{"id":"0fabb77a-e899-435b-8160-cf77f1df878f","arxiv_id":"2607.22437","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"TrimTenRidge uses ridge regression plus coefficient thresholding to screen millions of SNPs for association with tensor-valued facial shape, with a consistency proof and an application reporting new loci.","lead":"This paper introduces TrimTenRidge, a thresholded ridge-regression screening method for finding which genetic variants affect 3D facial shape when the outcome is a high-dimensional tensor. It claims consistent selection without sparsity assumptions and reports new facial-shape loci, though the real-data findings have not been independently validated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.2 establishes support recovery for the projected coefficient B=QQᵀA, not the SNP-level effect A; real-data 'novel loci' claims require an identifiability condition that is neither stated nor checked.","rationale":"The reader's CONDITIONAL verdict and weakest_assumption align with my reading: the projected coefficient B is not the original A, so the paper's theory and simulations (both framed around B) cannot support the GWAS claim. I considered whether the proof's apparent h^{-2} tail bound is the more fundamental issue; it is not, because even a corrected tail bound would still leave the B/A mismatch. The paper is not internally inconsistent: the mathematical object being selected is B, and the theorems are about B. The mismatch is between that object and the applied interpretation. This is not a matter of disagreeing with a consensus; it is an unstated identifiability assumption. Because the algorithm itself is well-defined and the simulations demonstrate something real, I would not reject the paper. I would keep CONDITIONAL, with the condition being either (a) state and verify an assumption under which support(B)=support(A) or a clear redefinition of the screening target, and (b) validate the real-data findings through replication or an LD-aware analysis. The proposed LD simulation is the minimal check that would show whether the concern lands.","tokens_in":18743,"tokens_out":6610,"duration_ms":80987,"concrete_test":"Simulate from model (1) with sparse A: n=500, p=2000, one causal SNP k with A_{k,1,1}=1 (or 2) and all other A entries 0; construct X with an LD block where a null SNP j has correlation r²≈0.8 with k (e.g., X_{·,j}=0.9 X_{·,k}+√0.19 Z). Run the paper's full TrimTenRidge pipeline (§2.4): GCV-chosen h, 100 CV repetitions for the threshold. Across 100 replications, record the selected support of B̂: does it contain the causal k, the null j, or both, and is the causal SNP ever missed? If null partners are systematically selected or causal SNPs are missed, then Theorem 2.2's 'selection consistency' is only for B, not SNP-level association; the real-data 2,391-SNP list needs independent replication before being called novel loci.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing gap is that the method's target is B = ⟨QQᵀ, A⟩_{2,1}, the projection of A onto the row space of X, not A itself. In §2.2 the authors state that 'it will be sufficient to transfer model (1) into model (2)', but this sufficiency holds only for the conditional mean XA; it does not extend to variable selection. Because rank(QQᵀ)=r≤n=2,342 while p=6,322,729, B is a low-dimensional mixture of all SNP coefficients. A null SNP in LD with a causal SNP will generally have a nonzero entry in B, and a causal SNP whose effect vector is partly orthogonal to R(X) can have a small or zero B entry. Theorem 2.2's high-probability statement is literally about M_{B,a_n u_n} and M_{B,a_n/u_n}, the thresholded support of B. Corollary 2.3 only says that with mn=√(d1d2)O(n^τ), the support defined by B_{kij}≠0 is recovered exactly; it does not assert support(A)=support(B). No assumption is stated that A∈R(X), nor any beta-min or sparse-eigenvalue condition linking support(A) to support(B). The simulations in §3 generate B directly from model (2), so they validate recovery of B, not of A. The real-data section then interprets 2,391 nonzero B entries as detected SNPs and calls them 'novel genetic loci'; without an identifiability guarantee, these entries may be LD proxies, and the list is not a validated set of genetic associations. This is a correctness risk for the central applied claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TrimTenRidge, a ridge-regularized feature screening method for tensor responses with ultrahigh-dimensional matrix predictors. The core idea is to project the coefficient tensor A onto the row space of X, yielding B = ⟨QQ^T, A⟩_{2,1}, and then threshold the ridge estimate of B to select important predictor-response component combinations. The authors state that no sparsity assumption is needed, prove a selection-consistency theorem for the thresholded support of B, and apply the method to a facial-shape GWAS with a 2,342×7,160×3 tensor response and 6,322,729 SNPs, reporting both confirmatory and novel genetic loci.","tokens_in":19080,"tokens_out":13835,"duration_ms":136667,"significance":"If the proof gaps are repaired, the method would be a valuable screening tool for tensor-response regressions in ultrahigh