Pith. sign in

REVIEW 4 major objections 4 minor 44 references

Symmetrized Sinkhorn-Gibbs Inference for Oscillatory Inverse Problems

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that comparing a signed signal and its negation through a symmetrized Sinkhorn divergence gives Gibbs posteriors with smoother risk landscapes, stronger noise robustness, and better population recovery than Euclidean or tra

desk verdict The symmetrized Sinkhorn construction is a genuinely useful empirical idea, but the robustness theorems rely on an unproved H^{-1} Lipschitz step, so the paper's theoretical claims are conditional; worth peer review but needs a fix. read the letter →

arxiv 2607.26264 v1 pith:KK2XGVQH submitted 2026-07-28 math.OC

Gabriel Huerta , Mohammad Motamed This is my paper
classification math.OC MSC 49Q2262F15
keywords oscillatoryinverseproblemsGibbsposteriorSinkhorndivergenceoptimaltransportcycleskippinguncertaintyquantificationseismicinversionsignedsignalnormalization
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 tries to establish that for oscillatory inverse problems, where pointwise losses suffer from cycle skipping and spurious local minima, a Gibbs posterior built on a symmetrized Sinkhorn divergence—comparing a softplus-normalized signal and its normalized negation—produces more reliable uncertainty quantification than Euclidean or trace-wise Wasserstein losses. It proves that the loss is smooth and positive definite, that the resulting Gibbs posterior is well defined, and that the posterior is robust to perturbations of both the data and the forward model when those perturbations are measured in the negative Sobolev norm. Numerical experiments on a seismic benchmark show the symmetrized loss yields less oscillatory risk landscapes, more accurate posterior inference, and population-level recovery close to the oracle distribution.

What carries the argument

The key object is the symmetrized normalized squared Sinkhorn loss D^2_{σ,ε}(f,g) = S^2_{σ,ε}(f,g) + S^2_{σ,ε}(-f,-g), where S_{σ,ε} is the Sinkhorn divergence—entropically regularized optimal transport—applied to softplus-normalized zero-mean functions. The softplus map σ_δ(z)=ln(1+e^{δz}) turns signed oscillations into positive probability densities, and the negation term captures the complementary negative phase. This loss supplies the empirical risk in a Gibbs posterior Π(dθ) ∝ exp(-λ Φ(θ;g)) Π0(dθ), and its smoothness and Lipschitz bounds in the negative Sobolev norm carry the well-definedness and robustness theorems.

What would settle it

Take a one-dimensional oscillatory signal with a deep negative trough such that δ times the maximum negative amplitude is large, compute the softplus-normalized density on a fine grid, and check whether its minimum value falls below, say, 1e-8; if so, the positive-lower-bound assumption of Proposition 5 fails, and a direct numerical test of the claimed Hellinger-in-H^{-1} bound should show the linear robustness estimate is violated for suitable small perturbations.

Watch

Extended reading notes

Core claim

The central claim is that the nonconvexity of oscillatory inverse problems can be tamed by comparing both a signed signal and its negation in transport space. After mean-centering, the paper maps a signal and its negation through a softplus normalization to probability densities; the loss is the sum of the squared Sinkhorn divergences between normalized prediction and data and between normalized negated prediction and negated data. This symmetrization preserves information from both phases of oscillation, restores asymptotic convexity as the softplus sharpness tends to infinity, and makes the Gibbs posterior Lipschitz stable in the negative Sobolev norm. The paper argues that these propertie

Load-bearing premise

The load-bearing premise, flagged by the paper only for asymptotic convexity in Remark 3, is that softplus-normalized densities stay bounded above and below by positive constants, which is not guaranteed for signals with deep negative or positive excursions because normalized mass can approach zero exponentially.

Editorial extensions

If this is right

  • If correct, the framework provides a model-agnostic empirical risk that can be coupled with any forward model, since the loss itself does not depend on the PDE or simulator structure.
  • Posterior accuracy should improve rather than stagnate as data resolution increases, because the symmetrized loss avoids the wrong-mode concentration seen with Euclidean risks.
  • The H^{-1} robustness bounds imply that low-frequency perturbations of the data or forward model have controlled influence on the posterior in Hellinger distance, which is directly relevant to uncertainty quantification under model error.
  • Population-level inference from multiple noisy events should track the oracle empirical distribution closely, enabling recovery of the underlying parameter distribution.
  • The adaptive sampler treats the inverse temperature as an unknown, so the method can be deployed without manual calibration of the learning rate in multi-parameter settings.

