Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

For Gaussian measures whose precision matrices share a sparse graph, the cluster moment relaxation computes the squared Wasserstein-2 cost with error at most C d^{3/2} ρ^{-h}, and exactly when the graph is complete.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 21:42 UTC pith:7AIERS53

load-bearing objection A real but narrow Gaussian convergence theorem, wrapped in an abstract that promises more than the manuscript delivers; worth a serious referee, but needs a scope-correction revision. the 3 major comments →

arxiv 2511.13847 v3 pith:7AIERS53 submitted 2025-11-17 math.OC cs.NAmath.NAstat.CO

Convex relaxation approaches for high-dimensional optimal transport

Yuehaw Khoo , Tianyun Tang This is my paper
classification math.OC cs.NAmath.NAstat.CO MSC 49Q2290C22
keywords optimal transportsemidefinite programmingmoment relaxationWasserstein distanceGaussian graphical modelscorrelative sparsitytransport mapsgenerative modeling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper attempts to make optimal transport (OT) tractable in high dimensions by relaxing the full coupling measure to a sparse collection of low-order marginals or cluster moments, yielding semidefinite programs that give lower bounds on the OT cost. Its main provable result is for Gaussian measures whose precision matrices share a sparse graph: the cluster moment relaxation with reference graph G_h approximates the squared Wasserstein-2 distance with error at most C d^{3/2} ρ^{-h}, decaying exponentially in the correlation radius h, and becomes exact when the reference graph is complete. The dual of the relaxation is a sum-of-squares approximation of the Kantorovich problem, so an approximate transport map can be read off from a dual potential. The authors show numerically on Gaussian, Ising, and Ginzburg–Landau models that the approach scales roughly linearly in dimension and is robust for non-Gaussian distributions. If the theorem holds, OT between locally correlated high-dimensional distributions becomes a small convex program rather than an exponentially large one.

Core claim

In the Gaussian setting, the paper's central discovery is Theorem 3.4: if both covariances have spectra in [a,b] and both precision matrices lie in the same sparsity class S(G,0), then the degree-one cluster moment relaxation (35) is exact when G is complete and satisfies |opt_{G_h} − W_2^2(µ,ν)| < C d^{3/2} ρ^{-h} for the power graph G_h, where C and ρ>1 depend only on a and b. The proof works by taking the closed-form covariance coupling that achieves the Wasserstein-2 cost, zeroing its entries outside G_h, and perturbing by multiples of the identity to restore feasibility of the dual; the size of the perturbation is controlled by the off-support norm of the coupling matrices. Because the

What carries the argument

The object doing the work is the cluster moment relaxation: replace the joint plan π by a block moment matrix M whose blocks M_k, M_{ij} are indexed by local cluster bases Φ_k, require M ⪰ 0 and consistency, and fix the entries determined by prescribed marginals of µ and ν. In the Gaussian case the relaxation collapses to the SDP (35) over second-moment blocks Z1, Z2, Y with sparsity enforced by [·]_{G,0}; the reference graph G_h encodes correlation radius h. The proof of exponential convergence uses polynomial approximation: the map t ↦ t^{1/2} (and its inverse) is approximated on the spectral interval [b^{-2}, a^{-2}] by Chebyshev polynomials, and because the precision matrices are sparse,

Load-bearing premise

Both Gaussian precision matrices must be supported on the same sparse graph and have spectra bounded away from zero and infinity; if correlations are long-range or the two sparsity patterns differ, the exponential-decay proof no longer applies and the relaxation may need a dense graph.

