Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Federated Calculation of the Free-Support Transportation Barycenter by Single-Loop Dual Decomposition

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single-loop dual subgradient algorithm computes free-support Wasserstein barycenters without solving any optimal transport subproblem.

desk verdict A genuinely new dual-decomposition technique for free-support Wasserstein barycenters, but the paper's central claim is ahead of its proof: the algorithm provably converges to the LP relaxation, and the binary support recovery is heuristic. read the letter →

arxiv 2507.19627 v1 pith:RRD6MJ6H submitted 2025-07-25 cs.LG math.OCstat.ML

Zhengqi Lin , Andrzej Ruszczyński This is my paper
classification cs.LGmath.OCstat.ML MSC 49Q2290C2590C1590C11
keywords Wassersteinbarycenterfree-supportdualdecompositionsubgradientmethodfederatedlearningprivacy-preservingoptimizationoptimaltransportmixed-integerprogramming
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 a federated algorithm for computing the Wasserstein barycenter of several probability distributions when the barycenter's support is free to be chosen rather than fixed in advance. Its central claim is that the free-support barycenter can be found by a single-loop dual subgradient method that never solves an optimal transport (mass transportation) subproblem: each iteration uses only closed-form updates with cost $O(NK)$, where $N$ is the number of local distributions and $K$ is the number of candidate support points. The method is designed for privacy-preserving settings: local devices send only highly aggregated scores, and the coordinator never sees raw particles, distances, or local dual variables. If the claim is right, it would make large-scale and federated barycenter computation practical in applications such as fair insurance pricing and preprocessing data for fairness.

What carries the argument

The machinery is a partial Lagrangian dual of the mixed-integer program (12): the mass-balance constraints (12d) and the cardinality constraint (12e) are dualized with multipliers $\theta_{si}$ and $\theta_0$, while the coupling constraints $\sum_{i\in I_s}\beta_{sik}=\gamma_k$ stay in the primal feasible set $\Gamma$. Because of this choice, the dual function (13) splits into $K$ closed-form subproblems indexed by candidate points $\zeta^k$; for each $k$ the optimal $\beta$ is a winner-take-all assignment to the particle minimizing $w_s d_{sik}-\theta_{si}$, and $\gamma_k$ is decided by the threshold rule (14), which compares the aggregated score $\sum_{s=1}^N T_{sk}$ with $\theta_0$. The quantities $T_{sk}$ in (15) are the only information each device must report, which is what gives the method its privacy and its $O(NK)$ per-iteration cost.

What would settle it

Take a small instance, for example $N=3$ devices, $K=50$ candidate points, and $M=10$; compute the true mixed-integer optimum of (12) with an exact solver, and run the federated dual subgradient algorithm to its stopping criterion. If the returned binary support has an objective value strictly above the mixed-integer optimum while the dual value equals the linear programming relaxation value, then the output is not a solution to (8) and the central claim is refuted.

Watch

Extended reading notes

Core claim

The authors reformulate the discrete free-support barycenter problem as a mixed-integer linear program (12), in which binary variables $\gamma_k$ select the support of a uniform barycenter measure from a large candidate set $Z$, and scaled transportation plans $\beta$ couple it to each local empirical distribution. They then partially dualize: Lagrange multipliers $\theta_{si}$ are attached to the local mass-balance constraints and $\theta_0$ to the cardinality constraint, while the assignment constraints $\sum_{i\in I_s}\beta_{sik}=\gamma_k$ are kept in the primal feasible set. This Lagrangian decomposes into $K$ independent single-candidate subproblems with closed-form solutions, and the rule (14) sets each $\gamma_k$ to 1 exactly when an aggregated threshold comparison holds. The resulting single-loop federated subgradient algorithm iterates the global dual variable $\theta_0$ at the coordinator and local dual variables at the devices, with no matrix-vector operations and no transportation problem solves; the authors state that to their knowledge it is the first efficient methodology for the Wasserstein barycenter without repeated subproblem solution and under privacy preservation.

