{"id":"85751802-036b-4527-a63b-df34e9fbbceb","arxiv_id":"1908.08095","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper develops a paired-sample test for matrix Gaussian graphical models with a variance correction for before/after correlations, and proves asymptotic false discovery rate control.","lead":"A new statistical test detects changes in brain connectivity networks before and after an intervention when the same subjects are scanned twice, explicitly accounting for the correlation between the paired scans. It gives researchers a way to identify which brain-region links change while controlling false discoveries, a setting current tests do not handle for matrix-valued imaging data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.3's proof invokes an independence-only large deviation bound (Lemma B.2) for paired residual differences, so the central FDR-control guarantee is not established for dependent samples.","rationale":"The paper proposes a genuinely new paired-testing procedure for matrix graphs and provides a plausible variance correction with simulations supporting its use in the studied diagonal-P regimes. The reader's verdict of CONDITIONAL is reasonable; the matrix-normal and separable-covariance assumption is a legitimate concern, and the real-data example falls outside the asymptotic conditions. My independent scan of the proof, however, identifies a more specific and more load-bearing gap: the proof of Theorem 3.3 repeatedly relies on large deviation bounds for independent samples, most visibly Lemma B.2, when the paired design induces dependence between the two samples. The denominator in Lemma B.2 is the sum of marginal variances, which is not the variance of the paired difference; for negative within-pair correlation the actual tail probability can be larger. In addition, when P_{T1,2} is non-diagonal, the test statistics are sums over time points of dependent residuals, and the transferred proof from Xia and Li (2019) does not handle this temporal dependence. Since the paper's simulations only use a diagonal P_{T1,2} (or perturbations of the off-diagonal Kronecker block with the same diagonal P), they cannot reveal whether the claimed FDR guarantee extends to the non-diagonal P allowed by the model. This is not a claim that the theorem is false; it is a claim that the argument as written is incomplete. A targeted simulation or a re-derivation of the tail bound would settle whether the central FDR-control claim actually holds under the stated assumptions. Because the reader already issued CONDITIONAL and my concern reinforces that a condition should be attached to the theorem or the proof repaired, I do not change the verdict.","tokens_in":35278,"tokens_out":32855,"duration_ms":266428,"concrete_test":"Simulate the paired test under model (2.1) with n=15, p=200, q=50, the same spatial structures as Table 1, and a non-diagonal P_{T1,2}, e.g., an AR(1) matrix with autocorrelation 0.5 (normalized to have bounded singular values so that (A1) holds), with Σ_{S1,2}=γΣ_{S1} and γ=0.6. Compute the empirical FDR of the variance-corrected procedure over 100 replications. If the FDR exceeds the nominal 1% by a margin similar to the uncorrected test in Setting II (e.g., >2%) while the diagonal-P case remains near 1%, the missing dependence treatment in the proof is load-bearing. As a complementary analytical check, re-derive the tail bound for b_{i,j} in Appendix C.4 using the true paired variance, including the covariance term, and verify whether P(max_{i,j} b_{i,j}^2 ≥ 2 log p) → 0 under (A1)–(A5); if it does not, the proof of Theorem 3.3 is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 3.3 (Appendix C.4), the authors bound the contribution of the set Aτ by showing P(max_{Aτ} W_{i,j}^2 ≥ 4 log p − log log p + h) ≤ |Aτ|(P(V^2 ≥ log p/8) + P(b_{i,j}^2 ≥ 2 log p)) and then assert that the last probability is o(1) as 'a direct result of Lemma B.2'. Lemma B.2, however, gives a large deviation bound for the difference of sample covariances from two independent samples: its denominator is var{(X_{k,i}−μ_{1,i})(X_{k,j}−μ_{1,j})}/n1 + var{(Y_{k,i}−μ_{2,i})(Y_{k,j}−μ_{2,j})}/n2, which omits the covariance term that appears when X_k and Y_k are paired. In this paper the residuals ε^{(1)}_{k,i,l} and ε^{(2)}_{k,j,l} from the same subject are correlated through Σ_{S1,2}⊗P_{T1,2}, so the variance of the difference of sample variances of the squared residuals is not the sum of the two marginal variances. For negative cross-correlations (as in Setting II), the true variance is larger than the bound's denominator, and P(b_{i,j}^2 ≥ 2 log p) need not be o(1). The proof also transfers 'the proof of Theorem 1 in Xia and Li (2019)' to the redefined V_m without addressing the temporal dependence of Z_{k,m,l} across l, which is nonzero whenever P_{T1,2} is non-diagonal. The paper's simulations only use a diagonal P_{T1,2}, so they do not exercise this gap. Thus, as written, the proof does not establish the claimed asymptotic FDR control for the general dependence structure allowed by (2.1).