{"id":"4d870594-1bd7-4a4d-8144-758e1e42eb5b","arxiv_id":"2507.11160","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ECCAR reformulates sparse CCA as a Lasso-style regression on the identity matrix, yielding fast computation with proved error bounds and one-sided support recovery.","lead":"This paper introduces ECCAR, a sparse canonical correlation analysis method that estimates shared structure between two high-dimensional datasets by solving a penalized regression with an identity target. It claims to be faster than existing theoretically grounded sparse CCA methods while providing consistency and support-selection guarantees.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Support-recovery guarantee is only one-sided inclusion, so the 'sparsistent' claim overstates; it also rests on an unverifiable irrepresentable condition that can fail in practice.","rationale":"The reader's weakest assumption identified the irrepresentable condition and λ_r^-2 scaling. I agree that the support-recovery claim is the most load-bearing weakness: the theorem gives only one-sided inclusion, and the population condition is unverifiable. This does not refute the consistency of the estimated directions, but it means the advertised 'sparsistent' and 'support recovery guarantees' are not established. The direction-consistency bound (Theorem 3.2) appears structurally sound. One additional concrete issue emerged while auditing the appendix: Theorem F.1 (group-sparse) claims ∥∆∥_F ≲ ρ√s with s the number of groups, but the proof bounds the active-group term by ρ√s instead of ρ√(∑_{g∈G0} T_g); this is only valid when all groups have size 1. The group-sparse extension therefore lacks a correct proof as stated. Since this concerns a secondary extension rather than the ℓ1 ECCAR central claim, it does not change the verdict beyond the reader's CONDITIONAL.","tokens_in":38883,"tokens_out":39360,"duration_ms":427373,"concrete_test":"Re-run the Section 4 simulation (n=400, p=q=1000, su=sv=15, λ=0.9) and compute the recall of the support of \\hat U and \\hat V, i.e., the fraction of rows of U* and V* with nonzero loadings that are selected by ECCAR, alongside precision. If recall is substantially below 1 while subspace distance is small, the one-sided inclusion guarantee does not deliver exact support recovery and the 'sparsistent' claim is overstated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's advertised contribution includes support recovery and 'sparsistency' (Introduction, Contribution (2); Conclusion). The actual guarantee, Theorem 3.3 and Corollary 3.3.1, establishes only supp(\\hat B) ⊆ Su × Sv, i.e., no false positives. It does not show that all true support variables are recovered, nor that supp(\\hat U)=Su or supp(\\hat V)=Sv. Section 6's 'sparsistent' claim is therefore not supported by the theorems. Corollary 3.3.1 also depends on the population irrepresentable-type condition Eq. 12, which is unverifiable from sample data and fails when in-support and out-of-support variables are correlated; in that regime, no support guarantee is given. This is load-bearing because the paper's novelty over prior sparse CCA work (e.g., Gao et al. 2017) is explicitly tied to these support-recovery guarantees, while the direction-consistency result (Theorem 3.2) is not affected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ECCAR, a sparse canonical correlation analysis method that estimates the population matrix B* = U*Λ*V*^T by solving the convex Lasso-type problem in Eq. (8), then extracts canonical directions through a rank-r SVD and normalization (Algorithm 1). The main theoretical results are Theorem 3.1, giving a Frobenius-norm error bound of order sqrt(s_u s_v log(p+q)/n) for the estimate of B*; Theorem 3.2, giving a direction-estimation error bound of order (1/λ_r^2) sqrt(s_u s_v log(p+q)/n); and Theorem 3.3 with Corollary 3.3.1, providing support-recovery guarantees under an irrepresentable-type condition. The paper also reports extensive simulations and applications to AUD genetics, ABIDE neuroimaging, and LLM embedding interpretability, and it makes an R package available.","tokens_in":39094,"tokens_out":6903,"duration_ms":86120,"significance":"If the central consistency results are correct, ECCAR is a useful addition to the sparse CCA toolbox: it avoids Fantope projections, does not require sample splitting, and the reported simulations suggest substantial computational gains over existing theory-based methods. The proof of Theorem 3.1 follows standard Lasso techniques and is plausible. However, the paper's advertised support-recovery and 'sparsistency' contribution is materially stronger than what the theorems prove, and the proof of the high-probability support-inclusion corollary contains a gap. The direction-consistency rate in Theorem 3.2 is also acknowledged to be slower than minimax rates, so the significance rests mainly on computation and practical performance rather than on optimal statistical efficiency.","major_comments":[{"comment":"The paper advertises support recovery and 'sparsistency', but Theorem 3.3 