dimensions. The computational trick reducing the p×p inverse to an n×n inverse is correct and practically important. The application to a large 3D facial-shape tensor is of broad interest, and the emphasis on polygenic, non-sparse effects is timely. However, the current manuscript has two load-bearing problems: (1) the theory recovers the support of the projected coefficient B, not the SNP-level coefficient A, so the real-data 'novel loci' claims are not supported; and (2) the main theorem's proof contains an algebraic error in the sub-Gaussian tail bound, inflating the claimed rate. These issues are fixable in principle but require a substantial revision.","major_comments":[{"comment":"The screening target is B = ⟨QQ^T, A⟩_{2,1}, the projection of the coefficient tensor A onto the row space of X. Models (1) and (2) are equivalent in conditional mean, so the transfer is sufficient for prediction, but not for variable selection. Theorem 2.2 and Corollary 2.3 only concern the thresholded support of B. The real-data section interprets nonzero entries of the estimator as 2,391 detected SNPs and calls them 'novel genetic loci.' Since B is a mixture of all entries of A through QQ^T, a null SNP in LD with a causal SNP can have a large B entry, and a causal SNP whose effect is partly orthogonal to R(X) can have a small or zero B entry. No identifiability condition (e.g., A ∈ R(X), or a suitable sparse-eigenvalue/irrepresentable condition) is stated or checked. The simulations generate B directly from model (2), so they validate recovery of B, not of A. The applied conclusion is","section":"§2.2 and §4"},{"comment":"The tail bound is algebraically incorrect. With a_k^T = e_k^T (X^T X + hI)^{-1} X^T, a_k^T a_k = e_k^T (X^T X + hI)^{-1} X^T X (X^T X + hI)^{-1} e_k ≤ e_k^T (X^T X + hI)^{-1} e_k ≤ h^{-1}. Thus the linear combination a_k^T E_{:ij} has sub-Gaussian variance proxy σ²a_k^T a_k ≤ σ²h^{-1}. The proof instead uses |a_k^T a_k|² ≤ h^{-2} as the variance factor, which inflates the exponent by a factor of h. With h = M₂a_n^{-2}(log log n)n^θ and (u_n−1)a_n ~ a_n/log log n, the correct exponent is at most of order n^θ/(d₁d₂ log log n) (up to constants and union bounds), not n^{2θ+2α}. The stated rate log p = o(n^{2θ+2α}) is therefore unsupported. Please correct the proof and revise the theorem statement and its consequences accordingly.","section":"Appendix, Proof of Theorem 2.2"},{"comment":"The corollary asserts exact support recovery P(Î = I*) → 1 under only mn = ∑_{I*}|B_{kij}| = √(d₁d₂)O(n^τ). No beta-min condition is imposed. Since a_n → 0, nonzero entries of B that are smaller than a_n u_n will be trimmed with probability not tending to zero (and asymptotically entirely if they are o(a_n)). A condition such as min_{I*}|B_{kij}| ≫ a_n is needed for Theorem 2.2 to imply exact recovery. As written, the corollary is not a consequence of Theorem 2.2.","section":"Corollary 2.3"}],"minor_comments":[{"comment":"The text calls '1-specificity' the 'false negative rate'; this is actually the false positive rate (or 1-specificity). Please correct the terminology.","section":"§3"},{"comment":"The text says 'in the following Tables 4-6' but the tables are labeled 2, 3, and 4. Please fix the numbering.","section":"Supplementary Material"},{"comment":"The proof jumps from E∥bias∥ ≤ ... to E∥bias∥² = ... without justification; since the bias is deterministic given X, this should be stated. Also, the second term in part 2 appears to be h²λ₁^{-2}∥B∥² = O(h²n^{-2(η-τ)}), while the displayed rate is h²n^{-(1+η-2τ)}. The discrepancy should be resolved.","section":"Proof of Lemma 2.1"},{"comment":"The statement that rs6109993 and rs1479927 are 'ranked the highest by the TrimTenRidge approach' is not explained; no ranking procedure is defined in the methods section.","section":"§4"},{"comment":"The paper does not compare with any existing tensor-response method. Since Sun and Li (2017) and Li and Zhang (2017) address the same model class, a numerical comparison, even on the non-sparse settings, would strengthen the empirical contribution.","section":"§5"},{"comment":"The notation ⟨QQ^T, A⟩_{2,1} should be explicitly defined, since QQ^T is a matrix and A is a 3-way tensor; the definition is only implicit from the general inner-product notation.","section":"§2.2"},{"comment":"The condition λ₁^{-1}=O_P(n^{-η}) is not verified for the real data. Since the rates in Theorem 2.2 depend on η, at least a heuristic check or discussion of its plausibility for the GWAS design matrix is needed.","section":"Condition (C1)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising idea and a substantial real-data application, but the two load-bearing issues—the target-of-selection gap (B vs A) and the tail-bound error—must be resolved. If the authors can correct the proof and carefully reframe the claims (either by proving support recovery for A under suitable conditions or by clearly limiting the conclusions to the projected coefficient B), the paper could become publishable. The current version overstates the biological findings."