pith. sign in

arxiv: 2605.27883 · v1 · pith:WCODFZIBnew · submitted 2026-05-27 · 🧮 math.OC · math.AP· math.FA

Stability of Quadratically Regularized Optimal Transport

Alberto Gonz\'alez-Sanz , Marcel Nutz This is my paper

Pith reviewed 2026-06-29 11:20 UTC · model grok-4.3

classification 🧮 math.OC math.APmath.FA
keywords optimal transportquadratic regularizationstabilitydual potentialssupport stabilityHausdorff distanceLipschitz continuity
0
0 comments X

The pith

Quadratically regularized optimal transport has uniform L^∞ stability for dual potentials and support under marginal perturbations.

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

The paper develops a quantitative stability theory for quadratically regularized optimal transport when the marginal distributions, the cost function, or the regularization parameter are perturbed. It establishes an L^∞-stability result for the dual potentials by starting from a marginal-dependent L² error bound and upgrading it via a self-bound property of the potentials to a local L^∞-Lipschitz bound that holds uniformly over marginals. This bound implies stability of the optimal coupling and, for quadratic costs, local Lipschitz continuity of the support in Hausdorff distance. A reader would care because such guarantees ensure that small data perturbations produce controlled changes in the transport plan, making the method reliable for applications with estimated or noisy inputs.

Core claim

Quadratically regularized optimal transport admits an L^∞-stability result for the dual potentials. Starting from an L² error bound that varies with the marginals and invoking a self-bound property of the potentials, the analysis derives a local L^∞-Lipschitz bound that is uniform over marginals. This bound yields stability of the optimal coupling and of its support. In particular, for the quadratic transport cost the support of the optimal coupling is locally Lipschitz in Hausdorff distance under perturbations of the marginals.

What carries the argument

The self-bound property of the dual potentials, which upgrades a marginal-dependent L² error bound into a uniform local L^∞-Lipschitz bound.

If this is right

  • The optimal coupling varies continuously with perturbations of the marginals.
  • For quadratic costs the support of the optimal coupling is locally Lipschitz continuous in Hausdorff distance.
  • Stability also holds under perturbations of the cost function and the regularization parameter.
  • This supplies the first stability result for the optimal support in regularized optimal transport.

Where Pith is reading between the lines

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

  • The upgrade technique from L² to L^∞ via self-bounds may extend to other regularizers if analogous self-bound properties can be verified.
  • Error bounds derived this way could guide the construction of numerical schemes that remain accurate when marginals are estimated from finite samples.
  • Similar stability statements for non-quadratic costs would follow if the self-bound property is shown to hold more generally.

Load-bearing premise

An initial L² error bound that varies with the marginals exists together with a self-bound property of the dual potentials that permits upgrading to a uniform L^∞ Lipschitz bound.

What would settle it

A sequence of marginal perturbations where the Hausdorff distance of the supports grows faster than linearly with the marginal perturbation size would falsify the claimed local Lipschitz stability of the support.

read the original abstract

Quadratically regularized optimal transport (QOT) is a sparse alternative to entropic optimal transport. We develop a quantitative stability theory for QOT under perturbations of the marginals, the transport cost function, and the regularization parameter. The centerpiece is an $L^\infty$-stability result for the dual potentials. Starting from an error bound in an $L^2$-space that varies with the marginals, we use a self-bound for the potentials to derive a local $L^\infty$-Lipschitz bound that is uniform over marginals. This bound also yields stability of the optimal coupling and of its support. In particular, we show that for the quadratic transport cost, the support of the optimal coupling is locally Lipschitz in Hausdorff distance under perturbations of the marginals. To the best of our knowledge, this is the first stability result for the optimal support in regularized optimal transport.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper develops a quantitative stability theory for quadratically regularized optimal transport (QOT) under perturbations of the marginals, the transport cost, and the regularization parameter. The central result is an L^∞-stability bound on the dual potentials, obtained by upgrading a marginal-dependent L² error bound via a self-bound property of the potentials; this yields local L^∞-Lipschitz continuity uniform over marginals, from which stability of the optimal coupling and (for quadratic cost) local Hausdorff-Lipschitz stability of its support are deduced. The authors claim this is the first such support-stability result in regularized OT.

