Pith. sign in

REVIEW 3 major objections 4 minor 10 references

Moment Relaxations for Data-Driven Wasserstein Distributionally Robust Optimization

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Moment relaxations of Wasserstein DRO preserve the empirical problem's asymptotic consistency.

desk verdict A genuinely new and useful idea—moment relaxations for Wasserstein DRO with O(r) consistency theorems—but the proof has a concrete error in Lemma 3.5 and an N-independence gap in Theorem 1.1(i) that a referee should push to fix. read the letter →

arxiv 2505.19278 v1 pith:RN73FNPM submitted 2025-05-25 math.OC

Shixuan Zhang , Suhan Zhong This is my paper · ORCID
classification math.OC MSC 90C2390C2290C1590C17
keywords DistributionallyrobustoptimizationWassersteindistancemomentrelaxationpolynomialsemidefinitetwo-stagestochasticprogrammingO(r)-consistency
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 proposes replacing the inner maximization in data-driven Wasserstein distributionally robust optimization (DRO) by a moment relaxation: rather than searching over all probability measures in a Wasserstein ball, one searches over truncated pseudo-moment sequences, turning the inner problem into a semidefinite program (a convex matrix-inequality problem). The central claim is that this replacement preserves the known $O(r)$-consistency of Wasserstein DRO: for bounded decision sets, when the relaxation order $k$ is high enough and either the Wasserstein exponent $p$ is at least the degree of the polynomial cost or the uncertainty set satisfies a quadratic-module ball condition, every optimal decision of the relaxed problem is an $O(r)$-optimal solution of the empirical stochastic optimization problem. The same statement is proved for two-stage problems with linear recourse, provided the recourse dual variables are bounded, and a strengthened relaxation is proposed for the unbounded case. A reader should care because the original Wasserstein DRO is generally NP-hard, while the relaxed problems are convex and semidefinite representable; the theorems say the tractable surrogates inherit the asymptotic consistency that makes data-driven Wasserstein DRO attractive.

What carries the argument

The central object is the truncated pseudo-moment sequence (tms): a finite vector $y=(y_\alpha)_{\alpha\in\mathbb{N}^l_{2k}}$ with $y_0=1$ whose moment matrix $M_k[y]$ and localizing matrices $L^{(k)}_{h_j}[y]$ are positive semidefinite. The set $\bar S[h]_{2k}$ of such sequences is the semidefinite stand-in for the set of probability measures on $\Xi=\{h\ge0\}$; the two-stage version $\bar S[g,h]_{2k}$ does the same on the joint $(\xi,u)$ region. The argument that the relaxation is tight to within $O(r)$ compares the relaxed primal value $V_{p,k}$ with the empirical value $V_p(x;0)$ through the Lagrangian dual pair (3.3)--(3.5), then uses moment-matrix positivity and norm inequalities to show that the truncated moments of any feasible relaxed measure stay close to the point mass at the corresponding sample. Degree conditions ensure the cost polynomial is read off exactly from the truncated moments, and the quadratic-module ball condition supplies the boundedness needed when the cost degree exceeds $p$.

What would settle it

Take a bounded decision set $X$, a fixed polynomial $F$ with $\deg(F_x)\le p$, and samples $\hat\xi^{(i)}$ in an unbounded set $\Xi$ with $\|\hat\xi^{(i)}\|\to\infty$; compute $C_1=\max\{\|F(x,\cdot+\hat\xi^{(i)})\|_\infty: x\in X,\ i\le N\}$. If $C_1\to\infty$ as $N\to\infty$, the uniform-constant premise of Theorem 3.7 fails, and one can directly test whether $v_{p,k}(x;r^p)-v_p(x;r^p)$ has a slope in $r$ that grows with $N$; that would show the consistency rate is not uniform in the sample size as claimed.

