Pith. sign in

REVIEW 4 major objections 4 minor 11 references

Robust Mean Estimation With Auxiliary Samples

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read One scalar weight sets the optimal blend of true and auxiliary samples.

desk verdict 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. read the letter →

arxiv 2501.18095 v2 pith:46BZLIKZ submitted 2025-01-30 math.ST stat.TH

Barron Han , Danil Akhtiamov , Reza Ghane , Babak Hassibi This is my paper · ORCID
classification math.STstat.TH MSC 62C2062F3562G05
keywords meanestimationauxiliarysamplesWasserstein-2distancedistributionallyrobustoptimizationminimaxestimatordataaugmentationGaussianlocationmodellinear
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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 $R^*_F$.

What would settle it

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 $N$.

Watch

Extended reading notes

Core claim

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 $\delta^2$.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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 $R^*_F$.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

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.

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 (4)
  1. [IV, Eqs. (18)–(19)] 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.
  2. [Theorem 1, scalar counterexample] 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.
  3. [Appendices A and B] 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.
  4. [Theorem 1, Eq. (7)] 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.
minor comments (4)
  1. [IV, around Eq. (22)] 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.
  2. [V, Figure 1] 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.
  3. [II, Problem formulation] 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.
  4. [Abstract and Section II] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the minimax derivation is self-contained; the N→∞ substitution in the proof is a correctness gap, not a circular reduction.

full rationale

The paper's claimed minimax values come from solving (3a)–(3c) directly. After Assumption 1 restricts attention to linear estimators, Lemma 1 uses Gelbrich's W2 formula (Theorem 2.1 of [11]) to justify a reduction to Gaussian laws; this is an external optimal-transport result, not a self-citation or an assumption that contains the target conclusion. The proof then performs an explicit optimization over A, Σ_X, and Σ_Z. No parameter is fitted to data, and no empirical output is later relabeled as a prediction; the constants s and R* depend only on the declared problem parameters n, N, d, ε, and δ. The numerical section evaluates these closed-form expressions rather than fitting them. The most load-bearing step, replacing the stationarity solution (18) by its N→∞ limit (19), is a mathematical approximation/error in the derivation: it is not a reduction of the announced result to an input or to the paper's own assumptions. The paper even states earlier that N≫1 is assumed, though the theorem omits that condition and gives no error bound; that is a correctness and rigor concern, not a circularity concern. Because the derivation chain is self-contained against external benchmarks and contains no fitted inputs, self-citation chains, or definitional equivalences, the circularity score is 0.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new entities; the only added structure beyond the problem inputs is the linear estimator class and the normalization constant delta. The heavy lifting is done by Gelbrich's theorem and an unproven large-N covariance alignment.

free parameters (1)
  • delta = user-specified lower bound on ||Sigma_X||_F
    introduced in (2) to normalize the objective and keep it finite; the weight s and risk R* depend on delta^2/epsilon^2.
assumptions (3)
  • standard math Gelbrich's theorem: W2(P,Q) >= W2(N(mu_P,Sigma_P), N(mu_Q,Sigma_Q))
    used in Lemma 1 to reduce worst-case distributions to Gaussians with matched first and second moments; external reference [11].
  • domain assumption Assumption 1: the estimator is restricted to the linear form f(X,Z)=A*(sum X_i/n) + B*(sum Z_j/N)
    the paper does not prove that this linear class contains the optimal estimator over all measurable functions; the 'optimality' is conditional on this restriction.
  • ad hoc to paper Large-N covariance alignment: the optimal auxiliary covariance \tilde{Sigma_Z} in (18) is replaced by Sigma_X as N -> infinity
    used between (18) and (19); no theorem condition or error bound accompanies this substitution, and it is invalid for finite N.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Mean Estimation With Auxiliary Samples." pith.science (2026). https://pith.science/paper/46BZLIKZ

@misc{pith2026250118095,
  author       = {Pith},
  title        = {Pith review of: Robust Mean Estimation With Auxiliary Samples},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/46BZLIKZ}},
  note         = {Machine review of arXiv:2501.18095}
}
read the original abstract

In data-driven learning and inference tasks, the high cost of acquiring samples from the target distribution often limits performance. A common strategy to mitigate this challenge is to augment the limited target samples with data from a more accessible "auxiliary" distribution. This paper establishes fundamental limits of this approach by analyzing the improvement in the mean square error (MSE) when estimating the mean of the target distribution. Using the Wasserstein-2 metric to quantify the distance between distributions, we derive expressions for the worst-case MSE when samples are drawn (with labels) from both a target distribution and an auxiliary distribution within a specified Wasserstein-2 distance from the target distribution. We explicitly characterize the achievable MSE and the optimal estimator in terms of the problem dimension, the number of samples from the target and auxiliary distributions, the Wasserstein-2 distance, and the covariance of the target distribution. We note that utilizing samples from the auxiliary distribution effectively improves the MSE when the squared radius of the Wasserstein-2 uncertainty ball is small compared to the variance of the true distribution and the number of samples from the true distribution is limited. Numerical simulations in the Gaussian location model illustrate the theoretical findings.

