Pith. sign in

REVIEW 3 major objections 3 minor 80 references

Optimal Transport under Group Fairness Constraints

T0 review · 3 major / 3 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Fair matching plans are one Sinkhorn projection away

desk verdict A novel fair-OT framework with plausible relaxation bounds, but FairSinkhorn's convergence is unproved and the experiments use an invalid target matrix. read the letter →

arxiv 2601.07144 v3 pith:ZDTHQIIG submitted 2026-01-12 stat.ML cs.LGmath.STstat.TH

Linus Bleistein , Mathieu Dagréou , Francisco Andrade , Thomas Boudou , Aurélien Bellet This is my paper
classification stat.MLcs.LGmath.STstat.TH
keywords groupfairnessoptimaltransportentropicregularizationSinkhornalgorithmconstraintscostlearningsamplecomplexitymatching
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

This paper introduces a group-fairness criterion for optimal transport (OT): a transport plan is fair when the probability mass it sends from each sensitive group on one side to each group on the other matches a planner-specified target. It shows that the fair plan has the same kernel form as entropic OT, with one extra blockwise multiplier per group pair, and gives a modified Sinkhorn algorithm, FairSinkhorn, that alternates the usual row-column scalings with updates of these multipliers. Because exact fairness can be expensive, it also studies two relaxations: adding a convex penalty on fairness violations (with an O(n^-1/2) sample-complexity bound), and learning the ground cost so that the standard OT solution is fair (with a bound on fairness deviation on unseen samples). Experiments on synthetic school-assignment examples show the trade-off between cost and fairness and that learned costs generalize to new batches.

What carries the argument

FairSinkhorn, the paper's Algorithm 1, iterates three updates: u <- n^{-1} 1 / ((K ⊙ T)v), v <- m^{-1} 1 / ((K ⊙ T)^T u), and L <- F / Phi(u,v), where Phi computes the group-pair mass of the current scaled plan and T is the blockwise sum of L_sw B_sw. The block multiplier update is the added mechanism that enforces the fairness target; the same block structure underlies the penalized relaxation (a strongly convex objective whose unique minimizer can be found by generalized conditional gradient) and the bilevel cost-learning formulation.

What would settle it

Take a small two-group instance where the unique F-fair plan is known by direct enumeration, run FairSinkhorn for many iterations, and check whether the iterate's group-pair masses approach F; rows with zero intermediate Phi entries that stall, or iterates that oscillate, would refute the claim that Algorithm 1 computes the exact solution.

Watch

Extended reading notes

Core claim

The paper's central discovery is structural: the solution to entropic optimal transport with linear equality constraints on group-pair mass remains a product of diagonal scalings, the kernel K=e^{-C/epsilon}, and a block matrix H = sum_{s,w} e^{h_sw/epsilon} B_sw, where B_sw selects entries with source group s and target group w. The group constraints enter only through the multipliers h_sw, so the exact F-fair plan can in principle be obtained by a Sinkhorn-like alternating iteration: rescale rows and columns to match marginals, then rescale each group block to match the fairness target. On this basis the paper proposes FairSinkhorn, proves existence and uniqueness of fair plans under compa

Load-bearing premise

FairSinkhorn's alternating updates converge to the unique fair plan: the paper proves the form of the solution and existence of the fair plan, but does not prove the iteration's convergence or handle zero entries in Phi, so the exact-fairness results depend on this unverified step.

Editorial extensions

If this is right

  • If FairSinkhorn converges as claimed, a central planner can enforce exact matching quotas or anti-homogamy targets without changing the entropic OT machinery, only adding a group-pair rescaling step.
  • The penalized relaxation preserves the standard n^{-1/2} sample complexity of entropic OT, so enforcing fairness approximately does not degrade statistical efficiency.
  • The cost-learning bound implies a learned cost can be reused: matching fresh samples with vanilla OT keeps fairness deviations controlled as long as sample sizes are large and the cost family is bounded.
  • The convex penalization and the nonconvex cost-learning approach occupy different trade-off positions, with penalized OT more flexible and cost learning faster at inference because it reduces fairness to a one-time geometric reshaping.