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a paired test of matrix graphs for comparing the spatial partial correlation matrices of two groups of matrix-variate samples that are paired (e.g., before and after a stimulus). The data are modeled as jointly matrix-normal with separable between-sample covariance Σ_S1,2 ⊗ Σ_T1,2. The test statistic is based on residuals from node-wise regressions after temporal whitening, with a bias-corrected estimator of the partial correlation difference and a new variance correction that accounts for the between-sample spatial and temporal dependence (Proposition 2.1). A multiple testing procedure based on the corrected statistic is shown to control FDR asymptotically (Theorem 3.3) under conditions A1–A7. The method is evaluated in simulations with autoregressive and moving-average temporal structures and several spatial graphs, and is applied to an ADNI dataset of 23 subjects converting from MCI to AD.","tokens_in":35706,"tokens_out":6588,"duration_ms":57779,"significance":"If the theoretical claims are correct, this is the first paired test for matrix graphs in the high-dimensional setting, filling a genuine gap between one-sample and independent two-sample procedures. The variance correction is a nontrivial extension of the authors' earlier work, and the paper provides concrete simulation evidence that the correction matters: the uncorrected two-sample test loses power under positive cross-correlation and inflates FDR under negative cross-correlation. The paper also ships reproducible R code and makes the ADNI-derived data publicly available, and the sensitivity analyses that deviate from the matrix-normal and Kronecker assumptions are a useful practical check. The value of the contribution depends on whether the asymptotic FDR guarantee holds for the general dependence structure in (2.1); the proof as written has a gap for non-diagonal P_T1,2, as detailed below.","major_comments":[{"comment":"The bound on P(b_i,j^2 ≥ 2 log p) is claimed to be a 'direct result of Lemma B.2', but Lemma B.2 is a large deviation bound for the difference of sample covariances from two independent samples, with denominator var{(X_k,i−μ1,i)(X_k,j−μ1,j)}/n1 + var{(Y_k,i−μ2,i)(Y_k,j−μ2,j)}/n2. In the paired setting of this paper, the relevant difference is (nq)^{-1}Σ_{k,l}(ε^{(1)}_{k,i,l})^2 − (nq)^{-1}Σ_{k,l}(ε^{(2)}_{k,i,l})^2, whose variance is var((ε^{(1)}_{k,i,l})^2 − (ε^{(2)}_{k,i,l})^2)/(nq) and contains the cross-covariance term −2 cov((ε^{(1)})^2,(ε^{(2)})^2) that is absent from Lemma B.2. When this cross-covariance is negative, the true variance is larger than the denominator in Lemma B.2, and the conclusion P(b_i,j^2 ≥ 2 log p) = o(1) does not follow from that lemma. The proof needs a dedicated large deviation bound for paired samples that includes the covariance term.","section":"Appendix C.4 (proof of Theorem 3.3)"},{"comment":"The redefined variables V_m = (nq Θ_m)^{-1/2} Σ_{k=1}^n Σ_{l=1}^q Z_{k,m,l} are used with the assertion that the argument follows 'the proof of Theorem 1 in Xia and Li (2019)'. However, when P_T1,2 is non-diagonal, the summands Z_{k,m,l} are correlated across l through the between-sample temporal covariance, so var(Σ_{k,l} Z_{k,m,l}) is not nq Θ_m, and the independence-based proof of Xia and Li (2019) does not transfer without an additional argument. The simulations in Section 4 use only diagonal P_T1,2, so this dependence structure is not exercised empirically, and the claimed FDR control for general P_T1,2 allowed by (2.1) is not established by the proof as written.","section":"Appendix C.4 (proof of Theorem 3.3)"}],"minor_comments":[{"comment":"There is a typo: 'Neuroimaing Initiative' should be 'Neuroimaging Initiative'.","section":"Abstract"},{"comment":"The data example has p = 116, q = 130, n = 23, but condition A3 requires q = o(np/(log p)^2), and with these values np/(log p)^2 ≈ 118, so A3 is not satisfied. The paper does not verify this or discuss its implication for the applicability of Theorem 3.3 to the data analysis.","section":"Section 5 (ADNI analysis)"},{"comment":"The banded covariance estimator for Σ_Tt is stated to be adopted, but no description is given of how the banding