and Corollary 3.3.1 only prove supp(hat B) ⊆ S_u × S_v, i.e., absence of false positives. They do not establish supp(hat B) = S_u × S_v, nor do they guarantee that all true support variables are recovered. The paragraph after Corollary 3.3.1 similarly only gives inclusion for hat U and hat V. Since Section 6 defines 'sparsistent' as coming with provable guarantees on identified variables, the conclusion that ECCAR is 'the first estimator ... that provably meets all three criteria' overstates the result. Please either prove exact support recovery under an additional signal-strength (beta-min) condition or reframe the contribution as false-positive control and support inclusion.","section":"§1 Contribution (2); §3.2; §6"},{"comment":"The high-probability version of the irrepresentable condition is not established by the cited lemmas. The proof asserts, immediately after invoking Lemma D.3, that ||(hat Σ_X)_{S_u^c S_u}||_{2,∞} ≤ ||(Σ_X)_{S_u^c S_u}||_{2,∞} + a√s_u √(log(p+q)/n) (and similarly for Y). Lemma D.3 controls the entrywise infinity norm of hat Σ_XY - hat Σ_X B* hat Σ_Y, and Lemma D.4 controls sparse eigenvalues; neither directly gives the stated 2,∞ block-norm deviation. A separate concentration argument for block row norms is needed. As written, the proof of Corollary 3.3.1 is incomplete, which is load-bearing because the corollary is the main support-recovery guarantee.","section":"Appendix F.1.3, proof of Corollary 3.3.1"},{"comment":"Even if the proof gap is repaired, the support guarantee is conditional on the population irrepresentable condition in Eq. (12), which is not checkable from the data and fails when in-support and out-of-support variables are highly correlated. The paper does not provide a data-dependent verification procedure or a fallback guarantee for that regime. Because the support-recovery claim is one of the three advertised contributions and is used in the conclusion to claim that ECCAR 'provably meets all three criteria', this limitation should be stated explicitly in Section 3.2 and the conclusion, and the claims should be softened accordingly.","section":"Eq. (12) and §3.2"}],"minor_comments":[{"comment":"There is a notation mismatch: the main-text Theorem 3.2 uses λ_r^{*2} in the sample-size condition, while the appendix statement of the same theorem uses ρ_r^{*2}. Also, within the proof of Theorem 3.2 in Appendix F.1.4, the text says 'Following the same logic as in the proof of Theorem 3.2', which appears to be a self-reference; this should likely refer to Theorem 3.1.","section":"Theorem 3.2 and Appendix F.1.4"},{"comment":"The concentration display for the sample covariance blocks contains a likely typo: the sequence lists σmin(bΣY) twice where one instance should presumably be σmax(bΣY), since the intended result is a two-sided bound on both restricted eigenvalues.","section":"Appendix F.1.3, proof of Corollary 3.3.1"},{"comment":"The row for Gao and Ma (2023) contains 'r(su+sv)rlog(p+q)' with a duplicated 'r'; the displayed rate should be cleaned up for readability.","section":"Table 3"},{"comment":"The main text refers to the 'sin Θ distance between the stacked (U*,V*) and the estimates', but no formal definition of this distance is given; please define it explicitly for reproducibility.","section":"§4, Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The support-recovery overclaim and the gap in the proof of Corollary 3.3.1 are the main obstacles. The direction-consistency results and computational claims appear plausible and could support publication after revision. I would ask the authors to either add a beta-min condition to prove exact support recovery or rewrite the contribution statements to accurately describe one-sided support inclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful sparse CCA paper, and the core estimator and consistency theory deserve a serious referee. But the advertised support-recovery guarantee is weaker than claimed, and the authors need to fix the overstatement before publication.\n\nWhat's actually new: the identity-target regression objective (fitting X B Y^T to the identity) is a real twist on Donnat and Tuzhilina (2024), and the extension to both views being high-dimensional is a meaningful step beyond that prior work. The proof machinery is standard Lasso/ADMM, but the combination is clean, and the computational gains over Fantope-based methods are substantial and honestly documented. The simulations are extensive, the real-data applications are convincing enough, and the R package is a plus.\n\nWhere the soft spots are: the support-recovery theorem (Theorem 3.3 and Corollary 3.3.1) only shows inclusion, supp(B_hat) ⊆ Su × Sv. That is a one-sided guarantee — no false positives, but no promise that all truly relevant variables are recovered. The introduction says the method 'recovers the correct support' and the conclusion calls it 'sparsistent.' That is overstating what the theorem delivers. The stress-test note is right about this, and it matters because part of the claimed novelty over Gao et al. (2017) is tied to support recovery. The fix is straightforward: soften the language to 'no false positives' and explicitly state that exact support recovery is not guaranteed.