Pith. sign in

REVIEW 3 major objections 5 minor 40 references

Unregularized limit of stochastic gradient method for Wasserstein distributionally robust optimization

T0 review · 3 major / 5 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 Entropic smoothing and sampling produce a faithful surrogate for unregularized WDRO: gradients and critical points converge to the nonsmooth objective's Clarke subdifferential and critical set as $\beta\to0$ and $m\to\infty$, with SGD…

desk verdict Promising paper with an important intended result, but the main gradient-convergence theorem is not proven as written: Assumption 1 does not imply the required nonpositive growth condition, so the central proof has a load-bearing gap. read the letter →

arxiv 2506.04948 v2 pith:I5NNE2TL submitted 2025-06-05 math.OC stat.ML

Tam Le (LPSM (UMR_8001) , UPCité) This is my paper
classification math.OCstat.ML MSC 90C1568Q3262F35
keywords distributionallyrobustoptimizationentropicregularizationWassersteindistancestochasticgradientmethodsClarkesubdifferentialgraphicalconvergencesmoothingweakconvexity
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 proves that the entropically smoothed, Monte-Carlo-sampled objective used in practice for Wasserstein distributionally robust optimization is a faithful surrogate for the nonsmooth original. As the smoothing parameter $\beta$ goes to 0 and the number of approximation samples $m$ goes to infinity, gradients of $F^{\beta,m}$ converge uniformly to the Clarke subdifferential of the unregularized objective $F$, and the critical points of $F^{\beta,m}$ converge to critical points of $F$, on the restricted dual domain $\Theta\times\Lambda$. This licenses projected stochastic gradient methods with convergence guarantees: fixed-small-regularization runs land near the true critical set almost surely, and vanishing-regularization schedules achieve rates $O(\log N/\sqrt{N})$ up to sampling error. Under stronger assumptions the approximation is quantitative, with uniform error $O(\beta|\log\beta|)$ for the objective and a corresponding Hausdorff bound for gradient graphs.

What carries the argument

The engine is entropic smoothing of a maximum: $H^\beta(w)=\beta\log\mathbb{E}_{z\sim\pi_0}[\exp(h(w,z)/\beta)]$, whose gradient is the expectation of $\nabla_\theta h$ under the Gibbs measure $d\pi_{\beta,w}\propto e^{h(w,z)/\beta}d\pi_0(z)$. As $\beta\to0$ the Gibbs measure concentrates on the argmax set of $h(w,\cdot)$, so the Clarke envelope theorem identifies the limit of $\nabla H^\beta$ with $\partial_c H$; Lemmas 3.2–3.5 make this concentration and uniform integrability precise under a coercivity/upper-growth condition. Proposition 3.1 then converts a uniform $O(\beta|\log\beta|)$ objective gap into graphical convergence of subdifferentials using weak-convexity results. On the optimization side, stationarity is measured through the Moreau envelope gradient $\nabla\varphi_{1/\tau}(w)=\tau(w-\hat w)$, and the proofs combine projected-SGD descent inequalities with a model-error term $\bar b$ that absorbs the sampling approximation. A structural companion is the restriction to $\Lambda=[\lambda_{\min},\lambda_{\max}]$, justified by the asymptotic slope $\Gamma(\theta)>0$ (Proposition 2.1) so that no dual minimizer is lost at $\lambda=0$.

What would settle it

Take the closed-form linear-regression example of Proposition 2.3 with fixed data and $\rho,\kappa>0$. For $\beta=10^{-1},10^{-2},\dots$ and $m$ large enough that the finite-sample error in Lemma 2.1 is negligible, compute $\mathrm{dist}(\nabla F^{\beta,m}(\theta^*,\lambda^*),\partial_c F(\theta^*,\lambda^*)+\varepsilon\bar B)$ at the known minimizer $(\theta^*,\lambda^*)$; Theorem 2.2 predicts the distance goes to 0, so saturation above 0 would refute the central claim. For the quantitative version, a loss with $\nabla_z^2\ell$ near $(2\lambda_{\min}-\delta)I_d$ lets one measure the exponent in $\sup|F^\beta-F|$; a rate materially worse than $\beta|\log\beta|$ would refute Theorem 2.3.

