{"id":"9b8ed72d-cdf1-4f9b-afef-ff1caa026363","arxiv_id":"1908.06129","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A data-driven estimator is shown to asymptotically achieve the minimum mean squared error among all separable rules for estimating normal means with independent side information.","lead":"This paper derives the optimal rule for using a second, auxiliary dataset when estimating many unknown averages from a primary dataset, and gives a practical procedure that learns whether the auxiliary data help. It is a contribution for integrative genomics, where combining public datasets can improve estimates of gene-level effects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3 guarantees optimality for the estimator (11) over the fixed set T (10) with ρ>0, but the implemented and simulated procedure uses ρ=0 and a data-dependent search set \\hat T, a gap the paper explicitly leaves to future work.","rationale":"I read the paper in good faith. The mathematical core of Proposition 1, Theorem 1, and Theorem 3 appears coherent: the oracle δ* is the finite-sample optimal separable rule, the regularized oracle δ*_ρ is asymptotically equivalent under Assumption 1, and the SURE-based minimization argument in the proof of Theorem 3 correctly uses the fact that \\hat t minimizes sure(t) pointwise, so E sure(\\hat t) ≤ E sure(θ). Assumption 1 is a standard growth condition and the covering-number argument in Theorem 2 explicitly needs only η>0, not a stronger restriction. I did not find an internal inconsistency in the proofs I checked. The main soft spot is the acknowledged gap between the theoretical procedure (11) over T and the implemented procedure in Section 4.3: the latter uses ρ=0, a data-dependent search set \\hat T in place of the fixed T, and coordinate descent, none of which is covered by Theorems 2–3. This gap is not a fatal flaw in the theorem, but it means the paper's practical and empirical claims are not fully supported by the theory. The reader's CONDITIONAL verdict already captures this; my stress-test sharpens the concern but does not move the verdict. I also noted the probability calculation for \\hat T in Section 4.3 appears to be a typo (Φ(−M)^n is far too small), but this is not load-bearing for the central claim.","tokens_in":32836,"tokens_out":26928,"duration_ms":235198,"concrete_test":"Run the published cole implementation (ρ=0, \\hat T with M=5, coordinate descent) on the Section 5 simulation settings with increasing n = 250, 500, 1000, 2000, and compare its average loss against the oracle (4) and against a version of the theoretical procedure that uses a small positive ρ (e.g., 10^{-6}) and a fixed T large enough to contain the true θ. If the practical implementation's average loss does not approach the oracle risk, or does not track the theoretical procedure's loss, as n grows, the gap is consequential and the practical claims require a new theorem. As an analytic complement, re-derive Theorem 2 for ρ=0 and identify which step (e.g., (18) or the bound on term (34)) breaks when ρ=0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is Theorem 3: for \\hat t defined in (11) as the minimizer of sure(t) over the fixed set T in (10), Eℓ_n(\\hat t) ≤ R_n(θ,δ*)+o(1), proved for 0<ρ≤1 and for t restricted to the non-random set T whose bounds |t_jd|≤C n^{1/4−η} are assumed known. Section 4.3 then states that the R package cole runs the method with ρ=0 and, because the constants C and η are unknown, performs the minimization over the data-dependent set \\hat T=∏[X_id−Mσ_d, X_id+Mσ_d] with M=5, using coordinate descent. The paper explicitly says 'bridging the gap between the theoretical procedure and its practical implementation is an important direction for future work.' This is a genuine limitation: the uniform approximation of sure by the loss in Theorem 2, and hence the oracle inequality in Theorem 3, is established only over the fixed set T, not over the random \\hat T, and the proof uses ρ>0 in essential denominator bounds (e.g., (18) and the bounds after (34)). Therefore the implemented estimator, which is what the simulations and the genomics application actually use, is not covered by the central theorem. The abstract's claim that 'the data-driven rule' asymptotically achieves the optimal separable risk is literally true only for the (11)+T procedure with known constants and positive ρ.