Gradient Flow Sampler-based Distributionally Robust Optimization

Jia-Jie Zhu; Zusen Xu

REVIEW 3 major objections 5 minor 41 references

Gradient Flow Sampler-based Distributionally Robust Optimization

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The inner maximization in entropy-regularized Wasserstein DRO is solved by gradient-flow sampling.

desk verdict A useful sampler-perspective on Sinkhorn DRO with new algorithmic variants, but the headline guarantees rest on unverified LSI constants and one flawed finite-sample bound. read the letter →

arxiv 2510.25956 v3 pith:YD6PAK5T submitted 2025-10-29 math.OC math.APstat.ML

Zusen Xu , Jia-Jie Zhu This is my paper · ORCID

classification math.OCmath.APstat.ML MSC 90C1590C3049Q2265C05

keywords distributionallyrobustoptimizationgradientflowsWassersteindistanceentropyregularizationsamplingalgorithmsLangevindynamicsWasserstein-Fisher-RaoSteinvariationaldescent

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

This paper argues that the inner worst-case-distribution step of entropy-regularized Wasserstein distributionally robust optimization (DRO) can be treated as a sampling problem rather than an adversarial ascent. The worst-case conditional distribution is a Gibbs density, and any gradient flow whose energy is a KL divergence to that density—Wasserstein, Wasserstein–Fisher–Rao, Stein, or rejection-based—can be used inside an outer SGD loop to supply robust gradients. This yields practical algorithms for a broad class of DRO problems and recovers a classical worst-case risk minimization method as the zero-entropy limit. The paper backs the framework with an end-to-end complexity bound for the unadjusted-Langevin instantiation and with classification and adversarial-robustness experiments.

What carries the argument

The Gibbs conditional density ρ*_{Y|X=x}(y) ∝ exp(−(2τ/ε)(V(y)+ c(y,x)/(2τ))) — the entropic-JKO solution of the half-bridge formulation. The driving functional is F(ρ)=(ε/2τ)KL(ρ∥Gibbs), which combines loss, transport cost, and entropy; by choosing the dissipation geometry (Wasserstein, Wasserstein–Fisher–Rao, Stein/kernel, rejection) the same energy yields different samplers. Its role: it converts the inner maximization into a KL-minimization/sampling target, so convergence of the sampler is convergence of the inner problem.

What would settle it

Take a quadratic loss and compute the LSI constant λ_U of the Gibbs density exp(−(2τ/ε)((y−x)²/(4τ)−ℓ(θ,y))) as ε→0; if λ_U decays faster than ε, the claimed λ_U^{-3} inner-loop complexity is unattainable. Separately, estimate W1 between empirical finite-N ULA samples and the target: if it is controlled by N^{-1/d} rather than e^{-λ t}, Proposition 1's flow-time-only bound misses a finite-N error.

Watch

Extended reading notes

Core claim

Central claim: the inner maximization in entropy-regularized Wasserstein DRO is solved by sampling, not ascent. The worst-case marginal is a mixture of Gibbs conditionals ρ_{Y|X=x} ∝ exp(−(2τ/ε)(V+c/2τ)), and minimizing expected KL to them equals the inner problem. Any gradient flow with that KL as energy—Wasserstein, Wasserstein–Fisher–Rao, Stein/kernel, rejection—yields worst-case samples for an outer SGD loop. As ε→0, the Langevin update loses noise and reduces to the deterministic inner step of earlier worst-case risk minimization. For the Langevin variant, complexity is Õ(L_Φ L_U² L_f² d²/(λ_U³ ε_opt⁴)).

Load-bearing premise

The load-bearing premise is Assumption 4: for each θ and x the conditional Gibbs density is L_U-smooth and satisfies a log-Sobolev inequality with a known constant λ_U, which the paper neither verifies nor bounds and which may shrink as ε/τ changes.

Editorial extensions

If this is right

Any gradient-flow sampler with a target Gibbs density can be dropped into DRO; sampler mixing-time theory directly supplies DRO convergence rates.
The earlier WRM method becomes the ε=0 limit of the new Wasserstein-flow algorithm, and its inner gradient descent is the deterministic part of the ULA update.
For ε>0, the new flows produce continuous worst-case samples rather than single worst-case points, which enables richer adversarial training.
The ULA variant finds an ε_opt-stationary point in Õ(L_Φ L_U² L_f² d²/(λ_U³ ε_opt⁴)) total operations (Theorem 2).
The Wasserstein–Fisher–Rao flow dissipates the KL energy faster than the pure Wasserstein flow under warm starts, yielding practical speedups in the experiments.