Watch

Extended reading notes

Core claim

The central discovery is Theorem 2.2: under Assumption 1, for every $\varepsilon>0$ there is $\bar\beta>0$ such that for any $\beta\in(0,\bar\beta]$ and all sufficiently large $m$, almost surely $\nabla F^{\beta,m}(\theta,\lambda)\in \partial_c F((\theta,\lambda)+\varepsilon\bar B)+\varepsilon\bar B$ for all $(\theta,\lambda)\in\Theta\times\Lambda$, and $\mathrm{crit}_{\Theta\times\Lambda}F^{\beta,m}$ lies in an $\varepsilon$-neighborhood of $\mathrm{crit}_{\Theta\times\Lambda}F$. The smoothed, sampled gradient oracle is therefore a legitimate substitute for the nonsmooth subgradient oracle of the true WDRO problem, and approximate critical points carry over. With additional regularity (quadratic transport cost, strong concavity in $z$, small-ball lower bounds on $\pi_0$), Theorem 2.3 quantifies the trade-off: $\sup|F^\beta-F|\le C\beta|\log\beta|$ and the graph of $\nabla F^\beta$ is within $\sqrt{2C\beta|\log\beta|/L}$ of the graph of $\partial_c F$ in the product Hausdorff distance. These approximation facts are turned into SGD guarantees in Proposition 2.5, with convergence in Moreau-envelope stationarity (and in function value under convexity) at rate $O(\log N/\sqrt{N})$ plus sampling error.

Load-bearing premise

Every gradient-convergence theorem is stated on the restricted domain $\Theta\times[\lambda_{\min},\lambda_{\max}]$ with $\lambda_{\min}>0$, and the paper's justification is an asymptotic slope condition $\Gamma(\theta)>0$; if $\Gamma(\theta)=0$ for some parameter, the true dual minimizer can sit at $\lambda=0$ and the smoothed-gradient approximation no longer describes the unrestricted WDRO problem the user is solving.

Editorial extensions

If this is right

  • Projected SGD run on the smoothed sampled objective with fixed small $\beta$ converges almost surely, up to any prescribed $\varepsilon$, to a critical point of the true unregularized WDRO problem under standard Robbins-Monro step sizes.
  • With $\beta_k=\beta_0/(k+1)$ and $\alpha_k=\alpha_0/\sqrt{k+1}$, the Moreau-envelope stationarity measure of the iterates converges at rate $O(\log N/\sqrt{N}+\bar b)$; if the loss is convex in $\theta$, expected function values converge at the same rate.
  • The entropic approximation error is $O(\beta|\log\beta|)$ uniformly for quadratic transport costs, giving an explicit trade-off between regularization bias and sampling variance.
  • For linear regression, the dual variable of every minimizer is bounded below by $\kappa^{-1}+\delta_\kappa^2$ and above by $(\hat\sigma_y+\sqrt{\rho/\kappa})^2/\rho$, so the restricted-domain assumption is checkable in closed form.
  • The graphical approximation of critical sets transfers from the smoothed objective to the true one without any convexity assumption on the loss, because the proof uses graph-closedness of the Clarke subdifferential and compactness rather than global convexity.

Reading between the lines

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

  • The $\lambda_{\min}>0$ restriction is doing real work: when $\Gamma(\theta)=0$ the true minimizer sits at $\lambda=0$, and the $\varepsilon$-inclusion describes a truncated problem, not the WDRO the user meant to solve. A practical implementation should detect and handle this boundary case.
  • The sampling-size condition (14) in Algorithm 2 is existential rather than constructive; the rates hold for sufficiently large $m_k$ but the paper gives no stopping rule, so turning the method into an implementable algorithm requires an adaptive sampling or confidence-bound procedure.
  • The Gibbs-concentration/envelope-theorem route is generic: it should give similar graphical convergence for any entropic smoothing of a coercive max-plus family, so the same proof pattern likely transfers to other nonsmooth min-max and bilevel objectives beyond WDRO.
  • The $\beta|\log\beta|$ function-gap rate and the uniform small-ball lower bound on $\pi_0$ suggest that the choice of reference distribution affects the hidden constant; distributions with thin tails or sparse support will need smaller $\beta$ for the same accuracy.
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 / 5 minor

Summary. The paper studies the entropic-smoothing regularization of the dual formulation of Wasserstein distributionally robust optimization (WDRO). For the joint objective F(θ,λ) over parameters and the dual variable, the authors define regularized objectives F^β and finite-sample versions F^{β,m}, and claim that as the regularization β tends to zero and the sampling size m tends to infinity, gradients of the smoothed objectives converge graphically to the Clarke subdifferential of the unregularized objective, with corresponding convergence of critical sets. Under stronger assumptions, they derive quantitative O(β|log β|) approximation rates and stochastic-gradient convergence rates of order O(log N/√N + bar b) for a vanishing-regularization scheme. The paper also proposes conditions under which the dual variable can be restricted to a compact interval [λ_min,λ_max] and gives an explicit lower bound for linear regression. The proofs rely on a general theory of entropic smoothing for pointwise maxima developed in Section 3.