Load-bearing premise

The load-bearing premise is that the binary support produced by the dual subgradient rule is an acceptable answer to the discrete barycenter problem, even though the dual method converges only to the value of the linear programming relaxation and the exact $M$ cardinality constraint is treated as soft; if that relaxation gap matters, the method solves a different problem than the one it claims to solve.

Editorial extensions

If this is right

  • The free-support Wasserstein barycenter can be approximated without repeated solutions of mass transportation problems, removing the main computational bottleneck of earlier variable-support methods.
  • The coordinator never sees local particles, distances, or local dual variables; only the aggregated scores $T_{sk}$ are transmitted, so the algorithm fits federated and privacy-restricted deployments.
  • Per-iteration cost is $O(NK)$ with no matrix-vector products, so the method scales linearly in the number of devices and candidate support points.
  • Convergence follows from standard subgradient arguments at rate $O(1/\sqrt{j})$, with primal recovery through ergodic averaging of the last iterates.
  • A stochastic variant samples a batch of candidate points each iteration, cutting per-iteration communication from $O(K)$ to $O(B)$.

Reading between the lines

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

  • The paper's headline claim is weaker than it appears: the dual method provably converges to the value of the linear programming relaxation of (12), and the final binary support from rule (14) is not guaranteed to have exactly $M$ points or to be optimal for the integer problem (8); the authors call the correction to exactly $M$ "not really necessary." A rigorous reader should treat the method as a
  • The same partial-dualization recipe could apply to other two-stage stochastic programs with binary first-stage variables and decomposable subproblems whose Lagrangian minimizers are closed-form; the Wasserstein structure is convenient but not essential.
  • Privacy here is computational, not cryptographic: the aggregated scores $T_{sk}$ can themselves carry information across iterations, especially when the mixture weights are skewed, as the authors themselves acknowledge; adding differential privacy or secure aggregation would be a natural testable extension.
  • An immediate empirical check is to measure the relaxation gap on small instances against an exact solver; if the gap is typically zero or tiny, the method's practical claims are stronger than its theory.
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

5 major / 6 minor

Summary. The paper proposes a federated dual decomposition algorithm for the discrete free-support Wasserstein barycenter problem. Starting from a mixed-integer linear programming reformulation (12), the authors derive the Lagrangian dual (16) with a closed-form subproblem minimization and the binary support-selection rule (14), and then present a single-loop subgradient algorithm (Algorithm 1) with per-iteration work O(NK) and no repeated optimal-transport subproblem solves. Numerical experiments on Gaussian mixture models compare the method against Bregman-projection baselines. The paper claims to be the first efficient privacy-preserving methodology for determining Wasserstein barycenters without repeated subproblem solution.

Significance. The algorithmic idea is attractive: dualizing the local mass-balance constraints yields closed-form updates and very cheap iterations, and the federated communication pattern is simple. The derivation of the dual function (13) and the threshold rule (14) are mathematically transparent, and the per-iteration complexity is a genuine strength. However, the central claim that the algorithm computes a free-support barycenter is not established. The dual method optimizes the LP relaxation of the mixed-integer problem, the output support need not be feasible for the discrete barycenter problem (8), and the momentum-accelerated updates and stopping criterion lack theoretical support. With a substantial reframing as a heuristic and additional validation against exact solutions on small problems, the core idea could be of interest, but the current guarantees are not adequate for the claims made.