parameter is chosen in the simulations; Appendix D only describes tuning for the Lasso parameter λ_n,i. A brief statement of the banding choice would improve reproducibility.","section":"Section 4.1 and Appendix D"},{"comment":"In the Sensitivity I power block for (p,q) = (800,200), the row '6' appears twice; the second occurrence should likely be '8'.","section":"Table 3"},{"comment":"In the sentence following (2.8), 'we show in Section 3 that ˆΘ_i,j in (2.8) provides an accurate estimation of Θ_i,j' should say 'of Θ_i,j' (subscript consistency with θ_i,j elsewhere is fine, but the notation should be uniform).","section":"Section 2.3, Eq. (2.8)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a direct extension of the authors' own prior work (Xia and Li 2017, 2019), which is appropriate given the problem, but the incremental novelty is moderate. The central new theoretical claim is the FDR guarantee for paired samples, and the proof gap for non-diagonal P_T1,2 is the main obstacle. I believe the gap is likely fixable by adding a paired large-deviation lemma and handling the temporal dependence in the V_m variables, so I do not recommend rejection at this stage. The data application is somewhat cursory and does not verify the regularity conditions; this should be addressed in a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague: The paper is worth knowing about, but the main theorem has a gap that needs attention. It develops the first paired test for matrix-variate graphical models, with a variance correction that explicitly accounts for the cross-covariance between before/after samples. The construction is natural: extend Xia and Li's two-sample test, add a correction term involving rho_S1,2 and P_T1,2, and estimate it by pooling across spatial locations and time. The proposition giving the expression for Theta_ij is clean, and the simulations are extensive—both for FDR control and for sensitivity to non-normality and Kronecker perturbations. They also shipped code and data. The problem is in the proof of Theorem 3.3. The proof bounds the contribution of the set A_tau by invoking Lemma B.2, which is a large deviation result for the difference of sample covariances from two independent samples. Its denominator is the sum of the marginal variances of the two groups. But the residuals here are paired, so the variance of the difference of squared residuals includes a cross-covariance term. When that term is negative, the true variance is larger, and the bound as written doesn't follow. The paper's own simulations in Setting II create exactly that kind of negative cross-correlation for about half the pairs, yet the test still controls FDR—so the method may well be correct. But the proof as it stands doesn't cover it, and the simulations use a diagonal P_T1,2, so they don't exercise the temporal dependence either. The ADNI example is also outside the stated asymptotics (n=23, p=116, q=130 fails condition A3), so the real-data analysis should be read as an illustration, not a validation. Minor: the first p-value in Table 4 is reported as exactly 0, which is likely a numerical artifact. Net: the idea is right and the empirical work is credible, but the central FDR guarantee is not proven as written. This is the kind of paper a good referee should engage with—it deserves the time—but it needs a corrected proof or a clear statement of an additional condition (e.g., on the cross-covariance of squared residuals) before the headline claim can stand. I'd send it to review, and I'd ask for the proof fix; I wouldn't cite the guarantee yet.","headline":"A useful paired test for matrix graphs with strong simulations, but the main FDR proof leans on an independent-sample bound that doesn't apply to paired residuals; fix the proof and it's publishable.","tokens_in":36184,"tokens_out":3154,"would_cite":false,"duration_ms":29614,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H15","62H20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper develops a paired test of matrix graphs that compares brain connectivity networks between correlated samples and asymptotically controls the false discovery rate.","keywords":["paired test","matrix graph","brain connectivity","matrix normal distribution","partial correlation","variance correction","false discovery rate","multiple testing"],"falsifier":"Generate paired samples from a matrix-normal model but replace the cross-covariance $\\Sigma_{S1,2}\\otimes\\Sigma_{T1,2}$ with a non-separable matrix of the same norm, keep all marginal variances unchanged, and run the procedure at $\\alpha=1\\%$ with $n=15$, $p=200$, $q=50$ over many