{"id":"c2c2187c-8143-4895-8de5-88e735c356a7","arxiv_id":"2501.18095","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper derives a linear shrinkage estimator for mean estimation with auxiliary samples under a Wasserstein-2 constraint, but the claimed exact minimax risk is only an asymptotic-in-N approximation and is contradicted for finite N.","lead":"This paper asks whether cheap samples from a similar, auxiliary distribution can improve mean estimation of a target distribution when the two distributions are within a Wasserstein-2 distance. It proposes a closed-form weighted-averaging estimator, but the claimed exact worst-case error formula is only a large-N approximation and fails for finite auxiliary sample sizes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's finite-N minimax claim is false: the proof replaces the stationarity solution for Σ_Z (eq. 18) by its N→∞ limit Σ_X (eq. 19), omitting a covariance-mismatch term; for d=1, n=1, N=10, ε=δ=1 the claimed value 11/21 is below the true worst-case 10/19.","rationale":"The reader's verdict is REJECT, and the stress-test agrees. The Gaussian reduction (Lemma 1) and the derivation up to (18) are sound; the problem is that (18) is not used as a finite-N stationarity condition but is immediately replaced by its N→∞ limit. This is not a cosmetic approximation: for A = sI the exact stationarity solution is Σ_Z = N^2/(N-1)^2 Σ_X, which differs from Σ_X by O(1/N). In the W2 constraint, substituting Σ_Z=Σ_X frees the covariance-mismatch term (sqrt(Σ_X)−sqrt(Σ_Z))^2 for use as a mean shift, underestimating the worst-case MSE. The scalar counterexample is decisive because it is within the paper's own model: P=N(0,1), Q with mean sqrt(80/81) and variance 100/81, W2(P,Q)=1, ||Σ_X||_F=1. For the paper's optimal s=11/21, the MSE is 11/21, but the adversary can do better: at s=10/19 and Q as above, MSE=10/19 > 11/21. Since the theorem claims R*_F is the worst-case optimal value, it is false as stated. The paper's simulations use N=1000, which is large enough that the discrepancy is invisible, and the N≫1 remark in Section II hints at a large-N intent, but the theorem statements and abstract do not qualify the formulas as asymptotic. A conditional acceptance would be possible if the authors restated the results with explicit O(1/N) error bounds or solved the finite-N problem exactly; as written, REJECT is appropriate.","tokens_in":9464,"tokens_out":8832,"duration_ms":79523,"concrete_test":"Evaluate the scalar minimax problem exactly: set d=1, n=1, N=10, ε=δ=1 in (14) and maximize over μ_X−μ_Z and Σ_Z subject to W2(P,Q)≤1 and Σ_X≥1, then minimize over the scalar A. Use the closed-form worst-case sqrt(Σ_Z)=N/(N-1), which follows from (18). If the resulting minimum is 10/19 rather than 11/21, Theorem 1 is false for finite N and the large-N substitution is the cause.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is the transition from equation (18) to equation (19) in the proof of Theorem 1. After solving the stationarity condition for the auxiliary covariance, the authors note that as N → ∞ the solution converges to Σ_X and then substitute Σ_Z = Σ_X into the objective. This removes from the W2 budget the term Tr(Σ_X + Σ_Z − 2(Σ_X^{1/2}Σ_ZΣ_X^{1/2})^{1/2}), which for finite N is positive and reduces the budget available for the mean shift. For the optimal diagonal structure A = sI, the stationarity solution (18) is tilde Σ_Z = N/(N-1) Σ_X, i.e. Σ_Z = N^2/(N-1)^2 Σ_X, not Σ_X. Keeping this term changes both the optimal weight s and the minimax value. In the scalar case (d=1, n=1, N=10, ε=δ=1), the theorem yields s=11/21 and R*=11/21 ≈ 0.5238, while the exact worst-case optimization over Σ_Z and the mean shift gives R*=10/19 ≈ 0.5263, achieved at s=10/19 and sqrt(Σ_Z)=10/9. Thus the stated formula is not an upper bound on the true risk; the announced exact minimax value fails for finite N. The theorems are plausible large-N asymptotics but are not exact as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies minimax linear estimation of the mean of a target distribution from n target samples and N auxiliary samples, when the target and auxiliary distributions are within Wasserstein-2 distance ε and the estimator is restricted to be linear in the two sample means. It claims exact formulas for the optimal weighting between target and auxiliary samples and for the minimax MSE under three normalizations (Frobenius, trace, and operator norm of the target covariance). The main result, Theorem 1, states that the optimal estimator is A = sI, B = I − A and gives an explicit finite-N expression for the worst-case MSE.","tokens_in":9673,"tokens_out":16650,"duration_ms":146727,"significance":"If the claims were correct, the paper would provide a clean, closed-form minimax characterization of a natural two-sample mean-estimation problem with a distribution-shift constraint, complementing the existing Wasserstein distributional-robustness literature. The paper has real strengths: Lemma 1's reduction to Gaussian measures via Gelbrich's theorem is sound and useful; the