Significance. If the main theorems were correct, the paper would provide a useful theoretical justification for practical entropic regularization in WDRO and would add to the literature on smoothing of nonsmooth max functions with explicit rates. The general graphical-convergence framework for Gibbs measures in Section 3 is potentially of independent interest, and the attempts to give verifiable conditions for the dual restriction are commendable. However, the paper's central application to WDRO contains serious gaps: the key verification that the hypotheses of the abstract theory hold under the paper's Assumption 1 is false, and Assumption 1 itself is too weak to guarantee that the WDRO objective is finite. These flaws undermine Theorems 2.1, 2.2, 2.3, and the stochastic-gradient results in Section 2.4. The significance is therefore conditional on a substantial revision.

major comments (3)
  1. [Section 5.1 (proof of Theorem 2.1), Assumption 3.3] The proof states 'Under Assumption 1, Assumption 3 and Assumption 4 hold for each h_i.' This is incorrect. Assumption 3.3 requires h(w,z) ≤ −κ(w)ψ(z) with ψ:Z→[0,∞) coercive, ∫ψ^{1+η}dπ0<∞, and κ≥1, so h must be nonpositive everywhere and tend to −∞ at infinity. For h_i(θ,λ,z)=ℓ(θ,z)+λ(ρ−c(ξ_i,z)), at z=ξ_i one has h_i(θ,λ,ξ_i)=ℓ(θ,ξ_i)+λρ. For any nonnegative loss (e.g., squared loss, or even ℓ≡0), this value is strictly positive, so no nonnegative ψ can satisfy the inequality. The lemmas leading to Theorem 3.1 (Lemmas 3.2–3.5) rely on Assumption 3.3 to obtain nonemptiness of near-maximal sets, uniform integrability of the Gibbs measures, and concentration on the argmax set. Consequently, Corollary 3.1 cannot be invoked, and Theorem 2.1, together with the dependent results Theorems 2.2, 2.3, and the SGD guarantees, is unsupported as written.
  2. [Section 2.1, Assumption 1] Assumption 1 does not ensure that the unregularized objective F in (10) is finite on Θ×Λ. The growth condition ℓ(θ,z) ≤ Ψ(z)+λ̄c(ξ_i,z) with λ̄<λ_min and Ψ coercive but otherwise arbitrary allows Ψ to dominate c. For instance, take Z=R, c(x,y)=|x−y|^2, ℓ(θ,z)=z^4, Ψ(z)=z^4+1, λ̄=0; then Assumption 1.3 holds, but h_i(θ,λ,z)=z^4+λ(ρ−|z−ξ_i|^2)→∞ for every λ>0, so F≡∞ and the Clarke subdifferential and the regularized objectives in Theorem 2.1 are not even well-defined. The restriction to Λ=[λ_min,λ_max] with λ_min>0 does not fix this because the term λc is of lower order than Ψ. A growth condition such as ℓ(θ,z) ≤ a(θ)+b c(ξ_i,z) with b<λ_min, or an equivalent bound guaranteeing a finite log-partition, is needed. This is a load-bearing gap in the statement of Theorem 2.1, independent of the proof issue in the previous comment.
  3. [Sections 2.1 and 2.4, definition of H_i^{β,m}] The sampling approximation H_i^{β,m} is defined inconsistently across the paper. In the introduction, equation (6) defines the finite-sample smoothed objective with the term λρ included, i.e., exp(β^{-1}(ℓ(θ,z_j)+λ(ρ−c(ξ_i,z_j)))). However, in the 'On sampling approximation' paragraph of Section 2.1 and again in Section 2.4, the definition is H_i^{β,m}(θ,λ)=β log(1/m Σ exp(β^{-1}(ℓ(θ,z_j)−λ c(ξ_i,z_j)))), omitting λρ. This changes the objective: the λ-derivative of H_i^{β,m} differs by ρ from the derivative of the true smoothed dual objective, so Algorithms 1 and 2, as written, do not minimize the WDRO objective (10) but a different function. Since the proofs in Section 5.2 set tilde H_i^k = H_i^{β_k,m_k} with this definition, the convergence results are for a different problem unless the definition is corrected. Please reconcile the definitions and, if the omission is a typo, state explicitly that λρ is included in all sampling approximations.
minor comments (5)
  1. [Throughout] There are several typos and minor stylistic issues: 'developped' in Section 3, 'quantative' in Section 3.3, 'reinjecting' in the proof of Proposition 4.2, and 'assum' in the same proof. A careful copyedit is needed.
  2. [Section 3.1, Definition 3.1] In the definition of Graph D, the text reads 'Graph D := {(v,w) ∈ R^q × R^q : w ∈ R^q, v∈ D(w)}'; the condition 'w ∈ R^q' is redundant and the ordering of (v,w) is inconsistent with the later notation. Please clarify the convention used for graph coordinates.
  3. [Section 4, Assumption 7] Assumption 7.1 refers to H_i^{β_k} without prior definition; only F^{β_k} is defined immediately below. Please define H_i^{β_k} explicitly.
  4. [Section 2.4, after eq. (14)] The paper acknowledges that the sample-size rule m_k is only theoretical and no practical selection rule is provided. This is an honest limitation, but it should be highlighted in the introduction or abstract so that readers are not misled about the applicability of Algorithm 2.
  5. [Proposition 2.3] In the proof of the lower bound for λ, the text derives λ⋆ > ∥θ⋆∥^2+κ^{-1} and then states λ⋆ ≥ κ^{-1}+δ_κ^2. Since δ_κ is only a lower bound on ∥θ⋆∥, the strict inequality implies the non-strict bound; this is fine but could be stated more clearly to avoid the appearance of a gap.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the gradient-convergence and SGD-rate results are derived from explicit assumptions using independent smoothing and weakly-convex machinery; the single self-citation to [25] is contextual, not load-bearing.