Reading between the lines

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

  • If the convergence gap in FairSinkhorn is closed, the same block-multiplier update likely extends to other linear constraints on transport plans (e.g., sparsity patterns or capacity bounds), giving a general constrained-Sinkhorn template rather than a fairness-specific one.
  • The cost-learning view suggests a testable practical recipe: pretrain an embedding with the fairness penalty switched off, then fine-tune the cost; the paper's experiments initialize the neural cost this way, and one could systematically study initialization as a lever for the nonconvex bilevel problem.
  • The exponential dependence on 1/epsilon in the generalization bound suggests that very small entropic regularization could hurt fairness transfer; a natural extension is to anneal epsilon during training or to design costs that remain fair across a range of epsilon.
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 / 3 minor

Summary. The paper proposes a group-fairness notion for entropic optimal transport in which a central planner prescribes target group-pair matching probabilities F. It introduces FairSinkhorn, an alternating-scaling algorithm intended to solve the exact F-fair entropic OT problem; a penalized OT relaxation with a claimed O(n^{-1/2}) sample complexity; and a bilevel cost-learning method with a uniform bound on the expected fairness deviation on unseen samples. Experiments on synthetic school-assignment data compare exact, penalized, and cost-learning approaches in a cost--fairness plane.

Significance. The fairness notion is natural and more flexible than previous OT-based fairness criteria, and the two relaxations address a real practical issue: exact fairness can be too expensive. If the theoretical results are completed and verified, the paper would make a useful contribution: the penalized formulation is convex, the sample-complexity rate matches unconstrained entropic OT, and the cost-learning bound is uniform over a parametric family and thus supports reusing a learned cost. The experiments clearly illustrate the intended cost--fairness trade-off and the inference-time advantage of cost learning. However, the central algorithmic claim for FairSinkhorn is not established, and the exact-constraint formulation has a finite-sample feasibility gap that the paper does not address. These issues are load-bearing and must be fixed before the claims can be accepted.

major comments (3)
  1. [§3.3, Algorithm 1] The abstract and §3.3 claim that FairSinkhorn computes perfectly fair transport plans efficiently. Proposition 3.3/B.1 only gives the structural form of any solution to Problem (3); it does not show that the cyclic updates in Algorithm 1 converge to that solution. After the L-update in Step 7, the row and column constraints are violated again, so the standard two-marginal Sinkhorn contraction argument does not apply. The paper provides no fixed-point analysis, no Lyapunov function, and no appeal to iterative-proportional-fitting convergence for the three constraint families. Step 7 also performs 0/0 when F_sw = Φ_sw = 0. The empirical perfect-fairness claim in Fig. 3a therefore rests on an unverified convergence property. A convergence theorem, or an explicit repositioning of FairSinkhorn as a heuristic, is needed.
  2. [§3.2, Problem (3)] The exact fairness constraints are only feasible when F is a coupling of the empirical group marginals, not merely of the population marginals. Summing the constraints of (3) over w gives row sums of F equal to the empirical proportion of group s in the sample, and summing over s gives column sums equal to the empirical proportion of group w. But F is defined in §3.2 as a coupling of the population marginals p and q, which generally differ from the observed group frequencies. Hence the feasible set of Problem (3) can be empty for finite samples. Proposition 3.2 is a population-level existence result and does not cover this case; the balanced group draws in the experiments hide the issue. The paper should either define a sample-dependent target F_n, impose exact constraints only for group pairs with positive empirical mass, or prove conditions under which the population F is feasible for
  3. [§C.2, Theorem 4.2] The proof of the claimed O(n^{-1/2}) sample-complexity bound is not fully checkable. It invokes 'Theorem C.4' for sub-Gaussian concentration and again near Eq. (21), but no Theorem C.4 is stated; the appendix contains Lemma C.4 but no theorem with that label. The proof also relies on 'Theorem 18 in [32]' for the sample complexity of entropic OT with a modified cost, without verifying that the modified cost \(\hat c = c + 2\lambda\sum_{s,w}\langle\xi_{sw},\pi^*_\infty\rangle\xi_{sw}\) satisfies the hypotheses of that result. Since this is one of the paper's stated theoretical contributions, the proof needs to be completed or the external results need to be stated with their exact conditions.
minor comments (3)
  1. [Eq. (12)] The displayed rate in Eq. (12) reads inconsistently as \(O(1/\sqrt{n}^{-1})\); it should be \(O(1/\sqrt{n})\).
  2. [§5.1] The experiment section refers to 'Theorem 4.1' and 'Theorem 4.2' for the Mahalanobis and neural cost parameterizations; these are Examples 4.1 and 4.2 and should be cited accordingly.
  3. [§C.1, Lemma C.1] The construction in Lemma C.1 divides by p(S=s) and q(W=w). If a group has zero probability, the formula is undefined; while the zero-mass case can be handled by convention, the proof should state this explicitly.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the fairness constraint in Algorithm 1 is enforced by construction (normal constraint satisfaction, not a fitted prediction), the statistical bounds use external black-box results without fitted parameters, and the only self-citations support the bilevel solver rather than the fairness result. The main gap is an unproved convergence theorem, which is a correctness issue