linear-estimator parametrization is natural; and the algebra after equation (19) is internally consistent. The numerical experiment, while simple, illustrates the qualitative benefit of auxiliary samples. However, the central finite-N minimax claim is not supported by the proof, and a scalar counterexample shows the stated formula is not an upper bound on the true worst-case risk. The exactness of Theorems 1–3 for finite N is therefore the paper's main weakness, not a presentation issue.","major_comments":[{"comment":"The proof of Theorem 1 replaces the stationarity solution for the auxiliary covariance, \\tildeΣ_Z in equation (18), by its limit Σ_X as N → ∞ (stated after (18)), and then optimizes equation (19) as if \\tildeΣ_Z = Σ_X exactly. No finite-N error bound or large-N condition is stated in Theorem 1, yet the theorem claims an exact finite-sample minimax value. The omitted term Tr(Σ_X + Σ_Z − 2(Σ_X^{1/2}Σ_ZΣ_X^{1/2})^{1/2}) in the Wasserstein-2 constraint is positive for finite N, so this substitution removes a genuine budget term and changes the finite-N trade-off. This step is load-bearing because the remaining proof optimizes a different problem and never establishes that the original objective is bounded above by the claimed value.","section":"IV, Eqs. (18)–(19)"},{"comment":"The claimed finite-N optimal value is not merely unproved; it is false. For d=1, n=1, N=10, ε=δ=1, equation (18) gives \\tildeΣ_Z = (10/9)Σ_X, so the covariance-mismatch term is not negligible. Optimizing the original worst-case objective in (16) over Σ_X and the mean shift with A = sI yields worst-case risk 10/19 ≈ 0.5263 at s = 10/19, whereas Theorem 1 reports 11/21 ≈ 0.5238. Since 10/19 exceeds 11/21, the stated R*_F is not an upper bound on the true worst-case risk, contradicting the announced minimax optimality for finite N.","section":"Theorem 1, scalar counterexample"},{"comment":"The proofs of Theorems 2 and 3 explicitly reuse 'all steps in the proof of Theorem 1 up to equation (19)' (Appendix A, first sentence; Appendix B likewise). They therefore inherit the same covariance-substitution approximation, and the same finite-N gap applies to the trace-normalized and operator-norm-normalized objectives. A defensible version of the paper would need to either state a quantitative large-N regime with error bounds or solve the finite-N optimization directly; neither is present.","section":"Appendices A and B"},{"comment":"The displayed formula for R*_F in Theorem 1 is algebraically inconsistent with the value derived in the proof, equation (28). The proof's expression simplifies to √d(ε²N/δ² + √d)/(ε²nN/δ² + √d(n+N)), which is equivalent to √d(δ²√d/ε² + N)/(δ²√d(n+N)/ε² + nN), not to the printed √d(δ²√d/ε² + N)/(δ²√d(n+N)/ε² + ε²nN). The last denominator term appears to contain an erroneous ε²; unless ε=δ, the theorem's formula and the proof's formula disagree. This is a central stated result and the mismatch should be corrected immediately.","section":"Theorem 1, Eq. (7)"}],"minor_comments":[{"comment":"In the first case of the SVD analysis, the text writes 's_d = 1 + N/n = s' where the intended expression is likely 's_d = 1/(1+N/n) = s'; the missing fraction bar makes the case analysis difficult to follow.","section":"IV, around Eq. (22)"},{"comment":"The caption of Figure 1 does not report the number of Monte Carlo trials or whether error bars are shown; adding this information would support the comparison between the three estimators.","section":"V, Figure 1"},{"comment":"The assumption 'N ≫ 1' is stated informally, but Theorems 1–3 present exact finite-N formulas; the manuscript should either formalize a large-N threshold or explicitly qualify the theorems as asymptotic.","section":"II, Problem formulation"},{"comment":"The phrase 'with labels' in the abstract is not defined in the main text; consider replacing it with 'when the origin of each sample (target vs. auxiliary) is known' for clarity.","section":"Abstract and Section II"}],"recommendation":"reject","confidential_remarks":"The scalar counterexample is decisive for the finite-N claim, and the same gap propagates to Theorems 2 and 3. The paper might be salvageable as an asymptotic contribution with quantitative error bounds, or as a finite-N solution of the correct optimization, but the current version's central exactness claims are false. No concerns about novelty or citation practices."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the question is good, the Gaussian reduction is real, and the qualitative answer is probably right. But the main theorem as stated is false for finite N, because the proof replaces the stationarity solution for Sigma_Z with its N-to-infinity limit and gives no error bounds.\n\nWhat's new: the two-sample labeled setting under a W2 ball, with the linear estimator restricted to A and B = I - A, and the closed-form expressions for s and R*. The reduction to Gaussians via Gelbrich is correct, and the algebra after equation (19) checks out. The paper is honest that it assumes N >> 1 in the problem setup, but the theorem statements don't carry that qualification, and the proof doesn't quantify the approximation.