full rationale

The paper's central derivation chain is self-contained. Section 3 develops graphical convergence of entropic-regularized gradients (Theorem 3.1) from Assumptions 3–4 through the Gibbs-measure estimates in Lemmas 3.2–3.5 and the Clarke envelope theorem; Section 5.1 reduces Theorem 2.1 to this result via Corollary 3.1, and Theorem 2.2 adds only a standard uniform-law-of-large-numbers sampling estimate (Lemma 2.1). The quantitative result Theorem 2.3 is obtained by verifying Assumptions 5–6 under Assumption 2 and invoking the published external graph-convergence result [13, Theorem 5.1], not by assuming the target conclusion. Section 4's rates are conditional on Assumption 7, whose condition 3 is verified through Theorem 2.3, and the final bounds explicitly retain the sampling/approximation error term, so no fitted parameter is relabelled as a prediction. The only self-citation is [25] (Le & Malick, involving the present author), used as background on lower bounds for the dual variable; the restriction to Θ×Λ is independently justified in Propositions 2.1–2.3, so the self-citation is not load-bearing. The skeptical observation that Assumption 1 may not imply Assumption 3.3 for nonnegative losses (since h_i(θ,λ,ξ_i)=ℓ(θ,ξ_i)+λρ>0 for nonnegative ℓ, while Assumption 3.3 requires h_i≤−κψ≤0) is a mathematical correctness concern about the proof as written, not a circularity: the theorem is not equivalent by construction to its inputs, merely potentially under-supported by the stated assumptions. Therefore no circular step can be exhibited, and the circularity score is low.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

The paper introduces no fitted free parameters and no new physical or mathematical entities. The Gibbs measures πβ,w are standard tools from entropic regularization, not invented objects. The central theoretical results rest on domain assumptions about compactness, growth, smoothness, and reference-measure mass, all stated explicitly.

