{"id":"ab8a27b5-8518-4dba-92f6-cbb3ea139d16","arxiv_id":"2502.05969","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 proves that Gaussian knockoffs based on first-two-moment matching achieve asymptotic FDR control for two-moment-based knockoff statistics, under strong signal and distributional conditions.","lead":"This statistics paper proves conditions under which approximate knockoffs built from a misspecified distribution, in particular Gaussian knockoffs from first two moments, still control false discovery rate asymptotically. It provides a unified theoretical framework and claims the first formal justification for the widely used Gaussian knockoffs generator.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FDR control is only proved when a_n→∞ strong signals exist; all-null/weak-signal regimes are excluded, so the abstract's unconditional claim is not supported.","rationale":"The reader's weakest_assumption is the same condition: the strong-signal coverage requirement a_n→∞. My analysis of Lemma 2 and Theorem 5 confirms that (C3) is not a minor technical convenience. Without it, the proof cannot choose an α_n that is simultaneously small enough for the moderate-deviation approximation and large enough to contain T_q. In the all-null case, which any valid FDR procedure must handle, the proof is silent. This is a substantive limitation on the central claim, and it supports the reader's CONDITIONAL verdict. I did not find the concern fatal: the theorem is correctly stated as a conditional result, and the independent-features and regression-based arguments are plausible. The secondary gaps noted by the reader—the correlated-features marginal-correlation derivation and the absence of a formal theorem for estimated moments—are also present but do not change the verdict. Since the reader already conditioned on the strong-signal restriction and related issues, my stress-test does not alter the recommended verdict.","tokens_in":55050,"tokens_out":21329,"duration_ms":215686,"concrete_test":"Re-derive the localization step in Lemma 2 with H1=∅: determine whether there exists any α_n satisfying α_n=o(n^{1/7}) and pα_n→∞ such that P(T_q≤α_n)→1 under (C1)-(C2). If the minimizer T_q is, instead, of order max_j |W_j| (about √log p), then the moderate-deviation range of Theorem 5 is exceeded and conditions (1)-(3) of Theorem 1 cannot be verified, confirming that (C3) is essential.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The asymptotic guarantee in Theorem 2 is confined to a strong-signal regime. Condition (C3) (and its analogues (O4), (D3)) requires a_n = #{j∈H1 : w_j ≥ Cδ_n} to diverge, and (C4)/(O5) require signals to avoid large negative values. These assumptions enter Lemma 2, Lemma 8, and Lemma 11 to localize T_q below α_n = P^{-1}(q a_n/(2p)) and to ensure pα_n→∞ for the indicator approximation. If a_n is bounded, as in the all-null or weak-signal case, the proof offers no usable α_n: the natural α_n would be too large for the moderate-deviation approximations in Theorem 5, whose validity is limited to t up to n^{1/7}, or α_n would be infinite. Consequently, conditions (1)-(3) of Theorem 1 cannot be verified in exactly the regime where FDR control is expected of any valid procedure. The abstract states 'asymptotic FDR control' without this strong-signal qualification, so the headline robustness claim is conditional on an unstated regime restriction. This does not invalidate the theorem as formally stated, but it makes the abstract's phrasing misleading and leaves the practically important null/weak-signal case unproved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the asymptotic FDR of the model-X knockoffs procedure when knockoff variables are generated from a misspecified working distribution, focusing on the Gaussian knockoffs generator that matches only the first two moments. It first states a general sufficient-conditions framework (Theorem 1): if the null knockoff statistics satisfy approximate symmetry, indicator-function concentration, and threshold localization on a common interval, then limsup FDR is bounded by q. It then verifies these conditions for two-moment-based statistics: marginal correlation differences (Theorem 2), OLS coefficient differences (Theorem 3), and debiased Lasso coefficient differences (Theorem 4). The verification is carried out under explicit moment, sparsity, signal-strength, and dimensionality conditions. The paper also presents simulations and an HIV drug-resistance application illustrating the finite-sample behavior of the method.","tokens_in":55316,"tokens_out":11401,"duration_ms":120879,"significance":"If