Significance. If the uniformity of the L^∞ bound holds, the result would be a meaningful advance: support stability for the optimal plan is a natural but previously unavailable quantitative property in regularized OT, with direct implications for robustness in applications that rely on the geometry of the coupling. The self-bound technique for norm upgrading may also be reusable in other regularized variational problems.

major comments (2)
  1. [Abstract / §4] Abstract and §4 (main stability theorem): the upgrade from a marginal-dependent L² error bound to a uniform-over-marginals L^∞-Lipschitz constant is asserted to follow from the self-bound property of the dual potentials, yet the argument does not explicitly control how the self-bound constants themselves vary with the marginals. If those constants scale with the same quantities that drive the L² error, uniformity fails and the subsequent claims on coupling and support stability do not follow.
  2. [§5] §5 (support stability for quadratic cost): the local Hausdorff-Lipschitz continuity of the support is derived from the L^∞ potential bound; because the potential bound itself rests on the unverified uniformity step above, the support result is not yet load-bearing.
minor comments (2)
  1. [§3] Notation for the self-bound constant is introduced without a displayed equation number; adding an explicit label would clarify subsequent references.
  2. [§4] The statement that the L² error bound 'varies with the marginals' is used repeatedly but never given an explicit functional dependence; a displayed inequality would help the reader track constants.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and the constructive comments, which help clarify the presentation of the uniformity argument. We address each major comment below and will revise the manuscript accordingly to make the control of constants explicit.

read point-by-point responses

  1. Referee: [Abstract / §4] Abstract and §4 (main stability theorem): the upgrade from a marginal-dependent L² error bound to a uniform-over-marginals L^∞-Lipschitz constant is asserted to follow from the self-bound property of the dual potentials, yet the argument does not explicitly control how the self-bound constants themselves vary with the marginals. If those constants scale with the same quantities that drive the L² error, uniformity fails and the subsequent claims on coupling and support stability do not follow.

    Authors: We agree that an explicit verification of uniformity is needed. The self-bound constants arise from the quadratic regularization term and a priori L^∞ bounds on the potentials, which in turn depend on the diameter of the supports and the Lipschitz constant of the cost; these quantities are fixed and continuous with respect to marginal perturbations in the local neighborhood under consideration. To address the concern directly, we will insert a short lemma in §4 that quantifies the dependence of the self-bound constants on the marginals and shows they remain bounded uniformly under the local perturbations assumed in the theorem. This will make the passage from the marginal-dependent L² bound to the uniform L^∞-Lipschitz bound fully rigorous. revision: yes

  2. Referee: [§5] §5 (support stability for quadratic cost): the local Hausdorff-Lipschitz continuity of the support is derived from the L^∞ potential bound; because the potential bound itself rests on the unverified uniformity step above, the support result is not yet load-bearing.

    Authors: The support-stability statement in §5 is obtained by combining the L^∞ potential bound with the quadratic cost structure and a standard argument relating potential differences to Hausdorff distance of supports. Once the uniformity of the L^∞ bound is made explicit via the added lemma in §4, the derivation in §5 becomes valid. We will add a cross-reference in §5 to the strengthened statement in §4 and, if needed, include a brief remark confirming that all constants remain controlled locally. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation relies on independent functional-analytic bounds

full rationale

The paper derives L^∞ stability from an L² error bound (varying with marginals) plus a self-bound property of dual potentials, then upgrades to uniform Lipschitz control and support stability. No quoted step reduces the claimed result to a fitted parameter, self-definition, or load-bearing self-citation; the self-bound is invoked as an external property rather than constructed from the target stability statement itself. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; the central claim rests on an unstated L^2 error bound and a self-bound property whose precise formulation is not given.

pith-pipeline@v0.9.1-grok · 5682 in / 1119 out tokens · 35162 ms · 2026-06-29T11:20:44.029592+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

33 extracted references · 27 canonical work pages · 4 internal anchors

  1. [1]

    Stability and sample complexity of diver- gence regularized optimal transport.Bernoulli31.1 (2025), pp

    Bayraktar, E., Eckstein, S., and Zhang, X. Stability and sample complexity of diver- gence regularized optimal transport.Bernoulli31.1 (2025), pp. 213–239.doi:10.3150/ 24-BEJ1725

    2025

  2. [2]

    Smooth and Sparse Optimal Transport.Proceed- ings of the Twenty-First International Conference on Artificial Intelligence and Statis- tics

    Blondel, M., Seguy, V., and Rolet, A. Smooth and Sparse Optimal Transport.Proceed- ings of the Twenty-First International Conference on Artificial Intelligence and Statis- tics. Ed. by A. Storkey and F. Perez-Cruz. Vol. 84. Proceedings of Machine Learning Research. PMLR, 2018, pp. 880–889

    2018

  3. [3]

    Abbe,Community Detection and Stochastic Block Models: Recent Developments, Journal of Machine Learning Research, 18 (2018), pp

    Bonnans,J.F.andShapiro,A.Perturbation Analysis of Optimization Problems.Springer Series in Operations Research. New York: Springer-Verlag, 2000, pp. xviii+601.doi: 10.1007/978-1-4612-1394-9. 28

    work page doi:10.1007/978-1-4612-1394-9 2000

  4. [4]

