genuine contributions. An R package with pre-tabulated quantiles is a practical strength. The consistency theorems would be a solid theoretical contribution if the load-bearing pseudo-MLE rates under alternatives were established or clearly demoted; as written they remain conditional.","major_comments":[{"comment":"Theorems 2 and 3 (and the separation arguments that follow Assumptions (A5)/(B5)) treat as given the high-dimensional rates (A4) 1/√(pq)‖Σ̂1⊗̃Σ̂2−Σ1⊗̃Σ2‖_F=o_p(1) and the stronger operator-norm (B4). The manuscript states these rates under non-separable alternatives are open (Discussion §7; comparison to Franks et al. 2026, which covers only the separable case) and supports them only by numerical box-plots in Appendix A.1. Without (A4)/(B4) the inheritance of dense separation after whitening can fail, so the consistency claims are conditional on an unproved estimation result that is load-bearing for the central theoretical contribution. Either prove the rates, replace them by weaker verifiable conditions, or restate Theorems 2–3 as conditional on (A4)/(B4) and move the unconditional claim to finite-sample level plus empirical power.","section":"§3.2 Assumptions (A4)–(A5); §5 (B4)–(B5); Theorems 2–3; Discussion §7; Appendix A.1"},{"comment":"The abstract and introduction state that the paper proves high-dimensional consistency under dense alternatives without flagging that the proofs rest on open rates for the pseudo-true separable factors. That overstates what is established. Soften the abstract/intro claims to match the conditional status of Theorems 2–3, or complete the rate analysis.","section":"Abstract; §1; Theorems 2–3"}],"minor_comments":[{"comment":"Abstract is truncated mid-sentence (“of the prop”). Fix before resubmission.","section":"Abstract"},{"comment":"Notation for the tensor product ˜b versus Kronecker product is introduced carefully but still dense; a short display equating vec((Σ1˜bΣ2)X)=(Σ2⊗Σ1)vec(X) early in §2 would help readers who work only with Kronecker notation.","section":"§2 Preliminaries"},{"comment":"Figure 3 panels for df=5 and small m show residual size inflation for the angular test; the text acknowledges this but could quantify how large N must be before Gaussian calibration is reliable under matrix-t.","section":"§6.1.2; Figure 3"},{"comment":"Comparison to Sung and Hoff (2026) is useful; make explicit in the main text (not only §4) which of their statistics coincides with the elliptical test after the latest revision, to avoid reader confusion.","section":"§4"},{"comment":"Proposition 1’s sample-size condition N≥⌈p/q+q/p⌉+2 is cited from Drton et al. (2021); a one-sentence reminder that this is almost necessary would help practitioners.","section":"Proposition 1"}],"recommendation":"major_revision","confidential_remarks":"The finite-sample level theory and the angular robustness idea are publishable contributions even if consistency remains conditional. I would not reject solely on the open rates, but the abstract currently overclaims; major revision that either proves (A4)/(B4) or clearly conditions the theorems and softens the abstract is the right bar. Overlap with Sung and Hoff is acknowledged and the angular test plus dense-alternative consistency angle differentiate the work sufficiently for this journal."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing to know is that this is a practical, well-executed method for testing covariance separability when p and q grow with N. After MMLE whitening they reduce the problem to a classical sphericity statistic; John, Nagao and Ledoit–Wolf collapse to the same functional, Monte Carlo gives exact finite-sample level under matrix-affine equivariance, and they add an angular (spatial-sign) version that keeps level under heavy-tailed elliptical data when you still calibrate with Gaussians. That angular step is the piece I would actually use.\n\nWhat is new relative to the concurrent Sung–Hoff whitening preprint is the unification after whitening, the dense-alternative consistency theorems, the angular robustness argument, and the careful positioning. They disclose the overlap at length and do not over-claim priority. The finite-sample theory (Theorem 1, Corollaries, Props. 1–2) is clean and standard. Simulations are extensive and match the theory: under Gaussian data both tests sit on level and have power; under matrix-t the elliptical version inflates while the angular version largely holds. The acoustic application is a sensible illustration.\n\nThe soft spot is exactly the one the stress-test flags. Consistency (Theorems 2–3) inherits separation from the dense Frobenius alternative only after the whitening estimator tracks the pseudo-true factors at rates (A4)/(B4). Those rates are assumed, not proved; the authors say so in the discussion and only show numerical box-plots under alternatives. That is a genuine gap, but it is load-bearing only for the asymptotic power claim, not for the level control or the practical procedure. I would not call the theorems empty—they are conditional results written carefully—but a referee will correctly ask for either a proof or a clearer demotion of the claim.\n\nPackage claim is incomplete in the text; if the R package with tabulated quantiles is real it should be released with the paper.