Pith. sign in

REVIEW 5 major objections 5 minor 58 references

Neural Estimation for Scaling Entropic Multimarginal Optimal Transport

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that entropic multimarginal optimal transport can be estimated by neural dual potentials trained on mini-batches, achieving parametric error $n^{-1/2}$ at per-epoch cost $O(nk^2)$ instead of Sinkhorn's $O(n^k)$.

desk verdict A genuinely useful neural estimator for entropic MOT with real speedups, but the formal guarantee covers only a global maximizer of a nonconvex objective, not the SGD algorithm that is actually run. read the letter →

arxiv 2506.00573 v1 pith:RPJM5ZKQ submitted 2025-05-31 cs.LG stat.ML

classification cs.LGstat.ML MSC 49Q2268T07
keywords entropicmultimarginaloptimaltransportneuralestimationSinkhornalgorithmparametricconvergencerateGromov-Wassersteinmini-batchoptimizationdualpotentialsplan
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 NEMOT, a neural estimator for entropic multimarginal optimal transport (EMOT). Instead of forming the $k$-way coupling tensor that makes Sinkhorn's algorithm cost $O(n^k)$, NEMOT optimizes the dual objective of EMOT over neural-network potentials using mini-batch gradient ascent, so the per-epoch cost scales polynomially in $n$ and $k$. The paper proves non-asymptotic error bounds of order $(1+\epsilon^{-(\lfloor d/2\rfloor+1)})k^{p(G_c)}(m^{-1/2}+n^{-1/2})$ for shallow ReLU networks, which it interprets as a sharp parametric $n^{-1/2}$ rate and minimax optimality. It also bounds the KL error of the induced neural transport plan by $\epsilon^{-1}$ times the cost error. If correct, NEMOT computes EMOT costs and plans at dataset sizes and numbers of marginals where Sinkhorn cannot complete a single iteration.

What carries the argument

The load-bearing object is the dual representation of entropic MOT: $\mathrm{MOT}_{c,\epsilon}(\mu^k) = \sup_{\phi_1,\dots,\phi_k} \sum_{i=1}^k \int \phi_i\,d\mu_i - \epsilon \int \exp((\oplus_{i=1}^k \phi_i - c)/\epsilon)\,d(\otimes_i \mu_i) + \epsilon$, an unconstrained concave maximization over $k$ dual potentials. NEMOT replaces each $\phi_i$ by a neural network $f_{\theta_i}$, approximates the integrals by sample means over mini-batches, and maximizes the objective by gradient ascent. The proof machinery splits the error into an approximation term (the optimal potentials are shown to be smooth, so $m$-neuron ReLU networks approximate them at rate $m^{-1/2}$) and an estimation term (empirical-process maximal inequalities give rate $n^{-1/2}$). The cost graph $G_c$, recording which pairwise terms appear in $c$, sets the polynomial order $k^{p(G_c)}$ in the bound and the per-epoch cost $O(n|E|)$.

What would settle it

On a full-cost EMOT problem with $k=3$ where multimarginal Sinkhorn is tractable, train NEMOT from many random initializations and learning rates, and test whether the final objective value reliably matches Sinkhorn within the Theorem 3.1 bound as $n$ grows; if seed-to-seed variation or a plateau in the loss exceeds the bound, the global-maximizer premise fails. A second check is to replace the smooth pairwise cost with a non-smooth cost such as $|x_i-x_j|$ and see whether the $n^{-1/2}$ rate persists.

Watch

Extended reading notes

Core claim

The central claim is that the computational bottleneck of entropic multimarginal optimal transport can be shifted from the dataset size to the network architecture: by parameterizing the $k$ dual potentials in the EMOT dual representation by neural nets and optimizing with mini-batches, the estimator achieves the same $n^{-1/2}$ statistical rate as the empirical plug-in while avoiding the $O(n^k)$ tensor operations of multimarginal Sinkhorn. The paper shows this through Theorem 3.1, which bounds the expected absolute error of the NEMOT cost estimate by a term proportional to $k^{p(G_c)}(m^{-1/2}+n^{-1/2})$ with a constant depending on the regularization $\epsilon$ and dimension, and Theorem 3.2, which transfers that bound to KL divergence between the true entropic plan and the neural plan up to a factor $1/\epsilon$. The error analysis separates approximation error (optimal dual potentials are smooth, so shallow ReLU networks approximate them at rate $m^{-1/2}$) from estimation error (empirical means over $n$ samples converge at $n^{-1/2}$). The paper further claims orders-of-magnitude runtime speedups over Sinkhorn in experiments, including circle-cost problems with $k=10$ and full-cost problems with $n=10^4$, and extends the estimator to multimarginal entropic Gromov-Wasserstein via alternating optimization over auxiliary matrices.