the formal results hold, this is a useful contribution: it provides the first modular sufficient conditions under which the practically ubiquitous Gaussian knockoffs construction can achieve asymptotic FDR control despite covariate-distribution misspecification. The proof of Theorem 1 is transparent and short, and the supplementary material contains detailed moderate-deviation and concentration arguments rather than leaving the verification as a black box. The empirical sections support but do not replace the theory. The main caveat is that the headline claim is conditional on a strong-signal regime that is not stated in the abstract, and one of the illustrative corollaries appears to conflict with its own moment assumption.","major_comments":[{"comment":"The abstract and introduction claim asymptotic FDR control for the Gaussian knockoffs generator without qualification, but the actual proof requires the strong-signal condition (C3): a_n = #{j ∈ H1 : \\hat w_j ≥ C δ_n} → ∞. This condition enters the localization step in Lemma 2 through α_n = P^{-1}(q a_n/(2p)) and p α_n → ∞, and its analogues (O4) and (D3) play the same role in Lemmas 8 and 11. In the all-null or weak-signal regime a_n does not diverge, so the proof provides no localization of T_q and conditions (1)–(3) of Theorem 1 are not verified; yet a valid procedure is expected to control FDR in that regime as well. The formal theorems are stated conditionally, so this is fixable, but the abstract's unqualified \"asymptotic FDR control\" is misleading and should be revised to state the strong-signal restriction or supplemented by an argument covering the null/weak-signal case.","section":"§2.2–§3.1 (Theorem 1, Theorem 2, Lemma 2)"},{"comment":"Condition (C1') requires the entries of Q to have finite q-th moments for some q ≥ 3, while Corollary 1, part 2, claims coverage for X_i i.i.d. t-distributed with q degrees of freedom and q ≥ 3. A t-distribution with q degrees of freedom does not have a finite q-th moment, since E|X|^r < ∞ only for r < q. Thus the stated corollary is inconsistent with its own assumption. This can be repaired by requiring t-distributed covariates with more than q degrees of freedom, or by rephrasing the moment condition with a separate exponent, but the mismatch should be corrected.","section":"§3.1, condition (C1') and Corollary 1"},{"comment":"Condition (2) of Theorem 1 is stated as a uniform approximation over all t ∈ (0, α_n). In the subsequent verification, the width of this interval is ultimately tied to the signal count a_n through α_n = P^{-1}(q a_n/(2p)). This means that the theorem, as applied, cannot certify FDR control when the number of detectable signals is bounded, even if the null statistics themselves are well behaved. This is the same issue as the strong-signal restriction above, but it is worth emphasizing that the restriction is structural to the framework, not merely a technical tightening in one lemma: without a_n → ∞, the interval (0, α_n) on which conditions (1)–(3) are verified is not constructed at all.","section":"§2.2 (Theorem 1, condition 2)"}],"minor_comments":[{"comment":"The notation section states \"For any positive integer m ∈ Z+, let [m] ≡ {1, ..., n}\"; the right-hand side should be {1, ..., m}.","section":"§1.1 (Notation)"},{"comment":"In Proposition 3, the vector r is described as r ∈ R^n, but r in the Gaussian knockoff construction is a p-dimensional diagonal vector; it should be r ∈ R^p.","section":"§3.4 (Proposition 3)"},{"comment":"In the proof of Theorem 1, the notation \"1{#\\hat W_j ≥ t}\" appears where # is used as a placeholder for the sign; this is not defined in the main text and should be replaced by explicit positive and negative indicators.","section":"§2.2 (Theorem 1 proof)"},{"comment":"The abstract and introduction repeatedly describe the Gaussian generator as \"arguably the most popularly used\" method; the qualifier is fine, but the phrasing is informal for a journal paper and could be made more precise by citing specific prevalence in the software ecosystem.","section":"§1 (Introduction)"},{"comment":"Table 1 appears in the Introduction before the data description in Section 4.2; this is unconventional and may confuse readers, though the results are eventually explained.","section":"§4 (Numerical studies)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper delivers the first asymptotic FDR control guarantee for Gaussian knockoffs that match only the first two moments, for the standard marginal-correlation, OLS, and debiased-Lasso knockoff statistics. The unified Theorem 1 is a clean and useful way to organize robustness arguments. But the abstract sells more than the theorems deliver: the proof requires a growing number of strong signals (a_n → ∞), never handles the all-null regime, and the estimated-moments case is described but not actually proved.