\n\nThis is for people who actually fit separable models on matrix data and need a level-controlled check before they do. It deserves a serious referee. I would engage with it, cite the angular test, and treat the consistency statements as conditional until the rates are settled.","headline":"Solid high-d separability test with clean finite-sample level and a useful angular fix; consistency is real but conditional on open pseudo-MLE rates the authors flag themselves.","tokens_in":37817,"tokens_out":542,"would_cite":true,"duration_ms":7037,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H15","62H12","62F05","62E20"],"pacs":[],"model":"grok-4.5","headline":"Separability of a matrix covariance can be tested by whitening with the separable MLE and checking sphericity, with an angular version that resists heavy tails.","keywords":["matrix-variate data","separable covariance","Kronecker covariance","sphericity test","high-dimensional testing","elliptical distributions","matrix normal MLE","angular statistics"],"falsifier":"Construct a dense non-separable family (fixed positive Frobenius gap after scaling by 1/√(pq)) for which the separable MLE fails to converge in Frobenius or operator norm when N ~ pq, or show that the angular test's empirical level collapses under a non-elliptical law while the Gaussian Monte Carlo quantiles stay fixed.","tokens_in":37723,"feed_emoji":"📊","tokens_out":1046,"duration_ms":16915,"temperature":0.7,"pith_summary":"When each observation is a matrix, an unstructured covariance is a huge tensor that is often impossible to estimate or invert. Separability reduces that tensor to a product of two ordinary matrices, but imposing the structure blindly can bias later analysis. This paper shows how to test the structure without ever forming the full covariance: estimate the separable factors by maximum likelihood, whiten the sample, and test whether the whitened data are spherical. Monte Carlo under the whitened null gives exact finite-sample level when the calibration law is correct, and the authors prove high-dimensional consistency against dense departures from separability. Projecting each whitened observation onto the unit sphere yields an angular version that keeps nearly the same power while staying closer to nominal level under heavy-tailed elliptical data calibrated as Gaussian. The procedures are cubic in sample size and come with pre-tabulated quantiles, so the cost of checking separability stays comparable to fitting the separable model itself.","feed_headline":"Separability tests reduce to sphericity after MLE whitening","feed_subtitle":"An angular version keeps level under heavy tails without losing power when N scales with pq.","key_machinery":"Matrix whitening by the flip-flop MLE of the separable factors, after which John's, Nagao's and Ledoit–Wolf's sphericity statistics coincide (up to a constant absorbed by Monte Carlo); the angular version replaces the sample covariance of the whitened data by the spatial-sign covariance obtained after unit-sphere projection.","core_discovery":"Within the elliptical family, covariance separability is equivalent to sphericity after whitening by a unique matrix-affine-equivariant separable estimator (the matrix-normal MLE). The resulting Monte Carlo test based on the unified John/Nagao/Ledoit–Wolf statistic therefore has exact finite-sample level under the null, and is consistent in the high-dimensional regime N proportional to pq against dense alternatives measured in Frobenius distance. An angular refinement that radially normalizes the whitened observations inherits the same level property and consistency under slightly stronger conditioning assumptions, while empirically reducing sensitivity to radial misspecification.","pith_inferences":["The same whitening-plus-sphericity reduction is likely usable for other Kronecker-type structural hypotheses (e.g., multi-way separability or core-shrinkage models) once an equivariant estimator under the null is available.","Closing the open high-dimensional rate for the pseudo-MLE under alternatives would immediately upgrade both consistency theorems from conditional to fully rigorous.","Because the angular statistic discards only radial scale, it may remain approximately pivotal for mild departures from ellipticity that still preserve spherical directions after whitening—an empirical regime worth mapping.","The eigenvalue-dispersion view of the statistic suggests natural sparse-alternative competitors (largest eigenvalue after whitening) that the paper already contrasts and that could be hybridized with the angular projection."],"forward_implications":["Separability can be checked at roughly the cost of one separable fit, without ever storing the full pq-by-pq covariance.","When N is proportional to pq, dense non-separability is asymptotically detectable after MMLE whitening.","The angular test can be used under unknown elliptical tails with only a Gaussian Monte Carlo table, retaining nearly full power.","Pre-tabulated quantiles for each (p,q,N) make repeated testing and software packaging immediate.","Downstream matrix-variate procedures that rely on separability gain a reliable pre-test that respects the same computational budget."],"fun_headline_variants":["Whitening by separable MLE reduces separability to sphericity","High-dim covariance separability tested via post-whitening sphericity","Angular refinement robustifies the whitened sphericity test","Monte Carlo sphericity after MLE whitening controls finite-sample level","Dense nonseparability detected consistently when N scales with pq"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"Consistency rests on the unproven claim that the separable pseudo-MLE still converges, at a high-dimensional rate, when the true covariance is not separable.","fun_headline_variants_meta":{"raw":{"variants":["Whitening by separable MLE reduces separability to sphericity","High-dim covariance separability tested via post-whitening sphericity","Angular refinement robustifies the whitened sphericity test","Monte Carlo sphericity after MLE whitening controls finite-sample level","Dense nonseparability detected consistently when N scales with pq"]},"model":"grok-4.5","effort":"low","cost_usd":0.004752,"raw_usage":{"total_tokens":1334,"prompt_tokens":717,"num_sources_used":0,"completion_tokens":93,"cost_in_usd_ticks":47520000,"prompt_tokens_details":{"text_tokens":717,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":524,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":717,"tokens_out":93,"duration_ms":4995,"temperature":1.0,"reasoning_tokens":524,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T08:15:58.829109+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Construct a dense non-separable family (fixed positive Frobenius gap after scaling by 1/√(pq)) for which the separable MLE fails to converge in Frobenius or operator norm when N ~ pq, or show that the angular test's empirical level collapses under a non-elliptical law while the Gaussian Monte Carlo quantiles stay fixed.","supporting_citations":[],"review_version":1}