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies simultaneous estimation of n normal means θ_i1 from primary observations X_i1, with auxiliary observations X_i2 used as side information. The estimation problem is formulated as a compound decision problem over separable rules of the form δ_i(x1,x2)=f(x1,x2). The author derives an oracle separable rule f* (4), introduces a regularized oracle f*_ρ (6), and proposes a fully data-driven estimator obtained by minimizing an unbiased SURE-type risk estimate sure(t) over a feasible set T (10). The main theoretical results are Theorem 1 (regularized oracle is asymptotically equivalent to the oracle), Theorem 2 (uniform approximation of the loss by sure(t) over T), and Theorem 3 (the data-driven estimator asymptotically attains the optimal separable risk). The paper also reports simulations and an integrative genomics classification application, and it describes an R package implementation called cole.","tokens_in":33094,"tokens_out":8305,"duration_ms":86607,"significance":"If Theorems 1–3 hold as stated, the paper makes a useful contribution to compound decision theory and integrative analysis: it provides a direct risk-minimization route to the oracle separable rule with side information, avoiding the intermediate f- or g-modeling steps, and it gives detailed proofs using SURE and empirical process techniques. The oracle derivation in Proposition 1 is elegant, and the numerical results suggest practical potential. However, the significance is currently limited by a substantial gap between the theoretical procedure analyzed in the theorems and the procedure that is actually implemented and evaluated in the simulations and data analysis.","major_comments":[{"comment":"Theorem 3 is proved only for the exact minimizer of sure(t) over the fixed, non-random set T in (10), with 0<ρ≤1 and the constants C and η of Assumption 1 assumed known. The implementation described in §4.3, and every simulation and data analysis in Sections 5 and 6, instead uses ρ=0 and minimizes over the data-dependent set \\hat T = ∏_i [X_id−Mσ_d, X_id+Mσ_d] with M=5. The proof of Theorem 2 requires ρ>0 in the denominator bounds, for example (18) and the bounds following (34), and the uniform approximation of sure(t) by the loss is established only over the fixed set T, not over the random set \\hat T. The manuscript itself states in §4.3 that 'bridging the gap between the theoretical procedure and its practical implementation is an important direction for future work.' As a consequence, the abstract's claim that the data-driven rule asymptotically achieves the optimal separable risk, and the simulation evidence for the implemented rule, are not covered by the theorems as stated. The paper should either extend the theory to \\hat T and ρ=0, or sharply qualify the claims so that the theorem and the implementation are clearly distinguished.","section":"§4.3 and Theorem 3"},{"comment":"The estimator in Theorem 3 is \\hat t = argmin_{t∈T} sure(t), an exact global minimizer. The implementation instead uses coordinate descent with K=10 equally spaced candidate values per coordinate and a tolerance of 10^{-5}; sure(t) is nonconvex, and no result is given showing that this procedure reaches the global minimizer. Therefore the numerical comparisons in Section 5 and the genomics application in Section 6 evaluate an approximate, possibly local-minimum version of the estimator, and they cannot be interpreted as direct verifications of Theorem 3. This is a second gap between theory and implementation that should be stated explicitly, together with either a convergence guarantee for the optimization algorithm or a clear caveat that the theoretical result applies to the exact global minimizer.","section":"§4.3, Eq. (11)"}],"minor_comments":[{"comment":"The statement 'lim_{n→∞} {Eℓ_n(\\hat t) − R_n(θ,δ*)} ≤ 0' is not a well-formed limit statement; the proof actually establishes limsup ≤ 0. Moreover, because δ_{\\hat t}_ρ is a separable rule, Proposition 1 implies its risk is never below R_n(θ,δ*), so the correct conclusion is lim_{n→∞} {Eℓ_n(\\hat t) − R_n(θ,δ*)} = 0.","section":"Theorem 3"},{"comment":"The definition of the set A2 is printed as A2 = {(x1,x2) : −C n^{1/4−η} < x1 ≤ −C n^{1/4−η}}, which is empty; the later integration from −C n^{1/4−η} to C n^{1/4−η} indicates the intended upper endpoint is +C n^{1/4−η}.","section":"Appendix C, proof of Theorem 1"},{"comment":"The sentence 'it is known from Theorem 1 that E{sure(t)} achieves a global