\n\nWhat is genuinely good: Theorem 1 is elementary and correct, and the moderate-deviation results behind the independent-features marginal-correlation and the regression-coefficient theorems are serious technical work. The paper honestly flags the heavy-tail limitation (p log n = o(√n)) and gives an instructive comparison with the coupling approach of Fan et al. (2025), showing why moments matching needs a different proof. The simulations and real-data table are fine, not overclaimed.\n\nWhere the soft spots are, in proportion: the stress-test concern is on target. Condition (C3) requires a_n → ∞; without strong signals, Lemma 2 has no usable α_n, so the proof gives nothing in the all-null or weak-signal case. Many asymptotic knockoff proofs have signal-strength conditions, but they typically treat the null case separately or state it as a limiting case. Here the omission is not acknowledged, and the abstract's unconditional phrasing is misleading.\n\nThe estimated-moments subsection is also thinner than the introduction implies. Proposition 3 is only a coupling bound; the actual FDR theorem is asserted to follow under \"mild additional assumptions\" with details omitted. The reader's suspicion about the correlated-features marginal-correlation extension in Supplement D looks plausible to me: the proof is a sketch, and the random Y-normalization changes the effective covariance away from the target Σ block, which the brief argument does not explicitly handle. That is likely fixable, but a referee should ask for it.\n\nThe reliance on conditions from Fan et al. (2025) is heavy but not disqualifying; this paper extends that framework rather than merely repackaging it.\n\nBottom line: the core OLS and debiased-Lasso results are probably correct and worth publishing after revision. I would send this to serious referees, but they should push hard on the abstract, the null-signal regime, and the two sketched parts (Supplement D and Section 3.4).","headline":"A real first formal justification for moments-matching Gaussian knockoffs, but the abstract overclaims: the proof excludes the all-null/weak-signal regime and the estimated-moments case is only sketched.","tokens_in":55828,"tokens_out":3502,"would_cite":true,"duration_ms":37151,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62E20","62J15","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the Gaussian knockoffs generator, matching only the first two moments of the covariate distribution, still controls FDR asymptotically for two-moment-based statistics, a first formal justification of its success.","keywords":["model-X knockoffs","Gaussian knockoffs generator","moments matching","asymptotic FDR control","approximate knockoffs","moderate deviations","high-dimensional variable selection","debiased Lasso"],"falsifier":"Run the approximate knockoffs procedure with Gaussian knockoffs and the marginal correlation statistic on $t$-distributed covariates with three degrees of freedom and $p\\log n$ at or above $\\sqrt{n}$, so the dimensionality condition of the heavy-tailed case in Theorem 2 fails while signal conditions hold; empirical FDR staying at or below $q$ would show that boundary is not sharp, while systematic inflation above $q$ confirms it is real. A complementary check targets the signal-coverage assumption: in the all-null model, where $a_n = 0$ and condition (C3)/(O4) fails, systematic FDR inflation would show the assumption is genuinely load-bearing, whereas clean control would show the stated conditions are sufficient but not necessary.","tokens_in":54846,"feed_emoji":"🎯","tokens_out":15083,"duration_ms":128359,"temperature":0.7,"pith_summary":"Model-X knockoffs only come with a finite-sample FDR guarantee when knockoff variables are drawn from the true covariate distribution, which is almost never known. This paper asks when substituting a user-specified working distribution still keeps FDR under control, and answers with a unified framework: three sufficient conditions on the approximate knockoff statistics that any working distribution can be checked against. The main application is the most popular practical construction — the Gaussian knockoffs generator, which matches only the first two moments of the covariates. The paper proves that for two-moment-based knockoff statistics (marginal correlation differences and regression coefficient