Load-bearing premise

The error bounds in Theorems 3.1 and 3.2 hold only for a global maximizer of the nonconvex NEMOT objective; the implemented mini-batch gradient ascent has no optimization-error analysis, and the theory also assumes the cost is a sum of continuously differentiable pairwise terms even though the method is advertised for arbitrary costs.

Editorial extensions

If this is right

  • At dataset sizes and marginal counts where multimarginal Sinkhorn is infeasible, NEMOT returns cost estimates whose errors, under the idealized optimizer, decay at the parametric $n^{-1/2}$ rate.
  • NEMOT provides a neural transport plan with KL error $O(\epsilon^{-1} n^{-1/2})$, so downstream tasks that need the coupling rather than just the cost can use the plan.
  • Because the objective is evaluated on mini-batches, NEMOT can be inserted as a differentiable loss or regularizer in larger learning pipelines without materializing the $k$-way tensor.
  • For sparse cost graphs such as circle or tree structures, the per-epoch cost drops from $O(nk^2)$ to $O(nk)$, widening the feasible $(n,k)$ regime further.
  • The variational form of multimarginal entropic Gromov-Wasserstein as an infimum over EMOT problems yields a neural MGW solver by alternating gradient steps on auxiliary matrices with NEMOT updates.

Reading between the lines

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

  • The paper does not analyze the optimization error of mini-batch gradient ascent; a natural extension is to certify or bound that error, which would turn NEMOT into a fully verified differentiable EMOT layer.
  • The stated $O(nk^2)$ per-epoch cost counts only the objective evaluation, while backpropagation adds $O(nkLN^2)$; the practical speedup over Sinkhorn therefore depends on keeping the networks small relative to the cost graph size.
  • The author-conjectured refinement in Remark A.4, replacing the largest marginal dimension by the second largest in the rate, would make NEMOT substantially cheaper for problems with one very high-dimensional marginal.
  • Because NEMOT is trained rather than solved, it could amortize across repeated EMOT problems by fine-tuning one network as $\epsilon$ or the cost changes, a mode of use Sinkhorn does not offer.
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 / 5 minor

Summary. The paper proposes NEMOT, a neural estimator for entropic multimarginal optimal transport. It parameterizes the k dual potentials in the Carlier–Laborde dual formulation by shallow ReLU networks, replaces population expectations by sample means, and optimizes the resulting empirical objective with mini-batch Adam. The main theoretical results are Theorem 3.1, an O((1+ε^{-(d/2+1)}) k^{p(G_c)} (m^{-1/2}+n^{-1/2})) bound on the cost estimation error for a global maximizer of the empirical objective, and Theorem 3.2, a KL bound on the induced neural plan. The paper also develops a neural entropic Gromov-Wasserstein extension via an alternating scheme, analyzes complexity, and reports experiments showing orders-of-magnitude speedups over multimarginal Sinkhorn on synthetic and image data.

Significance. If the results held for the implemented algorithm, the contribution would be substantial: a parametric n^{-1/2} rate with O(n k^2) per-epoch complexity would let EMOT be applied in regimes inaccessible to Sinkhorn. The paper's derivation is principled: the error decomposition into approximation and estimation terms is explicit, the approximation relies on the smoothness of the Schrödinger-system potentials, the estimation bound uses empirical process maximal inequalities, and the sparse-cost U-statistic formulation is a useful addition. The empirical evaluation supports the practical speedups. However, the theoretical claims are conditional on a global maximizer of a nonconvex objective and on a cost structure that is narrower than advertised; the minimax-optimality statement is also not supported by a lower bound. These gaps are fixable and should be addressed before the guarantees can be taken at face value.

