Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

This paper recasts Wasserstein barycenter computation as a gradient flow over probability measures, yielding mini-batch algorithms, plug-in regularizers, and label-aware costs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A mini-batch Wasserstein gradient-flow algorithm computes scalable and label-aware Wasserstein barycenters, with empirical gains on domain adaptation.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Useful mini-batch barycenter solver with a nice label-aware formalization, but the convergence theorem is not proved as stated and should be fixed or reclassified. the 3 major comments →

arxiv 2510.04602 v4 pith:FJH6K3CO submitted 2025-10-06 stat.ML cs.AIcs.LG

Wasserstein Gradient Flows for Scalable and Regularized Barycenter Computation

Eduardo Fernandes Montesuma , Yassir Bendou , Mike Gartrell This is my paper
classification stat.ML cs.AIcs.LG MSC 49Q2265K10
keywords Wasserstein barycentergradient flowsoptimal transportmini-batchdomain adaptationGaussian mixture modelsPolyak–Lojasiewicz inequalitylabel-aware transport
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Wasserstein barycenters average probability measures while respecting the geometry of the space they live in, but existing solvers are either stuck with full-sample discrete methods or complicated neural-network training that cannot use labels. This paper claims that viewing the barycenter as the steady state of a gradient flow in the space of measures fixes both problems at once. The flow minimizes a functional made of the barycenter objective plus internal, potential, and interaction energies; time-discretizing that flow gives a simple particle update that only needs mini-batches of each input measure. The same view extends to Gaussian mixture measures, and by embedding labels into the ground cost the flow produces labeled barycenters that dominate unlabeled ones across five domain-adaptation benchmarks. If correct, this makes regularized, label-aware barycenter computation as cheap as a few hundred mini-batch OT solves.

Core claim

The central claim is that the Wasserstein barycenter problem can be solved by following the Wasserstein gradient flow of the functional F(P)=B_Q(P)+G(P)+V(P)+U(P), where B_Q is the barycenter objective and G, V, U are internal, potential, and interaction energies. The paper shows that a forward-Euler discretization of this flow reproduces the classical fixed-support solver of Cuturi and Doucet as a special case, generalizes it to arbitrary differentiable ground costs, and yields two new algorithms: one for empirical measures with mini-batch sampling, one for Gaussian mixtures. It proves convergence of both algorithms under a measure-theoretic Polyak–Lojasiewicz inequality, with an exponentia

What carries the argument

The key object is the Wasserstein gradient of the composite functional F(P)=B_Q(P)+G(P)+V(P)+U(P). For the barycenter term, the gradient is the sum of displacement vectors 2Σ_k λ_k (Id − T^*_{P→Q_k}) built from optimal transport maps to each input measure; the update rule z_{τ+1,i} = z_{τ,i} − α ∇F(P_τ) is a forward-Euler discretization of the continuity equation. Labeled measures enter through the ground cost on the joint feature-label space with one-hot labels relaxed through a softmax parametrization, and the labeled-GMM distance is justified by Proposition 3.1. The convergence argument is carried by a measure-theoretic Polyak–Lojasiewicz inequality ∥∇B(P)∥_{L2(P)} ≥ C_PL(B(P)−B*), which

Load-bearing premise

The convergence theorems assume the barycenter objective satisfies a strong Polyak–Lojasiewicz inequality — the squared norm of its gradient must be at least a constant times its optimality gap — and this is neither proved for the barycenter functional nor checked on any dataset; the proof also applies a continuous-time exponential-decay bound to a discrete mini-batch algorithm without a step-size condition.

What would settle it

Take K=2 Gaussian measures with distinct means and covariances, where the true barycenter is known in closed form; run the empirical flow at a fixed learning rate and measure (B(P_τ)−B*)/∥∇B(P_τ)∥²_{L2(P)}. If the ratio is not bounded below by a positive constant as P_τ approaches the barycenter, or if the sequence error does not decay exponentially for any step size, the PL-based bound in Theorem 3.1 does not hold in the regime the algorithm operates in.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Barycenter computation becomes a mini-batch algorithm: each iteration only needs a batch of samples from each input measure, so the per-iteration cost drops from O(K n^3 log n) to O(K n_b^3 log n_b) in the empirical case.
  • Regularization is modular: entropy, repulsion, or target-alignment terms can be added as V and U without changing the core solver, giving a principled way to impose structure on the barycenter.
  • Label information can be integrated into the ground cost via a metric on one-hot or soft labels, and the labeled barycenter consistently outperforms the unlabeled one in domain adaptation.
  • The framework unifies previous algorithms (Cuturi–Doucet fixed-support barycenters, label-aware dictionary learning, GMM barycenters) as special cases of one gradient flow.
  • If the PL assumption holds, convergence is exponential down to the empirical-approximation error floor, with rate independent of the number of input measures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the solver's representation of the barycenter is a fixed-size particle system or GMM regardless of K, the same algorithm should scale to hundreds of input measures; the ISRUC experiment with 90 source domains hints at this, but the paper does not test beyond 90.
  • The softmax label parametrization suggests a direct test: measure classification accuracy of the barycenter's soft labels against the argmax labels, which would isolate how much of the gain comes from label geometry rather than from the flow itself.
  • The PL inequality is the bridge between a continuous-time proof and a discrete algorithm; a natural check is whether the exponential term in Theorem 3.1 actually matches observed convergence for a single pair of Gaussians, where B_Q is known in closed form.
  • The same gradient-flow construction could be applied to other OT-type objectives (e.g., unbalanced or weak OT barycenters) by swapping the displacement term, since the flow only needs differentiability of the ground cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes a Wasserstein gradient-flow framework for computing Wasserstein barycenters. The barycenter objective B_Q is augmented with internal, potential, and interaction energies, and two algorithms are given: Algorithm 1 for empirical measures using mini-batch optimal transport, and Algorithm 2 for Gaussian mixture measures with Cholesky parametrized covariances. A label-aware ground cost is introduced, and Proposition 3.1 justifies the resulting labeled-GMM MW2 formula. The paper claims convergence guarantees in Theorems 3.1 and 3.2 under a measure-theoretic Polyak–Łojasiewicz inequality. Experiments on a Swiss-roll toy example and five multi-source domain adaptation benchmarks (Office31, Office-Home, TEP, BCI-CIV-2a, ISRUC) report state-of-the-art accuracy, with code provided.