full rationale

We walked the derivation chain. The structural form of the fair entropic OT solution (Prop. 3.3/B.1) follows from standard first-order conditions of the Lagrangian; it is not assumed to equal the output. The sample-complexity bound (Thm. 4.2) is proved by linearizing the convex fairness penalty, reducing to the sample complexity of entropic OT via external results [32,52], and controlling the remainder with first-order optimality (Lemma C.2); no parameter is fitted to make the theorem true. The cost-learning bound (Thm. 4.4) is uniform over the parameter family, with constants depending only on the assumed bound R_Theta, and extends Rigollet-Stromme [64]; again no fitted input. The self-citations [18,19] are cited only as possible bilevel solvers; the fairness analysis does not depend on them. The 'perfect fairness' of FairSinkhorn in Fig. 3a is a direct consequence of the algorithm's own update L <- F / Phi(u,v): substituting T = sum L B into the returned plan makes each group-block sum equal F_sw identically. This is constraint enforcement, not a prediction, so we do not count it as circular; it would be circular only if the paper claimed to learn or predict fairness from data, which it does not. We do flag a substantial correctness gap: no convergence proof is given for the alternating updates among the three constraint families (row, column, group-block sums), and degenerate blocks with F_sw = Phi_sw = 0 produce 0/0 in Step 7. This affects the 'efficiently' claim in the abstract, but it is a missing proof, not a circular derivation. Overall low circularity burden; score 1 for minor non-load-bearing self-citations.

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

The paper's theoretical claims rely on standard assumptions (compact support, smooth/bounded costs) and external sample-complexity results. The user-specified fairness target and hyperparameters λ, ε are not hidden degrees of freedom in the derivations. No new entities are postulated.

free parameters (4)
  • fairness target F = [0.20 0.30; 0.28 0.22] in experiments
    User-specified desired group matching probabilities; not fitted to data but chosen by the planner. The experimental F is not a valid coupling (col sums 0.48/0.52), which breaks the problem's feasibility.
  • fairness penalty λ = grid logspace(0,3,80) etc.
    In penalized OT and bilevel cost learning, λ controls the fairness–cost trade-off; chosen by hand/validation. Theorems hold for any λ>0, so it's not fitted to derive the bounds.
  • entropic regularization ε = 1 in most experiments, varied in vanilla OT
    Chosen by hand; the theoretical bounds scale with exp(5R_Θ/ε). Not fitted to data.
  • learned cost parameters θ (Mahalanobis M or MLP weights) = learned from training data
    In cost learning, θ is optimized to minimize fairness loss; the bound in Thm 4.4 holds uniformly over θ∈Θ, so the fitted value is not part of the theoretical claim.
assumptions (6)
  • domain assumption µ and η are compactly supported (Assumption 3.1)
    Used to ensure existence of limits and boundedness in sample complexity proofs; standard in statistical OT [64].
  • domain assumption The ground cost c is infinitely differentiable (Assumption 4.1)
    Invoked in Theorem 4.2 to apply smoothness-based sample complexity results for entropic OT.
  • domain assumption Cost family is uniformly bounded: sup_θ ∥c_θ∥_∞ < R_Θ (Assumption 4.3)
    Needed for the uniform deviation bound in Theorem 4.4; holds for Lipschitz networks or bounded parameter sets.
  • standard math External entropic OT sample complexity bounds (Genevay et al. [33], Rigollet & Stromme [64] Theorem 6)
    The proofs of Theorems 4.2 and 4.4 directly extend these external results; the paper does not re-derive them.
  • standard math Sinkhorn–Knopp convergence for standard entropic OT
    Assumed for the inner solver used in GCG and cost learning; standard and cited.
  • standard math Differentiability of Sinkhorn outputs for iterative/implicit differentiation (Pauwels & Vaiter [59], Dagréou et al. [18,19])
    The cost-learning bilevel optimization relies on these existing results to compute gradients.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimal Transport under Group Fairness Constraints." pith.science (2026). https://pith.science/paper/ZDTHQIIG