major comments (5)
  1. [Section 3.1, Theorems 3.1–3.2, Algorithm 1] The main error bounds are stated for a global maximizer θ* of the empirical NEMOT objective (5), and the text explicitly defers the optimization error to future work. Algorithm 1, however, solves (5) by mini-batch Adam over nonconvex neural-network parameters, and no result shows that its iterates approach a near-global maximizer or that a local maximum has a controlled suboptimality gap. Consequently, Theorems 3.1 and 3.2 do not apply to the output of the implemented algorithm, and the abstract's 'formal guarantees on the accuracy of NEMOT' are not currently established for the method as run. Please either supply an optimization-error analysis for the actual routine, or empirically validate near-global maximization (e.g., multiple restarts and certified upper bounds on small instances) and explicitly restate the guarantees as conditional on the optimizer.
  2. [Appendix A.6, Eq. (26)] Equation (26) concludes S2_ϵ = sup_{A∈D_M} inf_{π∈Π(µk)} [32||A||^2_F + ∫ c_A dπ + εDKL(π∥µ)]. This sign is inconsistent with Proposition 4.2 and Eq. (11), where S2_ϵ is an infimum over A of the EMOT value (itself an infimum over π). As written, the sup/inf order reverses the optimization and, if used in the NEMGW estimator (12), would change the objective. Please correct the sign and, if an inf/sup exchange is intended, justify it explicitly.
  3. [Lemma A.3 and proof of Theorem 3.1] The proof of Lemma A.3 derives an estimation bound proportional to a times (1+ε^{1-s})(deg G_c)^s and then says 'by setting a = \bar c_{b,d}' the displayed lemma bound follows. The lemma statement itself contains no explicit a, while Theorem 3.1 sets a = C(1+ε^{-(⌊d/2⌋+3)}). The dependence of the final ε exponent in (7) on this choice of a is therefore not checkable from the text. Please spell out the substitution and the constants so that the exponent on ε in (7) follows from the proof.
  4. [Theorem 3.1 and Section 5] The constant in (7) has the form (1+ε^{-(⌊d/2⌋+1)}) k^{d+6}, which is exponential in the dimension d. The experiments report accurate estimates for d up to 1000, where this bound is vacuous; the theory therefore does not explain the observed high-dimensional behavior. The statement 'this implies minimax optimality of NEMOT' also does not follow from an upper bound of parametric order unless a matching lower bound for the EMOT estimation problem is provided or cited. Please add the lower-bound reference or argument and clarify that the theoretical rate is for fixed dimension.
  5. [Theorem 3.2 and Eq. (6)] The plan bound in Theorem 3.2 is stated for the population-level density π^{θ*}_ϵ defined in (6), which uses the true marginal product ⊗_i μ_i. Algorithm 1 and the experiments output a normalized empirical tensor. The theorem does not account for the error incurred by replacing the true product measure with the empirical one, so the KL guarantee does not cover the plan that is actually produced. Please either state a plan bound for the empirical plan or clarify precisely what object the theorem controls.
minor comments (5)
  1. [Section 1.1] The phrase 'estimation erros' in the contributions paragraph is a typo and should read 'estimation errors'.
  2. [Appendix A.3] The name 'Faa di Bronu' should be 'Faa di Bruno'.
  3. [Appendix A.1, Eq. (14)] The Lipschitz constant for the exponential map is written as e^M/ε, but for exp((∑φ_i - c)/ε) on a set where the exponent is bounded by M, the constant should involve e^{M/ε}/ε. Please clarify the definition of M and the resulting constant.
  4. [Figure 5(c)] The caption says 'Estimated EMOT vs. k' but the y-axis is labeled 'OT loss', and for k>3 no Sinkhorn baseline is available; the claimed linear scaling is not independently verified in that panel.
  5. [Section 3.2, Table 1] The NEMOT complexity entry O(n k^2) for the full cost refers to cost evaluation only; the accompanying backpropagation cost O(n k L N^2) can dominate, as the text later notes. It would be clearer to state the total per-epoch complexity in the table or its caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: NEMOT's guarantees are derived from the independent EMOT dual formulation and empirical process bounds; reliance on same-group prior results is not load-bearing circularity, and the unaccounted optimization error is a validity gap, not a circular reduction.