Watch

Extended reading notes

Core claim

At the level of the paper's own claims, the discovery is a moment-relaxation version of the consistency theorem for Wasserstein DRO. For single-stage problems with polynomial cost $F$ and even Wasserstein order $p$, the relaxed value $v_{p,k}(x;r^p)$ satisfies $v_{p,k}(x;r^p)-v_p(x;r^p)=O(r)$ for small $r>0$, whenever $2k\ge \max\{\deg(F_x),\deg(h),p\}$ and either (i) $p\ge\deg(F_x)$ or (ii) $R-\|\xi\|^2\in Q[h]_{2k_1+2}$ with $k\ge\deg(F_x)-p/2+k_1$. Therefore any optimal decision $x$ of the single-stage relaxation (1.5) is an $O(r)$-optimal solution of the empirical problem (1.2). For two-stage linear recourse, Theorem 4.4 gives the analogous statement under the bounded-dual assumption $R_0-\|u\|^2\in Q[g,h]_{2k_0}$, and Section 4.2 adds a strengthened relaxation for unbounded dual variables. Examples 3.8 and 4.5 demonstrate that the degree and boundedness conditions are necessary in general: without them the relaxed problem can be unbounded or fail to be $O(r)$-consistent.

Load-bearing premise

The load-bearing premise is that the coefficient norms of the translated polynomials $F(x,\cdot+\hat\xi^{(i)})$ (and of the two-stage recourse data) remain bounded uniformly over all samples $i$ and all $x\in X$; if the sample cloud can drift outward as $N$ grows, this uniform bound can fail and the claimed $N$-independent $O(r)$ constant is not established.

Editorial extensions

If this is right

  • For single-stage polynomial costs with $p\ge\deg(F_x)$, any relaxation order $k\ge d_1$ yields decisions that are $O(r)$-optimal for the empirical problem, even when the uncertainty set $\Xi$ is unbounded.
  • When $\Xi$ is compact or carries a quadratic-module ball constraint $R-\|\xi\|^2\in Q[h]$, increasing $k$ beyond $\deg(F_x)-p/2$ restores $O(r)$-consistency for high-degree costs.
  • For two-stage linear recourse with bounded recourse dual variables, the standard relaxation is $O(r)$-consistent; in the common case $\deg(G_x),\deg(g)\le2$, one can take $p=2$ and $k=1$, giving semidefinite relaxations whose representation is quadratic in the dimensions $n_0$ and $n_2$.
  • If recourse dual variables are unbounded, the strengthened relaxation with the perturbed constraints $g^{(i)}_\epsilon$ preserves $O(r)$-consistency under a nondegenerate optimal-basis condition.
  • Because the relaxed in-sample cost dominates the true Wasserstein DRO value, the out-of-sample performance guarantee of Wasserstein DRO transfers to the moment relaxation.

Reading between the lines

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

  • A testable extension the paper leaves implicit is whether the $O(r)$ bound stays uniform in $N$ when the samples are not confined to a bounded set: the proof needs a uniform coefficient-norm bound for the translated cost polynomials, so without a growth condition on the sample cloud the same argument does not give an $N$-independent constant.
  • The proof structure suggests the same degree-versus-radius tradeoff will appear in any moment-based surrogate for Wasserstein DRO: terms of $F_x$ above the $p$-th moment are only controlled by boundedness of the uncertainty set, so matching $p$ to the cost degree is the robust choice.
  • The strengthened two-stage relaxation of Section 4.2 has not been made numerically stable, so a natural follow-up is to test whether facial-reduction or regularization techniques make it practical while preserving the $O(r)$ guarantee.
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

3 major / 4 minor

Summary. This paper proposes moment relaxations for data-driven Wasserstein DRO problems with polynomial costs and even-order Wasserstein distances. For single-stage problems, relaxation (1.5) is claimed to be O(r)-consistent under either p >= deg(Fx) or a quadratic-module ball constraint on Xi; for two-stage linear recourse problems, relaxation (1.7) is claimed to be O(r)-consistent under a bounded dual variable condition and analogous degree conditions. The paper also gives examples illustrating necessity of some assumptions and a numerical study of a two-stage production problem.

Significance. If the main theorems were correct, the paper would make a useful contribution: it provides explicit SDP-based surrogates with parallelizable subgradient formulas, extends consistency results beyond the usual p=1 Wasserstein setting, and includes a reproducible numerical implementation. The proof strategy via moment duality is largely self-contained and the examples are informative. However, Theorem 1.1(i) as stated is false for unbounded Xi, and Lemma 3.5 is false as written; these issues affect the central consistency claims and must be repaired before the results can be relied upon.