minimum at t_{jd}=θ_{jd}' is not a consequence of Theorem 1 for the regularized family with ρ>0; Theorem 1 only compares the risk of δρ* with the oracle risk and does not identify θ as the minimizer over all t. For ρ=0, t=θ makes δ^θ_0 exactly the oracle f*, but ρ=0 is excluded from Theorem 2. The statement should be corrected or proved.","section":"§4.3, paragraph on \\hat T"},{"comment":"The paper would benefit from a concise summary table that distinguishes the theoretical estimator (11) over T with 0<ρ≤1 from the practical estimator over \\hat T with ρ=0, and from an explicit statement about which claims are asymptotic, which are finite-sample, and which are only supported by simulations.","section":"General presentation"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the theory-implementation gap: the central theorems cover a procedure that the paper itself does not implement, and the implemented procedure is explicitly left to future work. The reviewer believes the central derivation is sound and the gap is fixable, either by extending the uniform approximation argument to the data-dependent set \\hat T (possibly with M growing slowly) and to ρ=0, or by carefully narrowing the paper's claims. The nonconvex optimization gap should also be addressed or explicitly acknowledged."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid contribution to the normal-means-with-side-information literature. What is genuinely new is the direct SURE-minimization strategy: instead of estimating the oracle rule through f- or g-modeling, the author proposes choosing t by minimizing an unbiased risk estimate, and proves a uniform approximation theorem and an oracle inequality for the resulting separable rules. The oracle rule itself is derivative—as the paper acknowledges, it follows from the fundamental theorem of compound decision problems—but the estimator and its risk analysis are the novel part. The proofs in the appendix are detailed and follow standard empirical-process machinery; I did not find a red flag in the mathematical core. The simulations are honest and the method does appear to approach oracle performance when the side information is informative. The genomics illustration is what it is: an illustrative application, not a rigorous benchmark, and the lack of error bars there is a minor weakness.\n\nThe main soft spot is exactly what the stress-test note says. Theorem 3 is proved for the estimator (11) that minimizes sure(t) over the fixed, non-random set T in (10), with known constants C and η and with ρ > 0. The implementation in Section 4.3 uses ρ = 0, a data-dependent search set \\hat T, and coordinate descent, with the author explicitly saying that bridging the theory-implementation gap is future work. So the abstract's claim that the data-driven rule asymptotically achieves the optimal separable risk is literally true only for the theoretical procedure, not for the procedure that actually runs in simulations and the data example. I do not think this is fatal—the regularized estimator is a legitimate object, and the paper is admirably transparent about the gap—but it is a real limitation, and it should be surfaced in the abstract or at least the introduction rather than resting in the implementation section.\n\nAssumption 1, the uniform growth bound |θ_id| ≤ C n^{1/4−η}, is restrictive but comparable to weak-moment conditions used elsewhere in this literature; I would call it a conditioning assumption rather than a defect. The citation pattern is appropriate, with proper credit to Robbins, Jiang, Zhang, and others. No self-citation inflation.\n\nWho gets value from this paper: statisticians working on compound decision theory, empirical Bayes, and integrative genomics methods. It deserves serious refereeing. I would send it out, with the request that the author clarify the scope of the theorem versus the implementation and, if possible, add a theoretical justification for the data-dependent feasible set or tone down the abstract accordingly.","headline":"A careful compound-decision paper whose main theorem covers a regularized, ρ>0 procedure on a fixed set, while the implemented and tested version uses ρ=0 on a data-dependent grid—a real gap, but one the author openly flags.","tokens_in":775,"tokens_out":1048,"would_cite":false,"duration_ms":30436,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62C25","62C12","62F10","62H12"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that, in