Significance. If the theoretical results were correct, the paper would provide a useful unifying view of discrete and GMM barycenter solvers, with scalable mini-batching and modular regularizers, and a clean justification of label-aware GMM barycenters. The experimental effort is broad and reproducible: five benchmarks, multiple neural and discrete baselines, ablations, running-time tables, and public code. Proposition 3.1 is a sound and useful contribution. However, the convergence guarantees — a central advertised contribution — are not established as written: the proof of Theorem 3.1 misuses the stated PL inequality, and the continuous-time bound is transferred to the discrete stochastic algorithm without a valid descent lemma. The PL assumption itself is neither proved for B_Q nor checked experimentally. These issues are load-bearing for the paper's central claim.

major comments (3)
  1. [§3.4 / App. D.3, Eq. (47)–(48)] The stated PL inequality is first-order: ||∇B(P)||_{L2(P)} ≥ C_PL (B(P)−B*). From the continuity equation, d/dt (B(P_t)−B*) = −||∇B(P_t)||^2, so the stated assumption implies dφ/dt ≤ −C_PL^2 φ^2, i.e. polynomial decay, not exponential decay. The proof instead writes ∂_t φ ≤ −C_PL φ (or similar), which requires the stronger condition ||∇B||^2 ≥ C_PL (B−B*), an unstated squared-gradient PL inequality. Therefore the exponential term e^{−C_PL τ} in Theorem 3.1 is not a consequence of the stated assumption.
  2. [App. D.3, Eq. (53)] The proof treats B_Q(P_τ)−B*_Q as deterministic because it 'does not involve sampling from the measures in Q.' But P_τ is itself generated by Algorithm 1 using random mini-batches drawn from Q. The expectation on the left-hand side is over exactly this sampling randomness, so E[|B_Q(P_τ)−B*_Q|] is not equal to |B_Q(P_τ)−B*_Q|. No discrete-time lemma bounds E[φ(τ+1)] from E[φ(τ)] for the actual update, no step-size restriction is stated, and the mini-batch noise is only added as a final independent constant. Thus Theorem 3.1 does not give a convergence guarantee for Algorithm 1.
  3. [Theorem 3.2 / App. D.4] Theorem 3.2 inherits the PL proof defect of Theorem 3.1. In addition, the theorem is stated for GMMs but no argument is given that B_Q(P)=Σ λ_k MW2(P,Q_k)^2 satisfies the PL inequality under the bounded-parameter and empirical-convergence hypotheses. Since the theorem's conclusion is an exponential decay in the optimization gap, this missing verification is load-bearing. The conditional statement 'if PL holds then...' is acceptable only if the PL-to-decay derivation is corrected, and the result still needs a discrete-time analysis for Algorithm 2.
minor comments (4)
  1. [Algorithm 2] The pseudocode input/output and step 7 refer to 'barycenter support' and update z^{(P)}_{τ+1,i}, while the algorithm is for GMM parameters θ=(π,μ,L). This is confusing and should be rewritten in terms of θ updates.
  2. [Abstract and §1] There are small typos, e.g. 'discrete methods are not because as they assume,' and the main text refers to 'equation 27' for the ground cost while it is introduced as Eq. (19). Please harmonize numbering and proofread.
  3. [Theorem 3.1 statement] The theorem statement uses expected empirical gap E[hat B_Q(P_τ)−hat B*_Q], while the proof bounds |B_Q(P_τ)−B*_Q| and then adds approximation terms twice. The notation hat B vs B and the placement of expectations should be made consistent, ideally with a fully discrete statement.
  4. [Table 2] The caption states bold is 'best barycenter method,' but GMM-DaDiL, a non-barycenter baseline, is bold in some columns. Clarify which methods are barycenter solvers and what the bold/underlined conventions mean for non-barycenter baselines.