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this paper has a real new idea—feature screening for tensor responses without a sparsity assumption—and the region-localization output is genuinely useful. But the main applied claim does not follow from the theory. The target is B = QQ^T A, the projection of the coefficient tensor onto the row space of X, and the thresholded support of B is not the support of A. Without an identifiability or LD-structure assumption, null SNPs correlated with causal SNPs can have nonzero B entries, and causal SNPs orthogonal to R(X) can be missed. The simulations generate B directly from model (2), so they validate recovery of B, not A. The real-data list of 2,391 SNPs should not be called 'detected genetic loci' without addressing this.\n\nWhat's good: the paper fills a gap—no prior feature screening method handles tensor responses. The trimmed ridge extension is straightforward but new; the response-region mapping is a nice practical output. The proof of selection consistency for B is plausible in spirit, and they clearly credit Shao et al. (2012) for the basic ridge-threshold approach.\n\nWhere it's soft: the proof of Theorem 2.2 contains a tail-bound error—they use h^{-2} where the sub-Gaussian variance proxy gives h^{-1}, which inflates the rate from something like n^theta/(log log n) to n^{2theta+2alpha}. That is a real mathematical slip, though the result may survive with a slower rate. Corollary 2.3 asserts exact support recovery without a beta-min condition; the theorem only gives thresholded support within a factor. Minor issues: Simulation Setting 3 has inconsistent sample sizes across text, figures, and tables; there are no comparisons to existing methods, and the real-data analysis has no multiple-testing control or external validation, so 'novel' is premature.\n\nWho it's for: statisticians working on tensor regression or high-dimensional screening. The facial-shape application is interesting but the biological claims need more careful casting.\n\nRecommendation: send it to peer review—the gap is real and the idea is worth developing—but the revision needs to fix the projection/identifiability issue, correct the proof, and reframe the real-data results as exploratory rather than confirmed loci.","headline":"Novel tensor-response screening method with a genuine gap: it selects on the projected coefficient B, not the SNP effect A, so the real-data loci claims and the stated theorem rate both need major revision.","tokens_in":19629,"tokens_out":5739,"would_cite":false,"duration_ms":66711,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62J07","62F12","62P10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a thresholded ridge estimator recovers the support of the projected coefficient tensor in ultrahigh-dimensional tensor regression without sparsity, and that applying it to a 2,342x7,160x3 facial-shape tensor flags 2,39","keywords":["tensor response","feature screening","ridge regression","trimmed estimator","selection consistency","ultrahigh-dimensional data","facial shape GWAS","non-sparse effects"],"falsifier":"Use a design with two perfectly correlated SNP columns and set $A$ nonzero only for the first SNP. TrimTenRidge assigns identical nonzero coefficients to both columns, so it flags a null SNP as strongly as the causal one, demonstrating that the recovered support is the correlation block's support, not the causal SNP's. For the theorem itself, simulate the Theorem 2.2 setup at the rate boundary with $B$ entries just above $a_n u_n$: if the empirical inclusion probability fails to approach the stated lower bound over repeated samples, the bound is wrong.","tokens_in":18584,"feed_emoji":"🧬","tokens_out":11229,"duration_ms":125281,"temperature":0.7,"texified_at":"2026-08-05T21:41:38.299619+00:00","pith_summary":"TrimTenRidge is a feature-screening procedure for regression where the response is a multi-dimensional array (a tensor) and the predictors are ultrahigh-dimensional. The paper's central claim is that a ridge fit—least squares with an L2 penalty—followed by a threshold that zeroes small coefficients consistently separates strong from weak entries of the projected coefficient tensor, without any sparsity assumption. The proof delivers an explicit probability bound: with mild eigenvalue and boundedness conditions on the design and coefficient fibers, the thresholded support lies between the supports at slightly larger and slightly smaller thresholds with probability tending to 1, and the predictor dimension may grow at an exponential rate governed by the threshold. This matters for genome-wide facial-shape studies, where many SNPs have small nonzero effects and polygenic architecture makes sparsity unrealistic. Applied to a 2,342-by-7,160-by-3 tensor of facial shape from 2,342 people, the method flags 2,391 SNPs and associates each with facial regions.