Reading between the lines

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

  • A natural extension the paper leaves implicit is that symmetrization could transfer to any signed feature data where both phases carry information, such as medical imaging or quantum state population measurements, not just seismic wavefields.
  • The asymptotic convexity result suggests an adaptive annealing strategy for the softplus sharpness δ during MCMC—starting at moderate δ and increasing it—which the paper does not explore.
  • The paper's H^{-1} framing indicates the method is best suited to low-frequency or smooth model errors; testing on high-frequency model perturbations would probe the boundary of the stated guarantee.
  • A testable practical extension is automatic calibration of the Sinkhorn regularization ε and softplus sharpness δ, which are currently set manually in the experiments.
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

4 major / 4 minor

Summary. The paper proposes a Gibbs-posterior inference framework for oscillatory inverse problems, built on a symmetrized Sinkhorn divergence. Signed signals are mean-centered, mapped to probability densities via a softplus-type normalization, and compared through the sum of Sinkhorn divergences of the normalized signals and their negations. The authors claim smoothness, asymptotic convexity, well-definedness of the Gibbs posterior, and H^{-1}-Lipschitz robustness to data and forward-model perturbations. Numerical experiments on a synthetic seismic benchmark compare the proposed loss with Euclidean and trace-wise Wasserstein losses, reporting smoother risk landscapes, more accurate posterior means, and better population-level recovery.

Significance. If the theory and experiments hold, the framework would be a useful contribution: it is forward-model-agnostic, provides a principled way to apply transport-based divergences to signed oscillatory data, and includes an adaptive MCMC scheme that learns the inverse temperature. The numerical study is carefully structured, with single-event, noise-robustness, and population-level experiments, and the results are consistent with the claimed qualitative advantages. However, the theoretical robustness results are presently conditional on assumptions that are neither verified for the chosen normalization nor shown to hold for the benchmark signals. The paper does not provide code or data, and the benchmark is synthetic. The central contribution is therefore promising but not yet fully supported at the level claimed in the abstract.

major comments (4)
  1. [Section 4.1, Proposition 4] The proof of the second inequality asserts that Tσ is locally Lipschitz in L∞ and then concludes a Lipschitz bound in Ḣ^{-1}. This does not follow: Lipschitz continuity in L∞ controls pairings with L1 test functions, and hence gives a bound in terms of ||f−f'||_{L1}, not the weaker Ḣ^{-1} norm. Multiplication by σ'(f) is not in general a bounded operator on Ḣ^{-1} without additional regularity on f. Since Proposition 5, Corollary 2, and Theorems 2–3 all rely on this step, the robustness estimates are not established as written.
  2. [Section 4.3, Proposition 5] The assumption that the normalized densities Tσf, Tσg, Tσg' are bounded below by a positive constant is not guaranteed for the softplus function σδ(z)=log(1+e^{δz}) used throughout the experiments. For signals taking sufficiently negative values on parts of the domain, Tσf is exponentially small in δ and can be arbitrarily close to zero. The W2-to-Ḣ^{-1} equivalence from [44] requires such lower bounds, and the constant Cε in Proposition 5 depends on them. The manuscript should either impose and verify a quantitative lower-bound condition for the chosen normalization, or restrict the robustness claims to regimes where that condition provably holds for the benchmark signals.
  3. [Theorems 2–3 and Corollary 2] The Hellinger robustness estimates are presented as unconditional guarantees, but they inherit the gaps in Proposition 4 and Proposition 5. In particular, Theorem 3 applies Corollary 2 with the first argument perturbed; this is formally justified by symmetry of D^2, but it still depends on the same unproved Ḣ^{-1} Lipschitz step. The manuscript should explicitly state the regularity and density lower-bound conditions as hypotheses of Theorems 2–3 and discuss whether they are satisfied by typical oscillatory signals. Otherwise the abstract's claim of 'greater robustness to observational noise' is supported only empirically.
  4. [Propositions 2–3 and Remark 3] The convexity results are asymptotic as δ→∞, and Remark 3 correctly concedes that uniform convexity may fail for finite δ. This is not a logical error, but it weakens the theoretical support for the claimed 'favorable convexity behavior' in the abstract and Contribution 3. The manuscript should more clearly state that finite-δ convexity is not established and that the reduced spurious-minima behavior rests on numerical evidence.
minor comments (4)
  1. [Data availability] The paper states that code and data are available from the corresponding author upon request and will be made public upon acceptance. For a numerical paper whose central claims are empirical, a permanent repository link at submission time would materially aid reproducibility.
  2. [Section 5.1] The benchmark is synthetic and designed by the authors, and no comparison is made with a fully Bayesian likelihood-based posterior. This should be acknowledged as a limitation in Section 5.1 or the conclusions.
  3. [Notation, Eq. (10) vs. Eq. (17)] The discrete loss is denoted Sσ and the continuous loss Sσ,ε; the dependence on ε in the discrete setting is implicit (appearing in the Sinkhorn divergence definition). A brief remark would help avoid confusion.
  4. [Section 4.1, proof of Proposition 2] The pointwise limit of σδ(f) for f<0 is stated as 0, which is correct, but the convergence is not uniform near f=0. This nonuniformity is the source of the Remark 3 caveat and deserves a sentence in the proof rather than only in the remark.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the construction is built from standard external transport results and explicit tuning choices, not from its own conclusions.