Circularity Check

0 steps flagged

No significant circularity: the derivation is conditional on an unproved PL assumption, but no fitted quantity is repackaged as a prediction and no load-bearing self-citation chain is used.

full rationale

The paper's central derivations are not circular. Proposition 3.1 follows by construction from the definition of MW2 (Eq. 6) and the product structure of labeled GMMs; it is a formal justification of a formula previously used heuristically in [12], not a fitted result presented as a prediction. Theorems 3.1 and 3.2 are conditional guarantees: they assume a measure-theoretic PL inequality (Eq. 22/47) and then import external empirical/GMM approximation bounds from [42], [73], and [43]. The PL inequality is not verified for B_Q, and the proof of Theorem 3.1 appears to apply the first-order PL bound as if it were a squared-gradient bound. That is a correctness/support gap, not a circular one: an unproved or misused assumption makes a theorem unsupported, but it does not make the claimed conclusion identical to a fitted input. The paper also re-derives the Cuturi-Doucet update as a special case of its gradient-flow discretization (Appendix D.1/C.2), but this is a generalization/reformulation with mini-batching and extra functionals, not a renaming of a known result as a new prediction. Self-citations to [10,11,12] are used as prior work, baselines, and starting points; none of the convergence claims is justified solely by those self-citations, and the new experiments and code provide independent empirical support. Hence the circularity score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The central algorithm depends on several free hyperparameters and on strong analytical assumptions. The most consequential is the PL inequality, which is simply assumed. The empirical approximation bounds are imported from the literature. No new physical or mathematical entities are invented.

free parameters (7)
  • beta (label weight in ground cost) = not reported
    Equation (19) introduces beta to balance feature and label distances; no value or tuning protocol is given in the hyperparameter table.
  • margin in repulsion functional U_R = 1 for Office-Home; not listed for other benchmarks
    The hinge margin in Equation (21) controls how far same-label points must be; chosen per benchmark without a reported validation procedure.
  • learning rate alpha = 1.0, 1.0, 0.1, 1.0, 0.1 for Office31, OfficeHome, TEP, BCI, ISRUC
    Per-benchmark learning rates in Table 10; no sensitivity analysis is reported.
  • batch size n_b = 1550, 650, 290, 200, 1000
    Mini-batch sizes in Table 10; the algorithm's convergence proof does not explicitly depend on batch size and no stochastic approximation analysis is given.
  • number of barycenter samples n = 1550, 1300, 870, 800, 1000
    Support sizes in Table 10, chosen roughly as 50 samples per class.
  • number of GMM components = 124, 975, 145, 400, 40
    Model complexity for WGF-GMM; fixed per benchmark.
  • choice of V and U functionals = W2(P,Q_T) and/or Repulsion/Entropy per benchmark
    Table 10 selects different regularizers for different benchmarks; the paper does not describe a validation protocol for these choices.
axioms (4)
  • ad hoc to paper Strong PL inequality: ||grad B(P)||_{L2(P)} >= C_PL (B(P)-B*) for B_Q on P_2,ac(B(0,R))
    Assumed in Theorem 3.1/3.2 (Eq. 22/47). Not proved for the barycenter functional, not verified experimentally, and the proof uses it to get exponential decay.
  • domain assumption Input measures are supported on a bounded ball B(0,R); GMM parameters are bounded with fixed R_mu and R_Sigma
    Used in Theorem 3.1 and Proposition D.3 to invoke finite-radius empirical approximation bounds.
  • domain assumption Empirical approximation bounds of [73] and GMM approximation bounds of [43] hold with the stated constants
    Theorems rely on E sup |B - \hat B| <= C_p,R sqrt(C_d,m/n) and on GMM parameter convergence rates; these are cited external results, not proved.
  • domain assumption Using a mini-batch empirical measure \hat Q_k in W_p(P,\hat Q_k) gives a valid descent direction for the true W_p(P,Q_k)
    Algorithm 1 replaces Q_k by a sampled empirical measure; W_p(P,\hat Q_k) is not generally an unbiased estimator of W_p(P,Q_k), and the convergence proof does not account for this bias.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Wasserstein Gradient Flows for Scalable and Regularized Barycenter Computation." pith.science (2026). https://pith.science/paper/FJH6K3CO

@misc{pith2026251004602,
  author       = {Pith},
  title        = {Pith review of: Wasserstein Gradient Flows for Scalable and Regularized Barycenter Computation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FJH6K3CO}},
  note         = {Machine review of arXiv:2510.04602}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Wasserstein barycenters provide a principled approach for aggregating probability measures, while preserving the geometry of their ambient space. Existing discrete methods are not because as they assume access to the complete set of samples from the input measures. Meanwhile, neural network approaches do scale well, but rely on complex optimization problems and cannot easily incorporate label information. We address these limitations through gradient flows in the space of probability measures. Through time discretization, we achieve a scalable algorithm that i) relies on mini-batch optimal transport, ii) accepts modular regularization through task-aware functions, and iii) seamlessly integrates supervised information into the ground-cost. We empirically validate our approach on domain adaptation benchmarks that span computer vision, neuroscience, and chemical engineering. Our method establishes a new state-of-the-art Wasserstein barycenter solver, with labeled barycenters consistently outperforming unlabeled ones. Our code at https://github.com/SigmaNova/barycentric-gradient-flows

Figures

Figures reproduced from arXiv: 2510.04602 by Eduardo Fernandes Montesuma, Mike Gartrell, Yassir Bendou.

Figure 1
Figure 1. Figure 1: In (a), we show the usual static notion of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Location-scatter family generated by a [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison between Wasserstein barycenter solvers. Colored scatter plots indicate labeled barycenters. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: t-SNE [57] visualization of the barycenter [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Two moons samples alongside the fitted GMMs. Stars and ellipses denote the means and covariance matrices. In this experiment, similarly to the Swiss-roll example we explored in the main paper, we deal with Location-Scatter families of measures. The idea is that we are given Q0 ∈ P2,ac, and we define a finite family of measures, Q = {Tk,♯Q0 : Tk(x) = Akx + bk, Ak ∈ PD(d)} K k=1 (57) where PD(d) denotes the … view at source ↗
Figure 6
Figure 6. Figure 6: From left to right, comparison of the empirical barycenter with and without an interaction energy term, [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Imbalanced Gaussian blobs for the class imbalance example. We draw [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Discrete (a), and mini-batch Wasserstein barycenters (b and c). Overall, the discrete algorithm of [11] [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Provably convergent stochastic fixed-point algorithm for free-support Wasserstein barycenter of continuous non-parametric measures

    math.OC 2025-05 conditional novelty 7.0

    Develops the first provably convergent stochastic fixed-point algorithm for free-support 2-Wasserstein barycenters of continuous measures under Caffarelli regularity, using a modified entropic OT map estimator.