major comments (5)
  1. [Section 3, Eqs. (12)-(16)] The central claim that Algorithm 1 computes a free-support barycenter is not supported because the dual method solves the LP relaxation of the mixed-integer problem, not the integer problem. The paper itself acknowledges that the optimal value of (16) may be strictly below that of (12). Rule (14) produces binary gamma, but it does not enforce sum_k gamma_k = M, and even when the count matches, the associated beta need not satisfy the mass-balance constraint (12d); in fact Section 6 states that the couplings may be infeasible. The proposed fix of zeroing the gamma_k with the smallest effect on the bracketed expression in (13) is a heuristic, and the claim that 'this correction is not really necessary' is unsubstantiated. Replacing M by the number of selected points changes the problem; it does not produce a solution of (8).
  2. [Section 4, Algorithm 1, lines 16-22 and 34-35] The convergence certificate is missing. The stopping test in line 17 compares consecutive values of the dual function, which is not a valid optimality certificate for a nonsmooth subgradient method; the dual function can stagnate while the dual variables are far from optimal. Furthermore, the momentum-modified updates in lines 21-22 and 34-35 are not covered by the cited O(1/sqrt(j)) subgradient convergence results ([23, 53]) or by the ergodic primal convergence result [32], which concern plain (stochastic) subgradient methods. No convergence or rate is proved for Algorithm 1 as written.
  3. [Section 4, approximate primal recovery] The averaging argument is applied to a nonconvex primal problem. Because the binary constraints (12b) make the feasible set nonconvex, a convex combination of the observed Lagrangian minimizers (gamma^(j), beta^(j)) is generally infeasible for (12), and randomized rounding of gamma-bar to a binary support has no approximation guarantee. Thus the statement that gamma-bar can be accepted as an approximate solution of the barycenter problem is not justified by Ref. [32].
  4. [Section 5, Table 1] The numerical experiments do not validate that the returned support is a good solution of the discrete problem (8). The 'Barycenter Value' of the dual method (4.44) is not better than the Bregman projection baseline with regularization 0.1 (4.40), and there is no comparison against an exact optimum of the discrete problem or a lower bound. Without such a reference, the claim of convergence to a high-quality solution is not established.
  5. [Section 5, privacy discussion] The privacy claim needs qualification. The authors themselves observe in Section 5 that 'the global device may capture the essence of the data distributions of local devices with minimal errors' when the weights are skewed (Figures 1-2). Since the final barycenter and the reported quantities T_sk may reveal information about local data, the blanket statement in the Conclusion that the method 'protects data privacy' is stronger than what is demonstrated.
minor comments (6)
  1. [Algorithm 1] In line 17, LD(theta^(j)) is a scalar, so the norm notation ||.|| is inappropriate; use absolute value.
  2. [Section 4] The parameter a for the cardinality tolerance is described in the text but never appears in Algorithm 1 or in the experimental description; the experiments use a +/-10% rule, so the pseudocode and the experiment should be reconciled.
  3. [Section 3, Eq. (13)] The notation beta_..k and Gamma_k is introduced without a precise definition; specify the domain of Gamma_k (the projection of Gamma onto the coordinates of candidate k) explicitly.
  4. [Algorithm 1] In lines 26-32, the pseudocode should explicitly set beta_sik = 0 for all k with gamma_k = 0, since the current conditional only assigns values when gamma_k = 1.
  5. [References] References [19] and [20] are duplicates of the same paper (Dvurechensky et al., NIPS 2018) with different spellings; one should be removed.
  6. [Abstract] The phrase 'the absence of any matrix-vector operations' is imprecise; the local step involves O(|I_s| * K) comparisons, which is a componentwise operation rather than a matrix product.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: derivation is self-contained and not equivalent to its inputs by construction.

full rationale

The paper's derivation chain is internally consistent and does not reduce any prediction to a fitted input. Starting from the mixed-integer formulation (12), it forms the Lagrangian dual (13), derives the closed-form support rule (14), and iterates subgradient updates in Algorithm 1. No parameter is fitted to a target quantity and then reported as a prediction. The evaluation in Table 1 uses an external Bregman-projection baseline and reports barycenter objective values, not the dual function being optimized. The only author self-citation, Ref. [36], is contextual and does not supply any load-bearing unverified result; the integrated transportation distance there is cited as background for the special case used here. The paper also candidly states a genuine limitation: the dual optimal value equals the LP relaxation value and the selected support count may differ from M, with the correction called 'not really necessary.' This is a feasibility and approximation-quality concern about primal recovery, not circularity. The stopping rule based on dual-function change and selected-count proximity is self-monitoring rather than self-definitional. Therefore no circular step meeting the evidentiary standard is present, and the score is 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on standard convex duality, sample-based approximation of the input measures, the adequacy of the preselected candidate support Z, and unproved convergence of the momentum-modified subgradient updates. No new physical or conceptual entities are introduced; the free parameters are algorithm hyperparameters and problem sizes, not fitted constants.