\n\nThe soft spot: the transition from (18) to (19) is load-bearing. The stationarity condition gives tilde-Sigma_Z proportional to Sigma_X with a factor depending on N, not equal to Sigma_X. Dropping the covariance-mismatch term changes the finite-N trade-off. The scalar counterexample in the stress test is compelling: d=1, n=1, N=10, epsilon=delta=1 gives claimed R* = 11/21 but the true worst-case risk is 10/19. So Theorem 1 as written is not an upper bound on the true minimax risk. This is a derivation gap, not a circularity problem; there are no fitted parameters or self-citation issues. The fix is feasible: either solve the finite-N optimization exactly or state a clean asymptotic theorem with explicit convergence rates. The qualitative message—auxiliary data helps when epsilon^2/delta^2 is small and n is limited—may well survive the correction.\n\nWho it's for: people working on Wasserstein DRO or data augmentation theory. It's a useful problem statement and a plausible starting point, but the main result needs work. I'd send it to referees, but with a clear flag that the proof's gap is the central issue.","headline":"Good question, correct Gaussian reduction, but the main theorem is not true as stated for finite N because the proof swaps in a large-N limit without error bounds.","tokens_in":10324,"tokens_out":1989,"would_cite":false,"duration_ms":22225,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62C20","62F35","62G05"],"pacs":[],"model":"deepseek-v4-flash","headline":"One scalar weight sets the optimal blend of true and auxiliary samples.","keywords":["mean estimation","auxiliary samples","Wasserstein-2 distance","distributionally robust optimization","minimax estimator","data augmentation","Gaussian location model","linear estimator"],"falsifier":"Compute the finite-$N$ min-max value of the objective in (14) for $d=1$, small $n$ and $N$, and $\\epsilon = \\delta = 1$, optimizing over the auxiliary covariance rather than substituting the target covariance; if the value differs from (7), the closed-form formula is not exact for finite $N$.","tokens_in":9129,"feed_emoji":"🎯","tokens_out":7960,"duration_ms":74161,"temperature":0.7,"pith_summary":"The paper asks when extra samples from a cheap auxiliary distribution actually improve estimating the mean of a target distribution, and by how much. Modeling the gap between the two distributions as a Wasserstein-2 ball of radius $\\epsilon$, it claims a complete minimax answer for linear estimators: the optimal rule is a weighted average of the true-sample mean and the auxiliary-sample mean, controlled by a single scalar $s$. The paper derives $s$ and the worst-case mean squared error explicitly in terms of dimension $d$, target sample count $n$, auxiliary sample count $N$, uncertainty radius $\\epsilon$, and covariance lower bound $\\delta^2$. If correct, this gives a provable trade-off: auxiliary data help substantially only when $\\epsilon^2$ is small compared to the target variance and $n$ is limited. The argument proceeds by reducing arbitrary distributions to Gaussians with matching first two moments.","feed_headline":"One formula sets the optimal mix of true and auxiliary samples","feed_subtitle":"Worst-case mean estimation under Wasserstein-2 uncertainty reduces to one scalar weight s with an explicit error bound.","key_machinery":"The proof rests on four moves. First, Lemma 1 reduces arbitrary distributions to Gaussians with the same mean and covariance, using the fact that the Wasserstein-2 distance between the Gaussian proxies is no larger than between the original distributions. Second, the estimator is restricted to linear combinations of the two sample averages, and the relation $A+B=I$ is forced by unboundedness of the objective. Third, an SVD of $A$ diagonalizes the optimization, collapsing the matrix problem to a scalar $s$. Fourth, a stationarity condition on the auxiliary covariance is solved, and the paper substitutes the target covariance for the auxiliary covariance on the strength of an $N \\to \\infty$ convergence argument; this substitution is what turns the finite-$N$ trade-off into the closed-form $s$ and $R^*_F$.","core_discovery":"On the paper's own terms, the central discovery is that the distributionally robust mean estimation problem in (3a) has an explicit solution: the optimal linear estimator is $A = sI$ and $B = I - A$, with $s = (\\sqrt{d}/N + \\epsilon^2/\\delta^2)/(\\sqrt{d}/N + \\sqrt{d}/n + \\epsilon^2/\\delta^2)$, and the worst-case normalized mean squared error is $R^*_F = \\sqrt{d}((\\delta^2/\\epsilon^2)\\sqrt{d} + N)/ ((\\delta^2/\\epsilon^2)\\sqrt{d}(n+N) + nN)$. Equivalent results hold when the objective is normalized by trace or operator norm; only the dimension's role changes. The paper claims these formulas are exact minimax answers for the labeled auxiliary-sample setting, where the auxiliary distribution is chosen adversarially within $W_2$ distance $\\epsilon$ of the target, subject to the target covariance norm