@misc{pith2026260107144,
  author       = {Pith},
  title        = {Pith review of: Optimal Transport under Group Fairness Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDTHQIIG}},
  note         = {Machine review of arXiv:2601.07144}
}
read the original abstract

Ensuring fairness in matching algorithms is a key challenge in allocating scarce resources and positions. Focusing on Optimal Transport (OT), we introduce a novel notion of group fairness requiring that the probability of matching two individuals from any two given groups in the OT plan satisfies a predefined target. We first propose a modified Sinkhorn algorithm to compute perfectly fair transport plans efficiently. Since exact fairness can significantly degrade matching quality in practice, we then develop two relaxation strategies. The first one involves solving a penalized OT problem, for which we derive novel finite-sample complexity guarantees. Our second strategy leverages bilevel optimization to learn a ground cost that induces a fair OT solution, and we establish a bound on the deviation of fairness when matching unseen data. Finally, we present empirical results illustrating the performance of our approaches and the trade-off between fairness and transport cost.

Figures

Figures reproduced from arXiv: 2601.07144 by the authors.

Figure 1
Figure 1. Illustration of optimal transport under group fairness constraints. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Simulated datasets: Gaussians (left) and Circles (right). Fairness target. We aim to assign approximately 60% of underprivileged students to elite schools, which corresponds to the fairness target F = [ 0.20 0.30 0.28 0.22 ]. Approaches. We compare our three approaches: FairSinkhorn , penalized OT and cost learn￾ing. For cost learning, we evaluate two parameterizations: a Mahalanobis cost (Theorem 4.1) and a neural … view at source ↗
Figure 3
Figure 3. a. FairSinkhorn achieves perfect fairness with a high transport cost while Sinkhorn achieves low transport cost with low fairness. The penalized OT interpolates between Sinkhorn and FairSinkhorn . b. Cost-fairness trade-off of our penalized and cost-learning ap￾proaches on both datasets for varying fairness penalties. c. Vanilla entropic OT with different values of ε is included as a baseline. We also display the fa… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Left: Inference time of penalized OT and cost-learning approaches, highlighting the much faster inference of the cost-learning methods once the cost function is learned. Right: Fairness levels achieved on new samples using the learned cost function, with vanilla OT sho…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

80 extracted references · 14 linked inside Pith

  1. [32]

    Genevay.Entropy-regularized optimal transport for machine learning

    A. Genevay.Entropy-regularized optimal transport for machine learning. PhD thesis, Université Paris sciences et lettres, 2019. 7, 20

    2019

  2. [1]

    J. M. Altschuler, J. Niles-Weed, and A. J. Stromme. Asymptotics for semidiscrete entropic optimal transport.SIAM Journal on Mathematical Analysis, 54(2):1718–1741, 2022. 9

    2022

  3. [2]

    Andrade, G

    F. Andrade, G. Peyré, and C. Poon. Sparsistency for inverse optimal transport.arXiv preprint arXiv:2310.05461, 2023. 7

    arXiv 2023

  4. [3]

    Ashlagi and A

    I. Ashlagi and A. E. Roth. Free riding and participation in large scale, multi-hospital kidney exchange.Theoretical Economics, 9(3):817–863, 2014. 2, 3, 16

    2014

  5. [4]

    Barocas, M

    S. Barocas, M. Hardt, and A. Narayanan.Fairness and Machine Learning: Limitations and Opportunities. MIT Press, 2023. 1, 2

    2023

  6. [5]

    Bertsimas, V

    D. Bertsimas, V. F. Farias, and N. Trichakis. The price of fairness.Operations Research, 59 (1):17–31, 2011. doi: 10.1287/opre.1100.0865. 5

    arXiv 2011

  7. [6]

    Blondel, V

    M. Blondel, V. Seguy, and A. Rolet. Smooth and sparse optimal transport. InInternational Conference on Artificial Intelligence and Statistics (AISTATS), pages 880–889. PMLR, 2018. 3, 17

    2018

  8. [7]

    Bolte, E

    J. Bolte, E. Pauwels, and S. Vaiter. One-step differentiation of iterative algorithms. InAdvances in Neural Information Processing Systems (NeurIPS), volume 36, pages 77089–77103, 2023. 7 11