full rationale

The derivation chain is self-contained. The symmetrized Sinkhorn-Gibbs loss is explicitly constructed from a softplus normalization, the Sinkhorn divergence, and a symmetrization over sign flips (Eqs. 9, 12, 17, 18). Its smoothness, positive definiteness, and asymptotic convexity are attributed to external results on Sinkhorn divergences [30], and the robustness estimates rely on external Wasserstein-to-H-1 comparison results [44] and standard Bayesian stability arguments [1,3]. No parameter is fitted to a subset of data and then renamed as a prediction: the hyperparameters δ, ε, and λ are either calibrated or jointly inferred and are openly discussed as tuning/learning parameters in Sections 3.3 and 5.4. The self-citations (Motamed & Appelö 2019; Motamed 2020) are background citations for prior transport-based inversion and computational acceleration; they are not load-bearing for the paper's central claims. The reviewer-identified gap in Proposition 4 — the step from L8 Lipschitzness of the normalization to H-1 Lipschitzness — is a possible mathematical correctness issue in the proof, but it is not circularity: the paper does not define the loss in terms of the claimed robustness bound, and the bound is not an assumption smuggled into the construction. Accordingly, no circular step is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard transport theory (Sinkhorn divergence properties, W2-Ḣ^{-1} equivalence) and on regularity assumptions about the normalized densities. The free hyperparameters δ and ε, and the inverse temperature λ, are tuned by calibration rather than derived, and the theory's strongest guarantees hold only in the δ→∞ limit or under bounded-below density assumptions.

free parameters (3)
  • Softplus scale δ = intermediate calibrated value (not specified)
    Controls the sharpness of the normalization; the paper states values too small/large degrade performance and an intermediate value is used (Section 5.4).
  • Inverse temperature λ = calibrated per method; learned adaptively in multi-parameter
    In single-event experiments λ is calibrated separately per risk 'to ensure comparable posterior concentration' (Section 5.4); in multi-parameter experiments it is inferred jointly (Section 5.6).
  • Entropic regularization ε = not specified
    Used in the discrete Sinkhorn divergence (Eqs. 7-8); its value or schedule is not reported in the numerical experiments.
assumptions (4)
  • standard math The debiased Sinkhorn divergence Sε is symmetric, smooth, positive definite, and convex in the measure pair (Theorem 1 of Feydy et al., 2019).
    Invoked in Propositions 1–3 as the basis for smoothness, positive definiteness, and asymptotic convexity.
  • domain assumption On compact domains, the quadratic Wasserstein distance is equivalent to the negative Sobolev norm Ḣ^{-1} for probability densities bounded above and below by positive constants (Peyre, 2018).
    Used in Propositions 4–5 and Theorems 2–3 to turn transport stability into Ḣ^{-1} Lipschitz bounds; requires a uniform positive lower bound on normalized densities.
  • domain assumption The forward map f: Θ → F(X) is continuous into bounded zero-mean functions and the parameter space Θ is compact.
    Needed for well-definedness of the Gibbs posterior (Theorem 1), following Stuart's Bayesian inverse problem framework.
  • standard math The softplus normalization σδ(z)=log(1+e^{δz}) is one-to-one, smooth and strictly increasing.
    Used in the construction of normalized densities (Section 3.1, Eq. 9).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Symmetrized Sinkhorn-Gibbs Inference for Oscillatory Inverse Problems." pith.science (2026). https://pith.science/paper/KK2XGVQH