replications; if the empirical FDR moves well above the nominal level, for instance above $3\\%$, while the uncorrected test behaves differently, the variance-correction identity (2.6) is missing dependence structure.","tokens_in":35103,"feed_emoji":"🧠","tokens_out":9674,"duration_ms":88439,"temperature":0.7,"pith_summary":"The paper claims that brain-connectivity change between paired scans of the same subjects, before and after a stimulus or disease conversion, can be tested simultaneously across all region pairs while asymptotically controlling the false discovery rate at a pre-specified level. The target of inference is the spatial partial-correlation matrix, whose entries encode the connectivity graph. The authors construct a statistic for each pair of regions from a bias-corrected estimate of the difference in partial correlations, then add a variance correction that accounts for the correlation between the two scans. This correction is what makes the paired setting work; without it, a two-sample test either loses power or inflates false discoveries depending on the sign of the before-after correlation. Simulations and an Alzheimer's Disease Neuroimaging Initiative analysis illustrate the method.","feed_headline":"Paired brain-network test controls false discoveries","feed_subtitle":"A variance-corrected statistic handles paired before/after scans; the FDR stays at the nominal level.","key_machinery":"The central object is the paired test statistic $W_{i,j}=(\\hat{\\rho}_{S1,i,j}-\\hat{\\rho}_{S2,i,j})/\\hat{\\Theta}_{i,j}^{1/2}$, where $\\hat{\\rho}_{St,i,j}$ are bias-corrected partial-correlation estimates obtained from node-wise regressions of one brain region on the others. The load-bearing identity is the variance-correction formula (2.6): $\\Theta_{i,j}$ subtracts a between-sample term built from products of spatial cross-correlations and the squared Frobenius norm of the normalized between-sample temporal covariance $P_{T1,2}$; this term is estimated by pooling residuals across all regions and time points, which keeps the estimation error small enough for the FDR proof. The multiple-testing step then rejects pairs with $|W_{i,j}|$ above an estimated threshold $\\hat{h}_\\alpha$ that conservatively targets the false discovery proportion.","core_discovery":"Under the assumption that the paired $p\\times q$ matrices jointly follow a matrix-normal distribution with separable covariance, including a Kronecker-product cross-covariance $\\Sigma_{S1,2}\\otimes\\Sigma_{T1,2}$, the paper derives a closed-form expression for the variance of the difference between two estimated spatial partial correlations. The variance equals the sum of the two within-sample variances minus a cross term that factorizes into a spatial part $\\rho_{S1,2,i,i}\\rho_{S1,2,j,j}+\\rho_{S1,2,i,j}\\rho_{S1,2,j,i}$ and a temporal part $\\|P_{T1,2}\\|_F^2/q$. The test statistic $W_{i,j}$ uses a plug-in estimate of this variance with error $o_p(1/\\log p)$, and Theorem 3.3 proves that the resulting thresholding procedure satisfies $\\mathrm{FDR}(\\hat{h}_\\alpha)/(\\alpha\\ell_0/\\ell)\\to 1$ and $\\mathrm{FDP}(\\hat{h}_\\alpha)/(\\alpha\\ell_0/\\ell)\\to_p 1$ as $(nq,p)\\to\\infty$, giving asymptotic false-discovery control for dependent paired matrix samples.","pith_inferences":["The variance-correction identity could carry over to paired tests of covariance or precision matrices in other longitudinal designs, such as case-crossover gene-expression studies, wherever two measurements are taken on the same unit.","A practical extension the paper leaves implicit is irregular or subject-specific time grids: the estimator of $P_{T1,2}$ pools columns across a common time course, so applying the test to unequally spaced or varying-length recordings would require a new temporal-dependence estimator.","A nonparanormal copula version of the test, which the authors mention as future work, would be worth testing empirically; if marginal transformations preserve the separable cross-covariance structure, the same variance-correction formula should hold approximately."],"forward_implications":["For paired scans of the same subjects, the proposed procedure rejects edges whose spatial partial correlation changes, while the false discovery rate converges to the targeted level as $(nq,p)\\to\\infty$.","Ignoring the before-after correlation is not just inefficient: for positively correlated scans the uncorrected two-sample test is overly conservative, and for negatively correlated scans it over-rejects; the corrected statistic removes both distortions.","The asymptotic guarantee holds both when the temporal covariances $\\Sigma_{T1}$, $\\Sigma_{T2}$, and $\\Sigma_{T1,2}$ are known and when they are estimated from the data, with the same $o_p(1/\\log p)$ error rate for the variance estimator.","Sensitivity simulations indicate the procedure keeps FDR near the nominal level when the noise is $t$-distributed and when the true cross-covariance deviates from the Kronecker assumption, although the formal guarantee is proved under the matrix-normal model.","The ADNI analysis demonstrates the procedure's use in practice, identifying cerebellum-heavy connectivity changes after conversion from mild cognitive impairment to Alzheimer's disease."],"supporting_citations":[{"why":"The independent two-sample matrix graph test that this paper extends; its proof scheme is reused with redefined variables in the FDR theorem.","marker":"Xia and Li (2019)"},{"why":"Establishes the node-wise regression bias correction and the matrix-normal framework, including lemmas used for residual estimators and temporal-covariance estimation.","marker":"Xia and Li (2017)"},{"why":"Provides the earlier variance-correction idea for a single matrix-valued group that the paired correction builds on and contrasts with.","marker":"Chen and Liu (2018)"},{"why":"Supplies the large-deviation bound used to control the negligible set in the FDR proof.","marker":"Cai and others (2013)"},{"why":"Gives the normal regression representation linking precision-matrix entries to residual covariances, the basis for constructing the test statistic.","marker":"Anderson (2003)"},{"why":"The banded covariance estimator adopted for the unknown temporal covariance in the data-driven procedure.","marker":"Bickel and Levina (2008)"}],"fun_headline_variants":["FDR-control for paired brain network changes","Paired brain connectivity test with honest FDR","A corrected test for correlated brain scans","Paired-scan network inference with FDR guarantee","Variance-corrected test for paired brain graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paired observations jointly follow a matrix-normal distribution with a separable covariance structure, and specifically the cross-covariance between the two scans is the Kronecker product $\\Sigma_{S1,2}\\otimes\\Sigma_{T1,2}$; if the true dependence is not separable, the variance-correction formula no longer holds and the false-discovery guarantee could fail.","fun_headline_variants_meta":{"raw":{"variants":["FDR-control for paired brain network changes","Paired brain connectivity test with honest FDR","A corrected test for correlated brain scans","Paired-scan network inference with FDR guarantee","Variance-corrected test for paired brain graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1698,"prompt_tokens":952,"completion_tokens":746,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":676}},"tokens_in":568,"tokens_out":746,"duration_ms":7333,"temperature":1.0,"reasoning_tokens":676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:51:49.727243+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate paired samples from a matrix-normal model but replace the cross-covariance $\\Sigma_{S1,2}\\otimes\\Sigma_{T1,2}$ with a non-separable matrix of the same norm, keep all marginal variances unchanged, and run the procedure at $\\alpha=1\\%$ with $n=15$, $p=200$, $q=50$ over many replications; if the empirical FDR moves well above the nominal level, for instance above $3\\%$, while the uncorrected test behaves differently, the variance-correction identity (2.6) is missing dependence structure.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The independent two-sample matrix graph test that this paper extends; its proof scheme is reused with redefined variables in the FDR theorem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the node-wise regression bias correction and the matrix-normal framework, including lemmas used for residual estimators and temporal-covariance estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the earlier variance-correction idea for a single matrix-valued group that the paired correction builds on and contrasts with."},{"cited_title":"Tony, Liu, Weidong and Xia, Yin","cited_arxiv_id":null,"evidence_quote":"Supplies the large-deviation bound used to control the negligible set in the FDR proof."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the normal regression representation linking precision-matrix entries to residual covariances, the basis for constructing the test statistic."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The banded covariance estimator adopted for the unknown temporal covariance in the data-driven procedure."}],"review_version":1}