major comments (3)
  1. [Theorem 3.7 / Theorem 1.1(i)] The proof of case (i) does not establish an N-independent O(r) bound when Xi is unbounded. The constant C1 := max{||F~_x^(i)||_infty : x in X, i in [N]} is a coefficient norm on the translated polynomials F(x, . + xi_hat^(i)); for unbounded Xi these coefficients grow with |xi_hat^(i)|, so C1 is not bounded independently of N. The claim is false as stated: take X={0}, Xi=R, N=1, p=2, k=1, F(x,xi)=xi^2, and xi_hat^(1)=M. Both the Wasserstein DRO (1.3) and the moment relaxation (1.5) have value M^2+2Mr+r^2 while the ESO value is M^2, so the gap is 2Mr+r^2 and any O(r) constant must be at least 2M. Choosing M=N shows no N-uniform constant exists. To restore the theorem one must either assume the samples lie in a fixed bounded set or adopt a data-dependent O(r) constant, which changes the advertised scope.
  2. [Lemma 3.5] The lemma is false as stated and its proof contains two algebraic mistakes. In the expansion (3.7) the coefficient of (Q^T xi)^{2 alpha} should be theta^alpha, not theta^{2 alpha}; and the subsequent lower bound sum_l <(Q^T xi)_l^p, y> >= sum_l <xi_l^p, y> is not valid for p>2 because the rows of Q need not preserve p-th powers. A counterexample to the statement is H=4I, n0=1, p=2, y the moment sequence of delta_1: then <||xi||^2,y>=4 but (theta_min)^p y_{2 e_1}=16. The correct inequality, obtained from ||xi||^p = (xi^T H xi)^{p/2} >= (theta_min)^{p/2} ||xi||_2^p >= (theta_min)^{p/2} sum_l xi_l^p, has exponent p/2 instead of p. Since Lemma 3.5 is used in Theorems 3.6, 3.7, and 4.4, the proofs need this correction.
  3. [Theorem 4.4 / Theorem 1.2(i)] The same N-uniformity problem occurs in the two-stage consistency proof. In the case deg(Gx) <= p, the bound uses max{||f_x^(i)||_infty : x in X, i in [N]} on the translated recourse data, and for unbounded Xi this coefficient norm grows with the sample points xi_hat^(i). The boundedness of the dual variable u through R0 - ||u||^2 in Q[g,h] does not control xi, so condition (i) of Theorem 4.4 does not yield an N-independent O(r) constant without an additional bounded-support or sample-moment assumption.
minor comments (4)
  1. [Section 4, after Lemma 4.2] The phrase "The proof of of the following lemma" contains a duplicated "of".
  2. [Example 4.5] The sentence "the inner supremum is is unbounded" contains a duplicated "is".
  3. [Theorem 4.4 proof] The symbol f_x^(i) used in the bound max{||f_x^(i)||_infty : x in X, i in [N]} is never defined; it should presumably be the translated recourse polynomial phi_x^(i).
  4. [Section 1.2] The definition of O(r) says the constant is independent of x or N and may depend polynomially on n0, but the proof constants also depend on X, F, h, and the Wasserstein order p; this should be stated precisely in the definition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the O(r)-consistency theorems are proven from stated assumptions via standard moment duality, with no fitted parameter or self-citation chain doing the work.

full rationale

The paper derives its main results (Theorems 1.1/3.7 and 1.2/4.4) from explicit assumptions on X, the polynomial data, the Wasserstein order p, and quadratic-module ball constraints. The proofs proceed through Lemmas 3.2–3.6 and 4.2–4.3, using weak/strong Lagrangian duality between the value functions vp, vp,k, Vp, Vp,k and their two-stage analogues. The constants appearing in the O(r) bounds (e.g., C0 from Lemma 3.5, C1 = max ||F~x^(i)||∞, R1 from the compact-quadratic-module estimate) are bounds derived from the hypotheses, not parameters fitted to the data whose value the theorems later 'predict'. No step in the derivation chain redefines the target O(r)-optimality gap as an input; the gap is bounded directly from the moment-relaxation feasibility and moment inequalities. The cited works [NZ25], [ZS22], [BS+24], [NY+23] are background references or pointers to related techniques and are not used to assume the main consistency results. There is no invoked uniqueness theorem by the authors, no ansatz smuggled in via citation, and no known empirical pattern merely renamed. The paper even flags its own open limitations, e.g., that the strengthened relaxations of Section 4.2 have not been implemented successfully and that no effective order bound is known for certain preorder-based relaxations; these are honest limitations, not circular reasoning. The only substantive concern in the manuscript is the N-uniformity of C1 for unbounded Xi in Theorem 3.7 case (i): the constant depends on the sample magnitudes through the translated polynomials, so the stated N-independence is not established. That is a correctness/assumption gap, not circularity, because the bound is not defined in terms of the target conclusion. Accordingly, the circularity score is 0.