full rationale

The central derivation is not circular. NEMOT is defined as the empirical dual EMOT objective (5), and Theorems 3.1 and 3.2 bound the gap between its global maximizer value and the population EMOT cost/plan via the standard decomposition into neural approximation error and empirical-process estimation error (Appendix A.1). The approximation step uses the smoothness of Schrödinger-system dual potentials (Lemma A.2) and a general ReLU approximation bound from [27, Prop. 10]; the estimation step uses covering/bracketing maximal inequalities. None of these bounds are fitted to the data or defined in terms of the target quantity. The paper's use of prior work by the same group ([27], [29], [49], [50], [52], [56]) is as citable theorems and techniques with their own stated assumptions, not as an unverified premise asserting the present result; the multimarginal extension is proven in the appendices. The acknowledged omission in Section 3.1 — 'leaving the account of the optimization error for future work' — is a substantive gap between the global-maximizer theorem and Algorithm 1's mini-batch SGD, but it is not a circular reduction: the theorem explicitly quantifies over a global maximizer θ* of (5), and the empirical speedups are not used to derive the bound. No fitted parameter is renamed as a prediction, and no uniqueness theorem is imported to force the choice. Consequently, no circular step is present.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. The free parameters are standard regularization and architecture choices, and the axioms are standard OT duality, smoothness assumptions, and empirical process theory. No entity is pulled from a hat.

free parameters (4)
  • regularization epsilon
    Free hyperparameter controlling entropic regularization; error bounds scale as epsilon^{-O(d)}. Value is not specified for most synthetic experiments (only MNIST cosine experiment states epsilon=0.1).
  • neural network size (m, L, N)
    Network width m and depth L are chosen by hand (3 layers, widths 10K with K=min(10d,80)). The theorem's rate requires n similar to m but experiments fix m.
  • optimization hyperparameters
    Learning rate 5e-5, Adam, gradient clipping 0.1, exponential decay, batch size b=64. These affect the unaccounted optimization error.
  • constant a in function class = C(1+epsilon^{-(floor(d/2)+3)})
    Parameter bound for neural nets chosen in Theorem 3.1; C is an unspecified universal constant.
assumptions (6)
  • standard math Dual representation of EMOT (Eq. 3)
    Follows from [39] Carlier-Laborde; the objective is the starting point for NEMOT.
  • domain assumption Optimal dual potentials belong to a Holder class with explicit derivative bounds (Lemma A.2)
    Assumed and proven under the condition that the pairwise cost c is continuously differentiable and marginals are compactly supported with finite moments.
  • standard math Function approximation bound for shallow ReLU networks (Prop A.1 from [27])
    Used to control the approximation error; the constant depends on the smoothness parameter.
  • standard math Empirical process bounds (covering and bracketing numbers, Rademacher maximal inequalities)
    Used to bound the estimation error in Lemma A.3.
  • domain assumption Sampling: n i.i.d. k-tuples from the product measure mu_1 x ... x mu_k
    The empirical objective in (5) uses these tuples as a Monte Carlo estimate; the data are observed as separate marginal samples, not necessarily aligned tuples.
  • domain assumption KL divergence to the product measure is finite and the Schrodinger system admits a solution
    Used in the proof of Theorem 3.2 and Lemma A.2; requires absolute continuity of the plan.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Estimation for Scaling Entropic Multimarginal Optimal Transport." pith.science (2026). https://pith.science/paper/RPJM5ZKQ

@misc{pith2026250600573,
  author       = {Pith},
  title        = {Pith review of: Neural Estimation for Scaling Entropic Multimarginal Optimal Transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RPJM5ZKQ}},
  note         = {Machine review of arXiv:2506.00573}
}
abstract