the Gaussian sequence model with paired auxiliary data, a data-driven rule found by minimizing an unbiased risk estimate asymptotically attains the best risk achievable by any separable estimator.","keywords":["normal means","side information","compound decision","separable decision rules","SURE","empirical Bayes","data integration","high-dimensional classification"],"falsifier":"Set one primary mean to $\\theta_{i1}=n^{1/3}$ with all other means zero and $\\theta_{i2}=0$, then compare the proposed estimator's average loss with the oracle separable risk as $n$ grows; the claimed asymptotic optimality fails if the gap does not shrink to zero, since Assumption 1 is violated.","tokens_in":32595,"feed_emoji":"🎯","tokens_out":7452,"duration_ms":71802,"temperature":0.7,"pith_summary":"The paper studies how to use a second, auxiliary Gaussian sequence to estimate the primary means of a first sequence. It derives the oracle optimal separable estimator, a weighted average of the primary means with joint-normal density weights, and then builds a practical version that picks its tuning parameters by minimizing an unbiased estimate of its own risk. The main theorem states that this data-driven rule is asymptotically as good as the best separable rule that knows the true means. The paper also shows that integrating auxiliary data can improve high-dimensional classification, and it demonstrates the gain in a breast-cancer gene-expression analysis. Its deeper point is that side information helps even when the auxiliary observations are statistically independent of the primary ones, as long as the two mean vectors are related.","feed_headline":"Auxiliary data sharpen mean estimates—new rule hits optimal risk","feed_subtitle":"Minimizing an unbiased risk estimate provably attains the best separable risk in Gaussian estimation with side information.","key_machinery":"The load-bearing object is the oracle kernel rule $f^\\star(x_1,x_2)=\\sum_j \\theta_{j1}p_j^0(x_1,x_2)/\\sum_j p_j^0(x_1,x_2)$, a separable posterior expectation under the empirical prior. A regularized version $f^{\\rho}$ adds a small positive constant to the denominator so the rule is well defined everywhere. The data-driven estimator substitutes the free parameter vector $t$ into the same form and chooses $t\\in T$ to minimize an unbiased risk estimate derived from Stein's lemma. The proof chains three results: the regularized oracle has the same asymptotic risk as the oracle, the SURE statistic approximates the true loss uniformly over $T$, and minimizing that approximation therefore yields the oracle risk. The uniform approximation is proved with empirical-process truncation, symmetrization, and a covering-number bound that approximates the empirical prior by a finite mixture with a controlled number of atoms.","core_discovery":"The central claim is Theorem 3: for independent pairs $X_{i1}\\sim N(\\theta_{i1},\\sigma_1^2)$ and $X_{i2}\\sim N(\\theta_{i2},\\sigma_2^2)$, the estimator $\\delta_{\\hat t}^{\\rho}$ that minimizes the unbiased risk estimate $\\operatorname{sure}(t)$ over a bounded set $T$ satisfies $\\lim_{n\\to\\infty}\\{E\\ell_n(\\hat t)-R_n(\\theta,\\delta^\\star)\\}\\le 0$, where $R_n(\\theta,\\delta^\\star)$ is the minimum risk among all separable decision rules. The oracle $\\delta^\\star$ is the posterior expectation of $\\theta_{i1}$ under the discrete empirical prior, and the data-driven estimator approximates it by replacing the unknown means with free tuning parameters selected by SURE. Thus the procedure achieves oracle-level separable risk without requiring any explicit model for how $\\theta_{i2}$ relates to $\\theta_{i1}$.","pith_inferences":["Editorial: the same SURE-minimization template should transfer to heteroscedastic sequences or unknown variances, where Stein-type unbiased risk estimates already exist; the main obstacle is computational rather than statistical.","Editorial: at small sample sizes with uninformative side data, the procedure initially underperforms single-sequence methods; an adaptive screening step that tests whether the auxiliary sequence is informative before integrating would likely close that gap.","Editorial: the finite-mixture approximation used in the proof suggests a practical alternative to coordinate descent: solve for a small number of support points and weights of the empirical prior, turning the nonconvex optimization into a lower-dimensional