differences) these Gaussian knockoffs achieve asymptotic FDR control, $\\limsup_{n\\to\\infty}\\text{FDR}\\le q$, the first formal justification of the method's practical success despite an evidently misspecified covariate model. The cost is explicit: enough true signals must clear the noise floor, and the heavy-tailed case carries a dimensionality restriction $p\\log n = o(\\sqrt{n})$.","feed_headline":"Two moments suffice: Gaussian knockoffs control false discoveries","feed_subtitle":"The paper proves the standard approximate knockoff still meets its false-discovery target under a wrong covariate model.","key_machinery":"The central object is the Gaussian knockoffs generator, which builds knockoffs from the first two moments of the covariates alone. Its load-bearing property is covariance exchangeability: for every swap permutation the joint covariance of $(X^\\top, \\tilde X^\\top)^\\top$ is invariant, so on null features the original/knockoff pair has swap-invariant covariance and the approximating Gaussian law satisfies $P_j(t) \\equiv P(|G_1| - |G_2| \\ge t) = P(|G_1| - |G_2| \\le -t)$. The argument is carried by ratio-based moderate deviation theorems (Theorems 5, 7, and 8 of the supplement): they show $P(\\pm \\hat W_j \\ge t)/P_j(t) \\to 1$ uniformly in $t$ up to the localization threshold, which is exactly the accuracy the FDR proof needs because the denominator probabilities can vanish. The three sufficient conditions of Theorem 1 — approximate symmetry, indicator concentration, and threshold localization — then yield $\\limsup_{n\\to\\infty}\\text{FDR}\\le q$, and a coupling proposition (Proposition 3) extends the result from population to estimated moments.","core_discovery":"The paper's central claim is that the distributional exchangeability of perfect model-X knockoffs can be relaxed, with no loss of asymptotic FDR control, to three conditions on the approximate knockoff statistics: asymptotic approximate symmetry of null statistics, ratio-consistent concentration of their empirical indicator sums, and localization of the knockoff threshold $T_q$. Verifying these conditions for the Gaussian knockoffs generator based on first-two-moment matching proves $\\limsup_{n\\to\\infty}\\text{FDR}\\le q$ for the marginal correlation difference statistic (Theorem 2), the OLS regression coefficient difference statistic (Theorem 3), and the debiased Lasso coefficient difference statistic (Theorem 4), under the explicit conditions (C1)-(C4), (O1)-(O5), and (L1). The theorems cover covariates as non-Gaussian as Rademacher signs and $t$-distributions with three degrees of freedom — cases where the earlier coupling-based robustness proof provably cannot work — and the paper states this is the first theoretical justification of the Gaussian knockoffs generator's effectiveness and robustness.","pith_inferences":["The paper's stated recommendation — match as many moments as the statistic uses — suggests a testable hierarchy: a statistic built from third or higher moments, such as the distance correlation difference used in the real-data example, should need higher-order moment matching for the same guarantee; the theorems here cover only first-two-moment statistics, and the distance-correlation results are ","Because the theorem is silent when the number of strong signals stays bounded, and yet FDR control is the expected behavior of a valid procedure in the all-null regime, the signal-coverage condition (C3)/(O4) is plausibly an artifact of the threshold-localization proof rather than a real phenomenon; a sharper localization argument could remove it.","The proof requires of the noise vector $Z$ in the generator only moment conditions, so the construction should transfer to non-Gaussian generators with the same moment profile, demoting Gaussianity to a computational convenience."],"forward_implications":["Practitioners using the Gaussian knockoffs generator with marginal-correlation or regression-coefficient-difference statistics receive a formal asymptotic guarantee: $\\limsup_{n\\to\\infty}\\text{FDR}\\le q$ under the stated conditions.","Any working distribution is certified the same way: it suffices to verify the three conditions of Theorem 1 on its knockoff statistics, giving a modular robustness check independent of the coupling construction.","The guarantee covers cases the coupling-based theory cannot, including Rademacher covariates and $t$-distributed covariates with as few as three degrees of freedom, under the stated dimensionality restrictions.","When the precision matrix is estimated from data, the sample-moment knockoff matrix inherits the