Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Designing Algorithms for Entropic Optimal Transport from an Optimisation Perspective

T0 review · 3 major / 5 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 The paper derives a unified family of Sinkhorn-like algorithms for entropic optimal transport from a semi-dual optimisation viewpoint, with non-asymptotic $1/N$ and $1/N^2$ rates under minimal assumptions.

desk verdict Interesting Phi-match framework undermined by a false identity-kernel identification in Section 4.1; the 1/N rate for plain SGA is unsupported, though k-SGA and the projected methods may survive. read the letter →

arxiv 2507.12246 v1 pith:2QKXGEUL submitted 2025-07-16 math.OC math.PRstat.ML

Vishwak Srinivasan , Qijia Jiang This is my paper · ORCID
classification math.OCmath.PRstat.ML MSC 49Q2290C25
keywords entropicoptimaltransportSinkhornalgorithmsemi-dualmaximummeandiscrepancymirrordescentSchrödingerbridgenon-asymptoticconvergenceacceleratedoptimization
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

Entropic optimal transport is usually solved with a matrix-scaling algorithm whose classical non-asymptotic guarantees degrade exponentially as the regularisation parameter shrinks. This paper tries to establish that the problem can be redesigned from scratch as infinite-dimensional optimisation, maximising a semi-dual objective whose gradient is the mismatch between the Y-marginal of a structured coupling and the target distribution. From this viewpoint it defines a general update, $\Phi$-match, that contains the Sinkhorn iteration and plain semi-dual gradient ascent as two natural instantiations, and it shows every iterate remains in the family of couplings that can be completed to the entropic plan. This yields non-asymptotic rates for the Y-marginal, and hence for the coupling, under no assumptions on the marginals: $1/N$ for MMD-based, signed, and projected variants, and $1/N^2$ for an accelerated projected variant. The same machinery transfers to path measures, giving drift-update algorithms for the dynamical Schr\"odinger bridge problem.

What carries the argument

The central object is the $\Phi$-match update together with the semi-dual $J$ of entropic optimal transport. The semi-dual is built from the map $\phi_+(x)=\log\int_Y \exp((\phi(y)-c(x,y))/\varepsilon)\,d\nu(y)$, and the joint density $\pi(\phi,\phi_+)\propto \exp((\phi(y)-\phi_+(x)-c(x,y))/\varepsilon)\,\mu\otimes\nu$; the operator $\Phi$ selects which discrepancy between $\pi_Y$ and $\nu$ the method minimises. The machinery shows that $\Phi$-match is simultaneously an alternating projection, a local greedy root-finding step, and a mirror-descent step, which is what makes non-asymptotic convergence proofs possible.

What would settle it

On $\mathbb{R}$ with $\nu$ Lebesgue-absolutely continuous, compute $m_{k_{\mathrm{Id}}}(\nu)(y)=\int \mathbf{1}\{y=y'\}\nu(y')\,dy'$: the integrand is nonzero only on a set of measure zero, so the kernel mean is $0$ almost everywhere and the claimed equality $m_k(\xi)=\xi$ fails. A valid bounded positive-definite kernel returns a smoothed version of $\xi$, so the $1/N$ bound for plain SGA depends on a missing evaluation kernel.

Watch

Extended reading notes

Core claim

The central discovery is that entropic-optimal-transport algorithms are best understood not as scaling procedures but as iterates of an optimisation method on the semi-dual $J(\phi)=\int_Y \phi\,d\nu - \int_X \phi_+\,d\mu$, whose first variation is $\delta J(\phi)=\nu - \pi(\phi,\phi_+)_Y$. The proposed class $M_{\Phi\text{-match}}(\phi;\eta)=\phi-\eta(\log\Phi(\pi(\phi,\phi_+)_Y)-\log\Phi(\nu))$ keeps each joint distribution inside the structured set $Q$, so any method that brings the Y-marginal to $\nu$ automatically produces the optimal coupling. For a bounded positive-definite kernel $k$, kernelised SGA satisfies $L_k(\pi_N^Y;\nu)\le \max\{2c_k,1\}\,d_{\mathrm{KL}}(\pi^*\|\pi_0)/N$; under a cost-dependent smoothness condition, projected SGA and its accelerated version attain $O(1/N)$ and $O(1/N^2)$ semi-dual gaps. The same recursive update, lifted to path measures, gives a Schr\"odinger-bridge solver whose identity-operator special case is iterative proportional fitting.