\n\nThe other soft spot is the irrepresentable-type condition (Eq. 12). It is a population condition, unverifiable from the sample, and it fails when in-support and out-of-support variables are highly correlated. Again, this is acknowledged in the proof structure but not in the main text, and the conclusion should warn practitioners. The direction-consistency result (Theorem 3.2) is not affected by either issue, so the main estimation claim stands.\n\nMinor point: the sample requirement n ≥ c s_u s_v log(p+q)/λ_r^2 grows sharply for weak canonical correlations, but that is a normal feature of CCA and is compared fairly with existing rates.\n\nMy recommendation: send this to peer review. The core is sound, the experiments are thorough, and the overclaim is correctable in revision. A referee should push on the support-recovery language and on whether the irrepresentable condition can be checked or relaxed in practice.","headline":"Solid sparse CCA paper with real computational gains, but the support-recovery guarantee is only one-sided inclusion and the text overstates it as 'sparsistency'.","tokens_in":39604,"tokens_out":1584,"would_cite":true,"duration_ms":22537,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H20","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"ECCAR estimates sparse canonical directions by solving a Lasso-type reduced-rank regression, and the paper proves consistency and support recovery under an irrepresentable condition.","keywords":["canonical correlation analysis","sparse CCA","reduced-rank regression","Lasso","high-dimensional statistics","support recovery","group sparsity","multi-omics analysis"],"falsifier":"Simulate Gaussian data with sparse true directions, set $\\lambda_r$ small (e.g., 0.3), $n=400$, $p=q=1000$, $s_u=s_v=15$, and add one out-of-support variable with correlation 0.95 to an in-support variable; if ECCAR with $\\rho = C\\sqrt{\\log(p+q)/n}$ frequently assigns nonzero weight outside the true support, or if $\\|\\hat B - B^*\\|_F$ exceeds the theorem's stated rate, the support-recovery and consistency claims are contradicted.","tokens_in":38678,"feed_emoji":"🔗","tokens_out":6804,"duration_ms":71661,"temperature":0.7,"pith_summary":"Sparse canonical correlation analysis asks for a few sparse linear combinations of two high-dimensional datasets that are maximally correlated. This paper claims that the entire CCA solution can be recovered from the single matrix $B^* = U^*\\Lambda^*V^{*\\top}$, and that $B^*$ can be estimated by a Lasso-type reduced-rank regression. The paper proves that, under Gaussian data and sparsity, the estimator is consistent, gives high-probability error bounds, and recovers the correct support when an irrepresentable condition holds. If true, this would remove the need for expensive Fantope projections in sparse CCA and make the method practical for large multimodal data.","feed_headline":"Sparse CCA solved as a Lasso gets consistency guarantees","feed_subtitle":"ECCAR replaces costly Fantope projections with a penalized regression plus SVD and runs far faster.","key_machinery":"The central object is the matrix identity $\\Sigma_{XY} = \\Sigma_X U^*\\Lambda^*V^{*\\top}\\Sigma_Y$, which implies $B^* = U^*\\Lambda^*V^{*\\top} = \\Sigma_X^{-1}\\Sigma_{XY}\\Sigma_Y^{-1}$; in the low-dimensional limit the unpenalized least-squares solution $\\hat B = \\hat\\Sigma_X^{-1}\\hat\\Sigma_{XY}\\hat\\Sigma_Y^{-1}$ converges to $B^*$. In high dimensions the paper replaces this with an $\\ell_{1,1}$-penalized regression, and Algorithm 1 recovers the directions by computing the rank-$r$ singular value decomposition of $\\hat\\Sigma_X^{1/2}\\hat B\\hat\\Sigma_Y^{1/2}$ followed by normalization. The proof machinery combines the Lasso basic inequality with sparse eigenvalue bounds and a Davis-Kahan/Wedin perturbation bound to pass from the error on $\\hat B$ to the error on the estimated singular vectors.","core_discovery":"The central claim is that the product matrix $B^* = U^*\\Lambda^*V^{*\\top}$ can be estimated directly from the regression loss $\\frac{1}{2}\\|\\frac{1}{n}XBY^\\top - I_n\\|_F^2$, turning sparse CCA into a convex $\\ell_{1,1}$-penalized regression problem. With $n \\gtrsim s_u s_v \\log(p+q)$, the Lasso estimate satisfies $\\|\\hat B - B^*\\|_F \\lesssim \\sqrt{s_u s_v \\log(p+q)/n}$, and after a rank-$r$ SVD and normalization the estimated directions recover $U^*$ and $V^*$ up to rotation with error of order $\\lambda_r^{-2}\\sqrt{s_u s_v \\log(p+q)/n}$. Under the irrepresentable condition (Eq. 12), the support of $\\hat B$ is contained in the true support with high probability, and because the algorithm derives the direction estimates from $\\hat B$, sparsity transfers to $\\hat U$ and $\\hat V$.","pith_inferences":["If the consistency rates are sharp, a debiasing or thresholding post-processing step on $\\hat B$ could plausibly recover the sharper minimax error of order $\\sqrt{s_u r\\log(p)/n}$ without