Figures

Figures reproduced from arXiv: 2501.18095 by the authors.

Figure 1
Figure 1. MSE of optimal estimator from Theorem 1. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 8 canonical work pages

  1. [1]

    Prediction-powered inference,

    A. N. Angelopoulos, S. Bates, C. Fannjiang, M. I. Jordan, and T. Zrnic, “Prediction-powered inference,” Science, vol. 382, no. 6671, pp. 669–674, 2023. [Online]. Available: https://www.science.org/doi/ abs/10.1126/science.adi6000

    doi:10.1126/science.adi6000 2023

  2. [2]

    Data augmentation using generative adversarial networks (CycleGAN) to improve generalizability in CT segmentation tasks,

    V . Sandfort, K. Yan, P. J. Pickhardt, and R. M. Summers, “Data augmentation using generative adversarial networks (CycleGAN) to improve generalizability in CT segmentation tasks,” Scientific Reports , vol. 9, no. 1, pp. 16 884–9, 2019

    work page 2019

  3. [3]

    A General Decision Theory for Huber's $\epsilon$-Contamination Model

    M. Chen, C. Gao, and Z. Ren, “A general decision theory for Huber’s ϵ - contamination model,” 2017. [Online]. Available: https://arxiv.org/abs/1511.04144

    work page Pith review arXiv 2017

  4. [4]

    A theoretical review of modern robust statistics,

    P.-L. Loh, “A theoretical review of modern robust statistics,” Annual Review of Statistics and Its Application , vol. 12, 2024

    work page 2024

  5. [5]

    Data-driven distributionally robust optimization using the Wasserstein metric: Performance guarantees and tractable reformulations,

    P. Mohajerin Esfahani and D. Kuhn, “Data-driven distributionally robust optimization using the Wasserstein metric: Performance guarantees and tractable reformulations,” Mathematical Programming, vol. 171, no. 1, pp. 115–166, 2018

    work page 2018

  6. [6]

    The performance of Wasserstein distributionally robust m-estimators in high dimensions,

    L. Aolaritei, S. Shafieezadeh-Abadeh, and F. Dörfler, “The performance of Wasserstein distributionally robust m-estimators in high dimensions,” arXiv:2206.13269, 2023

    arXiv 2023

  7. [7]

    Wasserstein Distributionally Robust Optimization: Theory and Applications in Machine Learning

    D. Kuhn, P. M. Esfahani, V . A. Nguyen, and S. Shafieezadeh- Abadeh, “Wasserstein distributionally robust optimization: Theory and applications in machine learning,” 2024. [Online]. Available: https://arxiv.org/abs/1908.08729

    work page Pith review arXiv 2024

  8. [8]

    Generalized resilience and robust statistics,

    B. Zhu, J. Jiao, and J. Steinhardt, “Generalized resilience and robust statistics,” The Annals of Statistics , vol. 50, no. 4, pp. 2256–2283, 2022

    work page 2022

Show all 11 references
  1. [9]

    Robust W-GAN-based estimation under Wasser- stein contamination,

    Z. Liu and P.-L. Loh, “Robust W-GAN-based estimation under Wasser- stein contamination,” Information and Inference: A Journal of the IMA , vol. 12, no. 1, pp. 312–362, 2023

    2023

  2. [10]

    Statistical estimation under distribution shift: Wasserstein perturbations and minimax theory,

    P. Chao and E. Dobriban, “Statistical estimation under distribution shift: Wasserstein perturbations and minimax theory,” arXiv preprint arXiv:2308.01853, 2023

    2023

  3. [11]

    On a formula for the L2 Wasserstein metric between measures on Euclidean and Hilbert spaces,

    M. Gelbrich, “On a formula for the L2 Wasserstein metric between measures on Euclidean and Hilbert spaces,” Mathematische Nachrichten, vol. 147, no. 1, pp. 185–203, 1990. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/mana.19901470121 APPENDIX A PROOF OF ...

    1990 doi

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.