@misc{pith2026260726264,
  author       = {Pith},
  title        = {Pith review of: Symmetrized Sinkhorn-Gibbs Inference for Oscillatory Inverse Problems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KK2XGVQH}},
  note         = {Machine review of arXiv:2607.26264}
}
read the original abstract

Oscillatory inverse problems often exhibit highly nonconvex discrepancy landscapes due to signal misalignment and cycle-skipping phenomena, posing significant challenges for uncertainty quantification. While Gibbs posteriors provide a flexible framework for incorporating problem-specific discrepancy measures, their performance depends strongly on the empirical risk landscape induced by the chosen discrepancy measure. Optimal transport accounts for the spatial and temporal structure of the underlying feature domain when comparing oscillatory signals. However, the direct application of classical optimal transport is precluded because observed and predicted signals are typically signed and therefore do not satisfy the positivity requirements of classical transport formulations. We introduce a symmetrized Sinkhorn-Gibbs inference framework for oscillatory inverse problems. The proposed approach combines a normalization procedure for signed signals with a symmetrized Sinkhorn loss that exploits complementary transport information from the normalized signals and their normalized negations. The resulting loss is incorporated into the Gibbs posterior framework, yielding a Gibbs inference methodology tailored to oscillatory data. We establish smoothness properties of the proposed loss, prove well-definedness of the resulting Gibbs posterior, derive robustness guarantees, and develop an adaptive sampling strategy for posterior computation. Numerical experiments demonstrate less oscillatory empirical risk landscapes with fewer spurious local minima, more accurate posterior inference, greater robustness to observational noise, and improved population-level recovery than Gibbs posteriors based on Euclidean and trace-wise Wasserstein losses.

Figures

Figures reproduced from arXiv: 2607.26264 by the authors.

Figure 1
Figure 1. Three representative benchmark events shown as top views of the noise-free observed wavefield [PITH_FULL_IMAGE:figures/full_fig_p023_1.png] view at source ↗
Figure 2
Figure 2. Empirical risk landscapes as functions of the source location [PITH_FULL_IMAGE:figures/full_fig_p026_2.png] view at source ↗
Figure 3
Figure 3. Posterior behavior as a function of data resolution in the noise-free single-event setting. Left: [PITH_FULL_IMAGE:figures/full_fig_p027_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Noise robustness in the single-event setting on the coarse observation grid ( [PITH_FULL_IMAGE:figures/full_fig_p028_4.png]
Figure 5
Figure 5. Figure 5: Population-level recovery of the posterior-mixture population estimator as a function of the [PITH_FULL_IMAGE:figures/full_fig_p030_5.png]
Figure 6
Figure 6. Figure 6: Marginal posterior densities for the multi-parameter event-level problem with true parameter [PITH_FULL_IMAGE:figures/full_fig_p032_6.png]
Figure 7
Figure 7. Figure 7: Marginal posterior densities for the multi-parameter event-level problem with true parameter [PITH_FULL_IMAGE:figures/full_fig_p032_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 5 linked inside Pith

  1. [44]

    ComparisonbetweenW 2 distanceand 9H´1 norm, andlocalizationofWassersteindistance

    R.Peyre. ComparisonbetweenW 2 distanceand 9H´1 norm, andlocalizationofWassersteindistance. ESAIM: COCV, 24:1489–1501, 2018. 38

    2018

  2. [1]

    A. M. Stuart. Inverse problems: A Bayesian perspective.Acta Numerica, 19:451–559, 2010

    2010

  3. [2]

    Dashti and A

    M. Dashti and A. M. Stuart. The Bayesian approach to inverse problems. In R. Graham, D. Higdon, and H. Owhadi, editors,Handbook of Uncertainty Quantification, pages 311–428. 2016

    2016

  4. [3]

    T. J. Sullivan. Well-posed Bayesian inverse problems and heavy-tailed stable quasi-Banach space priors.Inverse Problems and Imaging, 11:857–874, 2017

