acceptodds
Under review as a conference paper at ICLR 2027

A Stochastic Multi-Gradient Algorithm for Wasserstein Distributionally Robust Multi-objective Optimization

Abstract

Distributionally robust multi-objective optimization (DRMOO) has been proposed to improve the robustness of multi-task learning models under distribution shifts. However, existing DRMOO methods mainly construct ambiguity sets based on -divergence, which may have limited capability in capturing the geometric structure of the sample space and changes in the support of distributions. To address this limitation, we propose a Wasserstein distributionally robust multi-objective optimization algorithm, which leverages the geometry of the sample space to characterize distributional uncertainty more flexibly. Specifically, by applying the dual reformulation of Wasserstein distributionally robust optimization, each robust objective is reformulated as an expectation involving an inner adversarial maximization problem. Based on this, we develop a double-loop optimization algorithm, where the inner loop employs -step gradient ascent to approximately solve the inner maximization problem, and the outer loop adopts an MGDA-type method to construct a common descent direction among multiple objectives. Theoretically, under non-convex and generalized -smooth assumptions, we establish the convergence guarantees of the proposed algorithm, proving that it achieves an -Pareto stationary point with a total sample complexity of . Experimental results demonstrate the practical performance of the proposed algorithm.

Primary area:
optimization
Keywords:
Distributionally robust optimization, Multi-objective optimization, Stochastic Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ZmzGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the distributionally robust multi-objective optimization (DRMOO) problem. Instead of using the f-divergence framework, the authors claim that a more challenging and practical Wasserstein DRO is studied. One algorithm is proposed, and the corresponding theoretical analysis is provided. However, this paper only considers the penalized DRO formulation, with large penalty parameter gamma to make the function strongly concave.

Critical strengths

This paper considers distributionally robust multi-objective optimization (DR-MOO) and proposes the WDR-MOO algorithm. The authors establish high-probability convergence to an approximate Pareto stationary point, with a computational complexity of the same order as that of the corresponding non-DRO problem. The theoretical analysis further considers generalized smooth loss functions, relaxing the commonly used global smoothness assumption.

Critical weaknesses
  1. The motivation for the proposed formulation is not sufficiently clear. The paper studies multi-objective optimization in a distributionally robust setting, but it is unclear to me whether this particular combination is well motivated by practical applications. More importantly, why is each objective associated with its own distributional uncertainty, rather than considering a shared unknown environment across all objectives? The former formulation appears easier to handle theoretically, but the paper does not provide enough justification for when such a setting naturally arises in practice.

  2. The strong concavity of the inner problem is another concern. The paper considers penalized Wasserstein DRO, where the penalty parameter γ\gamma needs to be sufficiently large to ensure strong concavity of the inner problem. The authors should clarify how restrictive this condition on γ\gamma is in practice. If a very large γ\gamma is required, the resulting inner problem may become considerably easier, and the overall method can then be viewed as adding a well-behaved strongly concave inner optimization problem to MGDA.

  3. The lack of numerical validation makes it difficult to assess the practical value of the proposed formulation. The theoretical results alone do not demonstrate whether the proposed distributionally robust objectives lead to better Pareto trade-offs under distribution shift. In particular, the paper does not provide numerical experiments showing the advantage of WDR-MOO over standard MOO methods.

Questions for authors

Please answer all my concerns in the weakness.

Reviewer_erKTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies DRMOO with objective-specific distributional uncertainty measured by Wasserstein distance from the nominal distributions. Using a penalized Wasserstein formulation, the authors express each robust objective as an expectation of a sample-wise adversarial maximum and propose a double-loop algorithm combining inner gradient ascent with outer stochastic MGDA updates. The sample losses fi(x,z)f_i(x,z) are assumed to be generalized ℓ\ell-smooth in terms of model parameters xx and globally LwL_w-smooth in the data variable zz. With a quadratic transport penalty and γ>Lw\gamma>L_w, each inner objective is strongly concave in zz. Under these conditions, bounded fourth moments of the centered robust stochastic gradients, and a uniform inner initialization-gap bound, the paper claims high-probability convergence to ϵ\epsilon-Pareto stationarity with O(ϵ−4)O(\epsilon^{-4}) sample complexity.