Multimarginal optimal transport (MOT) is a powerful framework for modeling interactions between multiple distributions, yet its applicability is bottlenecked by a high computational overhead. Entropic regularization provides computational speedups via the multimarginal Sinkhorn algorithm, whose time complexity, for a dataset size $n$ and $k$ marginals, generally scales as $O(n^k)$. However, this dependence on the dataset size $n$ is computationally prohibitive for many machine learning problems. In this work, we propose a new computational framework for entropic MOT, dubbed Neural Entropic MOT (NEMOT), that enjoys significantly improved scalability. NEMOT employs neural networks trained using mini-batches, which transfers the computational complexity from the dataset size to the size of the mini-batch, leading to substantial gains. We provide formal guarantees on the accuracy of NEMOT via non-asymptotic error bounds. We supplement these with numerical results that demonstrate the performance gains of NEMOT over Sinkhorn's algorithm, as well as extensions to neural computation of multimarginal entropic Gromov-Wasserstein alignment. In particular, orders-of-magnitude speedups are observed relative to the state-of-the-art, with a notable increase in the feasible number of samples and marginals. NEMOT seamlessly integrates as a module in large-scale machine learning pipelines, and can serve to expand the practical applicability of entropic MOT for tasks involving multimarginal data.

Figures

Figures reproduced from arXiv: 2506.00573 by the authors.

Figure 1
Figure 1. Cost graphs with |E| edges. 3 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. NEMOT algorithm and system architecture. The diagram shows how each marginal dataset [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. NEMOT estimation performance across different datasets. NEMOT consistently presents [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison of NEMOT and Sinkhorn runtime vs. dataset size ( [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: NEMOT scalability analysis, unlocking new regimes of feasible [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: NEMOT results - MNIST dataset. elements. This corresponds to estimating an MOT-based matching score between the underlying distributions of the datasets . As in our previous experiment, we divide the MNIST dataset into k = 10 subsets, each corresponding to a digit (lab…
Figure 7
Figure 7. Figure 7: NEMOT-based similarity score between MNIST image datasets. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: CIFAR dataset, ResNet18-based cost. linear dependence of the EMOT w.r.t. k is maintained, as expected. Furthermore, [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: MNIST dataset, consisting of 28 × 28 gray-pixel images with 10 labels. Let C ∈ R n k be the explicit cost tensor and (Ci ) k i=1 ⊂ R n×n be the corresponding pairwise cost matrices. We begin with the representation of the unregularized cost. We have the following: ⟨C, …
Figure 10
Figure 10. Figure 10: CIFAR10 images. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 47 canonical work pages

  1. [1]

    Optimal transport: old and new, volume 338

    Cédric Villani et al. Optimal transport: old and new, volume 338. Springer, 2009

  2. [2]

    Optimal transport for applied mathematicians

    Filippo Santambrogio. Optimal transport for applied mathematicians. Birkäuser, NY, 55(58- 63):94, 2015

  3. [3]

    Convolutional wasserstein distances: Efficient optimal transportation on geometric domains

    Justin Solomon, Fernando De Goes, Gabriel Peyré, Marco Cuturi, Adrian Butscher, Andy Nguyen, Tao Du, and Leonidas Guibas. Convolutional wasserstein distances: Efficient optimal transportation on geometric domains. ACM Transactions on Graphics (ToG), 34(4):1–11, 2015

  4. [4]

    Wasserstein auto- encoders

    Ilya Tolstikhin, Olivier Bousquet, Sylvain Gelly, and Bernhard Schoelkopf. Wasserstein auto- encoders. arXiv preprint arXiv:1711.01558, 2017

  5. [5]

    Wasserstein generative adversarial networks

    Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214–223. PMLR, 2017

  6. [6]

    Vector quantile regression: an optimal transport approach

    Guillaume Carlier, Victor Chernozhukov, and Alfred Galichon. Vector quantile regression: an optimal transport approach. 2016

  7. [7]

    Monge–kantorovich depth, quantiles, ranks and signs

    Victor Chernozhukov, Alfred Galichon, Marc Hallin, and Marc Henry. Monge–kantorovich depth, quantiles, ranks and signs. 2017

  8. [8]

    Optimal transport methods in economics

    Alfred Galichon. Optimal transport methods in economics. Princeton University Press, 2018

Show all 58 references
  1. [9]

    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

  2. [10]

    Density functional theory and opti- mal transportation with coulomb cost

    Codina Cotar, Gero Friesecke, and Claudia Klüppelberg. Density functional theory and opti- mal transportation with coulomb cost. Communications on Pure and Applied Mathematics , 66(4):548–599, 2013