    2017

  5. [4]

    Virieux and S

    J. Virieux and S. Operto. An overview of full-waveform inversion in exploration geophysics.Geo- physics, 74:WCC127–WCC152, 2009

    2009

  6. [5]

    P. G. Bissiri, C. C. Holmes, and S. G. Walker. A general framework for updating belief distributions. J. R. Statist. Soc. B, 78:1103–1130, 2016

    2016

  7. [6]

    Syring and R

    N. Syring and R. Martin. Calibrating general posterior credible regions.Biometrika, 106(2):479–486, 2019

    2019

  8. [7]

    Zafar and G

    S. Zafar and G. K. Nicholls. Exploring learning rate selection in generalised Bayesian inference using posterior predictive checks.arXiv:2410.01475, 2024

    arXiv 2024

Show all 44 references
  1. [8]

    J. E. Lee, S. Liu, and G. K. Nicholls. Bayesian inference for the learning rate in generalised Bayesian inference.arXiv:2506.12532, 2025

    2025 arXiv

  2. [9]

    Engquist and B

    B. Engquist and B. D. Froese. Application of the Wasserstein metric to seismic signals.Commu- nications in Mathematical Sciences, 12(5):979–988, 2014

    2014

  3. [10]

    Y. Yang, B. Engquist, J. Sun, and B. F. Hamfeldt. Application of optimal transport and the quadratic Wasserstein metric to full-waveform inversion.Geophysics, 83(1):R43–R62, 2018

    2018

  4. [11]

    Engquist and Y

    B. Engquist and Y. Yang. Seismic inversion and the data normalization for optimal transport. Methods Appl. Anal., 26:133—-148, 2019

    2019

  5. [12]

    Yang and B

    Y. Yang and B. Engquist. Analysis of optimal transport and related misfit functions in full-waveform inversion.Geophysics, 83(1):A7—-A12, 2018

    2018

  6. [13]

    Engquist and Y

    B. Engquist and Y. Yang. Seismic Imaging and Optimal Transport.Communications in Information and Systems, 19(2):95–145, 2019

    2019

  7. [14]

    Motamed and D

    M. Motamed and D. Appelö. Wasserstein metric-driven Bayesian inversion with applications to signal processing.International J. for Uncertainty Quantification, 9:395–414, 2019

    2019

  8. [15]

    M. M. Dunlop and Y. Yang. Stability of Gibbs posteriors from the Wasserstein loss for Bayesian full waveform inversion.Inverse Problems, 38:085011, 2022

    2022

  9. [16]

    M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. InAdvances in Neural Information Processing Systems 26, pages 2292–2300, 2013

    2013

  10. [17]

    Solomon, F

    J. Solomon, F. De Goes, G. Peyré, M. Cuturi, A. Butscher, A. Nguyen, T. Du, and L. Guibas. Convolutional Wasserstein distances: efficient optimal transportation on geometric domains.ACM Transactions on Graphics, 34:66:1–66:11, 2015. 36

    2015

  11. [18]

    M. Motamed. Hierarchical low-rank approximation of regularized Wasserstein distance. arXiv:2004.12511, 2020

    2004 arXiv

  12. [19]

    Métivier, A

    L. Métivier, A. Allain, R. Brossier, Q. Mérigot, E. Oudet, and J. Virieux. Optimal transport for mitigating cycle skipping in full-waveform inversion: A graph-space transform approach.Geophysics, 83(5):R515–R540, 2018

    2018

  13. [20]

    Kaipo and E

    J. Kaipo and E. Somersalo.Statistical and Computational Inverse Problems. Springer, New York, 2005

    2005

  14. [21]

    D. M. Blei, A. Kucukelbir, and J. D. McAuliffe. Variational inference: A review for statisticians. Journal of the American Statistical Association, 112:859–877, 2017

    2017

  15. [22]

    T. Zhang. Information theoretical upper and lower bounds for statistical estimation.IEEE Trans- action on Information Theory, 52:1307–1321, 2006

    2006

  16. [23]

    O. Catoni. PAC-Bayesian supervised classification: The thermodynamics of statistical learning.IMS Lecture Notes Monograph Series, 56:1–163, 2007

    2007

  17. [24]

    Villani.Topics in Optimal Transportation, volume 58 ofGraduate Studies in Mathematics