reintroducing Fantope projections, though the paper does not pursue this.","The regression formulation extends naturally to more than two datasets by penalizing a multiway array of cross-covariances; the paper explicitly lists multi-omics settings as a future direction.","The support-recovery guarantee is conditional on a population condition that cannot be verified from the sample, so a practical diagnostic would be to compare ECCAR's selected support under subsampling or perturbation, since the theory alone does not certify the condition for a given dataset."],"forward_implications":["With $n \\geq c\\,s_u s_v \\log(p+q)$, ECCAR estimates the full product matrix $B^*$ at Frobenius error of order $\\sqrt{s_u s_v \\log(p+q)/n}$, so the hard regime is when the product of the true support sizes is large relative to the sample size.","The estimated canonical directions converge up to rotation at rate $O(\\lambda_r^{-2}\\sqrt{s_u s_v \\log(p+q)/n})$, meaning a small $r$-th canonical correlation is the main obstacle to accurate subspace recovery.","Under the irrepresentable condition (Eq. 12), the support of $\\hat B$ is contained in the true support with high probability, and because the algorithm derives $\\hat U$ and $\\hat V$ from $\\hat B$, sparsity transfers to the estimated directions.","The group-sparse extension achieves rate $\\sqrt{s\\log(p+q)/n}$ where $s$ is the number of groups, allowing structured collections of variables such as brain networks or gene groups to be selected as blocks.","The total computational complexity, of order $O(p^2n+q^2n+T(pn^2+pqn))$, avoids the cubic cost of Fantope projection methods and makes the procedure feasible at dimensions around $p=q=1000$ in the reported simulations."],"supporting_citations":[{"why":"Supplies the canonical pair model and the Fantope-based competitor whose rate and computational cost ECCAR is compared against.","marker":"Gao et al. (2017)"},{"why":"Provides the thresholded-gradient sparse CCA baseline and the comparable rate without sample splitting that ECCAR's guarantees are contrasted with.","marker":"Gao and Ma (2023)"},{"why":"Popular heuristic sparse CCA baseline whose diagonal-covariance simplification motivates respecting the full covariance structure.","marker":"Witten et al. (2009)"},{"why":"Reframes CCA as reduced-rank regression, the idea ECCAR extends to the setting where both datasets are high-dimensional.","marker":"Donnat and Tuzhilina (2024)"},{"why":"Supplies the Davis-Kahan variant used to convert the error on the estimated matrix $\\hat B$ into error bounds on the estimated singular vectors.","marker":"Yu et al. (2015)"},{"why":"Provides the ADMM solver that makes the algorithm projection-free and determines the reported computational complexity.","marker":"Boyd et al. (2011)"},{"why":"Introduces the Lasso penalty that the $\\ell_{1,1}$ penalty used in ECCAR is based on.","marker":"Tibshirani (1996)"}],"fun_headline_variants":["Sparse CCA made fast and consistent via Lasso","ECCAR: sparse CCA that skips Fantope, stays rigorous","Sparse CCA reformulated as convex regression for speed","Projection-free sparse CCA with provable error bounds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire theory assumes Gaussian data and requires an unverifiable population condition: variables outside the true sparse support must not be strongly correlated with variables inside it, which is exactly what breaks down in collinear high-dimensional data.","fun_headline_variants_meta":{"raw":{"variants":["Sparse CCA made fast and consistent via Lasso","ECCAR: sparse CCA that skips Fantope, stays rigorous","Sparse CCA reformulated as convex regression for speed","Projection-free sparse CCA with provable error bounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1258,"prompt_tokens":903,"completion_tokens":355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":285}},"tokens_in":519,"tokens_out":355,"duration_ms":4210,"temperature":1.0,"reasoning_tokens":285,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:15:59.142667+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate Gaussian data with sparse true directions, set $\\lambda_r$ small (e.g., 0.3), $n=400$, $p=q=1000$, $s_u=s_v=15$, and add one out-of-support variable with correlation 0.95 to an in-support variable; if ECCAR with $\\rho = C\\sqrt{\\log(p+q)/n}$ frequently assigns nonzero weight outside the true support, or if $\\|\\hat B - B^*\\|_F$ exceeds the theorem's stated rate, the support-recovery and consistency claims are contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the canonical pair model and the Fantope-based competitor whose rate and computational cost ECCAR is compared against."},{"cited_title":"and Ma, Z","cited_arxiv_id":null,"evidence_quote":"Provides the thresholded-gradient sparse CCA baseline and the comparable rate without sample splitting that ECCAR's guarantees are contrasted with."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ADMM solver that makes the algorithm projection-free and determines the reported computational complexity."}],"review_version":1}