What would settle it

Choose two centered Gaussians in R^50 whose precision matrices are both tridiagonal but with a single edge different, so they do not share the same graph; solve (35) with G_h for h=1,...,5 and compare with formula (34). If the error fails to decay geometrically in h even though each matrix is individually sparse, that confirms the shared-sparsity premise is load-bearing; conversely, if exponential decay persists, the theorem's assumption is sufficient but not necessary.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • For Gaussian OT with a tree-like correlation graph, the squared Wasserstein-2 cost can be approximated to high accuracy using only moments within correlation radius h, with computational cost linear in dimension rather than quadratic.
  • When the reference graph is complete, the relaxation reproduces the closed-form Wasserstein-2 distance exactly, matching the fact that Gaussian measures are determined by first and second moments.
  • The relaxations always give lower bounds on the true OT cost, so they can be used to certify how far any other approximate solution is from optimal.
  • Because only low-order marginals or moments are needed, sample complexity no longer scales exponentially with dimension, unlike empirical-discretization OT.
  • The dual SOS formulation supplies approximate Kantorovich potentials and hence explicit transport maps, offering a convex alternative to neural-network approaches in generative modeling.
  • Chordal sparsity of the reference graph decomposes the SDP into small multi-block problems, with block size controlled by treewidth rather than ambient dimension.

Where Pith is reading between the lines

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

  • An implication the paper does not spell out: the exponential-rate result is evidence for a locality principle in OT—if the two measures have short-range dependence, the optimal coupling also has short-range dependence at the level of low-order moments; one could try to prove a general transfer theorem replacing Gaussianity with a suitable mixing or decay condition.
  • When the two precision matrices do not share the same graph, one could still run the relaxation on the union graph; the theorem suggests the error would then be controlled by how well the union graph captures the actual decay, giving a testable way to choose the reference graph from data.
  • The transport-map extraction via dual potentials is only guaranteed to be an approximation of the Monge map for absolutely continuous measures; in non-Gaussian experiments it performs well, but a rigorous extension would need a quantitative bound on ∇(φ − φ̂), not just on the cost.
  • The observed linear scaling in dimension suggests the bottleneck is graph-treewidth rather than dimension itself; choosing a reference graph with small treewidth that still respects correlation range is a modeling task the paper leaves open.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes convex relaxations for high-dimensional optimal transport, in two variants: a marginal relaxation that couples clusters through low-order marginals, and a cluster-moment relaxation that replaces measures by low-degree moment matrices and yields a semidefinite program. For Gaussian measures with a common sparse precision graph G, it proves that the degree-one cluster-moment relaxation is exact when the reference graph is complete (Theorem 3.4(i)) and, for reference graphs G_h built from graph distance h, gives the exponential error bound |opt_{G_h}-W_2^2| < C d^{3/2} ρ^{-h} (Theorem 3.4(ii)). The proof uses a feasible dual perturbation, Demko-Moss-Smith decay, and Chebyshev approximation. The paper also describes chordal sparsity conversion for the SDP, extraction of approximate transport maps from the dual, and numerical experiments for Gaussians, Ising models, and a Ginzburg-Landau generative-modeling task.

Significance. If the claims are taken exactly as proved, the Gaussian theorem is a substantive contribution: it gives a provable dimension-reduction guarantee for W_2 between Gaussian graphical models, with accuracy exponential in the correlation radius, and a clean chordal reduction of the SDP to small blocks. The lower-bound direction is sound because the true Gaussian coupling's moments are feasible; the upper-bound construction via a perturbed dual solution is self-contained and does not appear circular. The numerical experiments are suggestive and cover the advertised Gaussian and non-Gaussian regimes. The main weaknesses are in claim alignment: several results advertised in the abstracts and introduction are not proved in the full text, and the central theorem's shared-sparsity hypothesis is under-emphasized.