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":8944,"prompt_tokens":891,"completion_tokens":8053,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":891,"completion_tokens_details":{"reasoning_tokens":7181}},"feed_headline":"No-sparsity screen links 2,391 SNPs to facial shape","feed_subtitle":"It screens ultrahigh-dimensional SNP pools against whole-face 3D tensors and shows which face regions each gene affects.","key_machinery":"The load-bearing machinery is the trimmed ridge estimator: $\\hat{B} = \\langle (X^T X + h I)^{-1} X^T, Y \\rangle_{2,1}$, with entries below a threshold $a_n$ set to zero. The paper's key identity is the projection $B = \\langle QQ^T, A \\rangle_{2,1}$, where $X = P D Q^T$ is the SVD: it replaces the unidentifiable $A$ with the identifiable part of the coefficient tensor lying in the row space of $X$. The computation avoids inverting a $p \\times p$ matrix by using $(X^T X + h I)^{-1} X^T = X^T (X X^T + h I)^{-1}$, reducing the inverse to $n \\times n$. Theorem 2.2 then ties the threshold $a_n$ and penalty $h$ to the eigenvalue rate $n^{-\\eta}$ and coefficient growth $n^{\\tau}$, producing a probability bound that permits ultrahigh-dimensional $p$.","core_discovery":"The discovery is that ridge regression, normally an estimation tool, can act as a consistent feature screener for tensor responses once its coefficients are thresholded. The paper works with $B = \\langle QQ^T, A \\rangle_{2,1}$, the part of the coefficient tensor $A$ that lies in the row space of the design matrix $X$; because $A$ itself is not identifiable when $p \\gg n$, $B$ is the identifiable target. With sub-Gaussian errors and conditions (C1)-(C2), choosing the ridge penalty $h$ and threshold $a_n$ in the rate window $0 < \\alpha < (\\eta - \\tau - \\theta)/3$ gives $P(M_{B,a_n u_n} \\subset M_{\\hat{B}, a_n} \\subset M_{B, a_n/u_n}) \\ge 1 - 4p \\exp\\{\\ldots\\}$, so the trimmed support is sandwiched between strong and weak signal sets and no false posit","pith_inferences":["Inference beyond the paper: the 2,391-SNP list should be read as linkage-disequilibrium-block membership, not individual causality. Because B projects A onto the row space of X, null SNPs in perfect LD with a causal SNP receive identical B entries, while a causal SNP whose genotype direction lies mostly outside the row space can have a small B entry.","Inference beyond the paper: a natural two-stage extension is to use TrimTenRidge as a screening stage and then run fine-mapping or conditional analysis inside each selected LD block to resolve actual causal variants; the paper does not perform this second stage.","Inference beyond the paper: the data-driven choice of a_n via prediction-error cross-validation selects the constant while the theory fixes the rate; an extension would derive the operating characteristics of that choice, since tiny nonzero effects are increasingly hard to distinguish as n grows and p grows faster."],"forward_implications":["Tensor-response screening no longer requires low-rank, CP, or Tucker sparsity structure, so small nonzero polygenic effects are not forced to zero.","Selected predictors come with a spatial signature: for each SNP, the nonzero coefficient entries identify which response regions—eyes, nose, lips, forehead, or whole face—are affected.","Jointly modeling all SNPs in a chromosome avoids the model misspecification of single-SNP GWAS caused by ignoring the polygenic background.","The facial-shape application yields a candidate set of 2,391 SNPs, including previously reported loci such as TASP1-related chin dimple and SRPK2-related eye/nose effects, plus new candidates for follow-up.","The consistency bound indicates that the maximum tolerable predictor dimension is log p = o(n^{2theta+2alpha}), giving an explicit growth limit for applied screening studies."],"fun_headline_variants":["Trimmed tensor ridge screens genes for face shape","Face-shape genes found without sparsity assumption","Consistent screening for facial shape from SNPs","Ridge regression spots genes tied to face regions"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"For the gene list, the load-bearing premise is that a SNP's importance can be read from the nonzero entries of $B$, the projection of the coefficient tensor onto the row space of $X$; with linkage disequilibrium this can inflate null SNPs and shrink causal ones, so support recovery for $B$ does not automatically recover causal SNPs.","fun_headline_variants_meta":{"raw":{"variants":["Trimmed tensor ridge screens genes for face shape","Face-shape genes found without sparsity assumption","Consistent screening for facial shape from SNPs","Ridge regression spots genes tied to face regions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1343,"prompt_tokens":784,"completion_tokens":559,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":501}},"tokens_in":528,"tokens_out":559,"duration_ms":7094,"temperature":1.0,"reasoning_tokens":501,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T04:45:43.070166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use a design with two perfectly correlated SNP columns and set $A$ nonzero only for the first SNP. TrimTenRidge assigns identical nonzero coefficients to both columns, so it flags a null SNP as strongly as the causal one, demonstrating that the recovered support is the correlation block's support, not the causal SNP's. For the theorem itself, simulate the Theorem 2.2 setup at the rate boundary with $B$ entries just above $a_n u_n$: if the empirical inclusion probability fails to approach the stated lower bound over repeated samples, the bound is wrong.","supporting_citations":[],"review_version":1}