Reading between the lines

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

If λ_U is hard to verify, the practical use of Theorem 2 relies on warm-starting; the WFR weight dynamics should help most when initial particles sit near the decision boundary—a testable prediction the experiments support without the theory quantifying it.
The finite-N empirical-measure error in Proposition 1 is not analyzed; adding it would produce a N^{-1/d} term, which matters in high dimensions and is a natural next step.
The equivalence with entropic proximal sampling means DRO inner problems inherit Gibbs-sampling tools such as annealing; sampling an annealed version of the worst-case Gibbs density may handle multimodal losses.
Since the framework reduces WRM to the ε→0 limit, any improvement in sampler convergence transfers to robust training; comparing Hamiltonian or underdamped Langevin variants against ULA would test whether the gradient-flow view yields algorithmic gains beyond the listed flows.

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.

Desk Editor's Note

read the letter

Here's my take. The paper's core idea — solving the inner DRO maximization by sampling from the conditional Gibbs density with whatever gradient-flow geometry you like — is genuinely useful. Lemma 1 is known, but the way it is packaged (Schrödinger half-bridge, then different samplers) gives a clean recipe for generating DRO algorithms. The reductions to WRM and to the proximal sampler are informative, and the four proposed variants (WGF/ULA, WFR, SVGD, RGO) are not mere exercises: they behave differently in the experiments, and the authors are honest about where each one fails (e.g., SVGD mode collapse, RGO's restrictions). They are also honest about not inventing new DRO formulations. They ship code, which is a real plus.

The main soft spots are two. First, Proposition 1's bound (32) in the appendix is wrong as stated. The left side is W1 between an N-particle empirical measure and the target marginal; the right side bounds the flow convergence of a single conditional sampler, so the N^{-1/d} Monte Carlo error is missing. The gradient-oracle bound does not follow. Because Proposition 1 is presented as an 'ideal' guideline and Theorem 2 uses the Vempala-Wibisono analysis instead, this flaw doesn't collapse the paper's central theoretical claim, but it should be fixed or relabeled as a heuristic.

Second, and more important, Theorem 2's complexity is conditional on an LSI constant λ_U of the conditional Gibbs density that is never derived or checked. For a nonconvex loss, that LSI can be arbitrarily small or may not hold, and λ_U appears cubed in the denominator. The experiments use T=100 inner steps, far fewer than the bound would require in any moderate regime, so the headline complexity is formally valid but has no demonstrated connection to the practical behavior of the algorithm. The authors do flag the WFR warm-start limitation in the appendix but nowhere give the same caveat for λ_U.

The empirical evaluation is a bit thin — no error bars, mostly small toy datasets — but as a proof of concept it does what it needs to: it shows the sampler-based variants work and where they struggle.

This paper deserves a serious referee. The framework is worth publishing, the algorithms are new, and the problems are addressable. A referee should push for a corrected Proposition 1, a more careful statement of Theorem 2's assumptions (ideally a bound on λ_U or an explicit note that the bound is not practical without isoperimetry), and a stronger experimental section.

Referee Report

3 major / 5 minor

Summary. The paper proposes a gradient-flow-sampler framework for the inner maximization in entropy-regularized Wasserstein DRO. It first shows (Lemma 1) that this inner problem is a Schrödinger half-bridge, with the worst-case conditional distribution given by a Gibbs density, then instantiates several samplers (Wasserstein gradient flow/ULA, Wasserstein-Fisher-Rao, Stein variational, rejection sampling) and embeds them in an outer SGD loop (Algorithms 2–6). The main theoretical claim is Theorem 2, an Õ(L_Φ L_U² L_f² d²/(λ_U³ ε_opt⁴)) total complexity for Algorithm 3 under Assumptions 1–4, and Proposition 1 gives an idealized continuous-time gradient-oracle error bound. Numerical experiments compare the proposed methods with WRM and the dual Sinkhorn DRO solver on synthetic and CIFAR-10 tasks.