being at least $\\delta^2$.","pith_inferences":["Beyond the paper, the $N \\to \\infty$ substitution in the proof suggests the closed form should be read as a large-$N$ characterization; an exact finite-$N$ formula would likely include a covariance-mismatch correction term.","Beyond the paper, the Gaussian reduction predicts that any two target and auxiliary distributions with the same first two moments and Wasserstein distance yield the same worst-case MSE, a claim testable with heavy-tailed or discrete distributions.","Beyond the paper, the same worst-case analysis could be applied to other statistics such as variance or regression coefficients, where labeled auxiliary data play the same augmentation role.","Beyond the paper, the framework supplies a worst-case baseline for practical data augmentation: if the true and auxiliary distributions are closer than $\\epsilon$ in $W_2$, the guaranteed improvement is at least the gap between the target-only sample mean and $R^*_F$."],"forward_implications":["When $\\epsilon^2/\\delta^2$ is large relative to $\\sqrt{d}/n$, $s$ approaches 1 and the auxiliary samples are essentially ignored, reverting the estimator to the target-sample mean.","When $\\epsilon^2/\\delta^2$ is small and $n$ is limited, $s$ approaches 0, so the optimal rule mostly averages the auxiliary samples.","Auxiliary samples never hurt in the worst-case sense, because $s=1$ recovers the target-only sample mean.","The worst case always saturates the Wasserstein constraint, so the mean shift between the target and auxiliary distributions is exactly $\\epsilon$ in the worst case.","The same scalar structure survives under trace and operator-norm normalizations, only with $d$ replaced by 1 or appearing as a factor."],"supporting_citations":[{"why":"Supplies the Wasserstein-2 identity for Gaussian measures and the monotonicity property used to justify the reduction to Gaussian distributions.","marker":"[11]"},{"why":"Establishes the earlier exact minimax risk for mean estimation under Wasserstein distribution shift, which this paper extends by adding labeled auxiliary samples.","marker":"[10]"},{"why":"Motivates the auxiliary-data model through prediction-powered inference, where a small true sample is supplemented by a large generated sample.","marker":"[1]"},{"why":"Provides a prior robust estimation method under Wasserstein contamination whose optimality this paper compares against.","marker":"[9]"},{"why":"Gives the prior generalized-resilience approach that asymptotically attains optimal error under Wasserstein or total-variation shifts.","marker":"[8]"}],"fun_headline_variants":["Optimal auxiliary-sample weight is a single explicit scalar","Exact minimax mean estimator reduces to one scalar s","Robust mean estimation: closed-form optimal weight for auxiliary data","Auxiliary samples help only when uncertainty ball is small","Explicit worst-case MSE for labeled auxiliary samples"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes that in the worst case the auxiliary distribution's covariance can be replaced by the target distribution's covariance, relying on the number of auxiliary samples being very large; the closed-form formulas depend on this substitution.","fun_headline_variants_meta":{"raw":{"variants":["Optimal auxiliary-sample weight is a single explicit scalar","Exact minimax mean estimator reduces to one scalar s","Robust mean estimation: closed-form optimal weight for auxiliary data","Auxiliary samples help only when uncertainty ball is small","Explicit worst-case MSE for labeled auxiliary samples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2772,"prompt_tokens":955,"completion_tokens":1817,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":1738}},"tokens_in":571,"tokens_out":1817,"duration_ms":12842,"temperature":1.0,"reasoning_tokens":1738,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:43:33.189939+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the finite-$N$ min-max value of the objective in (14) for $d=1$, small $n$ and $N$, and $\\epsilon = \\delta = 1$, optimizing over the auxiliary covariance rather than substituting the target covariance; if the value differs from (7), the closed-form formula is not exact for finite $N$.","supporting_citations":[{"cited_title":"Statistical estimation under distribution shift: Wasserstein perturbations and minimax theory,","cited_arxiv_id":null,"evidence_quote":"Establishes the earlier exact minimax risk for mean estimation under Wasserstein distribution shift, which this paper extends by adding labeled auxiliary samples."},{"cited_title":"Robust W-GAN-based estimation under Wasser- stein contamination,","cited_arxiv_id":null,"evidence_quote":"Provides a prior robust estimation method under Wasserstein contamination whose optimality this paper compares against."},{"cited_title":"Generalized resilience and robust statistics,","cited_arxiv_id":null,"evidence_quote":"Gives the prior generalized-resilience approach that asymptotically attains optimal error under Wasserstein or total-variation shifts."}],"review_version":1}