  3. [11]

    Generalized incompressible flows, multi-marginal transport and Sinkhorn algorithm

    Jean-David Benamou, Guillaume Carlier, and Luca Nenna. Generalized incompressible flows, multi-marginal transport and Sinkhorn algorithm. Numerische Mathematik, 142:33–54, 2019

  4. [12]

    Multi- marginal wasserstein gan

    Jiezhang Cao, Langyuan Mo, Yifan Zhang, Kui Jia, Chunhua Shen, and Mingkui Tan. Multi- marginal wasserstein gan. Advances in Neural Information Processing Systems, 32, 2019

  5. [13]

    Scalable bayes via barycenter in Wasser- stein space

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

  6. [14]

    Distribution- ally robust linear and discrete optimization with marginals

    Louis Chen, Will Ma, Karthik Natarajan, David Simchi-Levi, and Zhenzhen Yan. Distribution- ally robust linear and discrete optimization with marginals. Operations Research, 70(3):1822– 1834, 2022

  7. [15]

    The multimarginal optimal transport formulation of adversarial multiclass classification

    Nicolás García Trillos, Matt Jacobs, and Jakwang Kim. The multimarginal optimal transport formulation of adversarial multiclass classification. Journal of Machine Learning Research, 24(45):1–56, 2023

  8. [16]

    Efficient discrete multi-marginal optimal transport regularization

    Ronak Mehta, Jeffery Kline, Vishnu Suresh Lokhande, Glenn Fung, and Vikas Singh. Efficient discrete multi-marginal optimal transport regularization. 2023

  9. [17]

    Contrasting multiple representa- tions with the multi-marginal matching gap

    Zoe Piran, Michal Klein, James Thornton, and Marco Cuturi. Contrasting multiple representa- tions with the multi-marginal matching gap. arXiv preprint arXiv:2405.19532, 2024

  10. [18]

    Hardness results for multimarginal optimal transport problems

    Jason M Altschuler and Enric Boix-Adsera. Hardness results for multimarginal optimal transport problems. Discrete Optimization, 42:100669, 2021

  11. [19]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013

  12. [20]

    Sample com- plexity of Sinkhorn divergences

    Aude Genevay, Lénaic Chizat, Francis Bach, Marco Cuturi, and Gabriel Peyré. Sample com- plexity of Sinkhorn divergences. In The 22nd international conference on artificial intelligence and statistics, pages 1574–1583. PMLR, 2019

  13. [21]

    Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem

    Gonzalo Mena and Jonathan Niles-Weed. Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem. Advances in neural information processing systems, 32, 2019

  14. [22]

    On the complexity of approximating multimarginal optimal transport

    Tianyi Lin, Nhat Ho, Marco Cuturi, and Michael I Jordan. On the complexity of approximating multimarginal optimal transport. Journal of Machine Learning Research, 23(65):1–43, 2022

  15. [23]

    Multi- marginal optimal transport and probabilistic graphical models

    Isabel Haasler, Rahul Singh, Qinsheng Zhang, Johan Karlsson, and Yongxin Chen. Multi- marginal optimal transport and probabilistic graphical models. IEEE Transactions on Informa- tion Theory, 67(7):4647–4668, 2021. 12

  16. [24]

    Accelerating the Sinkhorn algorithm for sparse multi-marginal optimal transport via fast fourier transforms

    Fatima Antarou Ba and Michael Quellmalz. Accelerating the Sinkhorn algorithm for sparse multi-marginal optimal transport via fast fourier transforms. Algorithms, 15(9):311, 2022

  17. [25]

    Polynomial-time algorithms for multimarginal optimal transport problems with structure

    Jason M Altschuler and Enric Boix-Adsera. Polynomial-time algorithms for multimarginal optimal transport problems with structure. Mathematical Programming, 199(1):1107–1178, 2023

  18. [26]

    Mutual information neural estimation

    Mohamed Ishmael Belghazi, Aristide Baratin, Sai Rajeshwar, Sherjil Ozair, Yoshua Bengio, Aaron Courville, and Devon Hjelm. Mutual information neural estimation. In International Conference on Machine Learning, pages 531–540. PMLR, 2018

  19. [27]