Reference graph

Works this paper leans on

80 extracted references · 10 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Barycen- ters in the wasserstein space.SIAM Journal on Mathematical Analysis, 43(2):904–924, 2011

    Martial Agueh and Guillaume Carlier. Barycen- ters in the wasserstein space.SIAM Journal on Mathematical Analysis, 43(2):904–924, 2011

    2011

  2. [2]

    Springer, 2008

    C´ edric Villani et al.Optimal transport: old and new, volume 338. Springer, 2008

    2008

  3. [3]

    A geometric study of wasser- stein spaces: Euclidean spaces.Annali della Scuola Normale Superiore di Pisa-Classe di Scienze, 9(2):297–323, 2010

    Beno ˆ ıt Kloeckner. A geometric study of wasser- stein spaces: Euclidean spaces.Annali della Scuola Normale Superiore di Pisa-Classe di Scienze, 9(2):297–323, 2010

    2010

  4. [4]

    Model fu- sion via optimal transport.Advances in Neural Information Processing Systems, 33:22045–22055, 2020

    Sidak Pal Singh and Martin Jaggi. Model fu- sion via optimal transport.Advances in Neural Information Processing Systems, 33:22045–22055, 2020

    2020

  5. [5]

    A barycenter-based ap- proach for the multi-model ensembling of subsea- sonal forecasts.arXiv preprint arXiv:2310.17933, 2023

    Camille Le Coz, Alexis Tantet, R´ emi Flamary, and Riwal Plougonven. A barycenter-based ap- proach for the multi-model ensembling of subsea- sonal forecasts.arXiv preprint arXiv:2310.17933, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Interpolation for robust learning: Data augmen- tation on wasserstein geodesics

    Jiacheng Zhu, Jielin Qiu, Aritra Guha, Zhuolin Yang, XuanLong Nguyen, Bo Li, and Ding Zhao. Interpolation for robust learning: Data augmen- tation on wasserstein geodesics. InInternational conference on machine learning, pages 43129– 43157. PMLR, 2023

    2023

  7. [7]

    Dataset distillation via the wasserstein metric

    Haoyang Liu, Yijiang Li, Tiancheng Xing, Vibhu Dalal, Luwei Li, Jingrui He, and Haohan Wang. Dataset distillation via the wasserstein metric. arXiv preprint arXiv:2311.18531, 2023

    Pith/arXiv arXiv 2023

  8. [9]

    Wasser- stein dictionary learning: Optimal transport- based unsupervised nonlinear dictionary learning

    Morgan A Schmitz, Matthieu Heitz, Nicolas Bon- neel, Fred Ngole, David Coeurjolly, Marco Cu- turi, Gabriel Peyr´ e, and Jean-Luc Starck. Wasser- stein dictionary learning: Optimal transport- based unsupervised nonlinear dictionary learning. SIAM Journal on Imaging Sciences, 11(1):643– 678, 2018

    2018

  9. [10]

    Wasserstein barycenter for multi-source domain adaptation

    Eduardo Fernandes Montesuma and Fred Mau- rice Ngole Mboula. Wasserstein barycenter for multi-source domain adaptation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16785–16793, 2021

    2021

  10. [11]

    Multi-source domain adaptation through dataset dictionary learning in wasserstein space

    Eduardo Montesuma, Fred Maurice Ngole Mboula, and Antoine Souloumiac. Multi-source domain adaptation through dataset dictionary learning in wasserstein space. InECAI 2023, pages 1739–1746. IOS Press, 2023

    2023

  11. [12]

    Lighter, bet- ter, faster multi-source domain adaptation with gaussian mixture models and optimal transport

    Eduardo Fernandes Montesuma, Fred Ngol` e Mboula, and Antoine Souloumiac. Lighter, bet- ter, faster multi-source domain adaptation with gaussian mixture models and optimal transport. InJoint European Conference on Machine Learn- ing and Knowledge Discovery in Databases, pages 21–38. Springer, 2024

    2024

  12. [13]

    Scalable bayes via barycenter in wasserstein space.Journal of Machine Learning Research, 19(8):1–35, 2018

    Sanvesh Srivastava, Cheng Li, and David B Dun- son. Scalable bayes via barycenter in wasserstein space.Journal of Machine Learning Research, 19(8):1–35, 2018

    2018

  13. [14]

    Fast computa- tion of wasserstein barycenters

    Marco Cuturi and Arnaud Doucet. Fast computa- tion of wasserstein barycenters. InInternational conference on machine learning, pages 685–693. PMLR, 2014

    2014

  14. [15]

    Iterative bregman projections for regularized transporta- tion problems.SIAM Journal on Scientific Com- puting, 37(2):A1111–A1138, 2015