Significance. The half-bridge/Gibbs conditional perspective (Lemma 1) is correct and provides a clean connection between DRO and modern sampling theory. The framework is a useful organizational device: it recovers WRM as a special case and offers a principled way to design samplers for the worst-case distribution. The accompanying code and experiments are a plus. However, the paper's central theoretical guarantee is not established as stated: Theorem 2 depends on an LSI constant λ_U that is neither derived from the DRO data nor verifiable, and Proposition 1 omits the finite-particle Monte Carlo error. The paper itself acknowledges important limitations (unverifiable warm-start for WFR, restrictive τ for RGO). With revisions addressing these gaps, the paper could be a valuable contribution.

major comments (3)

[§5, Theorem 2; Assumption 4] The complexity bound Õ(L_Φ L_U² L_f² d²/(λ_U³ ε_opt⁴)) is conditional on an LSI constant λ_U that is never bounded or certified from the DRO data. For the conditional Gibbs density (10), λ_U depends on the nonconvex loss landscape, temperature τ, and ε; when τ ≥ 1/sup_y λ_max(∇²_y ℓ), the strong-convexity sufficient condition fails, and LSI can degrade exponentially in the barrier height. Since λ_U appears cubed, even a moderate decrease (e.g., from 1 to 0.1) increases the predicted inner-loop count by 1000×. The authors should either provide explicit sufficient conditions on ℓ, τ, ε ensuring a computable λ_U, or clearly state the result as conditional and refrain from claiming an end-to-end complexity guarantee.
[§C.2, Eq. (32); Proposition 1] The bound W1(1/N Σ δ_{y_i}, π_Y) ≤ E[W1(bρ_Y|X, ρ*)] ≲ e^{-λt}/√λ omits the unavoidable N^{-1/d} Monte Carlo error of the empirical measure. For finite N, running the flow longer cannot reduce this term, so the claim that t ≳ (1/λ) log(1/(√λ ε)) yields an ε-gradient estimate is unjustified. The finite-sample error must be incorporated, and the required t and N must be stated as a joint trade-off.
[§A.2 and §B.2 (WFR and RGO)] The paper proposes WFR and RGO as main algorithmic contributions, but in A.2 it concedes that the warm-start condition for WFR 'cannot be verified in our DRO problems,' and in Remark 4 it concedes that RGO's complexity requires τ=1/(Ld), which severely limits robustness. These admissions are in tension with the abstract's claim of 'practical algorithms with theoretical guarantees.' The authors should either close these gaps or substantially soften the claims to match the proved statements, and the experimental evaluation should state whether the tested regimes satisfy the required conditions.

minor comments (5)

[§4, Proposition 1] The target accuracy ε is denoted by the same symbol as the entropy regularization parameter ε in Eq. (1). Rename the accuracy, e.g., ε_acc, to avoid ambiguity.
[§C.3, Theorem 1 proof] The proof refers to 'using Lemma 1' when bounding the bias; this should be Lemma 6.
[§4, Proposition 1] The statement includes L in the rate t ≳ (1/λ) log(L√λ/ε), but the derivation in Eq. (33) gives t ≳ (1/λ) log(1/(√λ ε)) with no L. Reconcile the statement and the proof.
[Algorithm 4, line 8] The weight update w_{t+1} = (w_t)^{1-ϵη_w/(2τ)} e^{-η_w eV} is not derived and appears dimensionally inconsistent (the exponent mixes stepsize with the ratio ϵ/τ). Provide the discretization derivation from the WFR PDE (14) or correct the formula.
[Abstract and Introduction] SVGD, Wasserstein–Fisher–Rao flow, and rejection sampling are existing methods. The phrase 'invented novel algorithms such as Stein variational gradient, Wasserstein-Fisher-Rao gradient, rejection sampler-based algorithms' overstates novelty; recommend phrasing such as 'novel instantiations of existing gradient-flow samplers for DRO.'