free parameters (7)
  • M (support cardinality) = 500 in Figure 1, 250 in Figure 3
    User-specified number of selected candidate points; defines the discrete approximation (8). Not fitted, but its choice strongly affects the approximation quality.
  • K (candidate set size) = 2000 in Figure 1, 1000 in Figure 3
    Number of potential support locations; the solution quality depends on whether Z is dense enough to contain near-optimal support points.
  • alpha^(0) (initial learning rate)
    Initial subgradient step size in Algorithm 1; value not reported. Chosen by hand; affects convergence speed and stability.
  • kappa1, kappa2 (momentum decay factors)
    Exponential decay factors between 0 and 1 in the momentum updates; values not reported. Their choice determines the effective averaging of subgradients.
  • epsilon (dual stopping tolerance) = 1e-4 in experiments
    Threshold for relative change in the dual function; controls when the algorithm stops, not a convergence guarantee for the primal solution.
  • a (cardinality tolerance)
    Parameter ensuring the iteration concludes when the number of selected points is close to M; used in experiments but not specified numerically.
  • B (batch size in stochastic variant)
    Randomly chosen batch size B << K for the stochastic adaptation; value not reported.
assumptions (5)
  • standard math The dual function LD is concave and the subgradient method with diminishing step sizes converges to its maximum.
    Invoked in Section 4 to claim O(1/sqrt(J)) convergence; standard for convex concave duals, but the actual algorithm uses momentum, which is not covered by [53].
  • domain assumption The empirical distributions Q(x_s) formed by local samples closely approximate the true conditional distributions.
    Section 2 relies on sample-based approximation of the conditional measures; supported by citations [18,22] but not verified in the experiments.
  • domain assumption The candidate set Z contains points sufficiently close to the optimal free-support barycenter support.
    Section 2 states Z is a preselected set, e.g., a sample from the mixture or a grid; if Z misses relevant regions, the approximation error is uncontrolled. No bound is provided.
  • ad hoc to paper The momentum-accelerated subgradient updates (lines 21-22, 34-35) converge for the nonsmooth dual function.
    The authors cite standard subgradient rates but the algorithm adds momentum, which is not a valid subgradient method and has no convergence proof in this nonsmooth setting. This is an unproved assumption.
  • domain assumption The uniform-weight restriction on the barycenter support is a reasonable approximation of the free-support barycenter.
    Problem (8) restricts to uniform masses on M points, while problem (3) allows arbitrary weights. The error of this restriction is not quantified anywhere in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Federated Calculation of the Free-Support Transportation Barycenter by Single-Loop Dual Decomposition." pith.science (2026). https://pith.science/paper/RRD6MJ6H

@misc{pith2026250719627,
  author       = {Pith},
  title        = {Pith review of: Federated Calculation of the Free-Support Transportation Barycenter by Single-Loop Dual Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RRD6MJ6H}},
  note         = {Machine review of arXiv:2507.19627}
}
read the original abstract

We propose an efficient federated dual decomposition algorithm for calculating the Wasserstein barycenter of several distributions, including choosing the support of the solution. The algorithm does not access local data and uses only highly aggregated information. It also does not require repeated solutions to mass transportation problems. Because of the absence of any matrix-vector operations, the algorithm exhibits a very low complexity of each iteration and significant scalability. We illustrate its virtues and compare it to the state-of-the-art methods on several examples of mixture models.

Figures

Figures reproduced from arXiv: 2507.19627 by the authors.

Figure 1
Figure 1. The GMMs with 5 components and the selected barycenter distribution. [PITH_FULL_IMAGE:figures/full_fig_p013_1.png] view at source ↗
Figure 2
Figure 2. The GMMs with 10 components and the selected barycenter distribution. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Barycenters obtained by the different methods: dual subgradient method (top-left), Bregman [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Closing the Gap: Efficient Algorithms for Discrete Wasserstein Barycenters

    math.OC 2025-11 accept novelty 7.0 of 10

    A (1+α)-approximation PTAS for discrete Wasserstein barycenters is achieved by averaging t randomly or deterministically sampled input support points, with runtime polynomial in (nk)^{1/α} and d.