Load-bearing premise

The load-bearing premise is that the identity kernel acts as an evaluation map on densities ($m_{k_{\mathrm{Id}}}(\xi)=\xi$), a property that holds for atomic measures but not for the continuous Lebesgue-density setting in which the paper states its guarantees.

Editorial extensions

If this is right

  • For any bounded positive-definite kernel, kernelised SGA converges in squared MMD at rate $O(1/N)$: $L_k(\pi_N^Y;\nu)\le \max\{2c_k,1\}d_{\mathrm{KL}}(\pi^*\|\pi_0)/N$, and since each iterate lies in $Q$, this drives the full coupling toward the entropic optimal plan.
  • Signed SGA and projected SGA maximise the semi-dual at rate $O(1/N)$, while accelerated projected SGA reaches $O(1/N^2)$ using only a cost-dependent constant rather than assumptions on the marginals.
  • The rate's dependence on the regularisation parameter is polynomial rather than the exponential contraction factor $e^{-\|c\|_\infty/\varepsilon}$ seen in classical Sinkhorn analyses, so the methods target the small-regularisation regime.
  • Path-space $\Phi$-match turns the same recursion into SDE drift updates, making the dynamical Schr\"odinger bridge problem solvable by the same optimisation template, with iterative proportional fitting as one parameter choice.

Reading between the lines

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

  • Editorial extension: The cleanest domain for the $1/N$ theorem is kernelised SGA with a bounded positive-definite kernel; the paper's remarks on Gaussian and Laplace kernels suggest a particle implementation, and one could test whether finite-sample kernelised SGA beats Sinkhorn when $\varepsilon$ is small and costs are ill-conditioned.
  • Editorial extension: Accelerated projected SGA demonstrates the benefit of staying in the dual; an implicit corollary is that primal momentum on the non-convex set $Q$ may control only marginals, not couplings, because convex combinations of iterates can leave $Q$.
  • Editorial extension: Replacing $\Phi$ in path-$\Phi$-match yields a one-parameter family of Schr\"odinger-bridge samplers; a natural experiment is to compare $\Phi=\exp$ against $\Phi=\mathrm{id}$ on Brownian bridges to see whether the MMD-style drift correction converges faster in practice.
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

3 major / 5 minor

Summary. The paper develops a unified optimisation perspective on entropic optimal transport (eOT), introducing a class of methods called Phi-match that interpolates between SGA (semi-dual gradient ascent) and Sinkhorn. It then derives several interpretations: alternating projections on the set Q of joint distributions with prescribed X-marginal, a local greedy update, and a mirror-descent viewpoint. The main theoretical results are a non-asymptotic O(1/N) guarantee for a kernelised variant k-SGA in squared maximum mean discrepancy (Theorem 1), O(1/N) and O(1/N^2) guarantees for a projected semi-dual gradient ascent and its accelerated version (Theorems 3 and 4), and a O(1/N) rate for a signed semi-dual ascent (Theorem 2). The framework is extended to the dynamical Schroedinger bridge problem, with a path-space analogue of Phi-match and an SDE-drift implementation. The paper is primarily theoretical, with no numerical experiments, and relies on imported lemmas from prior work on mirror descent and Sinkhorn.

Significance. If the main results were fully correct, the paper would make a useful contribution by providing non-asymptotic convergence guarantees for a family of eOT algorithms that do not require strict tail or log-concavity assumptions on the marginals. The k-SGA theorem and the proj-SGA/acceleration proofs follow standard templates and appear internally consistent. The framework productively unifies primal projections, local greedy updates, and mirror-descent views of Sinkhorn, and the extension to path-space Schr\"odinger bridges is conceptually attractive. The paper also gives clear attribution to prior work and does not rely on fitted constants or circular reasoning. However, a central advertised claim---that plain SGA inherits the O(1/N) MMD rate via an identity kernel---is based on a false identification in Section 4.1, and Theorem 2 contains an unproved finiteness assumption. These issues materially weaken the paper's headline contributions as currently stated.