Assumptions & free parameters 0 free parameters · 9 assumptions · 0 invented entities

The paper introduces no new physical entities or fitted constants. The central theorems rely on standard polynomial-optimization machinery plus two boundedness conditions (one for xi, one for u) that the necessity examples show cannot be dropped. The hidden uniformity assumption about sample-dependent coefficient bounds is not listed among the axioms but is used in the proofs.

assumptions (9)
  • domain assumption The Wasserstein distance order p is an even integer greater than 1.
    Assumed throughout so that ||xi||^p is an SOS polynomial of degree p (Section 1.1).
  • domain assumption F, h, B, b, c, d are polynomial maps; for the single-stage problem F(x, .) is a polynomial.
    Required for the moment/SDP representation (Sections 1.1, 4).
  • domain assumption The set Xi has nonempty interior.
    Used for closedness of quadratic modules and strong duality (Slater); stated in Section 2.1.
  • domain assumption X is bounded.
    Used to bound coefficients/F uniformly over x; stated in Theorems 1.1 and 1.2.
  • domain assumption For two-stage problems, strong LP duality between (P) and (D) holds, i.e., the recourse problem is feasible and finite for each xi.
    Used to express F(x,xi) as the max over u in (1.6).
  • ad hoc to paper For Theorem 1.2/4.4, R0 - ||u||_2^2 is in Q[g,h]_{2k0} for some R0,k0 (bounded recourse dual feasible set).
    This is the new boundedness condition; Example 4.5 shows it is necessary for O(r)-consistency.
  • domain assumption For Theorem 1.1 case (ii) and 1.2 case (ii), R - ||xi||_2^2 is in the quadratic module Q[h] (compact Xi).
    Used to bound all moments and apply the case (ii) proof.
  • standard math Slater's condition holds for r > 0 in the moment relaxation dual pairs.
    Ensures strong duality; satisfied because Dirac measures at the samples are strictly feasible for the primal.
  • ad hoc to paper For the strengthened relaxation (Section 4.2), Assumption 4.7: an optimal basis remains feasible for epsilon-perturbations of xi (nondegeneracy).
    Specific assumption for Theorem 4.8; not needed for the main theorems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Moment Relaxations for Data-Driven Wasserstein Distributionally Robust Optimization." pith.science (2026). https://pith.science/paper/RN73FNPM

@misc{pith2026250519278,
  author       = {Pith},
  title        = {Pith review of: Moment Relaxations for Data-Driven Wasserstein Distributionally Robust Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RN73FNPM}},
  note         = {Machine review of arXiv:2505.19278}
}
read the original abstract

We propose moment relaxations for data-driven Wasserstein distributionally robust optimization problems. Conditions are identified to ensure asymptotic consistency of such relaxations for both single-stage and two-stage problems, together with examples that illustrate their necessity. Numerical experiments are also included to illustrate the proposed relaxations.

Figures

Figures reproduced from arXiv: 2505.19278 by the authors.

Figure 1
Figure 1. Comparison of Mean Costs for the Two-Stage Production Problem [PITH_FULL_IMAGE:figures/full_fig_p021_1.png] view at source ↗
Figure 2
Figure 2. Out-of-sample Performance for the Two-Stage Production Problem [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 9 canonical work pages

  1. [1]

    DSOS and SDSOS optimization: more tractable alternatives to sum of squares and semidefinite optimization

    [AM19] Amir Ali Ahmadi and Anirudha Majumdar. “DSOS and SDSOS optimization: more tractable alternatives to sum of squares and semidefinite optimization”. In: SIAM Journal on Applied Algebra and Geometry 3.2 (2019), pp. 193–230. 22 [ApS25] MOSEK ApS. MOSEK Optimizer API for Julia 11.0.20

    work page 2019

  2. [6]