Reference graph

Works this paper leans on

53 extracted references · 48 canonical work pages · cited by 1 Pith paper

  1. [32]

    Larsson, M

    T. Larsson, M. Patriksson, and A.-B. Strömberg. Ergodic, primal convergence in dual subgradient schemes for convex programming.Mathematical Programming, 86:283–312, 1999

    work page 1999

  2. [1]

    Agueh and G

    M. Agueh and G. Carlier. Barycenters in the Wasserstein space.SIAM Journal on Mathematical Analysis, 43(2):904–924, Jan. 2011

    work page 2011

  3. [2]

    Altschuler, J

    J. Altschuler, J. Niles-Weed, and P. Rigollet. Near-linear time approximation algorithms for optimal transport via Sinkhorn iteration. In I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vish- wanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30, Long Beach, 2017. Curran Associates, Inc

    work page 2017

  4. [3]

    P. C. Álvarez-Esteban, E. Del Barrio, J. Cuesta-Albertos, and C. Matrán. A fixed-point approach to barycenters in Wasserstein space.Journal of Mathematical Analysis and Applications, 441(2):744– 762, 2016

    work page 2016

  5. [4]

    Arjovsky, S

    M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein GAN, 2017. arXiv:1701.07875

    arXiv 2017

  6. [5]

    Backhoff-Veraguas and G

    J. Backhoff-Veraguas and G. Pammer. Applications of weak transport theory.Bernoulli, 28(1):370– 394, 2022

    work page 2022

  7. [6]

    Benamou, G

    J.-D. Benamou, G. Carlier, M. Cuturi, L. Nenna, and G. Peyré. Iterative Bregman projections for regularized transportation problems.SIAM Journal on Scientific Computing, 37(2):A1111–A1138, 2015

    2015

  8. [7]

    X. Bing, F. Bunea, and J. Niles-Weed. The sketched Wasserstein distance for mixture distributions,

    work page

Show all 53 references
  1. [8]

    J. R. Birge and F. V . Louveaux. A multicut algorithm for two-stage stochastic linear programs.Euro- pean Journal of Operational Research, 34(3):384–392, 1988

    1988

  2. [9]

    G. V . Carassai. Neural optimal transport at scale: Wasserstein barycenters for fair insurance. Master’s thesis, ETH Zürich, 2024

    2024

  3. [10]

    Charpentier, F

    A. Charpentier, F. Hu, and P. Ratz. Mitigating discrimination in insurance with wasserstein barycenters. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 161–177. Springer, 2023

    2023

  4. [11]

    Y . Chen, J. Ye, and J. Li. Aggregated Wasserstein distance and state registration for hidden Markov models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(9):2133–2147, 2020

    2020

  5. [12]

    Chzhen, C

    E. Chzhen, C. Denis, M. Hebiri, L. Oneto, and M. Pontil. Fair regression with wasserstein barycenters. Advances in Neural Information Processing Systems, 33:7321–7331, 2020

    2020

  6. [13]

    M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Weinberger, editors,Advances in Neural Information Processing Systems, volume 26, Address, 2013. Curran Associates, Inc

    2013

  7. [14]

    Cuturi and A

    M. Cuturi and A. Doucet. Fast computation of Wasserstein barycenters. InInternational conference on machine learning, pages 685–693. PMLR, 2014. 17

    2014

  8. [15]

    Cuturi and G

    M. Cuturi and G. Peyré. A smoothed dual approach for variational Wasserstein problems.SIAM Journal on Imaging Sciences, 9(1):320–343, 2016

    2016

  9. [16]

    Cuturi and G

    M. Cuturi and G. Peyré. Semidual regularized optimal transport.SIAM Review, 60(4):941–965, 2018

    2018

  10. [17]

    Delon and A

    J. Delon and A. Desolneux. A Wasserstein-type distance in the space of Gaussian mixture models. SIAM Journal on Imaging Sciences, 13(2):936–970, 2020

    2020

  11. [18]

    Dereich, M

    S. Dereich, M. Scheutzow, and R. Schottstedt. Constructive quantization: Approximation by empirical measures.Annales de l’IHP Probabilités et Statistiques, 49(4):1183–1203, 2013

    2013

  12. [19]

    Dvurechenskii, D

    P. Dvurechenskii, D. Dvinskikh, A. Gasnikov, C. Uribe, and A. Nedich. Decentralize and randomize: Faster algorithm for wasserstein barycenters.Advances in Neural Information Processing Systems, 31, 2018

    2018

  13. [20]

    Dvurechensky, D

    P. Dvurechensky, D. Dvinskikh, A. Gasnikov, C. Uribe, and A. Nedich. Decentralize and randomize: Faster algorithm for Wasserstein barycenters. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors,Advances in Neural Information Processin...