assumptions (7)
  • domain assumption The minimization problem is studied only on Θ×[λmin,λmax] with λmin>0; validity of this restriction is treated in Section 2.3.
    All gradient convergence results are stated on this compact dual interval. If the true dual minimizer is λ=0, the theorems do not apply.
  • domain assumption Assumption 1: supp(π0)=Z, continuity and differentiability of ℓ, and growth conditions ℓ(θ,z)≤Ψ(z)+λ̄c(ξi,z), ∥∇θℓ(θ,z)∥≤Ψ(z) with λ̄<λmin and integrable higher moments.
    This controls the far tails of the inner supremum and gives the qualitative gradient convergence Theorem 2.1.
  • domain assumption Assumption 2: compact Z, quadratic cost, L2<2λmin strong concavity in z, and the ball-mass condition π0(B(z,r))≥ār^d.
    These conditions produce the O(β|log β|) objective error and the quantitative subdifferential approximation in Theorem 2.3.
  • domain assumption Assumption 3: h(w,z)≤−κ(w)ψ(z) for a coercive ψ with finite higher-order π0 moments.
    This gives tightness, uniform integrability of Gibbs measures, and support concentration on maximizers in Lemmas 3.2 to 3.5.
  • standard math Clarke subdifferential calculus, the envelope theorem, and the weakly convex graphical convergence result of Davis and Drusvyatskiy [13, Theorem 5.1].
    These background tools convert function-approximation rates into subdifferential graphical convergence and are imported without proof.
  • standard math A uniform law of large numbers applies to the finite-sample approximation in Lemma 2.1.
    The paper states this informally and does not provide a full proof of the uniform LLN step.
  • domain assumption Generic qualification: F^{β,m}(crit_{Θ×Λ}F^{β,m}) has empty interior, justified by Morse-Sard type results in practice.
    Used in Proposition 2.4 to convert gradient convergence into asymptotic criticality of the projected SGD iterates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unregularized limit of stochastic gradient method for Wasserstein distributionally robust optimization." pith.science (2026). https://pith.science/paper/I5NNE2TL

@misc{pith2026250604948,
  author       = {Pith},
  title        = {Pith review of: Unregularized limit of stochastic gradient method for Wasserstein distributionally robust optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I5NNE2TL}},
  note         = {Machine review of arXiv:2506.04948}
}
read the original abstract

Wasserstein distributionally robust optimization offers a framework for model fitting in machine learning under potential shifts in the data distribution. We study a regularized variant of this problem in which entropic smoothing produces a sampled approximation of the original objective. We establish convergence of the approximate gradients to subgradients of the unregularized objective as the regularization parameter vanishes, enabling convergence guarantees for stochastic gradient methods. We obtain qualitative convergence results under general assumptions, then we provide convergence rates under additional regularity. In particular, we prove rates for the convergence of the unregularized objective values, up to sampling errors, when the regularization level is decreased across iterations. Our analysis yields byproducts of independent interest, including approximation results for smoothing of maximum functions subdifferentials and empirical lower bounds for dual solutions of Wasserstein distributionally robust optimization.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 36 canonical work pages

  1. [1]

    J. H. Alcantara and A. Takeda, Theoretical smoothing frameworks for general nonsmooth bilevel problems , arXiv preprint arXiv:2401.17852, (2024)

    work page Pith review arXiv 2024

  2. [2]

    Attouch, Convergence de fonctionnelles convexes , in Journ´ ees d’Analyse Non Lin´ eaire: Proceedings, Besan¸ con, France, June 1977, Springer, 2006, pp

    H. Attouch, Convergence de fonctionnelles convexes , in Journ´ ees d’Analyse Non Lin´ eaire: Proceedings, Besan¸ con, France, June 1977, Springer, 2006, pp. 1–40

    work page 1977

  3. [3]

    Azizian, F

    W. Azizian, F. Iutzeler, and J. Malick , Exact generalization guarantees for (regularized) wasserstein distributionally robust models, in Advances in Neural Infor- mation Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, eds., vol. 36, Curran Associates, Inc., 2023, pp. 14584–14596

    work page 2023

  4. [4]

    Azizian, F

    W. Azizian, F. Iutzeler, and J. Malick , Regularization for wasserstein dis- tributionally robust optimization , ESAIM: Control, Optimisation and Calculus of Variations, 29 (2023), p. 33

    work page 2023

  5. [5]

    Blanchet, J

    J. Blanchet, J. Li, S. Lin, and X. Zhang , Distributionally robust optimization and robust statistics , arXiv preprint arXiv:2401.14655, (2024)

    arXiv 2024

  6. [6]

    Blanchet and K

    J. Blanchet and K. Murthy , Quantifying distributional model risk via optimal transport, Mathematics of Operations Research, 44 (2019), pp. 565–600

    work page 2019

  7. [7]