    2023

Show all 80 references
  1. [8]

    E. Budish. The combinatorial assignment problem: Approximate competitive equilibrium from equal incomes.Journal of Political Economy, 119(6):1061–1103, 2011. 16

    2011

  2. [9]

    Buyl and T

    M. Buyl and T. D. Bie. Optimal transport of classifiers to fairness. InAdvances in Neural Information Processing (NeurIPS), volume 35, pages 33728–33740, 2022. 3, 17

    2022

  3. [10]

    Calders, F

    T. Calders, F. Kamiran, and M. Pechenizkiy. Building classifiers with independency constraints. In2009 IEEE international conference on data mining workshops, pages 13–18. IEEE, 2009. 1

    2009

  4. [11]

    Carlier, A

    G. Carlier, A. Dupuy, A. Galichon, and Y. Sun. Sista: learning optimal transport costs under sparsity constraints.Communications on Pure and Applied Mathematics, 76(9):1659–1677,

  5. [12]

    Castera, F

    R. Castera, F. Garrido-Lucero, M. Molina, S. Mauras, P. Loiseau, and V. Perchet. The price of fairness in bipartite matching.arXiv preprint arXiv:2403.00397, 2024. 3, 16

    2024

  6. [13]

    Chiappa, R

    S. Chiappa, R. Jiang, T. Stepleton, A. Pacchiano, H. Jiang, and J. Aslanides. A general approach to fairness with optimal transport. InThe Thirty-Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applications of Artificial Intelligence ...

    2020 doi

  7. [14]

    Chowdhary, G

    S. Chowdhary, G. De Pasquale, N. Lanzetti, A.-A. Stoica, and F. Dorfler. Fairness in social influence maximization via optimal transport. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, pages 10380–10413, 2024

    2024

  8. [15]

    Chzhen, C

    E. Chzhen, C. Denis, M. Hebiri, L. Oneto, and M. Pontil. Fair regression with wasserstein barycenters. InAdvances in Neural Information Processing Systems (NeurIPS), volume 33, pages 7321–7331, 2020. 3, 17

    2020

  9. [16]

    Courty, R

    N. Courty, R. Flamary, D. Tuia, and A. Rakotomamonjy. Optimal transport for domain adaptation.IEEE transactions on pattern analysis and machine intelligence, 39(9):1853–1865,

  10. [17]

    I. Csiszár. i-divergence geometry of probability distributions and minimization problems.The Annals of Probability, 3(1):146–158, 1975. ISSN 00911798, 2168894X. URLhttp://www.jstor. org/stable/2959270. 3, 22

    1975

  11. [18]

    Dagréou, P

    M. Dagréou, P. Ablin, S. Vaiter, and T. Moreau. A framework for bilevel optimization that enables stochastic and global variance reduction algorithms. InAdvances in Neural Information Processing Systems (NeurIPS), volume 35, pages 26698–26710, 2022. 7

    2022

  12. [19]

    Dagréou, T

    M. Dagréou, T. Moreau, S. Vaiter, and P. Ablin. A Lower Bound and a Near-Optimal Algorithm for Bilevel Empirical Risk Minimization. InInternational Conference on Artificial Intelligence and Statistics (AISTATS), 2024. 7

    2024

  13. [20]

    Devic, D

    S. Devic, D. Kempe, V. Sharan, and A. Korolova. Fairness in matching under uncertainty. In International Conference on Machine Learning (ICML), pages 7775–7794. PMLR, 2023. 2, 16

    2023

  14. [21]

    J. P. Dickerson, A. D. Procaccia, and T. Sandholm. Price of fairness in kidney exchange. InProceedings of the 2014 international conference on Autonomous agents and multi-agent systems, pages 1013–1020, 2014. 3, 16

    2014

  15. [22]

    Divol and S

    V. Divol and S. Gaucher. Demographic parity in regression and classification within the unawareness framework.arXiv preprint arXiv:2409.02471, 2024. 3, 17

    2024 arXiv

  16. [23]

    Eisenberger, A

    M. Eisenberger, A. Toker, L. Leal-Taixe, F. Bernard, and D. Cremers. A Unified Framework for Implicit Sinkhorn Differentiation. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 499–508, New Orleans, LA, USA, 2022. IEEE. 7