    Neural estimation of statistical divergences

    Sreejith Sreekumar and Ziv Goldfeld. Neural estimation of statistical divergences. Journal of Machine Learning Research, 23(126):1–75, 2022

  20. [28]

    Neural estimation and optimization of directed information over continuous spaces

    Dor Tsur, Ziv Aharoni, Ziv Goldfeld, and Haim Permuter. Neural estimation and optimization of directed information over continuous spaces. IEEE Transactions on Information Theory, 2023

  21. [29]

    Neural estimation of entropic optimal transport

    Tao Wang and Ziv Goldfeld. Neural estimation of entropic optimal transport. arXiv preprint arXiv:2405.06734, 2024

  22. [30]

    Gromov-wasserstein alignment of word embedding spaces

    David Alvarez-Melis and Tommi S Jaakkola. Gromov-wasserstein alignment of word embedding spaces. arXiv preprint arXiv:1809.00013, 2018

  23. [31]

    Gromov-wasserstein learning for graph matching and node embedding

    Hongteng Xu, Dixin Luo, Hongyuan Zha, and Lawrence Carin Duke. Gromov-wasserstein learning for graph matching and node embedding. In International conference on machine learning, pages 6932–6941. PMLR, 2019

  24. [32]

    Tree-based diffu- sion schrödinger bridge with applications to Wasserstein barycenters

    Maxence Noble, Valentin De Bortoli, Arnaud Doucet, and Alain Durmus. Tree-based diffu- sion schrödinger bridge with applications to Wasserstein barycenters. Advances in Neural Information Processing Systems, 36, 2024

  25. [33]

    Efficient and exact multimarginal optimal transport with pairwise costs

    Bohan Zhou and Matthew Parno. Efficient and exact multimarginal optimal transport with pairwise costs. Journal of Scientific Computing, 100(1):1–27, 2024

  26. [34]

    Large-scale optimal transport and mapping estimation

    Vivien Seguy, Bharath Bhushan Damodaran, Rémi Flamary, Nicolas Courty, Antoine Rolet, and Mathieu Blondel. Large-scale optimal transport and mapping estimation. arXiv preprint arXiv:1711.02283, 2017

  27. [35]

    Score-based generative neural networks for large- scale optimal transport

    Max Daniels, Tyler Maunu, and Paul Hand. Score-based generative neural networks for large- scale optimal transport. Advances in neural information processing systems, 34:12955–12965, 2021

  28. [36]

    Energy-guided entropic neural optimal transport

    Petr Mokrov, Alexander Korotin, Alexander Kolesov, Nikita Gushchin, and Evgeny Burnaev. Energy-guided entropic neural optimal transport. arXiv preprint arXiv:2304.06094, 2023

  29. [37]

    Multi-marginal Gromov–Wasserstein transport and barycentres

    Florian Beier, Robert Beinert, and Gabriele Steidl. Multi-marginal Gromov–Wasserstein transport and barycentres. Information and Inference: A Journal of the IMA, 12(4):2753–2781, 2023

  30. [38]

    Convergence rate of entropy-regularized multi-marginal optimal transport costs

    Luca Nenna and Paul Pegon. Convergence rate of entropy-regularized multi-marginal optimal transport costs. Canadian Journal of Mathematics, pages 1–21, 2023

  31. [39]

    A differential approach to the multi-marginal Schrödinger system

    Guillaume Carlier and Maxime Laborde. A differential approach to the multi-marginal Schrödinger system. SIAM Journal on Mathematical Analysis, 52(1):709–717, 2020

  32. [40]

    Diagonal equivalence to matrices with prescribed row and column sums

    Richard Sinkhorn. Diagonal equivalence to matrices with prescribed row and column sums. The American Mathematical Monthly, 74(4):402–405, 1967

  33. [41]

    Approximation by combinations of relu and squared relu ridge functions with ℓ1 and ℓ0 controls

    Jason M Klusowski and Andrew R Barron. Approximation by combinations of relu and squared relu ridge functions with ℓ1 and ℓ0 controls. IEEE Transactions on Information Theory, 64(12):7649–7656, 2018

  34. [42]

    Universal approximation bounds for superpositions of a sigmoidal function

    Andrew R Barron. Universal approximation bounds for superpositions of a sigmoidal function. IEEE Transactions on Information theory, 39(3):930–945, 1993