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the worst-case Gibbs target is derived from the DRO objective, and the sampler convergence theory is imported from external, conditional results.

full rationale

The claimed derivation chain is not circular. Lemma 1 derives the conditional worst-case density (10) from the Schrödinger half-bridge / Lagrangian form of the Sinkhorn DRO objective; this is a variational calculation, not an ansatz disguised as a prediction. Algorithm 2/3 then choose a gradient-flow energy F whose stationary measure is that Gibbs density, and the analysis uses standard external convergence results (Vempala-Wibisono for ULA, Salim et al. for SVGD, Lee et al./Chewi et al. for RGO). Theorem 2 is explicitly conditional on Assumption 4 (L_U-smoothness and λ_U-LSI of the target); the paper does not fit λ_U from data nor present the bound as a parameter-free prediction, so the unverified LSI constant is an assumption/correctness risk rather than a circular step. Proposition 1's bound (32) does omit the N^{-1/d} Monte Carlo error, but that is an incomplete proof step, not a definitional reduction. Self-citations (Mielke & Zhu 2025; Zhu et al. 2021) are used for background/geometry and for a negative result about global PL, and the paper explicitly notes the warm-start condition 'cannot be verified in our DRO problems' (A.2); RGO limitations are likewise disclosed (Remark 4). The WRM and proximal-sampler equivalences are honest special-case observations. No load-bearing step reduces to its own input.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper leverages a large body of standard gradient-flow and sampling results; no ad hoc parameters are introduced into the derivations. The main baggage lies in strong, unverified geometric assumptions (LSI, smoothness, warm-start) on the worst-case conditional distribution.

assumptions (5)

domain assumption Existence and uniqueness of the Schrödinger half-bridge solution with the Gibbs form (10) for the entropic OT problem (8).
Lemma 1 relies on standard entropic OT theory; the form of the optimal conditional is taken from literature (Wang et al. 2021; Chen et al. 2022). The paper proves the equivalence but treats the Gibbs solution as known.
standard math Gradient flow convergence results: KL decays e^{-2λt} under λ-displacement convexity / PL (Lemma 4) and Talagrand's inequality hold for the target Gibbs density.
Used in Proposition 1; these are standard results in sampling theory (Vempala-Wibisono, gradient flow literature).
domain assumption Assumption 4: For every θ and x, the conditional distribution ρ_{Y|X=x} is L_U-smooth and satisfies λ_U-LSI.
This assumption is the key to the ULA inner-loop complexity in Theorem 2; it is stated but not verified and its constant λ_U depends on τ, ε, and the loss in a way that is not analyzed.
domain assumption Assumption 3: The stochastic gradient estimator has bounded variance σ² with σ ≤ ε_opt.
Used in Theorem 1 to control outer loop noise; the relation of σ to the sample size m is not established, so the guarantee is conditional.
domain assumption The WFR/heat-flow speed-up requires a warm-start density-ratio lower bound that is not checkable in the DRO setting.
Appendix A.2 cites Lu et al. (2023) for the improved WFR rate, but explicitly notes the condition cannot be verified; thus the advertised speed-up is empirical only.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Gradient Flow Sampler-based Distributionally Robust Optimization." pith.science (2026). https://pith.science/paper/YD6PAK5T

@misc{pith2026251025956,
  author       = {Pith},
  title        = {Pith review of: Gradient Flow Sampler-based Distributionally Robust Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YD6PAK5T}},
  note         = {Machine review of arXiv:2510.25956}
}

read the original abstract

We propose a mathematically principled PDE gradient flow framework for distributionally robust optimization (DRO). Exploiting the recent advances in the intersection of Markov Chain Monte Carlo sampling and gradient flow theory, we show that our theoretical framework can be implemented as practical algorithms for sampling from worst-case distributions and, consequently, DRO. While numerous previous works have proposed various reformulation techniques and iterative algorithms, we contribute a sound gradient flow view of the distributional optimization that can be used to construct new algorithms. As an example of applications, we solve a class of Wasserstein and Sinkhorn DRO problems using the recently-discovered Wasserstein Fisher-Rao and Stein variational gradient flows. Notably, we also show some simple reductions of our framework recover exactly previously proposed popular DRO methods, and provide new insights into their theoretical limit and optimization dynamics. Numerical studies based on stochastic gradient descent provide empirical backing for our theoretical findings.

Figures

Figures reproduced from arXiv: 2510.25956 by the authors.