    Jean-David Benamou, Guillaume Carlier, Marco Cuturi, Luca Nenna, and Gabriel Peyr´ e. Iterative bregman projections for regularized transporta- tion problems.SIAM Journal on Scientific Com- puting, 37(2):A1111–A1138, 2015

    2015

  15. [16]

    Debiased sinkhorn barycenters

    Hicham Janati, Marco Cuturi, and Alexandre Gramfort. Debiased sinkhorn barycenters. In International Conference on Machine Learning, pages 4692–4701. PMLR, 2020

    2020

  16. [17]

    Gradient descent algo- rithms for bures-wasserstein barycenters

    Sinho Chewi, Tyler Maunu, Philippe Rigollet, and Austin J Stromme. Gradient descent algo- rithms for bures-wasserstein barycenters. InCon- ference on Learning Theory, pages 1276–1304. PMLR, 2020

    2020

  17. [18]

    Statistical inference for bures–wasserstein barycenters.The Annals of Ap- plied Probability, 31(3):1264–1298, 2021

    Alexey Kroshnin, Vladimir Spokoiny, and Alexandra Suvorikova. Statistical inference for bures–wasserstein barycenters.The Annals of Ap- plied Probability, 31(3):1264–1298, 2021

    2021

  18. [19]

    A wasserstein- type distance in the space of gaussian mixture models.SIAM Journal on Imaging Sciences, 13(2):936–970, 2020

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

    2020

  19. [20]

    Multi-marginal optimal transport: theory and applications.ESAIM: Mathematical Modelling and Numerical Analysis, 49(6):1771– 1790, 2015

    Brendan Pass. Multi-marginal optimal transport: theory and applications.ESAIM: Mathematical Modelling and Numerical Analysis, 49(6):1771– 1790, 2015

    2015

  20. [21]

    Continuous wasserstein-2 barycenter estimation without minimax optimiza- tion.arXiv preprint arXiv:2102.01752, 2021

    Alexander Korotin, Lingxiao Li, Justin Solomon, and Evgeny Burnaev. Continuous wasserstein-2 barycenter estimation without minimax optimiza- tion.arXiv preprint arXiv:2102.01752, 2021. Computing W asserstein Barycenters through Gradient Flows

    Pith/arXiv arXiv 2021

  21. [22]

    Input convex neural networks

    Brandon Amos, Lei Xu, and J Zico Kolter. Input convex neural networks. InInternational confer- ence on machine learning, pages 146–155. PMLR, 2017

    2017

  22. [23]

    Estimating barycenters of distributions with neural optimal transport.arXiv preprint arXiv:2402.03828, 2024

    Alexander Kolesov, Petr Mokrov, Igor Udovichenko, Milena Gazdieva, Gudmund Pammer, Evgeny Burnaev, and Alexander Ko- rotin. Estimating barycenters of distributions with neural optimal transport.arXiv preprint arXiv:2402.03828, 2024

    Pith/arXiv arXiv 2024

  23. [24]

    Robust barycenter estimation using semi-unbalanced neural optimal transport

    Milena Gazdieva, Jaemoo Choi, Alexander Kolesov, Jaewoong Choi, Petr Mokrov, and Alexander Korotin. Robust barycenter estimation using semi-unbalanced neural optimal transport. arXiv preprint arXiv:2410.03974, 2024

    Pith/arXiv arXiv 2024

  24. [25]

    Com- puting optimal transport maps and wasserstein barycenters using conditional normalizing flows

    Gabriele Visentin and Patrick Cheridito. Com- puting optimal transport maps and wasserstein barycenters using conditional normalizing flows. arXiv preprint arXiv:2505.22364, 2025

    Pith/arXiv arXiv 2025

  25. [26]

    Wasserstein itera- tive networks for barycenter estimation.Ad- vances in Neural Information Processing Systems, 35:15672–15686, 2022

    Alexander Korotin, Vage Egiazarian, Lingxiao Li, and Evgeny Burnaev. Wasserstein itera- tive networks for barycenter estimation.Ad- vances in Neural Information Processing Systems, 35:15672–15686, 2022

    2022

  26. [27]

    Springer Science & Business Media, 2008

    Luigi Ambrosio, Nicola Gigli, and Giuseppe Savar´ e.Gradient flows: in metric spaces and in the space of probability measures. Springer Science & Business Media, 2008

    2008

  27. [28]

    Filippo Santambrogio.{Euclidean, metric, and Wasserstein}gradient flows: an overview.Bul- letin of Mathematical Sciences, 7:87–154, 2017

    2017

  28. [29]

    Computa- tional optimal transport: With applications to data science.Foundations and Trends®in Ma- chine Learning, 11(5-6):355–607, 2019

    Gabriel Peyr´ e, Marco Cuturi, et al. Computa- tional optimal transport: With applications to data science.Foundations and Trends®in Ma- chine Learning, 11(5-6):355–607, 2019

    2019

  29. [30]

    Re- cent advances in optimal transport for machine learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(2):1161–1180, 2025

    Eduardo Fernandes Montesuma, Fred Mau- rice Ngol` e Mboula, and Antoine Souloumiac. Re- cent advances in optimal transport for machine learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(2):1161–1180, 2025

    2025

  30. [31]

    M´ emoire sur la th´ eorie des d´ eblais et des remblais.Histoire de l’Acad´ emie Royale des Sciences de Paris, 1781

    Gaspard Monge. M´ emoire sur la th´ eorie des d´ eblais et des remblais.Histoire de l’Acad´ emie Royale des Sciences de Paris, 1781