major comments (3)
  1. [Section 4.1] The claim that the identity kernel k_Id(y,y') = 1 iff y = y' satisfies m_k(\xi)(y) = \xi(y) is false for atomless probability measures on a continuous space. For any Lebesgue-absolutely-continuous measure \xi, the integral \int k_Id(y,y') d\xi(y') equals \xi({y}) = 0 for Lebesgue-almost every y, not the density \xi(y). The kernel as written is the Kronecker-delta kernel on an uncountable set, not the Dirac-delta distribution needed for point evaluation of densities. Consequently, the identity kernel is not characteristic on P(Y), the map V_Phi for SGA does not coincide with the first variation of L_{k_Id}, and plain SGA is not a special case of k-SGA. The sentence following Theorem 1 claiming that SGA inherits the 1/N rate is therefore unsupported. This is load-bearing because the advertised connection between SGA and MMD is the motivation for introducing k-SGA.
  2. [Theorem 2] The theorem states a rate depending on diam(T_{\phi_0,y_{\rm anc}}; L^\infty(Y)), but it never proves that this diameter is finite. The set T is a superlevel set of the semi-dual J intersected with the anchoring condition \phi(y_{\rm anc})=0, and without additional assumptions nothing prevents J(\phi) \ge J(\phi_0) from holding for functions of arbitrarily large L^\infty norm. If the diameter is infinite, the displayed bound is vacuous and the claimed non-asymptotic rate for sign-SGA is not meaningful. The proof in Appendix A.3.1 uses diam in a Young-inequality step without verifying its finiteness. The theorem needs either a proof that T has finite L^\infty diameter under the stated assumptions or an explicit assumption to that effect.
  3. [Appendix B, Lemma 13] The continuous-time acceleration result in Appendix B again uses the same invalid identity kernel k(y,y') = 1 iff y = y'. For atomless marginals, L_k(\hat\pi_t^Y, \nu) is identically zero under this kernel, so the claimed O(1/t^2) decay does not establish convergence of the Y-marginal in any meaningful metric. This appendix should be revised to use a bounded characteristic kernel or be clearly marked as heuristic.
minor comments (5)
  1. [Section 3.1.1] The two observations labelled 'Fact 1' and 'Fact 2' are unnumbered; numbering them would make cross-referencing in the proofs easier.
  2. [Section 4.1] The sentence listing 'the identity, Gaussian, and Laplace kernels' as examples of characteristic kernels should be corrected: the identity kernel is not characteristic on uncountable spaces, and its mean embedding does not act on densities in the claimed way.
  3. [Abstract and Introduction] The phrase 'under minimal assumptions' is used prominently, but Theorem 1 requires a bounded kernel, Theorem 3 requires bounded cost with \lambda(B)<\infty, and Theorem 2 requires (at least implicitly) a finite L^\infty diameter. The paper should qualify these statements carefully.
  4. [Theorem 1] After Theorem 1, the paper says that for characteristic kernels '\pi_n^Y approaches \nu'. Strictly, the theorem gives convergence of L_k(\pi_N^Y,\nu) to zero; if the kernel is characteristic this implies convergence in the MMD metric, and the paper should spell out that this is the intended sense.
  5. [Section 5.2] The path-space extension is described with high-level statements; Proposition 3 and Lemma 7 rely on external results (Girsanov, h-transforms, time reversal) without a full derivation of all integrability conditions. A short remark on the required regularity would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation is self-contained, with key analytic lemmas imported from independent prior work; the identity-kernel issue in §4.1 is a mathematical error, not a circular reduction.

full rationale

The paper's central claim is an algorithm-design and convergence-rate argument, not a fitted prediction. The Phi-match update is derived by direct computation from the semi-dual and the Sinkhorn/SGA formulas (Section 3.2), and the projection/greedy interpretations in Lemmas 3-5 are proven algebraically rather than assumed. Theorem 1 for k-SGA relies on Proposition 1, which is explicitly attributed to Aubin-Frankowski et al. (2022), an independent external work, and the proof in Appendix A.2.3 is a standard three-point identity plus relative-smoothness argument; no constant is fitted to data and no target quantity is used as an input. Theorems 2-4 follow from Lemmas 1, 2, 6, and 8, all proved in the appendix from first variation calculations, with no self-citation chain carrying the load. The path-space extension builds on independently attributed results such as Reza Karimi et al. (2024, Thm. 4.2) and Rueschendorf (1995), again without fitting. The most serious substantive issue is Section 4.1's identification of the kernel k_Id(y,y') = 1 iff y = y' with the property m_k(xi)(y) = xi(y). For continuous densities with respect to Lebesgue measure, this is false: m_k_Id(xi)(y) = xi({y}) = 0 for Lebesgue-almost every y, so plain SGA is not a genuine instance of k-SGA with that kernel and the advertised 1/N guarantee for SGA does not follow from Theorem 1 as stated. That is an invalid specialization or correctness bug, not a self-definition, fitted-input prediction, or self-citation reduction. The k-SGA rate for genuinely bounded positive-definite kernels remains independently derived, and the paper does not rename a known result as new; it generalizes and recovers earlier analyses. Accordingly, no circular step is present.