    2018

  14. [21]

    Dvurechensky, A

    P. Dvurechensky, A. Gasnikov, and A. Kroshnin. Computational optimal transport: Complexity by accelerated gradient descent is better than by Sinkhorn’s algorithm. In J. Dy and A. Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 o...

    2018

  15. [22]

    Fournier and A

    N. Fournier and A. Guillin. On the rate of convergence in Wasserstein distance of the empirical measure.Probability Theory and Related Fields, 162(3):707–738, 2015

    2015

  16. [23]

    Garrigos and R

    G. Garrigos and R. M. Gower. Handbook of convergence theorems for (stochastic) gradient methods,

  17. [24]

    Gaucher, N

    S. Gaucher, N. Schreuder, and E. Chzhen. Fair learning with Wasserstein barycenters for non- decomposable performance measures. InInternational Conference on Artificial Intelligence and Statis- tics, pages 2436–2459. PMLR, 2023

    2023

  18. [25]

    Genevay, M

    A. Genevay, M. Cuturi, G. Peyré, and F. Bach. Stochastic optimization for large-scale optimal trans- port. InProceedings of the 30th International Conference on Neural Information Processing Systems, NIPS’16, page 3440–3448, Red Hook, NY , USA, 2016. Curran Associates Inc

    2016

  19. [26]

    T. L. Gouic, J.-M. Loubes, and P. Rigollet. Projection to fairness in statistical learning.arXiv preprint arXiv:2005.11720, 2020

    2005 arXiv

  20. [27]

    Heinemann, A

    F. Heinemann, A. Munk, and Y . Zemel. Randomized wasserstein barycenter computation: Resampling with statistical guarantees.SIAM Journal on Mathematics of Data Science, 4(1):229–259, 2022

    2022

  21. [28]

    N. Ho, X. Nguyen, M. Yurochkin, H. H. Bui, V . Huynh, and D. Phung. Multilevel clustering via Wasserstein means, 2017. arXiv:1706.03883. 18

    2017 arXiv

  22. [29]

    Kolouri, G

    S. Kolouri, G. K. Rohde, and H. Hoffmann. Sliced Wasserstein distance for learning Gaussian mixture models. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018

    2018

  23. [30]

    Kone ˇcný, H

    J. Kone ˇcný, H. B. McMahan, D. Ramage, and P. Richtárik. Federated optimization: Distributed ma- chine learning for on-device intelligence. 2016

    2016

  24. [31]

    Kroshnin, N

    A. Kroshnin, N. Tupitsa, D. Dvinskikh, P. Dvurechensky, A. Gasnikov, and C. Uribe. On the com- plexity of approximating Wasserstein barycenters. In K. Chaudhuri and R. Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 ofPro...

    2019

  25. [33]

    T. Li, A. K. Sahu, A. Talwalkar, and V . Smith. Federated learning: Challenges, methods, and future directions.IEEE Signal Processing Magazine, 37(3):50–60, 2020

    2020

  26. [34]

    T. Lin, N. Ho, X. Chen, M. Cuturi, and M. I. Jordan. Computational hardness and fast algorithm for fixed-support Wasserstein barycenter.arXiv preprint arXiv:2002.04783, 2020

    2002 arXiv

  27. [35]

    T. Lin, N. Ho, and M. Jordan. On efficient optimal transport: An analysis of greedy and accelerated mirror descent algorithms.PMLR, 97:3982–3991, 09–15 Jun 2019

    2019

  28. [36]