  31. [32]

    On the transfer of masses (in russian)

    L Kantorovich. On the transfer of masses (in russian). InDoklady Akademii Nauk, volume 37, pages 227–229, 1942

    1942

  32. [33]

    Wasserstein geometry of gaussian measures.Osaka Journal of Mathematics, 2011

    Asuka Takatsu. Wasserstein geometry of gaussian measures.Osaka Journal of Mathematics, 2011

    2011

  33. [34]

    Les ´ el´ ements al´ eatoires de na- ture quelconque dans un espace distanci´ e

    Maurice Fr´ echet. Les ´ el´ ements al´ eatoires de na- ture quelconque dans un espace distanci´ e. In Annales de l’institut Henri Poincar´ e, volume 10, pages 215–310, 1948

    1948

  34. [35]

    Riemannian center of mass and mollifier smoothing.Communications on pure and applied mathematics, 30(5):509–541, 1977

    Hermann Karcher. Riemannian center of mass and mollifier smoothing.Communications on pure and applied mathematics, 30(5):509–541, 1977

    1977

  35. [36]

    A fixed- point approach to barycenters in wasserstein space.Journal of Mathematical Analysis and Applications, 441(2):744–762, 2016

    Pedro C ´Alvarez-Esteban, E Del Barrio, JA Cuesta-Albertos, and C Matr´ an. A fixed- point approach to barycenters in wasserstein space.Journal of Mathematical Analysis and Applications, 441(2):744–762, 2016

    2016

  36. [37]

    Optimal transport for multi-source domain adaptation under target shift

    Ievgen Redko, Nicolas Courty, R´ emi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. InThe 22nd International Conference on artificial in- telligence and statistics, pages 849–858. PMLR, 2019

    2019

  37. [38]

    Springer, 2015

    Filippo Santambrogio.Optimal transport for ap- plied mathematicians, volume 87. Springer, 2015

    2015

  38. [39]

    Dataset dynamics via gradient flows in probability space

    David Alvarez-Melis and Nicol` o Fusi. Dataset dynamics via gradient flows in probability space. InInternational conference on machine learning, pages 219–230. PMLR, 2021

    2021

  39. [40]

    Gradient methods for solving equations and inequalities.USSR Computational Mathematics and Mathematical Physics, 4(6):17– 32, 1964

    Boris T Polyak. Gradient methods for solving equations and inequalities.USSR Computational Mathematics and Mathematical Physics, 4(6):17– 32, 1964

    1964

  40. [41]

    Linear convergence of gradient and proximal- gradient methods under the polyak- lojasiewicz condition

    Hamed Karimi, Julie Nutini, and Mark Schmidt. Linear convergence of gradient and proximal- gradient methods under the polyak- lojasiewicz condition. InJoint European conference on machine learning and knowledge discovery in databases, pages 795–811. Springer, 2016

    2016

  41. [42]

    Statistical optimal transport.arXiv preprint arXiv:2407.18163, 3, 2024

    Sinho Chewi, Jonathan Niles-Weed, and Philippe Rigollet. Statistical optimal transport.arXiv preprint arXiv:2407.18163, 3, 2024

    Pith/arXiv arXiv 2024

  42. [43]

    Differentiable expectation-maximisation and applications to gaussian mixture model optimal transport.arXiv preprint arXiv:2509.02109, 2025

    Samuel Bo ¨ ıt´ e, Eloi Tanguy, Julie Delon, Agn` es Desolneux, and R´ emi Flamary. Differentiable expectation-maximisation and applications to gaussian mixture model optimal transport.arXiv preprint arXiv:2509.02109, 2025

    arXiv 2025

  43. [44]

    A review of domain adaptation without target labels.IEEE transactions on pattern analysis and machine in- telligence, 43(3):766–785, 2019

    Wouter M Kouw and Marco Loog. A review of domain adaptation without target labels.IEEE transactions on pattern analysis and machine in- telligence, 43(3):766–785, 2019. Eduardo F ernandes Montesuma, Y assir Bendou, Mike Gartrell

    2019

  44. [45]

    A survey of multi-source domain adaptation.Infor- mation Fusion, 24:84–92, 2015

    Shiliang Sun, Honglei Shi, and Yuanbin Wu. A survey of multi-source domain adaptation.Infor- mation Fusion, 24:84–92, 2015

    2015

  45. [46]

    A sur- vey on transfer learning.IEEE Transactions on knowledge and data engineering, 22(10):1345– 1359, 2009

    Sinno Jialin Pan and Qiang Yang. A sur- vey on transfer learning.IEEE Transactions on knowledge and data engineering, 22(10):1345– 1359, 2009

    2009

  46. [47]

    Optimal transport for do- main adaptation.IEEE transactions on pattern analysis and machine intelligence, 39(9):1853– 1865, 2016

    Nicolas Courty, R´ emi Flamary, Devis Tuia, and Alain Rakotomamonjy. Optimal transport for do- main adaptation.IEEE transactions on pattern analysis and machine intelligence, 39(9):1853– 1865, 2016

    2016

  47. [48]

    Deep residual learning for image recog- nition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog- nition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  48. [49]

    Cbramod: A criss-cross brain foun- dation model for eeg decoding.arXiv preprint arXiv:2412.07236, 2024

    Jiquan Wang, Sha Zhao, Zhiling Luo, Yangx- uan Zhou, Haiteng Jiang, Shijian Li, Tao Li, and Gang Pan. Cbramod: A criss-cross brain foun- dation model for eeg decoding.arXiv preprint arXiv:2412.07236, 2024

    arXiv 2024

  49. [50]