    C. Villani.Topics in Optimal Transportation, volume 58 ofGraduate Studies in Mathematics. American Mathematical Society, 2003

    2003

  18. [25]

    Villani.Optimal Transport: Old and New, volume 338 ofGrundlehren der mathematischen Wissenschaften

    C. Villani.Optimal Transport: Old and New, volume 338 ofGrundlehren der mathematischen Wissenschaften. Springer Verlag, 2009

    2009

  19. [26]

    Ambrosio and N

    L. Ambrosio and N. Gigli. A user’s guide to optimal transport. InModelling and Optimisation of Flows on Networks, pages 1–155. Springer, 2013

    2013

  20. [27]

    Santambrogio.Optimal transport for applied mathematicians

    F. Santambrogio.Optimal transport for applied mathematicians. Birkäuser, 2015

    2015

  21. [28]

    Peyré and M

    G. Peyré and M. Cuturi. Computational optimal transport.Foundations and Trends in Machine Learning, 11:355–607, 2019

    2019

  22. [29]

    V. M. Panaretos and Y. Zemel. Statistical aspects of Wasserstein distances.Annual Review of Statistics and Its Application, 6:405–431, 2019

    2019

  23. [30]

    Feydy, F.-X

    J. Feydy, F.-X. Vialard T. Séjourné, S. Amari, A. Trouvé, and G. Peyré. Interpolating between optimal transport and MMD using Sinkhorn divergences. InThe 22nd International Conference on Artificial Intelligence and Statistics, pages 2681–2690, 2019

    2019

  24. [31]

    Ramdas, N

    A. Ramdas, N. G. Trillos, and M. Cuturi. On Wasserstein two-sample testing and related families of nonparametric tests.Entropy, 19:47, 2017

    2017

  25. [32]

    Genevay, G

    A. Genevay, G. Peyré, and M. Cuturi. Learning generative models with Sinkhorn divergences. In The 21st International Conference on Artificial Intelligence and Statistics, pages 1608–1617, 2018

    2018

  26. [33]

    Salimans, H

    T. Salimans, H. Zhang, A. Radford, and D. Metaxas. Improving GANs using optimal transport. In International Conference on Learning Representations, 2018

    2018

  27. [34]

    Sanjabi, J

    M. Sanjabi, J. Ba, M. Razaviyayn, and J. D. Lee. On the convergence and robustness of training GANs with regularized optimal transport. InAdvances in Neural Information Processing Systems 31, 2018. 37

    2018

  28. [35]

    Chizat, P

    L. Chizat, P. Roussillon, F. Léger, F.-X. Vialard, and G. Peyré. Faster Wasserstein distance estima- tion with the Sinkhorn divergence.arXiv:2006.08172, 2020

    2006 arXiv

  29. [36]

    Sinkhorn

    R. Sinkhorn. A relationship between arbitrary positive matrices and doubly stochastic matrices. Annals of Mathematical Statististics, 35:876–879, 1964

    1964

  30. [37]

    Sinkhorn and P

    R. Sinkhorn and P. Knopp. Concerning nonnegative matrices and doubly stochastic matrices.Pacific Journal of Mathematics, 21:343–348, 1967

    1967

  31. [38]

    Altschuler, J

    J. Altschuler, J. Weed, and P. Rigollet. Near-linear time approximation algorithms for optimal transport via Sinkhorn iteration. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Sys...

    1961

  32. [39]

    Dvurechensky, A

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

    2018

  33. [40]

    Chizat, G

    L. Chizat, G. Peyré, B. Schmitzer, and F.-X. Vialard. Unbalanced optimal transport: Dynamic and Kantorovich formulation.J. Funct. Anal., 274:3090–3123, 2018

    2018

  34. [41]

    Gangbo, W

    W. Gangbo, W. Li, S. Osher, and M. Puthawala. Unnormalized optimal transport.J. Comput. Phys., 399, 2019

    2019

  35. [42]

    Séjourné, J

    T. Séjourné, J. Feydy, F.-X. Vialard, A. Trouvé, and G. Peyré. Sinkhorn divergences for unbalanced optimal transport.arXiv:1910.12958, 2021

    1910 arXiv

  36. [43]

    M. Vihola. On the ergodicity of the adaptive metropolis algorithm.Statistics and Computing, 30(6):1633–1650, December 2020

    2020

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.