major comments (3)
  1. [Theorem 3.4 / Appendix A] The exponential bound is established only under the assumption that both precision matrices Σ1^{-1} and Σ2^{-1} belong to the same sparsity class S(G,0). The support argument in Eq. (72) — that [p_k(Σ1^{-1}Σ2^{-1})Σ1^{-1}]_{G_h,c}=0 for h≥2k+1 — relies on both factors having support in the same graph G. If the two precision matrices have different sparsity patterns, their products can become dense after a few factors and the projection onto G_{h,c} need not vanish; the stated rate is not established. The theorem itself states the shared-graph assumption, but the abstract and introduction say only 'Gaussian measures with sparse correlations,' which implies a broader scope. Please state the shared-sparsity condition prominently, and fix the ambiguous notation 'If G=G_h' by denoting the reference graph separately, e.g., H_h built from the base graph G.
  2. [Abstract / Section 1.3] The abstract claims an improved statistical error bound, and the introduction claims 'we prove that our method reduces both computational and sample complexity.' Theorem 3.4 is deterministic: it assumes the relevant moments are known exactly and bounds |opt_{G_h}-W_2^2|. I could find no theorem in Sections 3-6 or Appendix A that controls the error when moments are estimated from N samples, and no finite-sample complexity bound. Section 5.1 estimates moments empirically, but that is an experiment, not a proof. Either add a formal statistical bound under the Gaussian graphical-model assumptions or remove the sample-complexity/statistical-error claims from the abstract and introduction.
  3. [Abstract / full text] The arXiv metadata abstract additionally states: 'We also establish approximation error bounds for the marginal relaxation when the reference measures are local perturbations of mean-field measures.' No such theorem appears in the full text. The only theoretical result for the marginal relaxation is Proposition 3.2, which concerns chordal conversion; the only non-Gaussian validation is the numerical Ising study in Section 5.2. This is a material mismatch between advertised and delivered content. The missing theorem should be added with proof, or the abstract must be corrected.
minor comments (4)
  1. [Definition D4] The example 'let (u2,u3) ∼ η, P(u1,u2)(η) is the marginal of η on u2' is garbled; it should be a projection onto u2, and the notation P_{u2} should be introduced consistently.
  2. [Eqs. (59)-(64)] The proof of Lemma A.1 implicitly mean-centers the variables (subtracting m1m1^T, m1m2^T, etc.) but keeps the same symbols Z1,Z2,Y. Renaming the centered variables, e.g., \tilde Z_1,\tilde Z_2,\tilde Y, would prevent confusion with the original SDP (35).
  3. [Section 2.3] The consistency condition 'We say M is consistent' is described only verbally. Because consistency is a load-bearing constraint in (OT_mom) and (OT2_mom), it would help to state explicitly that entries of M corresponding to the same monomial in ΦΦ^T are equated (or to give the linear map A_c used in Section 4).
  4. [Minor] There are several minor presentation issues: 'Through this this paper' in Definition D4, 'M. Aps' in reference [3] should be 'MOSEK ApS', and the table headers in Section 5.2 are hard to parse (the tuple (J,h,β) is visually ambiguous).

Circularity Check

0 steps flagged

No significant circularity: the Gaussian convergence bound is derived from the closed-form W2 value, and no fitted parameter or self-citation is disguised as a prediction.

full rationale

The central result, Theorem 3.4, is derived from a genuine relaxation and compared against the closed-form squared Wasserstein distance for Gaussians. The relaxation (35) imposes constraints that are necessary for any feasible coupling, so the inequality opt_G ≤ W2^2(μ,ν) holds by construction; the lower-bound side is proved in Lemma A.1 by constructing feasible dual variables from the analytic Gaussian optimal coupling and bounding the projection error. The exponential rate in Theorem 3.4(ii) is obtained from Chebyshev polynomial approximation of x^{-1/2} on a spectral interval determined by a,b (Eqs. (68)–(75)), with constants C and ρ depending only on a,b. No constant is fitted to the target value W2^2, and the decay argument does not assume the conclusion. The complete-graph exactness in Theorem 3.4(i) follows from the fact that Gaussian measures are determined by first and second moments and from the closed-form W2 expression, not from the relaxation value being inserted into the statement. The self-citations [9], [25], [26] are methodological or historical and are not load-bearing; [47] supplies chordal conversion for computational reformulation, but the Gaussian error theorem does not depend on it. The shared-sparsity assumption on the two precision matrices is a substantive hypothesis, and the proof step in Eq. (72) would indeed require it, but this is a scope condition rather than circularity. The numerical experiments compare against exact or ground-truth values and do not rename fitted parameters as predictions. Overall, no circular step satisfying the quoted-reduction standard was found.