    2022

  17. [24]

    G. Fack, J. Grenet, and A. Benhenda. L’impact des procédures de sectorisation et d’affectation sur la mixité sociale et scolaire dans les lycées d’île-de-france. Technical report, Institut des politiques publiques (IPP), 2014. 1 12

    2014

  18. [25]

    Feydy, T

    J. Feydy, T. Séjourné, F.-X. Vialard, S.-i. Amari, A. Trouvé, and G. Peyré. Interpolating between optimal transport and mmd using sinkhorn divergences. InInternational Conference on Artificial Intelligence and Statistics (AISTATS), pages 2681–2690. PMLR, 2019. 3

    2019

  19. [26]

    Flamary, N

    R. Flamary, N. Courty, A. Gramfort, M. Z. Alaya, A. Boisbunon, S. Chambon, L. Chapel, A. Corenflos, K. Fatras, N. Fournier, L. Gautheron, N. T. Gayraud, H. Janati, A. Rakotoma- monjy, I. Redko, A. Rolet, A. Schutz, V. Seguy, D. J. Sutherland, R. Tavenard, A. Tong, and T. Vayer...

  20. [27]

    Flamary, C

    R. Flamary, C. Vincent-Cuaz, N. Courty, A. Gramfort, O. Kachaiev, H. Quang Tran, L. David, C. Bonet, N. Cassereau, T. Gnassounou, E. Tanguy, J. Delon, A. Collas, S. Mazelet, L. Chapel, T. Kerdoncuff, X. Yu, M. Feickert, P. Krzakala, T. Liu, and E. Fernandes Montesuma. Pot pyth...

    2024

  21. [28]

    Franceschi, M

    L. Franceschi, M. Donini, P. Frasconi, and M. Pontil. Forward and Reverse Gradient-Based Hyperparameter Optimization. InInternational Conference on Machine Learning (ICML),

  22. [29]

    Gale and L

    D. Gale and L. S. Shapley. College admissions and the stability of marriage.The American mathematical monthly, 69(1):9–15, 1962. 2, 16

    1962

  23. [30]

    Galichon.Optimal transport methods in economics

    A. Galichon.Optimal transport methods in economics. Princeton University Press, 2018. 2

    2018

  24. [31]

    Galichon

    A. Galichon. The unreasonable effectiveness of optimal transport in economics.arXiv preprint arXiv:2107.04700, 2021. 2

    2021 arXiv

  25. [33]

    Genevay, L

    A. Genevay, L. Chizat, F. Bach, M. Cuturi, and G. Peyré. Sample complexity of sinkhorn divergences. InThe 22nd international conference on artificial intelligence and statistics, pages 1574–1583. PMLR, 2019. 3, 6, 17

    2019

  26. [34]

    Ghadimi and M

    S. Ghadimi and M. Wang. Approximation Methods for Bilevel Programming.arXiv preprint arXiv:1802.02246, 2018. 7

    2018 arXiv

  27. [35]

    C. D. Godsil. Matchings and walks in graphs.Journal of Graph Theory, 5(3):285–297, 1981. 3, 16

    1981

  28. [36]

    Gordaliza, E

    P. Gordaliza, E. Del Barrio, G. Fabrice, and J.-M. Loubes. Obtaining fairness using optimal transport theory. InInternational conference on machine learning, pages 2357–2365. PMLR,

  29. [37]

    T. L. Gouic, J.-M. Loubes, and P. Rigollet. Projection to fairness in statistical learning.arXiv preprint arXiv:2005.11720, 2020. 3, 17

    2005 arXiv

  30. [38]

    Hazard and T

    Y. Hazard and T. Kitagawa. Who with whom? learning optimal matching policies.arXiv preprint arXiv:2507.13567, 2025. 2

    2025 arXiv

  31. [39]

    Hiller and O

    V. Hiller and O. Tercieux. Choix d’écoles en france: une évaluation de la procédure affelnet,

  32. [40]

    F. Hu, P. Ratz, and A. Charpentier. Fairness in multi-task learning via wasserstein barycenters. InJoint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 295–312. Springer, 2023. 3, 17

    2023

  33. [41]

    Y. Jia, X. Liu, and W. Xu. When online dating meets nash social welfare: Achieving efficiency and fairness. InProceedings of the 2018 World Wide Web Conference, pages 429–438, 2018. 1