**Figure 1.** Robust Decision Boundaries with Biased Data. (a) Decision boundaries learned by different methods on the biased circle dataset. The training data are shown as orange (positive class) and blue (negative class) points. The final classification boundaries are shown for the WRM (green), WGF (purple), WFR (red), and Dual (brown) models. All models were trained for 40 epochs. We set the regularization parameter τ = 2.5 fo… view at source ↗

**Figure 2.** Decision boundary comparison for all methods on the two-moon classification task. For all DRO methods, [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Evolution of E[Vex,τ (z)]. We run all methods for 300 steps with a stepsize of 0.01. For RGO (blue), we run a rejection sampling procedure after solving the inner optimization problem. SVG 0.1 and SVG 0.2 denote initial distributions with a standard deviation of 0.1 and 0.2, respectively. Conversely, WRM and RGO fail to significantly increase the objective. The non-convex nature of the objective, evidenced by the in… view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Visualization of perturbed samples (gray-edged smaller circles) generated from original data (black-edged [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

**Figure 5.** Figure 5: Evolution of particle positions in the SVG method. The driving force guides initial convergence, while the [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

**Figure 6.** Figure 6: Experiment results of adversarial training on CIFAR-10 features under an [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Decision boundaries on the Biased Circle dataset with 2000 training samples. WFR still learns a more [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]

**Figure 8.** Figure 8: Test loss as a function of the perturbation level [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 11 linked inside Pith

[1]

Ambrosio, L., Gigli, N., and Savare, G. (2005). Gradient Flows: In Metric Spaces and in the Space of Probability Measures . Springer Science & Business Media

2005
[2]

A., Ghassemi, M., Sun, S., Suzuki, T., Wu, D., and Zhang, T

Ba, J., Erdogdu, M. A., Ghassemi, M., Sun, S., Suzuki, T., Wu, D., and Zhang, T. (2021). Understanding the variance collapse of svgd in high dimensions. In International Conference on Learning Representations

2021
[3]

Ben-Tal , A., den Hertog , D., De Waegenaere, A., Melenberg, B., and Rennen, G. (2013). Robust Solutions of Optimization Problems Affected by Uncertain Probabilities . Management Science , 59(2):341--357

2013
[4]

Blanchet, J. H. and Glynn, P. W. (2015). Unbiased monte carlo for optimization and functions of expectations via multi-level randomization. In 2015 Winter Simulation Conference (WSC) , pages 3656--3667. IEEE

2015
[5]

A., Chen, Y., Huang, D

Carrillo, J. A., Chen, Y., Huang, D. Z., Huang, J., and Wei, D. (2024). Fisher-rao gradient flow: geodesic convexity and functional inequalities. arXiv preprint arXiv:2407.15693

arXiv 2024
[6]

Chen, Y., Chewi, S., Salim, A., and Wibisono, A. (2022). Improved analysis for a proximal algorithm for sampling. In Conference on Learning Theory , pages 2984--3014. PMLR

2022
[7]

Z., Huang, J., Reich, S., and Stuart, A

Chen, Y., Huang, D. Z., Huang, J., Reich, S., and Stuart, A. M. (2023). Sampling via gradient flows in the space of probability measures. arXiv preprint arXiv:2310.03597

arXiv 2023
[8]

R., Lu, C., Le Gouic, T., and Rigollet, P

Chewi, S., Gerber, P. R., Lu, C., Le Gouic, T., and Rigollet, P. (2022). The query complexity of sampling from strongly log-concave distributions in one dimension. In Conference on Learning Theory , pages 2041--2059. PMLR

2022

Show all 41 references

[9]

Chewi, S., Niles-Weed, J., and Rigollet, P. (2024). Statistical optimal transport. arXiv preprint arXiv:2407.18163

2024 arXiv
[10]

E., Hoffman, F., Mazumdar, E., and Ratliff, L

Conger, L. E., Hoffman, F., Mazumdar, E., and Ratliff, L. J. (2023). Strategic Distribution Shift of Interacting Agents via Coupled Gradient Flows . In Thirty-Seventh Conference on Neural Information Processing Systems

2023
[11]

and Ye, Y

Delage, E. and Ye, Y. (2010). Distributionally robust optimization under moment uncertainty with application to data-driven problems. Operations research , 58(3):595--612