    Benchmarking domain adapta- tion for chemical processes on the tennessee east- man process

    Eduardo Fernandes Montesuma, Michela Mulas, Fred Ngol` e Mboula, Francesco Corona, and An- toine Souloumiac. Benchmarking domain adapta- tion for chemical processes on the tennessee east- man process. InML4CCE Workshop, 2024

    2024

  50. [51]

    Adapting visual category mod- els to new domains

    Kate Saenko, Brian Kulis, Mario Fritz, and Trevor Darrell. Adapting visual category mod- els to new domains. InEuropean conference on computer vision, pages 213–226. Springer, 2010

    2010

  51. [52]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. InProceedings of the IEEE Confer- ence on Computer Vision and Pattern Recogni- tion, pages 5018–5027, 2017

    2017

  52. [53]

    Bci competition 2008–graz data set a.Insti- tute for knowledge discovery (laboratory of brain- computer interfaces), Graz University of Technol- ogy, 16(1-6):34, 2008

    Clemens Brunner, Robert Leeb, Gernot M¨ uller- Putz, Alois Schl¨ ogl, and Gert Pfurtscheller. Bci competition 2008–graz data set a.Insti- tute for knowledge discovery (laboratory of brain- computer interfaces), Graz University of Technol- ogy, 16(1-6):34, 2008

    2008

  53. [54]

    Automatic sleep staging: A computer assisted approach for optimal combina- tion of features and polysomnographic channels

    Sirvan Khalighi, Teresa Sousa, Gabriel Pires, and Urbano Nunes. Automatic sleep staging: A computer assisted approach for optimal combina- tion of features and polysomnographic channels. Expert Systems with Applications, 40(17):7046– 7059, 2013

    2013

  54. [55]

    PhD thesis, Universit´ e Paris- Saclay, 2024

    Eduardo Fernandes Montesuma.Multi-Source Domain Adaptation through Wasserstein Barycenters. PhD thesis, Universit´ e Paris- Saclay, 2024

    2024

  55. [56]

    Multi-source do- main adaptation via weighted joint distributions optimal transport

    Rosanna Turrisi, R´ emi Flamary, Alain Rakotoma- monjy, and Massimiliano Pontil. Multi-source do- main adaptation via weighted joint distributions optimal transport. InUncertainty in artificial in- telligence, pages 1970–1980. PMLR, 2022

    1970

  56. [57]

    Vi- sualizing data using t-sne.Journal of machine learning research, 9(Nov):2579–2605, 2008

    Laurens van der Maaten and Geoffrey Hinton. Vi- sualizing data using t-sne.Journal of machine learning research, 9(Nov):2579–2605, 2008

    2008

  57. [58]

    Rates of estimation of optimal transport maps using plug-in estimators via barycentric projections.Advances in Neural Information Pro- cessing Systems, 34:29736–29753, 2021

    Nabarun Deb, Promit Ghosal, and Bodhisattva Sen. Rates of estimation of optimal transport maps using plug-in estimators via barycentric projections.Advances in Neural Information Pro- cessing Systems, 34:29736–29753, 2021

    2021

  58. [59]

    Estimating barycenters of measures in high dimensions.arXiv preprint arXiv:2007.07105, 2020

    Samuel Cohen, Michael Arbel, and Marc Pe- ter Deisenroth. Estimating barycenters of measures in high dimensions.arXiv preprint arXiv:2007.07105, 2020

    Pith/arXiv arXiv 2007

  59. [60]

    Optimal transport mapping via input convex neural networks

    Ashok Makkuva, Amirhossein Taghvaei, Sewoong Oh, and Jason Lee. Optimal transport mapping via input convex neural networks. InInterna- tional Conference on Machine Learning, pages 6672–6681. PMLR, 2020

    2020

  60. [61]

    Scalable computations of wasserstein barycenter via input convex neural networks

    Jiaojiao Fan, Amirhossein Taghvaei, and Yongxin Chen. Scalable computations of wasserstein barycenter via input convex neural networks. arXiv preprint arXiv:2007.04462, 2020

    Pith/arXiv arXiv 2007

  61. [62]

    Gen- erative adversarial nets.Advances in neural infor- mation processing systems, 27, 2014

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen- erative adversarial nets.Advances in neural infor- mation processing systems, 27, 2014

    2014

  62. [63]

    Polar factorization and monotone rearrangement of vector-valued functions.Com- munications on pure and applied mathematics, 44(4):375–417, 1991

    Yann Brenier. Polar factorization and monotone rearrangement of vector-valued functions.Com- munications on pure and applied mathematics, 44(4):375–417, 1991

    1991

  63. [64]

    Kantorovich duality for general transport costs and applications.Journal of Functional Analysis, 273(11):3327–3405, 2017

    Nathael Gozlan, Cyril Roberto, Paul-Marie Sam- son, and Prasad Tetali. Kantorovich duality for general transport costs and applications.Journal of Functional Analysis, 273(11):3327–3405, 2017

    2017

  64. [65]

    Existence, duality, and cycli- cal monotonicity for weak transport costs.Calcu- lus of Variations and Partial Differential Equa- tions, 58(6):203, 2019

    Julio Backhoff-Veraguas, Mathias Beiglb¨ ock, and Gudmun Pammer. Existence, duality, and cycli- cal monotonicity for weak transport costs.Calcu- lus of Variations and Partial Differential Equa- tions, 58(6):203, 2019. Computing W asserstein Barycenters through Gradient Flows