Critical strengths

The strongest strength is extending the Wasserstein DRO formulation to multi-objective settings, which has not been frequently investigated before (as far as the reviewer knows). By leveraging a sample-wise adversarial reformulation, the paper develops a stochastic MGDA framework with approximately solved inner maximization problems and provides a convergence analysis under generalized smoothness. Connecting transport-based robustness with multi-objective gradient balancing is a meaningful research direction with potential theoretical and practical value.

Critical weaknesses

The main weaknesses concern empirical support, technical novelty, the interpretation of the complexity comparison, and the scope of the robustness guarantees.

  1. Insufficient empirical evidence. Neither the main paper nor the supplementary material provides experiments, although the conclusion claims validation through adversarial-attack experiments. Consequently, the submission does not establish whether Wasserstein DRMOO offers meaningful benefits over existing f-divergence DRMOO formulations, or whether its theoretical sample-complexity advantage translates into practical efficiency. The authors should report convergence against iterations, fresh sample draws, or iteration-wise wall-clock time (as Yang et al. [1] did). Robustness comparisons with f-divergence DRMOO across multiple datasets, under comparable tuning and computational budgets, are particularly important for supporting the motivation for changing the uncertainty model.

  2. The technical contribution beyond existing methods is insufficiently distinguished. Objective-wise distributional robustness and MGDA-based optimization have already been studied in Yang et al. [1], while the independent cross-Gram estimator, regularized simplex update, and generalized-smooth MGDA analysis build on earlier work such as Zhang et al. [2]. The sample-wise Wasserstein adversarial formulation and its strongly concave inner optimization are also established techniques [3]. Extending these ingredients to Wasserstein DR-MOO is fine, but the manuscript should identify the new analytical obstacles and explain how its treatment goes beyond combining existing components. The current presentation does not convincingly establish a substantial methodological advance.

  3. The complexity comparison involves different problem structures and oracle assumptions. The proposed method solves a deterministic adversarial subproblem conditional on each sampled data point, with uniform strong concavity assumption. In contrast, the double-loop f-divergence method in [1] solves an expected objective that is convex in its dual variable using stochastic gradients, without a corresponding uniform strong-convexity assumption. Thus, the difference in inner-loop complexity cannot be attributed solely to an improved multi-objective algorithm. Strong convexity alone would also not guarantee logarithmic sample complexity for a noisy stochastic inner solver. The authors should distinguish the benefits arising from Wasserstein’s sample-wise reformulation and imposed structure from those arising from their algorithm.

  4. The inner-loop efficiency claim needs a more explicit characterization of computational cost and conditioning. Lemmas A.5–A.7 provide finite-K approximation bounds, but K=Θ(log⁡(1/ϵ))K=\Theta(\log(1/\epsilon)) suppresses important dependence on the inner condition number and initialization bound. For a standard ascent step size, achieving squared inner-solution error at most r2r^2 requires a bound of the form $ K=O\left( \frac{\gamma+L_w}{\gamma-L_w} \log\frac{C_h}{r^2} \right). $ This dependence can become substantial as γ\gamma approaches LwL_w. Moreover, although sample reuse permits the stated fresh-sample complexity, the gradient-evaluation cost includes KK inner steps per objective and outer iteration. The authors should distinguish these resource measures, expose the dependence on conditioning and ChC_h, and discuss the restriction imposed by γ>Lw\gamma>L_w. Ablations over inner-step would help establish when the claimed efficiency advantage is practically meaningful.

  5. The robustness discussion should distinguish inner optimization accuracy, population stationarity, and generalization. Solving the adversarial problem accurately for a sampled data approximates hi∗(x,w)h_i^*(x,w); it does not imply that this individual sample loss is close to the population objective Hi(x)=EPi[hi∗(x,W)]H_i(x)=\mathbb E_{P_i}[h_i^*(x,W)]. The stochastic optimization argument instead relies on sampling from PiP_i and controlling robust-gradient noise under the stated moment assumptions. The authors should clarify whether PiP_i represents a population distribution or an empirical training distribution. If the implementation samples from a fixed dataset, the optimization guarantee targets the corresponding empirical robust objective and does not establish robustness on an unseen population (In that case, generalization error analysis between finite-sample and expected objective is required). Broader claims about out-of-sample robustness therefore require empirical support or a separate statistical analysis.