    2018

  34. [42]

    Karni, G

    G. Karni, G. N. Rothblum, and G. Yona. On fairness and stability in two-sided matchings. arXiv preprint arXiv:2111.10885, 2021. 2, 16 13

    2021 arXiv

  35. [43]

    N. Keriven. Entropic optimal transport in random graphs.arXiv preprint arXiv:2201.03949,

    arXiv

  36. [44]

    D. P. Kingma. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    arXiv

  37. [45]

    Korman and R

    J. Korman and R. McCann. Optimal transportation with capacity constraints.Transactions of the American Mathematical Society, 367(3):1501–1521, 2015. 3, 17

    2015

  38. [46]

    Leite and C

    D. Leite and C. De Bacco. Urban transportation networks and optimal transport-based infrastructures: similarity and economy of scale.arXiv preprint arXiv:2209.06751, 2022. 2

    2022 arXiv

  39. [47]

    R. Li, X. Ye, H. Zhou, and H. Zha. Learning to match via inverse optimal transport.Journal of Machine Learning Research, 20(80):1–37, 2019. URLhttp://jmlr.org/papers/v20/18-700. html. 7

    2019

  40. [48]

    T. Liu, J. Puigcerver, and M. Blondel. Sparsity-constrained optimal transport.arXiv preprint arXiv:2209.15466, 2022. 3, 17

    2022 arXiv

  41. [49]

    E. Lobo, J. Payan, C. Cousins, and Y. Zick. Fair and welfare-efficient constrained multi- matchings under uncertainty. InAdvances in Neural Information Processing Systems (NeurIPS), volume 37, pages 74579–74616, 2025. 3, 16

    2025

  42. [50]

    Maclaurin, D

    D. Maclaurin, D. Duvenaud, and R. P. Adams. Gradient-based Hyperparameter Optimization through Reversible Learning. InInternational Conference on Machine Learning (ICML), 2015. 7

    2015

  43. [51]

    Mastrandrea, P

    R. Mastrandrea, P. Pagnottoni, N. Pecora, and A. Spelta. An optimal transport approach to model the community structure of the international trade network.Social Networks, 82: 111–133, 2025. 2

    2025

  44. [52]

    Mena and J

    G. Mena and J. Niles-Weed. Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem.Advances in neural information processing systems, 32, 2019. 6, 7

    2019

  45. [53]

    Nadjahi, A

    K. Nadjahi, A. Durmus, L. Chizat, S. Kolouri, S. Shahrampour, and U. Simsekli. Statistical and topological properties of sliced probability divergences. InAdvances in Neural Information Processing Systems (NeurIPS), volume 33, pages 20802–20812, 2020. 17

    2020

  46. [54]

    Nguyen, H

    K. Nguyen, H. Nguyen, and N. Ho. Towards marginal fairness sliced wasserstein barycenter. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singa- pore, April 24-28, 2025. OpenReview.net, 2025. URLhttps://openreview.net/forum?id= NQqJPPCesd. 3, 17

    2025

  47. [55]

    Noiry, V

    N. Noiry, V. Perchet, and F. Sentenac. Online matching in sparse random graphs: Non- asymptotic performances of greedy algorithm.Advances in Neural Information Processing Systems, 34:21400–21412, 2021. 3, 16

    2021

  48. [56]

    Panda, A

    A. Panda, A. Louis, and P. Nimbhorkar. Individual fairness under group fairness constraints in bipartite matching-one framework to approximate them all. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 175–183, 2024. 3, 16

    2024

  49. [57]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. PyTorch: An Imperative Style, High-Per...

    2019

  50. [58]

    Paty and M

    F.-P. Paty and M. Cuturi. Subspace robust wasserstein distances. InInternational Conference on Machine Learning (ICML), pages 5072–5081. PMLR, 2019. 3, 17

    2019

  51. [59]

    Pauwels and S

    E. Pauwels and S. Vaiter. The Derivatives of Sinkhorn–Knopp Converge.SIAM Journal on Optimization, 33(3):1494–1517, 2023. ISSN 1052-6234, 1095-7189. 7, 25 14

    2023

  52. [60]

    Peyré, M

    G. Peyré, M. Cuturi, et al. Computational optimal transport: With applications to data science.Foundations and Trends®in Machine Learning, 11(5-6):355–607, 2019. 3

    2019

  53. [61]