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

No free parameters are fitted to data. The central rates depend on standard domain assumptions and on one ad hoc kernel identification in Section 4.1 that is not valid on continuous spaces. The remaining assumptions are explicit in the theorems.

assumptions (6)
  • domain assumption mu and nu have densities with respect to the Lebesgue measure
    Imposed at the start of Section 3.1.1 to make density notation valid; excludes atomic marginals.
  • domain assumption Theorem 1 requires a bounded positive-definite kernel with c_k < inf
    Assumed in Proposition 1 and Theorem 1; the Gaussian and Laplace kernels satisfy this, the identity kernel does not.
  • domain assumption Lemma 6 and Theorems 3 and 4 require nonnegative cost with lambda(B) < inf, and Proposition 2 requires bounded cost ||c||_inf < inf
    Needed for the L2(nu) smoothness and for choosing B = 3||c||_inf / 2.
  • ad hoc to paper The identity kernel k_Id defined by 1 iff y=y' is treated as a bounded positive-definite kernel on continuous Y with m_k(xi)=xi
    Asserted in Section 4.1; mathematically unsupported and false for Lebesgue densities, so this is a load-bearing assumption.
  • domain assumption Existence of an optimal coupling with dKL(pi* || pi_ref) < inf
    Needed for the rate bounds in Theorem 1; the paper notes the bound is meaningful only when the initial KL is finite.
  • domain assumption Finite L_inf diameter of the anchored superlevel set T_{phi0,yanc} in Theorem 2
    The rate in Theorem 2 is vacuous if diam(T; L_inf) = inf; finiteness is not proved or stated as a condition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Designing Algorithms for Entropic Optimal Transport from an Optimisation Perspective." pith.science (2026). https://pith.science/paper/2QKXGEUL

@misc{pith2026250712246,
  author       = {Pith},
  title        = {Pith review of: Designing Algorithms for Entropic Optimal Transport from an Optimisation Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2QKXGEUL}},
  note         = {Machine review of arXiv:2507.12246}
}
read the original abstract

In this work, we develop a collection of novel methods for the entropic-regularised optimal transport problem, which are inspired by existing mirror descent interpretations of the Sinkhorn algorithm used for solving this problem. These are fundamentally proposed from an optimisation perspective: either based on the associated semi-dual problem, or based on solving a non-convex constrained problem over subset of joint distributions. This optimisation viewpoint results in non-asymptotic rates of convergence for the proposed methods under minimal assumptions on the problem structure. We also propose a momentum-equipped method with provable accelerated guarantees through this viewpoint, akin to those in the Euclidean setting. The broader framework we develop based on optimisation over the joint distributions also finds an analogue in the dynamical Schr\"{o}dinger bridge problem.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 43 canonical work pages

  1. [1]

    Near-linear time approximation algorithms for optimal transport via S inkhorn iteration

    Jason Altschuler, Jonathan Niles-Weed, and Philippe Rigollet. Near-linear time approximation algorithms for optimal transport via S inkhorn iteration. In Advances in Neural Information Processing Systems, volume 30, 2017

    work page 2017

  2. [2]

    urich. Birkh\

    Luigi Ambrosio, Nicola Gigli, and Giuseppe Savar\'e. Gradient flows in metric spaces and in the space of probability measures. Lectures in Mathematics ETH Z\"urich. Birkh\"auser Verlag, Basel, second edition, 2008

    work page 2008

  3. [3]

    Mirror D escent with R elative S moothness in M easure S paces, with application to S inkhorn and EM

    Pierre-Cyril Aubin-Frankowski, Anna Korba, and Flavien L\' e ger. Mirror D escent with R elative S moothness in M easure S paces, with application to S inkhorn and EM . In Advances in Neural Information Processing Systems, volume 35, pages 17263--17275, 2022

    work page 2022

  4. [4]