    Strong duality in conic linear programming: facial reduction and extended duals

    [Pat13] Gábor Pataki. “Strong duality in conic linear programming: facial reduction and extended duals”. In: Computational and Analytical Mathematics: In Honor of Jonathan Borwein’s 60th Birthday. Springer. 2013, pp. 613–634. [PP18] Frank Permenter and Pablo Parrilo. “Partial facial reduction: simplified, equivalent SDPs via approximations of the PSD cone...

    work page 2018

  3. [9]

    Distributionally robust optimization with polynomial densities: theory, models and algorithms

    [KKP20] Etienne de Klerk, Daniel Kuhn, and Krzysztof Postek. “Distributionally robust optimization with polynomial densities: theory, models and algorithms”. In: Mathematical Programming 181 (2020), pp. 265–296. [Las01] Jean B Lasserre. “Global optimization with polynomials and the problem of moments”. In: SIAM Journal on optimization 11.3 (2001), pp. 796...

    work page 2020

  4. [12]

    Distributionally Robust Optimization with Polynomial Robust Constraints

    [NZ25] Jiawang Nie and Suhan Zhong. “Distributionally Robust Optimization with Polynomial Robust Constraints”. In: Journal of Global Optimization (2025). DOI: 10.1007/s10898-025-01504-

    work page doi:10.1007/s10898-025-01504- 2025

  5. [2011]

    A nonlinear programming algorithm for solving semidefinite programs via low-rank factorization

    [BM03] Samuel Burer and Renato DC Monteiro. “A nonlinear programming algorithm for solving semidefinite programs via low-rank factorization”. In: Mathematical programming 95.2 (2003), pp. 329–357. [BMN21] Jose Blanchet, Karthyek Murthy, and Viet Anh Nguyen. “Statistical analysis of Wasserstein distributionally robust estimators”. In: Tutorials in Operatio...

    arXiv 2003

  6. [2017]

    Decomposition Algorithm for Distributionally Robust Optimization using Wasserstein Metric

    [LD+23] Miles Lubin, Oscar Dowson, Joaquim Dias Garcia, Joey Huchette, Benoît Legat, and Juan Pablo Vielma. “JuMP 1.0: Recent improvements to a modeling language for mathematical optimization”. In: Mathematical Programming Computation (2023). [LM17] Fengqiao Luo and Sanjay Mehrotra. “Decomposition algorithm for distributionally robust optimization using W...

    work page Pith review arXiv 2023

  7. [2019]

    Tractable reformulations of two-stage distributionally robust linear programs over the type-∞ Wasserstein ball

    URL: https://pretalx. com/juliacon2019/talk/QZBKAU/. [Xie20] Weijun Xie. “Tractable reformulations of two-stage distributionally robust linear programs over the type-∞ Wasserstein ball”. In: Operations Research Letters 48.4 (2020), pp. 513–523. [ZCN24] Suhan Zhong, Ying Cui, and Jiawang Nie. “Towards global solutions for nonconvex two- stage stochastic pr...

    work page arXiv 2020

  8. [2023]

    Polynomials non-negative on strips and half-strips

    [NP12] Ha Nguyen and Victoria Powers. “Polynomials non-negative on strips and half-strips”. In: Journal of Pure and Applied Algebra 216.10 (2012), pp. 2225–2232. [NY+23] Jiawang Nie, Liu Yang, Suhan Zhong, and Guangming Zhou. “Distributionally robust opti- mization with moment ambiguity sets”. In: Journal of Scientific Computing 94.1 (2023), p

    work page 2012

Show all 10 references
  1. [2024]

    Decomposition methods for Wasserstein-based data-driven distributionally robust prob- lems

    URL: https://www.gurobi. com. [GV+21] Carlos Andrés Gamboa, Davi Michel Valladao, Alexandre Street, and Tito Homem-de-Mello. “Decomposition methods for Wasserstein-based data-driven distributionally robust prob- lems”. In: Operations Research Letters 49.5 (2021), pp. 696–702. ...

    2021

  2. [2025]

    Sparse PSD approximation of the PSD cone

    URL: https://docs.mosek.com/ latest/juliaapi/index.html. [BD+22] Grigoriy Blekherman, Santanu S Dey, Marco Molinaro, and Shengding Sun. “Sparse PSD approximation of the PSD cone”. In: Mathematical Programming 191.2 (2022), pp. 981–1004. [BFK25] Geunyeong Byeon, Kaiwen Fang, an...

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.