Axiom & Free-Parameter Ledger

3 free parameters · 8 axioms · 0 invented entities

The central Gaussian theorem rests on standard closed-form OT facts, polynomial approximation, and the domain assumption of shared sparse precision. No constants are fitted in the proof. In experiments, hyperparameters h, n, ω and the Sinkhorn entropy parameter are chosen by hand rather than data-fit, but they do affect the reported comparisons.

free parameters (3)
  • Correlation radius h of reference graph G_h = h=5 in Gaussian experiments; variable in Theorem 3.4
    Algorithm hyperparameter; not fitted to the ground truth in the theory, but the exponential rate and the size of the SDP depend on it.
  • Relaxation degree n = n=10 in the Ginzburg-Landau experiment; n=1 in the Gaussian analysis
    User-set hyperparameter controlling the moment basis size; no theory is given for how to choose it for non-Gaussian targets.
  • Sinkhorn entropy regularization parameter = 0.01 times the mean cost matrix
    Used only for the vanilla-OT baseline; this changes what 'vanilla OT' actually computes and is not swept or reported in detail.
axioms (8)
  • standard math Gaussian W2 closed form: W2^2(µ,ν) = ||m1-m2||^2 + Tr(Σ1 + Σ2 - 2(Σ1^{1/2}Σ2Σ1^{1/2})^{1/2})
    Used as exact ground truth in Eq. (34) and throughout Appendix A.
  • standard math Demko-Moss-Smith exponential decay of inverses of banded matrices
    Used to show the optimal Gaussian coupling covariance has small off-diagonal entries outside the sparsity graph G_h; cited as [14].
  • standard math Chebyshev polynomial approximation of x^{-1/2} on [b^{-2}, a^{-2}]
    Used to construct the polynomial p_k with exponential error in Eq. (69).
  • standard math Brenier's theorem: the optimal transport map is T(x)=x - (1/2)∇φ for quadratic cost
    Used in Section 4 to extract a transport map from the dual SDP potentials.
  • domain assumption Gaussian measures are fully determined by first and second moments
    Justifies reducing the general cluster moment relaxation to the degree-1 SDP (35) in the Gaussian case.
  • domain assumption The cost decomposes additively over clusters: c(z) = Σ_k c_k(z_k)
    Required for the marginal relaxation objective in Eq. (18); holds for quadratic cost with single-variable clusters.
  • domain assumption The precision matrices Σ1^{-1}, Σ2^{-1} share the same sparsity graph G and aI ⪯ Σ_i ⪯ bI
    Stated explicitly in Theorem 3.4; the exponential decay argument collapses if the sparsity patterns differ or correlations are long-range.
  • ad hoc to paper Low-order local moments suffice for non-Gaussian distributions
    Supported only by the numerical experiments in Section 5; no theorem is proved for non-Gaussian targets, so this is an empirical assumption of the method's practical reach.

pith-pipeline@v1.3.0-alltime-deepseek · 22864 in / 19174 out tokens · 167872 ms · 2026-08-03T21:42:50.116721+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Convex relaxation approaches for high-dimensional optimal transport." pith.science (2026). https://pith.science/paper/7AIERS53

@misc{pith2026251113847,
  author       = {Pith},
  title        = {Pith review of: Convex relaxation approaches for high-dimensional optimal transport},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7AIERS53}},
  note         = {Machine review of arXiv:2511.13847}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Optimal transport (OT) is a powerful tool in mathematics and data science but faces severe computational and statistical challenges in high dimensions. We propose convex relaxation approaches based on marginal and cluster moment relaxations that exploit locality in the distributions. These methods approximate high-dimensional couplings using low-order marginals and sparse moment statistics, yielding semidefinite programs that provide lower bounds on the OT cost with greatly reduced complexity. For Gaussian measures with sparse correlations, we prove an exponential convergence rate for the cluster moment relaxation and an improved statistical error bound. We also establish approximation error bounds for the marginal relaxation when the reference measures are local perturbations of mean-field measures. In addition, we demonstrate how to extract transport maps from our relaxations, offering a simpler and interpretable alternative to neural networks in generative modeling. Extensive numerical experiments demonstrate strong empirical performance across a range of distributions. Our results suggest that convex relaxations can provide a promising path for dimension reduction in high-dimensional OT.