    Bolte, A

    J. Bolte, A. Daniilidis, and A. Lewis , A nonsmooth morse–sard theorem for subanalytic functions, Journal of mathematical analysis and applications, 321 (2006), pp. 729–740

    work page 2006

  8. [8]

    Borkar , Stochastic Approximation: A Dynamical Systems Viewpoint;second Edition, Texts and Readings in Mathematics Series, Hindustan Book Agency, 2022

    V. Borkar , Stochastic Approximation: A Dynamical Systems Viewpoint;second Edition, Texts and Readings in Mathematics Series, Hindustan Book Agency, 2022. 24

    work page 2022

Show all 40 references
  1. [9]

    C ´edric, Optimal transport : old and new / C´ edric Villani , Grundlehren der mathematischen Wissenschaften, Springer, Berlin, 2009

    V. C ´edric, Optimal transport : old and new / C´ edric Villani , Grundlehren der mathematischen Wissenschaften, Springer, Berlin, 2009

    2009

  2. [10]

    Chizat , Annealed sinkhorn for optimal transport: convergence, regularization path and debiasing , arXiv preprint arXiv:2408.11620, (2024)

    L. Chizat , Annealed sinkhorn for optimal transport: convergence, regularization path and debiasing , arXiv preprint arXiv:2408.11620, (2024)

    2024 arXiv

  3. [11]

    Clarke, Optimization and Nonsmooth Analysis , Classics in Applied Mathemat- ics, Society for Industrial and Applied Mathematics, 1990

    F. Clarke, Optimization and Nonsmooth Analysis , Classics in Applied Mathemat- ics, Society for Industrial and Applied Mathematics, 1990

    1990

  4. [12]

    Davis and D

    D. Davis and D. Drusvyatskiy , Stochastic model-based minimization of weakly convex functions, SIAM Journal on Optimization, 29 (2019), pp. 207–239

    2019

  5. [13]

    , Graphical convergence of subgradients in nonconvex optimization and learning , Mathematics of Operations Research, 47 (2022), pp. 209–231

    2022

  6. [14]

    Dellacherie and P.-A

    C. Dellacherie and P.-A. Meyer, Probabilities and potential, c: potential theory for discrete and continuous semigroups , vol. 151, Elsevier, 2011

    2011

  7. [15]

    Ermoliev and V

    Y. Ermoliev and V. Norkin , Stochastic generalized gradient method for noncon- vex nonsmooth stochastic optimization, Cybernetics and Systems Analysis, 34 (1998), pp. 196–215

    1998

  8. [16]

    F ang and S.-Y

    S.-C. F ang and S.-Y. Wu , Solving min-max problems and linear semi-infinite programs, Computers & Mathematics with Applications, 32 (1996), pp. 87–93

    1996

  9. [17]

    Florian, W

    V. Florian, W. Azizian, F. Iutzeler, and J. Malick , skwdro: a library for wasserstein distributionally robust machine learning , Journal of Machine Learning Research, 27 (2026), pp. 1–7

    2026

  10. [18]

    Gao, Finite-sample guarantees for wasserstein distributionally robust optimiza- tion: Breaking the curse of dimensionality , Oper

    R. Gao, Finite-sample guarantees for wasserstein distributionally robust optimiza- tion: Breaking the curse of dimensionality , Oper. Res., 71 (2022), pp. 2291–2306

    2022

  11. [19]

    R. Gao, X. Chen, and A. J. Kleywegt , Wasserstein distributionally robust optimization and variation regularization , Operations Research, (2022)

    2022

  12. [20]

    Gao and A

    R. Gao and A. Kleywegt , Distributionally robust stochastic optimization with wasserstein distance, Math. Oper. Res., 48 (2023), pp. 603–655

    2023

  13. [21]

    Genans, A

    F. Genans, A. Godichon-Baggioni, F.-X. Vialard, and O. Winten- berger, Decreasing entropic regularization averaged gradient for semi-discrete opti- mal transport, arXiv preprint arXiv:2510.27340, (2025)

    2025

  14. [22]

    Kosowsky and A

    J. Kosowsky and A. Yuille , Solving the assignment problem with statistical physics, in IJCNN-91-Seattle International Joint Conference on Neural Networks, vol. 1, IEEE, 1991, pp. 159–164

    1991

  15. [23]