problem.","Editorial: because the oracle rule is a conditional expectation, standard diagnostics such as binned observed-versus-estimated means and bootstrap intervals could be built to measure how much side information is being borrowed in a given dataset."],"forward_implications":["Integrative estimation becomes fully data-driven: no model for the relationship between the primary mean and the auxiliary observation is needed, so nonlinear or unknown relationships cause no misspecification problem.","Asymptotically, using an auxiliary sequence cannot hurt: when the side information is uninformative the rule learns to discard it, and when informative it approaches the oracle separable risk.","Plugging the estimated means into the Greenshtein-Park classification rule yields an integrative high-dimensional classifier; in the paper's breast-cancer analysis, the integrative rule achieved the lowest misclassification rate for treatment-response prediction.","The framework extends to correlated primary and auxiliary observations, where auxiliary data help even when the mean vectors are unrelated, because the correlation itself carries information about the primary means.","The estimator, although not designed for sparse means, matches or beats sparse-focused methods in simulations when the side information is strong and the sample size is large."],"supporting_citations":[{"why":"Supplies the compound-decision theorem that makes the oracle rule risk-minimal among separable rules.","marker":"(Robbins, 1951)"},{"why":"Gives the g-modeling estimator used as a baseline and provides related theory for asymptotically optimal separable rules.","marker":"(Jiang et al., 2009)"},{"why":"Characterizes the optimal separable posterior-expectation rule that the oracle extends to the side-information setting.","marker":"(Zhang, 2003)"},{"why":"Provides Stein's lemma used to construct the unbiased risk estimate sure(t).","marker":"(Stein, 1981)"},{"why":"Establishes the nonparametric empirical Bayes and compound-decision framework for normal means that the paper builds on.","marker":"(Brown and Greenshtein, 2009)"},{"why":"Supplies the moment-matching lemma used to approximate the empirical prior and bound covering numbers in the proof of Theorem 2.","marker":"(Ghosal and Van Der Vaart, 2001)"},{"why":"Provides the symmetrization and maximal-inequality machinery for the uniform approximation argument.","marker":"(van der Vaart and Wellner, 1996)"},{"why":"Links normal-means risk to high-dimensional classification, motivating the integrative classification application.","marker":"(Greenshtein and Park, 2009)"}],"fun_headline_variants":["Side information sharpens mean estimates to oracle-level risk","SURE-based rule achieves optimal separable risk in mean estimation","Oracle risk attainable without modeling side information","Data-driven rule matches oracle for simultaneous normal means","Side information lets mean estimates hit minimal separable risk"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All true means in both sequences must stay within $Cn^{1/4-\\eta}$ in absolute value for some positive constants $C$ and $\\eta$; if a nontrivial share of means grows faster than this, the regularized oracle is not guaranteed to match the oracle and the SURE approximation to the true loss can break down.","fun_headline_variants_meta":{"raw":{"variants":["Side information sharpens mean estimates to oracle-level risk","SURE-based rule achieves optimal separable risk in mean estimation","Oracle risk attainable without modeling side information","Data-driven rule matches oracle for simultaneous normal means","Side information lets mean estimates hit minimal separable risk"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00051,"raw_usage":{"total_tokens":2452,"prompt_tokens":887,"completion_tokens":1565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":1493}},"tokens_in":503,"tokens_out":1565,"duration_ms":12372,"temperature":1.0,"reasoning_tokens":1493,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:54:10.490849+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Set one primary mean to $\\theta_{i1}=n^{1/3}$ with all other means zero and $\\theta_{i2}=0$, then compare the proposed estimator's average loss with the oracle separable risk as $n$ grows; the claimed asymptotic optimality fails if the gap does not shrink to zero, since Assumption 1 is violated.","supporting_citations":[],"review_version":1}