2010
[12]

and Lebret, H

El Ghaoui, L. and Lebret, H. (1997). Robust solutions to least-squares problems with uncertain data. SIAM Journal on matrix analysis and applications , 18(4):1035--1064

1997
[13]

and Kleywegt, A

Gao, R. and Kleywegt, A. J. (2016). Distributionally Robust Stochastic Optimization with Wasserstein Distance . arXiv preprint arXiv:1604.02199

2016 arXiv
[14]

Garc \' a Trillos, C. A. and Garc \' a Trillos, N. (2024). On adversarial robustness and the use of wasserstein ascent-descent dynamics to enforce it. Information and Inference: A Journal of the IMA , 13(3):iaae018

2024
[15]

and Sanz-Alonso, D

Garc \' a Trillos, N. and Sanz-Alonso, D. (2018). Continuum limits of posteriors in graph bayesian inverse problems. SIAM Journal on Mathematical Analysis , 50(4):4020--4040

2018
[16]

and Lan, G

Ghadimi, S. and Lan, G. (2013). Stochastic first-and zeroth-order methods for nonconvex stochastic programming. SIAM journal on optimization , 23(4):2341--2368

2013
[17]

and Hong, L

Hu, Z. and Hong, L. J. (2013). Kullback-leibler divergence constrained distributionally robust optimization. Available at Optimization Online , 1(2):9

2013
[18]

Krizhevsky, A., Hinton, G., et al. (2009). Learning multiple layers of features from tiny images.(2009)

2009
[19]

Kuhn, D., Shafiee, S., and Wiesemann, W. (2025). Distributionally robust optimization. Acta Numerica , 34:579--804

2025
[20]

T., Shen, R., and Tian, K

Lee, Y. T., Shen, R., and Tian, K. (2021). Structured logconcave sampling with a restricted gaussian oracle. In Conference on Learning Theory , pages 2993--3050. PMLR

2021
[21]

C., and Sidford, A

Levy, D., Carmon, Y., Duchi, J. C., and Sidford, A. (2020). Large-scale methods for distributionally robust optimization. Advances in neural information processing systems , 33:8847--8860

2020
[22]

and Wang, D

Liu, Q. and Wang, D. (2016). Stein variational gradient descent: A general purpose bayesian inference algorithm. Advances in neural information processing systems , 29

2016
[23]

Lu, Y., Lu, J., and Nolen, J. (2019). Accelerating langevin sampling with birth-death. arXiv preprint arXiv:1905.09863

2019 arXiv
[24]

Lu, Y., Slep c ev, D., and Wang, L. (2023). Birth-death dynamics for sampling: Global convergence, approximations and their asymptotics. Nonlinearity , 36(11):5731--5772

2023
[25]

Mielke, A. (2023). An introduction to the analysis of gradients systems. arXiv preprint arXiv:2306.05026

2023 arXiv
[26]

Mielke, A. (2025). Some notes on the hellinger distance and various fisher-rao distances. arXiv preprint arXiv:2510.02537

2025
[27]

and Zhu, J.-J

Mielke, A. and Zhu, J.-J. (2025). Hellinger-kantorovich gradient flows: Global exponential decay of entropy functionals. arXiv preprint arXiv:2501.17049

2025 arXiv
[28]

and Kuhn, D

Mohajerin Esfahani, P. and Kuhn, D. (2018). Data-driven distributionally robust optimization using the wasserstein metric: Performance guarantees and tractable reformulations. Mathematical Programming , 171(1):115--166

2018
[29]

Otto, F. (1996). Double degenerate diffusion equations as steepest descent . Sonderforschungsbereich 256

1996
[30]

Salim, A., Sun, L., and Richtarik, P. (2022). A convergence theory for svgd in the population limit under talagrand’s inequality t1. In International Conference on Machine Learning , pages 19139--19152. PMLR

2022
[31]

and Mackey, L

Shi, J. and Mackey, L. (2023). A finite-particle convergence rate for stein variational gradient descent. Advances in Neural Information Processing Systems , 36:26831--26844

2023
[32]

Sinha, A., Namkoong, H., Volpi, R., and Duchi, J. (2017). Certifying some distributional robustness with principled adversarial training. arXiv preprint arXiv:1710.10571