    D. Kuhn, P. M. Esfahani, V. A. Nguyen, and S. Shafieezadeh-Abadeh , Wasserstein distributionally robust optimization: Theory and applications in ma- chine learning, in Operations research & management science in the age of analytics, Informs, 2019, pp. 130–166. 25

    2019

  16. [24]

    Kushner and G

    H. Kushner and G. G. Yin , Stochastic approximation and recursive algorithms and applications, vol. 35, Springer Science & Business Media, 2003

    2003

  17. [25]

    Le and J

    T. Le and J. Malick , Universal generalization guarantees for wasserstein distri- butionally robust models , in The Thirteenth International Conference on Learning Representations, 2025

    2025

  18. [26]

    A. B. Levy, R. Poliquin, and L. Thibault , Partial extensions of attouch’s theorem with applications to proto-derivatives of subgradient mappings , Transactions of the American Mathematical Society, 347 (1995), pp. 1269–1294

    1995

  19. [27]

    Li and S.-C

    X.-S. Li and S.-C. F ang, On the entropic regularization method for solving min- max problems with applications , Mathematical methods of operations research, 46 (1997), pp. 119–130

    1997

  20. [28]

    G.-H. Lin, M. Xu, and J. J. Ye , On solving simple bilevel programs with a nonconvex lower level program , Mathematical Programming, 144 (2014), pp. 277– 305

    2014

  21. [29]

    J. Liu, T. W ang, H. Lam, H. Namkoong, and J. Blanchet , Dro: A python library for distributionally robust optimization in machine learning , arXiv preprint arXiv:2505.23565, (2025)

    2025 arXiv

  22. [30]

    W. Liu, M. Khan, G. Mancino-Ball, and Y. Xu , A stochastic smoothing framework for nonconvex-nonconcave min-sum-max problems with applications to wasserstein distributionally robust optimization , arXiv preprint arXiv:2502.17602, (2025)

    2025 arXiv

  23. [31]

    Mohajerin Esfahani and D

    P. Mohajerin Esfahani and D. Kuhn , Data-driven distributionally robust opti- mization using the wasserstein metric: performance guarantees and tractable refor- mulations, Mathematical Programming, 171 (2018), pp. 115–166

    2018

  24. [32]

    Moreau, Proximit´ e et dualit´ e dans un espace hilbertien, Bulletin de la Soci´ et´ e math´ ematique de France, 93 (1965), pp

    J.-J. Moreau, Proximit´ e et dualit´ e dans un espace hilbertien, Bulletin de la Soci´ et´ e math´ ematique de France, 93 (1965), pp. 273–299

    1965

  25. [33]

    A. P. Morse, The behavior of a function on its critical set , Annals of Mathematics, 40 (1939), pp. 62–70

    1939

  26. [34]

    Peyr ´e and M

    G. Peyr ´e and M. Cuturi , Computational optimal transport: With applications to data science , Found. Trends Mach. Learn., 11 (2019), pp. 355–607

    2019

  27. [35]

    Sard, The measure of the critical values of differentiable maps , (1942)

    A. Sard, The measure of the critical values of differentiable maps , (1942)

    1942

  28. [36]

    Schechtman, The gradient’s limit of a definable family of functions is a conser- vative set-valued field , arXiv preprint arXiv:2402.08272, (2024)

    S. Schechtman, The gradient’s limit of a definable family of functions is a conser- vative set-valued field , arXiv preprint arXiv:2402.08272, (2024)

    2024

  29. [37]

    Shafieezadeh-Abadeh, P

    S. Shafieezadeh-Abadeh, P. M. Esfahani, and D. Kuhn , Distributionally robust logistic regression , in Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 1, NIPS’15, Cambridge, MA, USA, 2015, MIT Press, pp. 1576–1584. 26

    2015

  30. [38]

    Shikhman, Topological Aspects of Nonsmooth Optimization, vol

    V. Shikhman, Topological Aspects of Nonsmooth Optimization, vol. 64 of Nonconvex Optimization and Its Applications, Springer, 01 2012

    2012

  31. [39]

    W ang, R

    J. W ang, R. Gao, and Y. Xie , Sinkhorn distributionally robust optimization , 2023

    2023

  32. [40]

    Zhang, J

    L. Zhang, J. Yang, and R. Gao, A short and general duality proof for wasserstein distributionally robust optimization , Operations Research, (2024). 27

    2024

Pith tools

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