    A fast iterative shrinkage-thresholding algorithm for linear inverse problems

    Amir Beck and Marc Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM Journal on Imaging Sciences, 2 0 (1): 0 183--202, 2009

    2009

  5. [5]

    Schr\" o dinger Bridge Samplers

    Espen Bernton, Jeremy Heng, Arnaud Doucet, and Pierre E Jacob. Schr\" o dinger Bridge Samplers . arXiv preprint arXiv:1912.13170, 2019

    arXiv 1912

  6. [6]

    Caluya and Abhishek Halder

    Kenneth F. Caluya and Abhishek Halder. Wasserstein P roximal A lgorithms for the S chr\" o dinger B ridge P roblem: D ensity C ontrol W ith N onlinear D rift. IEEE Transactions on Automatic Control, 67 0 (3): 0 1163--1178, 2022

    work page 2022

  7. [7]

    On the linear convergence of the multimarginal sinkhorn algorithm

    Guillaume Carlier. On the linear convergence of the multimarginal sinkhorn algorithm. SIAM Journal on Optimization, 32 0 (2): 0 786--794, 2022

    work page 2022

  8. [8]

    Vixie, and Erik M

    Rick Chartrand, Brendt Wohlberg, Kevin R. Vixie, and Erik M. Bollt. A gradient descent solution to the M onge- K antorovich problem. Applied Mathematical Sciences, 3 0 (21-24): 0 1071--1080, 2009

    work page 2009

Show all 49 references
  1. [9]

    A semiconcavity approach to stability of entropic plans and exponential convergence of Sinkhorn's algorithm

    Alberto Chiarini, Giovanni Conforti, Giacomo Greco, and Luca Tamanini. A semiconcavity approach to stability of entropic plans and exponential convergence of Sinkhorn's algorithm . arXiv preprint arXiv:2412.09235, 2024

    2024

  2. [10]

    Faster W asserstein distance estimation with the S inkhorn divergence

    L\' e na\" c Chizat, Pierre Roussillon, Flavien L\' e ger, Fran c ois-Xavier Vialard, and Gabriel Peyr\' e . Faster W asserstein distance estimation with the S inkhorn divergence. In Advances in Neural Information Processing Systems, volume 33, pages 2257--2269. Curran Associa...

    2020

  3. [11]

    Sharper Exponential Convergence Rates for Sinkhorn's Algorithm in Continuous Settings

    L \'e na \" c Chizat, Alex Delalande, and Tomas Va s kevi c ius. Sharper Exponential Convergence Rates for Sinkhorn's Algorithm in Continuous Settings . arXiv preprint arXiv:2407.01202, 2024

    2024 arXiv

  4. [12]

    Projected Langevin dynamics and a gradient flow for entropic optimal transport

    Giovanni Conforti, Daniel Lacker, and Soumik Pal. Projected Langevin dynamics and a gradient flow for entropic optimal transport . arXiv preprint arXiv:2309.08598, 2023

    2023 arXiv

  5. [13]

    Sinkhorn distances: L ightspeed C omputation of O ptimal T ransport

    Marco Cuturi. Sinkhorn distances: L ightspeed C omputation of O ptimal T ransport. In Advances in Neural Information Processing Systems, volume 26, 2013

    2013

  6. [14]

    Semidual regularized optimal transport

    Marco Cuturi and Gabriel Peyr\'e. Semidual regularized optimal transport. SIAM Review, 60 0 (4): 0 941--965, 2018

    2018

  7. [15]

    A stochastic control approach to reciprocal diffusion processes

    Paolo Dai Pra. A stochastic control approach to reciprocal diffusion processes . Applied mathematics and Optimization, 23 0 (1): 0 313--329, 1991

    1991

  8. [16]

    Wasserstein mirror gradient flow as the limit of the Sinkhorn algorithm

    Nabarun Deb, Young-Heon Kim, Soumik Pal, and Geoffrey Schiebinger. Wasserstein mirror gradient flow as the limit of the Sinkhorn algorithm . arXiv preprint arXiv:2307.16421, 2023

    2023 arXiv

  9. [17]

    An optimal transport approach for the S chr\"odinger bridge problem and convergence of S inkhorn algorithm

    Simone Di Marino and Augusto Gerolin. An optimal transport approach for the S chr\"odinger bridge problem and convergence of S inkhorn algorithm. Journal of Scientific Computing, 85 0 (2), 2020

    2020

  10. [18]

    Computational O ptimal T ransport: C omplexity by A ccelerated G radient D escent I s B etter T han by S inkhorn’s A lgorithm

    Pavel Dvurechensky, Alexander Gasnikov, and Alexey Kroshnin. Computational O ptimal T ransport: C omplexity by A ccelerated G radient D escent I s B etter T han by S inkhorn’s A lgorithm. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Pro...