  35. [43]

    U-statistics: Theory and Practice

    A J Lee. U-statistics: Theory and Practice. Routledge, 2019

  36. [44]

    Gromov–Wasserstein distances and the metric approach to object matching

    Facundo Mémoli. Gromov–Wasserstein distances and the metric approach to object matching. Foundations of computational mathematics, 11:417–487, 2011

  37. [45]

    Manifold alignment for heterogeneous single-cell multi-omics data integration using pamona

    Kai Cao, Yiguang Hong, and Lin Wan. Manifold alignment for heterogeneous single-cell multi-omics data integration using pamona. Bioinformatics, 38(1):211–219, 2022. 13

  38. [46]

    Scot: single-cell multi-omics alignment with optimal transport

    Pinar Demetci, Rebecca Santorella, Björn Sandstede, William Stafford Noble, and Ritambhara Singh. Scot: single-cell multi-omics alignment with optimal transport. Journal of computational biology, 29(1):3–18, 2022

  39. [47]

    A spectral notion of gromov–wasserstein distance and related methods

    Facundo Mémoli. A spectral notion of gromov–wasserstein distance and related methods. Applied and Computational Harmonic Analysis, 30(3):363–401, 2011

  40. [48]

    Commander

    Clayton W. Commander. A survey of the quadratic assignment problem, with applications. Morehead Electronic Journal of Applicable Mathematics, 4:MATH–2005–01, 2005

  41. [49]

    Gromov– wasserstein distances: Entropic regularization, duality and sample complexity

    Zhengxin Zhang, Ziv Goldfeld, Youssef Mroueh, and Bharath K Sriperumbudur. Gromov– wasserstein distances: Entropic regularization, duality and sample complexity. The Annals of Statistics, 52(4):1616–1645, 2024

  42. [50]

    Entropic gromov-wasserstein distances: Stability and algorithms

    Gabriel Rioux, Ziv Goldfeld, and Kengo Kato. Entropic gromov-wasserstein distances: Stability and algorithms. Journal of Machine Learning Research, 25(363):1–52, 2024

  43. [51]

    Fused Gromov-Wasserstein distance for structured objects

    Titouan Vayer, Laetitia Chapel, Rémi Flamary, Romain Tavenard, and Nicolas Courty. Fused Gromov-Wasserstein distance for structured objects. Algorithms, 13(9):212, 2020

  44. [52]

    Non-asymptotic performance guaran- tees for neural estimation of f-divergences

    Sreejith Sreekumar, Zhengxin Zhang, and Ziv Goldfeld. Non-asymptotic performance guaran- tees for neural estimation of f-divergences. InInternational Conference on Artificial Intelligence and Statistics, pages 3322–3330. PMLR, 2021

  45. [53]

    Weak convergence

    Aad W Van Der Vaart, Jon A Wellner, Aad W van der Vaart, and Jon A Wellner. Weak convergence. Springer, 1996

  46. [54]

    Lower complexity adaptation for empirical entropic optimal transport

    Michel Groppe and Shayan Hundrieser. Lower complexity adaptation for empirical entropic optimal transport. Journal of Machine Learning Research, 25(344):1–55, 2024

  47. [55]

    A multivariate faa di bruno formula with applications

    Gregory Constantine and Thomas Savits. A multivariate faa di bruno formula with applications. Transactions of the American Mathematical Society, 348(2):503–520, 1996

  48. [56]

    Gromov- wasserstein distances: Entropic regularization, duality, and sample complexity

    Zhengxin Zhang, Ziv Goldfeld, Youssef Mroueh, and Bharath K Sriperumbudur. Gromov- wasserstein distances: Entropic regularization, duality, and sample complexity. arXiv preprint arXiv:2212.12848, 2022

  49. [57]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019

  50. [58]

    sup f ∈F mnn (a) ˆΓ( ˆf ⋆) − Γ( ˆf ⋆) # ≤ kX j=1 1√n E

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 14 Supplementary material for: Neural Estimation for Scaling Entropic Multimarginal Optimal Transport Table of contents: • Section A: Proofs of the main theorems, ...

Pith tools

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