    Pooladian and J

    A.-A. Pooladian and J. Niles-Weed. Entropic estimation of optimal transport maps.arXiv preprint arXiv:2109.12004, 2021. 3

    2021 arXiv

  54. [62]

    Z. Qu, A. Galichon, W. Gao, and J. Ugander. On sinkhorn’s algorithm and choice modeling. Operations Research, 2025. 2

    2025

  55. [63]

    Rakotomamonjy, R

    A. Rakotomamonjy, R. Flamary, and N. Courty. Generalized conditional gradient: analysis of convergence and applications.arXiv preprint arXiv:1510.06567, 2015. 3, 6, 7, 17, 22

    2015 arXiv

  56. [64]

    Rigollet and A

    P. Rigollet and A. J. Stromme. On the sample complexity of entropic optimal transport.arXiv preprint arXiv:2206.13472, 2022. 3, 6, 7, 8, 9, 17, 22, 23

    2022 arXiv

  57. [65]

    Samadi, U

    S. Samadi, U. Tantipongpipat, J. H. Morgenstern, M. Singh, and S. Vempala. The price of fair pca: One extra dimension. InAdvances in Neural Information Processing Systems (NeurIPS), volume 31, 2018. 2

    2018

  58. [66]

    Santambrogio

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

  59. [67]

    Simioni and P

    M. Simioni and P. Steiner.Comment ça matche: Une sociologie de l’appariement. Presses de Sciences Po, 2022. 1

    2022

  60. [68]

    Sinkhorn and P

    R. Sinkhorn and P. Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 21(2):343–348, 1967. ISSN 0030-8730, 0030-8730. 25

    1967

  61. [69]

    Steinhaus

    H. Steinhaus. Sur la division pragmatique.Econometrica: Journal of the Econometric Society, pages 315–319, 1949. 16

    1949

  62. [70]

    D. Weller. Fair division of a measurable space.Journal of Mathematical Economics, 14(1): 5–17, 1985. 16

    1985

  63. [71]

    R. Xian, L. Yin, and H. Zhao. Fair and optimal classification via post-processing. In A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett, editors,International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, volume 202...

    2023

  64. [72]

    Xiong, N

    Z. Xiong, N. Dalmasso, S. Sharma, F. Lecue, D. Magazzeni, V. Potluru, T. Balch, and M. Veloso. Fair wasserstein coresets.Advances in Neural Information Processing Systems, 37: 132–168, 2024. 3, 17

    2024

  65. [73]

    Zdeborová and M

    L. Zdeborová and M. Mézard. The number of matchings in random graphs.Journal of Statistical Mechanics: Theory and Experiment, 2006(05):P05003, 2006. 3, 16

    2006

  66. [74]

    Zhang, X

    M. Zhang, X. Dai, and L. Li. Fairness-aware organ exchange and kidney paired donation. arXiv preprint arXiv:2503.06431, 2025. 3, 16

    2025

  67. [75]

    Zhao and G

    H. Zhao and G. J. Gordon. Inherent tradeoffs in learning fair representations. InNeurIPS,

  68. [76]

    Y. Zhao, Y. Wang, Y. Zhang, P. Wisniewski, C. Aggarwal, and T. Derr. Leveraging opposite gender interaction ratio as a path towards fairness in online dating recommendations based on user sexual orientation. InProceedings of the AAAI Conference on Artificial Intelligence, page...

    2024

  69. [77]

    Zheng, T

    Y. Zheng, T. Dave, N. Mishra, and H. Kumar. Fairness in reciprocal recommendations: A speed-dating study. InAdjunct publication of the 26th conference on user modeling, adaptation and personalization, pages 29–34, 2018. 1 15 A Extended Related Work A.1 Fairness in Matching Mec...

    2018

  70. [79]

    analyze merit-based fairness of one-to-one matching mechanisms, a notion close to individual fairness, under partial uncertainty over individual’s merit. A second line of research introduces metrics of fairness closer to group fairness, with a specific focus on kidney paired d...

  71. [80]

    sandwich

    analyze a variant of optimal transport in which the amount of mass that can be transported between two units is upper bounded. Specifically, we build on Rakotomamonjy et al.[63], Genevay et al.[33] and Rigollet and Stromme[64] to derive finite sample guarantees for a penalized...

    2000

  72. [2021]

    URLhttp://jmlr.org/papers/v22/20-451.html. 25

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.