  6. The stochastic assumptions require stronger justification and qualification in the comparison with prior work. Assumption 3 imposes a uniform fourth-moment bound on the centered gradients of the exact robust objectives, which also implies uniformly bounded gradient variance. In related nonconvex DRO analyses, Jin et al. [4] derive generalized smoothness together with an affine gradient-variance bound, while Zhang et al. [5] establish a more precise partially generalized smoothness and partially affine-variance characterization. These analyses accommodate gradient-dependent noise, and Jin et al. provide an example where the DRO dual gradient variance is globally unbounded. The present uniform fourth-moment assumption therefore places a substantive restriction on the stochastic regimes covered by the analysis. Although these works studied different DRO reformulations and gradient oracles, the authors should explicitly compare their assumptions, provide sufficient conditions on the losses, nominal distributions, and transport cost that ensure Assumption 3, and discuss whether their analysis can accommodate gradient-dependent noise. This distinction should also be reflected in the claimed complexity advantages.

References:

[1] Yang et al., From Dual Tracking to Clipping: Provably Faster Distributionally Robust Multi-Objective Optimization.
https://arxiv.org/abs/2605.05660

[2] Zhang et al., MGDA Converges under Generalized Smoothness, Provably.
https://arxiv.org/abs/2405.19440

[3] Sinha et al., Certifying Some Distributional Robustness with Principled Adversarial Training.
https://arxiv.org/abs/1710.10571

[4] Jin et al., Non-convex Distributionally Robust Optimization: Non-asymptotic Analysis, NeurIPS 2021.
https://arxiv.org/abs/2110.12459

[5] Zhang et al., Revisiting Large-Scale Non-convex Distributionally Robust Optimization, ICLR 2025.
https://openreview.net/forum?id=JYwVijuNA7

Questions for authors

All concerns in the weakness section should be addressed before the manuscript can lead to acceptance.

Reviewer_KiFAOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes WDR-MOO for multi-objective optimization under Wasserstein distributional robustness. Wasserstein duality rewrites each robust objective as an expectation with an adversarial maximization. The algorithm approximates each inner maximizer by gradient ascent, uses independent samples to update regularized MGDA weights, and updates the shared model with a stochastic gradient. Theorem 4.1 assumes generalized smoothness in the model parameters, strong concavity of the inner problem, bounded fourth moments, and a bounded initialization gap. It claims an average squared Pareto-stationarity measure of order epsilon squared after order epsilon to the minus four outer iterations. Each iteration uses a constant stochastic batch and logarithmically many inner-gradient steps. The supplement develops a stopping-time analysis that controls objective gaps, stochastic-gradient errors, error products, and the local region required by generalized smoothness. Under fixed confidence, the claimed asymptotic sample complexity improves on existing cited DR-MOO methods. Total gradient-oracle complexity also includes the logarithmic inner-loop factor.

Critical strengths
  • Appropriate robustness model. Wasserstein ambiguity can represent geometric displacement and support changes that are difficult to express with an f-divergence ball.
  • The algorithm combines approximate adversarial maximization with independent-sample MGDA updates under generalized smoothness. Appendix B complements the algorithm by making the parameter dependencies explicit.
  • Table 1 separates sample complexity, batch size, and smoothness assumptions when comparing with existing stochastic MOO methods.
Critical weaknesses
  • Missing empirical evidence. The adversarial multi-task experiments promised in the conclusion are absent from the paper and supplement.
  • Possible proof gap. In Appendix B.3, Equation (B.49) applies Lemma B.1 to the random objective selected when the stopping event occurs. Lemma B.1 covers fixed weights, so this step needs a union or maximal argument.
  • The reported sample complexity suppresses the confidence dependence in Appendix B.1 and excludes the logarithmic inner-gradient work.
Questions for authors
  1. Equation (B.49) concerns the data-dependent objective index selected when the gap stopping event fires. Lemma B.1 is stated for fixed weights. What justifies this substitution, and would a repair introduce dependence on the number of objectives?
  2. Where are the adversarial multi-task experiments mentioned in the conclusion? Under matched compute, how does WDR-MOO compare with the f-divergence methods of Yang et al. (2026) on both geometric support shifts and ordinary reweighting shifts?
  3. Please state the total gradient-oracle complexity, including the inner-loop and confidence factors. Under which parameter regimes does this complexity improve on Double-Loop or Double-Clip MGDA?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.