    2018

  11. [19]

    Interpolating between O ptimal T ransport and MMD using sinkhorn divergences

    Jean Feydy, Thibault S\' e journ\' e , Fran c ois-Xavier Vialard, Shun-ichi Amari, Alain Trouve, and Gabriel Peyr\' e . Interpolating between O ptimal T ransport and MMD using sinkhorn divergences. In Proceedings of the Twenty-Second International Conference on Artificial Inte...

    2019

  12. [20]

    Random fields and diffusion processes

    Hans F \"o llmer. Random fields and diffusion processes. In \'Ecole d'\'Et\'e de P robabilit\'es de S aint- F lour XV -- XVII , 1985--87 , volume 1362 of Lecture Notes in Math., pages 101--203. Springer, Berlin, 1988

    1985

  13. [21]

    On the scaling of multidimensional matrices

    Joel Franklin and Jens Lorenz. On the scaling of multidimensional matrices. Linear Algebra and its Applications, 114-115: 0 717--735, 1989

    1989

  14. [22]

    Dimensionality reduction for supervised learning with reproducing kernel hilbert spaces

    Kenji Fukumizu, Francis R Bach, and Michael I Jordan. Dimensionality reduction for supervised learning with reproducing kernel hilbert spaces. Journal of Machine Learning Research, 5 0 (Jan): 0 73--99, 2004

    2004

  15. [23]

    Wilfrid Gangbo and Robert J. McCann. The geometry of optimal transportation. Acta Math., 177 0 (2): 0 113--161, 1996

    1996

  16. [24]

    Stochastic O ptimization for L arge-scale O ptimal T ransport

    Aude Genevay, Marco Cuturi, Gabriel Peyr\' e , and Francis Bach. Stochastic O ptimization for L arge-scale O ptimal T ransport. In Advances in Neural Information Processing Systems, volume 29, 2016

    2016

  17. [25]

    Sample C omplexity of S inkhorn D ivergences

    Aude Genevay, L\' e na\" i c Chizat, Francis Bach, Marco Cuturi, and Gabriel Peyr\' e . Sample C omplexity of S inkhorn D ivergences. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, volume 89 of Proceedings of Machine Lea...

    2019

  18. [26]

    A kernel method for the two-sample-problem

    Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Sch\" o lkopf, and Alex Smola. A kernel method for the two-sample-problem. In Advances in Neural Information Processing Systems, volume 19. MIT Press, 2006

    2006

  19. [27]

    A fast approach to optimal transport: the back-and-forth method

    Matt Jacobs and Flavien L\'eger. A fast approach to optimal transport: the back-and-forth method. Numerische Mathematik, 146 0 (3): 0 513--544, 2020

    2020

  20. [28]

    Accelerated M irror D escent in C ontinuous and D iscrete T ime

    Walid Krichene, Alexandre Bayen, and Peter L Bartlett. Accelerated M irror D escent in C ontinuous and D iscrete T ime. In Advances in Neural Information Processing Systems, volume 28, 2015

    2015

  21. [29]

    A gradient descent perspective on S inkhorn

    Flavien L\'eger. A gradient descent perspective on S inkhorn. Applied Mathematics and Optimization, 84 0 (2): 0 1843--1855, 2021

    2021

  22. [30]

    A survey of the S chr\"odinger problem and some of its connections with optimal transport

    Christian L\'eonard. A survey of the S chr\"odinger problem and some of its connections with optimal transport. Discrete and Continuous Dynamical Systems. Series A, 34 0 (4): 0 1533--1574, 2014

    2014

  23. [31]

    Tianyi Lin, Nhat Ho, and Michael I. Jordan. On the E fficiency of E ntropic R egularized A lgorithms for O ptimal T ransport. Journal of Machine Learning Research, 23 0 (137): 0 1--42, 2022

    2022

  24. [32]