    2019

  65. [66]

    Neural optimal transport.arXiv preprint arXiv:2201.12220, 2022

    Alexander Korotin, Daniil Selikhanovych, and Evgeny Burnaev. Neural optimal transport.arXiv preprint arXiv:2201.12220, 2022

    Pith/arXiv arXiv 2022

  66. [67]

    Optimal entropy-transport problems and a new hellinger–kantorovich distance between positive measures.Inventiones mathematicae, 211(3):969–1117, 2018

    Matthias Liero, Alexander Mielke, and Giuseppe Savar´ e. Optimal entropy-transport problems and a new hellinger–kantorovich distance between positive measures.Inventiones mathematicae, 211(3):969–1117, 2018

    2018

  67. [68]

    The ge- ometry of optimal transportation

    Wilfrid Gangbo and Robert J McCann. The ge- ometry of optimal transportation. 1996

    1996

  68. [69]

    Normalizing flows: An introduction and review of current methods.IEEE trans- actions on pattern analysis and machine intelli- gence, 43(11):3964–3979, 2020

    Ivan Kobyzev, Simon JD Prince, and Marcus A Brubaker. Normalizing flows: An introduction and review of current methods.IEEE trans- actions on pattern analysis and machine intelli- gence, 43(11):3964–3979, 2020

    2020

  69. [70]

    An overview of statistical learning theory.IEEE transactions on neural net- works, 10(5):988–999, 1999

    Vladimir N Vapnik. An overview of statistical learning theory.IEEE transactions on neural net- works, 10(5):988–999, 1999

    1999

  70. [71]

    Dataset shift in machine learning

    Joaquin Qui˜ nonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. Mit Press, 2022

    2022

  71. [72]

    Multi- source domain adaptation meets dataset distil- lation through dataset dictionary learning

    Eduardo Fernandes Montesuma, Fred Ngol` e Mboula, and Antoine Souloumiac. Multi- source domain adaptation meets dataset distil- lation through dataset dictionary learning. In ICASSP 2024-2024 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 5620–5624. IEEE, 2024

    2024

  72. [73]

    Sample complexity of optimal transport barycenters with discrete support.arXiv preprint arXiv:2505.21274, 2025

    L´ eo Portales, Edouard Pauwels, and Elsa Cazelles. Sample complexity of optimal transport barycenters with discrete support.arXiv preprint arXiv:2505.21274, 2025

    arXiv 2025

  73. [74]

    Optimal transport for domain adaptation through gaussian mixture models.Transactions on Ma- chine Learning Research, 2025

    Eduardo Fernandes Montesuma, Fred Mau- rice NGOLE MBOULA, and Antoine Souloumiac. Optimal transport for domain adaptation through gaussian mixture models.Transactions on Ma- chine Learning Research, 2025

    2025

  74. [75]

    Adap- tiope: A modern benchmark for unsupervised domain adaptation

    Tobias Ringwald and Rainer Stiefelhagen. Adap- tiope: A modern benchmark for unsupervised domain adaptation. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 101–110, 2021

    2021

  75. [76]

    An extended tennessee eastman simula- tion dataset for fault-detection and decision sup- port systems.Computers & chemical engineering, 149:107281, 2021

    Christopher Reinartz, Murat Kulahci, and Ole Ravn. An extended tennessee eastman simula- tion dataset for fault-detection and decision sup- port systems.Computers & chemical engineering, 149:107281, 2021. Eduardo F ernandes Montesuma, Y assir Bendou, Mike Gartrell Computing Wasserstein Barycenters through Gradient Flows Contents 1 Introduction 1 2 Backgro...

    2021

  76. [77]

    However, the softmax operation for getting the labels, i.e.,y= softmax(ℓ), is not invertible

    Normalizing flows require invertible transformations. However, the softmax operation for getting the labels, i.e.,y= softmax(ℓ), is not invertible

  77. [78]

    However, applying the softmax breaks the convexity requirement

    Using ICNNs, one can enforce the convexity ofz= (x, ℓ). However, applying the softmax breaks the convexity requirement. As a result, adapting neural net based methods for the jointX × Yspace is challenging and not straightforward. We leave this question for future work. Eduardo F ernandes Montesuma, Y assir Bendou, Mike Gartrell C Additional Background C....

  78. [79]

    The exponential convergence given by the PL inequality, ∥∇ ∇B(P)∥L2(P) ≥C PL(B(P)−B ⋆),(47)

  79. [80]

    These results are available, respectively, in [42] and [73]

    An uniform bound onP7→B Q(P) controlling the error of approximatingQ k with ˆQk. These results are available, respectively, in [42] and [73]. For completeness, we re-state and demonstrate them. We start with a modified version of [42, Corollary 5.17], Proposition D.2.(Measure theoretic Gronwall’s Lemma [42]) LetF:P 2,ac(Ω)→Rbe a functional that satisfies ...

    2055

  80. [81]

    Our goal here is to perform cross-subjectadaptation, namely, we use data from a given set of source subjects, and try to predict on a target subject

    Therefore, in this benchmark we have four classes. Our goal here is to perform cross-subjectadaptation, namely, we use data from a given set of source subjects, and try to predict on a target subject. Therefore, each domain is a subject, so there are nine domains in total. There are, on average, 560 samples per subject, with a total of 5088 samples. We sh...

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.