asymptotic guarantee by coupling to the population-moment version (Proposition 3).","The conditions make the signal requirement explicit: enough features ($a_n\\to\\infty$) must have strengths clearing the noise floor for the threshold-localization step to work."],"supporting_citations":[{"why":"Supplies the model-X knockoffs framework, the exchangeability condition (1), the Gaussian knockoffs construction in (2), and the threshold $T_q$ that the paper generalizes.","marker":"Candès et al., 2018"},{"why":"The prior coupling-based robustness proof whose conditions the Gaussian knockoffs provably fail (Example 1), and the source of the concentration and threshold-localization lemmas adapted here.","marker":"Fan et al., 2025"},{"why":"Earlier robustness guarantee via KL divergence, the baseline that the paper's distribution-free conditions improve upon.","marker":"Barber et al., 2020"},{"why":"Supplies the Wasserstein moderate-deviation theorem behind the ratio-based tail approximations (Theorems 5-8).","marker":"Fang and Koike, 2023"},{"why":"Defines the debiased Lasso estimator used to build the third two-moment knockoff statistic.","marker":"Zhang and Zhang, 2013"},{"why":"Smallest singular value bound used to control the sample precision matrix in the OLS argument (Lemma 15).","marker":"Rudelson and Vershynin, 2009"},{"why":"Concentration inequalities (Theorem 5.4) behind the convex concentration condition (O2) used in the OLS analysis.","marker":"Boucheron et al., 2013"}],"fun_headline_variants":["Moment matching suffices for knockoff FDR control","Gaussian knockoffs still control FDR with misspecified covariates","First two moments enough for asymptotic FDR control","Knockoff FDR control relaxed: moments beat exchangeability","Moments matching preserves false discovery control in knockoffs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee requires the number of true signals strong enough to clear the noise floor to grow without bound (condition (C3) and its analogue (O4)); in the all-null or weak-signal regime where that fails, the threshold-localization lemmas prove nothing and the theorem is silent.","fun_headline_variants_meta":{"raw":{"variants":["Moment matching suffices for knockoff FDR control","Gaussian knockoffs still control FDR with misspecified covariates","First two moments enough for asymptotic FDR control","Knockoff FDR control relaxed: moments beat exchangeability","Moments matching preserves false discovery control in knockoffs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000359,"raw_usage":{"total_tokens":1940,"prompt_tokens":941,"completion_tokens":999,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":933}},"tokens_in":557,"tokens_out":999,"duration_ms":9021,"temperature":1.0,"reasoning_tokens":933,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T17:12:51.004360+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the approximate knockoffs procedure with Gaussian knockoffs and the marginal correlation statistic on $t$-distributed covariates with three degrees of freedom and $p\\log n$ at or above $\\sqrt{n}$, so the dimensionality condition of the heavy-tailed case in Theorem 2 fails while signal conditions hold; empirical FDR staying at or below $q$ would show that boundary is not sharp, while systematic inflation above $q$ confirms it is real. A complementary check targets the signal-coverage assumption: in the all-null model, where $a_n = 0$ and condition (C3)/(O4) fails, systematic FDR inflation would show the assumption is genuinely load-bearing, whereas clean control would show the stated conditions are sufficient but not necessary.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The prior coupling-based robustness proof whose conditions the Gaussian knockoffs provably fail (Example 1), and the source of the concentration and threshold-localization lemmas adapted here."},{"cited_title":"F., Cand\\`es, E","cited_arxiv_id":null,"evidence_quote":"Earlier robustness guarantee via KL divergence, the baseline that the paper's distribution-free conditions improve upon."},{"cited_title":"and Koike, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the Wasserstein moderate-deviation theorem behind the ratio-based tail approximations (Theorems 5-8)."},{"cited_title":"and Zhang, S","cited_arxiv_id":null,"evidence_quote":"Defines the debiased Lasso estimator used to build the third two-moment knockoff statistic."},{"cited_title":"and Vershynin, R","cited_arxiv_id":null,"evidence_quote":"Smallest singular value bound used to control the sample precision matrix in the OLS argument (Lemma 15)."}],"review_version":1}