    A kernelized Stein discrepancy for goodness-of-fit tests

    Qiang Liu, Jason Lee, and Michael Jordan. A kernelized Stein discrepancy for goodness-of-fit tests . In International conference on machine learning, pages 276--284. PMLR, 2016

    2016

  25. [33]

    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. In Advances in Neural Information Processing Systems, volume 32, 2019

    2019

  26. [34]

    Online S inkhorn: O ptimal T ransport distances from sample streams

    Arthur Mensch and Gabriel Peyr\' e . Online S inkhorn: O ptimal T ransport distances from sample streams. In Advances in Neural Information Processing Systems, volume 33, pages 1657--1667, 2020

    2020

  27. [35]

    Sinkhorn algorithm as a special case of stochastic mirror descent

    Konstantin Mishchenko. Sinkhorn algorithm as a special case of stochastic mirror descent. arXiv preprint arXiv:1909.06918, 2019

    1909 arXiv

  28. [36]

    Sobolev D escent

    Youssef Mroueh, Tom Sercu, and Anant Raj. Sobolev D escent. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, volume 89 of Proceedings of Machine Learning Research, pages 2976--2985, 2019

    2019

  29. [37]

    A. S. Nemirovski and D. B. Yudin. Problem complexity and method efficiency in optimization. Wiley-Interscience Series in Discrete Mathematics. John Wiley & Sons, Inc., New York, 1983. Translated from the Russian and with a preface by E. R. Dawson, A Wiley-Interscience Publication

    1983

  30. [38]

    Lectures on C onvex O ptimization , volume 137 of Springer Optimization and Its Applications

    Yurii Nesterov. Lectures on C onvex O ptimization , volume 137 of Springer Optimization and Its Applications. Springer, Cham, second edition, 2018

    2018

  31. [39]

    Introduction to Entropic Optimal Transport

    Marcel Nutz. Introduction to Entropic Optimal Transport . Lecture notes, Columbia University, 2021

    2021

  32. [40]

    Entropic optimal transport: convergence of potentials

    Marcel Nutz and Johannes Wiesel. Entropic optimal transport: convergence of potentials. Probability Theory and Related Fields, 184 0 (1-2): 0 401--424, 2022

    2022

  33. [41]

    Pavliotis

    Grigorios A. Pavliotis. Stochastic processes and applications, volume 60 of Texts in Applied Mathematics. Springer, New York, 2014. Diffusion processes, the Fokker-Planck and Langevin equations

    2014

  34. [42]

    Computational optimal transport: With applications to data science

    Gabriel Peyr\' e and Marco Cuturi. Computational optimal transport: With applications to data science. Foundations and Trends® in Machine Learning, 11: 0 355--607, 2019

    2019

  35. [43]

    Sinkhorn Flow as Mirror Flow: A Continuous-Time Framework for Generalizing the S inkhorn Algorithm

    Mohammad Reza Karimi, Ya-Ping Hsieh, and Andreas Krause. Sinkhorn Flow as Mirror Flow: A Continuous-Time Framework for Generalizing the S inkhorn Algorithm . In Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, volume 238 of Proceeding...

    2024

  36. [44]

    Real and complex analysis, 3rd ed

    Walter Rudin. Real and complex analysis, 3rd ed. McGraw-Hill, Inc., USA, 1987. ISBN 0070542341

    1987

  37. [45]

    Convergence of the iterative proportional fitting procedure

    Ludger R \"u schendorf. Convergence of the iterative proportional fitting procedure. The Annals of Statistics, pages 1160--1174, 1995

    1995

  38. [46]

    Optimal transport for applied mathematicians, volume 87 of Progress in Nonlinear Differential Equations and their Applications

    Filippo Santambrogio. Optimal transport for applied mathematicians, volume 87 of Progress in Nonlinear Differential Equations and their Applications. Birkh\"auser/Springer, Cham, 2015. Calculus of variations, PDEs, and modeling

    2015

  39. [47]

    Concerning nonnegative matrices and doubly stochastic matrices

    Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21 0 (2): 0 343--348, 1967

    1967

  40. [48]

    Support vector machines

    Ingo Steinwart and Andreas Christmann. Support vector machines. Information Science and Statistics. Springer, New York, 2008

    2008

  41. [49]

    Topics in optimal transportation, volume 58 of Graduate Studies in Mathematics

    C\'edric Villani. Topics in optimal transportation, volume 58 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2003

    2003

Pith tools

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