    Displacement smoothness of entropic optimal transport.ESAIM Control Optim

    Carlier, G., Chizat, L., and Laborde, M. Displacement smoothness of entropic optimal transport.ESAIM Control Optim. Calc. Var.30 (2024), Paper No. 25, 24.doi:10. 1051/cocv/2024013

    work page arXiv 2024

  5. [5]

    and Laborde, M

    Carlier, G. and Laborde, M. A differential approach to the multi-marginal Schrödinger system.SIAM J. Math. Anal.52.1 (2020), pp. 709–717.doi:10.1137/19M1253800

    work page doi:10.1137/19m1253800 2020

  6. [6]

    A., Matrán, C., and Tuero-Díaz, A

    Cuesta-Albertos, J. A., Matrán, C., and Tuero-Díaz, A. Optimal transportation plans and convergence in distribution.J. Multivariate Anal.60.1 (1997), pp. 72–83.doi: 10.1006/jmva.1996.1651

    work page doi:10.1006/jmva.1996.1651 1997

  7. [7]

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport.Ad- vances in Neural Information Processing Systems 26

    Cuturi, M. Sinkhorn Distances: Lightspeed Computation of Optimal Transport.Ad- vances in Neural Information Processing Systems 26. Ed. by C. J. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Q. Weinberger. Curran Associates, Inc., 2013, pp. 2292–2300.doi:10.5555/2999792.2999868

    work page doi:10.5555/2999792.2999868 2013

  8. [8]

    and Mérigot, Q

    Delalande, A. and Mérigot, Q. Quantitative stability of optimal transport maps under variations of the target measure.Duke Math. J.172.17 (2023), pp. 3321–3357.doi: 10.1215/00127094-2022-0106

    work page doi:10.1215/00127094-2022-0106 2023

  9. [9]

    de, and Doucet, A

    Deligiannidis, G., Bortoli, V. de, and Doucet, A. Quantitative uniform stability of the iterative proportional fitting procedure.Ann. Appl. Probab.34.1A (2024), pp. 501–516. doi:10.1214/23-AAP1970

    work page doi:10.1214/23-aap1970 2024

  10. [10]

    Tight stability bounds for entropic Brenier maps.Int

    Divol, V., Niles-Weed, J., and Pooladian, A.-A. Tight stability bounds for entropic Brenier maps.Int. Math. Res. Not. IMRN2025.7 (2025), rnaf078.doi:10 . 1093 / imrn/rnaf078. arXiv:2404.02855

    work page arXiv 2025

  11. [11]

    and Nutz, M

    Eckstein, S. and Nutz, M. Quantitative stability of regularized optimal transport and convergence of Sinkhorn’s algorithm.SIAM J. Math. Anal.54.6 (2022), pp. 5922–5948. doi:10.1137/21M145505X

    work page doi:10.1137/21m145505x 2022

  12. [12]

    and Nutz, M

    Eckstein, S. and Nutz, M. Convergence rates for regularized optimal transport via quantization.Math. Oper. Res.49.2 (2024), pp. 1223–1240.doi:10.1287/moor.2022. 0245

    work page doi:10.1287/moor.2022 2024

  13. [13]

    Essid,M.andSolomon,J.Quadraticallyregularizedoptimaltransportongraphs.SIAM J. Sci. Comput.40.4 (2018), A1961–A1986.doi:10.1137/17M1132665

    work page doi:10.1137/17m1132665 2018

  14. [14]

    Sepasdar, M

    Ghosal, P., Nutz, M., and Bernton, E. Stability of entropic optimal transport and Schrödinger bridges.J. Funct. Anal.283.9 (2022), Paper No. 109622.doi:10.1016/j. jfa.2022.109622

    work page doi:10.1016/j 2022

  15. [15]

    On Hölder continuity-in-time of the optimal transport map towards measures along a curve.Proc

    Gigli, N. On Hölder continuity-in-time of the optimal transport map towards measures along a curve.Proc. Edinb. Math. Soc. (2)54.2 (2011), pp. 401–409.doi:10.1017/ S001309150800117X

    2011

  16. [16]

    Sample complexity of quadratically regularized optimal transport

    González-Sanz, A., Barrio, E. del, and Nutz, M.Sample Complexity of Quadratically Regularized Optimal Transport. 2025. arXiv:2511.09807

    work page arXiv 2025

  17. [17]

    Sharp local sparsity of regularized optimal transport

    González-Sanz, A., Gvalani, R. S., and Koch, L.Sharp local sparsity of regularized optimal transport. 2026. arXiv:2604.00843

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  18. [18]

    and Nutz, M

    González-Sanz, A. and Nutz, M.Sparsity of Quadratically Regularized Optimal Trans- port: Scalar Case. SIAM J. Math. Anal., forthcoming. 2024. arXiv:2410.03353. 29

    work page arXiv 2024

  19. [19]