2017 arXiv
[33]

and Wibisono, A

Vempala, S. and Wibisono, A. (2019). Rapid convergence of the unadjusted langevin algorithm: Isoperimetry suffices. Advances in neural information processing systems , 32

2019
[34]

and Chizat, L

Wang, G. and Chizat, L. (2022). An exponentially converging particle method for the mixed nash equilibrium of continuous games. arXiv preprint arXiv:2211.01280

2022 arXiv
[35]

Wang, J., Gao, R., and Xie, Y. (2021). Sinkhorn distributionally robust optimization. arXiv preprint arXiv:2109.11926

2021 arXiv
[36]

Wibisono, A. (2025). Mixing time of the proximal sampler in relative fisher information via strong data processing inequality. arXiv preprint arXiv:2502.05623

2025 arXiv
[37]

Xu, C., Lee, J., Cheng, X., and Xie, Y. (2024). Flow-based distributionally robust optimization. IEEE Journal on Selected Areas in Information Theory , 5:62--77

2024
[38]

V., and Jordan, M

Yu, Y., Lin, T., Mazumdar, E. V., and Jordan, M. (2022). Fast Distributionally Robust Learning with Variance-Reduced Min-Max Optimization . In Proceedings of The 25th International Conference on Artificial Intelligence and Statistics , pages 1219--1250. PMLR

2022
[39]

and Guan, Y

Zhao, C. and Guan, Y. (2018). Data-driven risk-averse stochastic optimization with Wasserstein metric. Operations Research Letters , 46(2):262--267

2018
[40]

Zhu, J.-J., Jitkrittum, W., Diehl, M., and Sch \"o lkopf, B. (2021). Kernel distributionally robust optimization: Generalized duality theorem and stochastic approximation. In International Conference on Artificial Intelligence and Statistics , pages 280--288. PMLR

2021
[41]

and Xie, Y

Zhu, L. and Xie, Y. (2024). Distributionally robust optimization via iterative algorithms in continuous probability spaces. arXiv preprint arXiv:2412.20556

2024 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Ambrosio, L., Gigli, N., and Savare, G. (2005). Gradient Flows: In Metric Spaces and in the Space of Probability Measures . Springer Science & Business Media

2005

[2] [2]

A., Ghassemi, M., Sun, S., Suzuki, T., Wu, D., and Zhang, T

Ba, J., Erdogdu, M. A., Ghassemi, M., Sun, S., Suzuki, T., Wu, D., and Zhang, T. (2021). Understanding the variance collapse of svgd in high dimensions. In International Conference on Learning Representations

2021

[3] [3]

Ben-Tal , A., den Hertog , D., De Waegenaere, A., Melenberg, B., and Rennen, G. (2013). Robust Solutions of Optimization Problems Affected by Uncertain Probabilities . Management Science , 59(2):341--357

2013

[4] [4]

Blanchet, J. H. and Glynn, P. W. (2015). Unbiased monte carlo for optimization and functions of expectations via multi-level randomization. In 2015 Winter Simulation Conference (WSC) , pages 3656--3667. IEEE

2015

[5] [5]

A., Chen, Y., Huang, D

Carrillo, J. A., Chen, Y., Huang, D. Z., Huang, J., and Wei, D. (2024). Fisher-rao gradient flow: geodesic convexity and functional inequalities. arXiv preprint arXiv:2407.15693

arXiv 2024

[6] [6]

Chen, Y., Chewi, S., Salim, A., and Wibisono, A. (2022). Improved analysis for a proximal algorithm for sampling. In Conference on Learning Theory , pages 2984--3014. PMLR

2022

[7] [7]

Z., Huang, J., Reich, S., and Stuart, A

Chen, Y., Huang, D. Z., Huang, J., Reich, S., and Stuart, A. M. (2023). Sampling via gradient flows in the space of probability measures. arXiv preprint arXiv:2310.03597

arXiv 2023

[8] [8]

R., Lu, C., Le Gouic, T., and Rigollet, P

Chewi, S., Gerber, P. R., Lu, C., Le Gouic, T., and Rigollet, P. (2022). The query complexity of sampling from strongly log-concave distributions in one dimension. In Conference on Learning Theory , pages 2041--2059. PMLR

2022