    Lin and A

    Z. Lin and A. Ruszczy ´nski. An integrated transportation distance between kernels and approximate dynamic risk evaluation in Markov systems.SIAM Journal on Control and Optimization, 61(6):3559– 3583, 2023

    2023

  29. [37]

    Y . Liu, Y . Gao, and W. Yin. An improved analysis of stochastic gradient descent with momentum. Advances in Neural Information Processing Systems, 33:18261–18271, 2020

    2020

  30. [38]

    H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Arcas. Communication-efficient learning of deep networks from decentralized data.arXiv, 2016

    2016

  31. [39]

    J. M. Mulvey and A. Ruszczy ´nski. A new scenario decomposition method for large-scale stochastic optimization.Operations research, 43(3):477–490, 1995

    1995

  32. [40]

    Pele and M

    O. Pele and M. Werman. Fast and robust earth mover’s distances. In2009 IEEE 12th International Conference on Computer Vision, pages 460–467, 2009

    2009

  33. [41]

    R. T. Rockafellar and R. J.-B. Wets. Scenarios and policy aggregation in optimization under uncer- tainty.Mathematics of operations research, 16(1):119–147, 1991

    1991

  34. [42]

    Rubner, C

    Y . Rubner, C. Tomasi, and L. J. Guibas. The earth mover’s distance as a metric for image retrieval. International Journal of Computer Vision, 40:99–121, 2000

    2000

  35. [43]

    Ruszczy ´nski

    A. Ruszczy ´nski. A regularized decomposition method for minimizing a sum of polyhedral functions. Mathematical Programming, 35(3):309–333, 1986. 19

    1986

  36. [44]

    Ruszczy ´nski and A

    A. Ruszczy ´nski and A. Shapiro.Stochastic Programming, Handbooks in OR and MS. Elsevier, Ams- terdam, 2003

    2003

  37. [45]

    Shapiro, D

    A. Shapiro, D. Dentcheva, and A. Ruszczy ´nski.Lectures on Stochastic Programming: Modeling and Theory. SIAM, Philadelphia, 2021

    2021

  38. [46]

    Smith, C.-K

    V . Smith, C.-K. Chiang, M. Sanjabi, and A. Talwalkar. Federated multi-task learning. 2017

    2017

  39. [47]

    Solomon, R

    J. Solomon, R. Rustamov, L. Guibas, and A. Butscher. Wasserstein propagation for semi-supervised learning. In E. P. Xing and T. Jebara, editors,Proceedings of the 31st International Conference on Machine Learning, volume 32 (1) ofProceedings of Machine Learning Research, pages...

    2014

  40. [48]

    R. M. Van Slyke and R. Wets. L-shaped linear programs with applications to optimal control and stochastic programming.SIAM journal on applied mathematics, 17(4):638–663, 1969

    1969

  41. [49]

    Villani.Optimal Transport: Old and New

    C. Villani.Optimal Transport: Old and New. Springer, Berlin, 2009

    2009

  42. [50]

    J. Wang, Z. Charles, Z. Xu, G. Joshi, H. B. McMahan, B. A. y. Arcas, M. Al-Shedivat, G. Andrew, S. Avestimehr, K. Daly, D. Data, S. Diggavi, H. Eichner, A. Gadhikar, Z. Garrett, A. M. Girgis, F. Hanzely, A. Hard, C. He, S. Horvath, Z. Huo, A. Ingerman, M. Jaggi, T. Javidi, P. ...

    2021

  43. [51]

    Y . Xie, X. Wang, R. Wang, and H. Zha. A fast proximal point method for computing exact Wasserstein distance. InUncertainty in artificial intelligence, pages 433–453. PMLR, 2020

    2020

  44. [52]

    Y . Yan, T. Yang, Z. Li, Q. Lin, and Y . Yang. A unified analysis of stochastic momentum methods for deep learning. InProceedings of the 27th International Joint Conference on Artificial Intelligence, pages 2955–2961, 2018

    2018

  45. [53]

    Zinkevich

    M. Zinkevich. Online convex programming and generalized infinitesimal gradient ascent. InProceed- ings of the 20th International Conference on Machine Learning (ICML-03), pages 928–936, 2003. 20

    2003

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.