    González-Sanz,A.,Nutz,M.,andRiverosValdevenito,A.Linear Convergence of Gradi- ent Descent for Quadratically Regularized Optimal Transport. 2025. arXiv:2509.08547

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  20. [20]

    Polyak-Lojasiewicz Inequality for Quadratically Regularized Optimal Transport

    González-Sanz, A., Nutz, M., and Riveros Valdevenito, A. Polyak–Łojasiewicz Inequal- ity for Quadratically Regularized Optimal Transport.SIAM J. Optim.(2026). Forth- coming. arXiv:2605.27175

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  21. [21]

    and Sheng, S.Linearization of Monge–Ampère Equations and Sta- tistical Applications

    González-Sanz, A. and Sheng, S.Linearization of Monge–Ampère Equations and Sta- tistical Applications. 2024. arXiv:2408.06534

    work page arXiv 2024

  22. [22]

    Grisvard,P.Elliptic Problems in Nonsmooth Domains.Vol.24.MonographsandStudies inMathematics.Boston,MA:PitmanAdvancedPublishingProgram,1985,pp.xiv+410

    1985

  23. [23]

    Gluing methods for quantitative stability of optimal transport maps.arXiv preprint arXiv:2411.04908, 2024

    Letrouit, C. and Mérigot, Q.Gluing methods for quantitative stability of optimal trans- port maps. 2024. arXiv:2411.04908

    work page arXiv 2024

  24. [24]

    A., Manns, P., and Meyer, C

    Lorenz, D. A., Manns, P., and Meyer, C. Quadratically regularized optimal transport. Appl. Math. Optim.83.3 (2021), pp. 1919–1949.doi:10.1007/s00245-019-09614-w

    work page doi:10.1007/s00245-019-09614-w 2021

  25. [25]

    HAL preprint, hal-05616391

    Mérigot, Q.Sharp Stability of Brenier Maps via Quantitative Regularity of Potentials. HAL preprint, hal-05616391. 2026

    2026

  26. [26]

    Tsallis Regularized Optimal Trans- port and Ecological Inference.Proceedings of the AAAI Conference on Artificial Intel- ligence31.1 (2017).doi:10.1609/aaai.v31i1.10854

    Muzellec, B., Nock, R., Patrini, G., and Nielsen, F. Tsallis Regularized Optimal Trans- port and Ecological Inference.Proceedings of the AAAI Conference on Artificial Intel- ligence31.1 (2017).doi:10.1609/aaai.v31i1.10854

    work page doi:10.1609/aaai.v31i1.10854 2017

  27. [27]

    Lecture notes, Columbia Uni- versity,https://www.math.columbia.edu/~mnutz/docs/EOT_lecture_notes.pdf

    Nutz, M.Introduction to Entropic Optimal Transport. Lecture notes, Columbia Uni- versity,https://www.math.columbia.edu/~mnutz/docs/EOT_lecture_notes.pdf. 2021

    2021

  28. [28]

    Quadratically regularized optimal transport: existence and multiplicity of po- tentials.SIAM J

    Nutz, M. Quadratically regularized optimal transport: existence and multiplicity of po- tentials.SIAM J. Math. Anal.57.3 (2025), pp. 2622–2649.doi:10.1137/24M1638136

    work page doi:10.1137/24m1638136 2025

  29. [29]

    Computational Optimal Transport: With Applications to Data Science , url =

    Peyré, G. and Cuturi, M. Computational Optimal Transport: With Applications to Data Science.Foundations and Trends in Machine Learning11.5–6 (2019), pp. 355– 607.doi:10.1561/2200000073

    work page doi:10.1561/2200000073 2019

  30. [30]

    and Thomsen, W

    Rüschendorf, L. and Thomsen, W. Note on the Schrödinger equation andI-projections. Statist. Probab. Lett.17.5 (1993), pp. 369–375.doi:10.1016/0167-7152(93)90257-J

    work page doi:10.1016/0167-7152(93)90257-j 1993

  31. [31]

    Stabilized sparse scaling algorithms for entropy regularized transport problems.SIAM J

    Schmitzer, B. Stabilized sparse scaling algorithms for entropy regularized transport problems.SIAM J. Sci. Comput.41.3(2019),A1443–A1481.doi:10.1137/16M1106018

    work page doi:10.1137/16m1106018 2019

  32. [32]

    Wiesel,J.andXu,X.SparsityofQuadraticallyRegularizedOptimalTransport:Bounds on Concentration and Bias.SIAM J. Math. Anal.57.6 (2025), pp. 6498–6521.doi: 10.1137/25M1723633

    work page doi:10.1137/25m1723633 2025

  33. [33]

    Yao, R., Kim, Y.-H., and Schiebinger, G.Quantitative Stability of Many-Marginal Schrödinger Bridge. 2026. arXiv:2604.15191. 30

    work page internal anchor Pith review Pith/arXiv arXiv 2026