Figures

Figures reproduced from arXiv: 2511.13847 by Tianyun Tang, YueHaw Khoo.

Figure 1
Figure 1. Figure 1: Step 1. Partition x and y into K clusters (x1; x2; · · · ; xK) and (y1; y2; · · · ; yK). Construct a local coupling within each cluster: zk = (xk, yk) ∼ πk. The mean-field approximation ⊗K k=1πk of π is exact if the zk’s are independent. See the following example: G : S(G, 0) :=      a b 0 b c d 0 d e      , S(G, c) :=      0 0 f 0 0 0 f 0 0      (8) We say that G is a sparsity pattern … view at source ↗
Figure 2
Figure 2. Figure 2: Step 2. Add pairwise couplings between correlated clusters: (zi , zj) ∼ πij , consistent with marginals Pzi (πij ) = πi ,Pzj (πij ) = πj and satisfying the PSD constraint (15). In this example, the reference graph G (Definition D3) is a triangle. The problem (OT) can be equivalently written as follows inf π π(c) (OT) s.t. Px(π) = µ, Py(π) = ν (1a) π ∈ P(Z). (1b) We will relax the above conditions (1a) and … view at source ↗
Figure 3
Figure 3. Figure 3: Cluster moment relaxation (35) for OT between Gaussian distributions. The correlation sparsity pattern G of Gaussian distributions is a path. The reference graph G is chosen as Gh for various correlation radius h (Definition D3) [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of cluster moment relaxation (OT2 mom) and the vanilla OT solver for (OT) between Gaussian distributions. The reference graph for (OT2 mom) is G = Gh (Definition D3), such that G is a path graph and the correlation radius h = 5 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of marginal relaxation and traditional OT solver for (OT) between Ising models with (J1, h1, β1) = (J2, h2, β2) = (−1, 0.2, 0.3). ω = 1 is the size of clusters in marginal relaxation defined in (54) and (55). The reference graph G (Definition D3) is a path graph. The first panel in [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of the cluster moment relaxation (OT2 mom) and a neural network method for generative modeling of the 1D Ginzburg–Landau model (d = 10, β = 1/8, L = 2.5, λ = 0.03). For the cluster moment relaxation, we set the number of clusters K = 10 and relaxation degree n = 10, and use a path graph as the reference graph G (Definition D3). The first and fourth panels show the empirical marginals of referenc… view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of the cluster moment relaxation (OT2 mom) and a neural network method for generative modeling of the 1D Ginzburg–Landau model (d = 50, β = 1/20, L = 2.5, λ = 0.01). For the cluster moment relaxation, we set the number of clusters K = 50 and relaxation degree n = 10, and use a path graph as the reference graph G (Definition D3). The first and fourth panels show the empirical marginals of referen… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Convex Relaxations for the Optimization of Markov Processes

    math.OC 2026-07 conditional novelty 6.0

    Sequential-coupling convex relaxations using local marginals and cluster moments solve high-dimensional Markov process optimization, recovering Benamou–Brenier dynamics and general kernels as special cases.

Reference graph

Works this paper leans on

55 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Agler, W

    J. Agler, W. Helton, S. McCullough, and L. Rodman. Positi ve semidefinite matrices with a given sparsity pattern. Linear algebra and its applications , 107:101–149, 1988

    1988

  2. [2]

    G. An. A note on the cluster variation method. Journal of Statistical Physics , 52(3):727–734, 1988

    1988

  3. [3]

    M. Aps. Mosek optimization toolbox for matlab. User’s Guide and Reference Manual Version, 4(1):116, 2019

    2019

  4. [4]

    Arjovsky, S

    M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gen erative adversarial networks. In International conference on machine learning , pages 214–223. PMLR, 2017

    2017

  5. [5]

    Balaji, R

    Y. Balaji, R. Chellappa, and S. Feizi. Robust optimal tra nsport with applications in generative modeling and domain adaptation. Advances in Neural Information Process- ing Systems , 33:12934–12944, 2020

    2020

  6. [6]

    Bonneel, J

    N. Bonneel, J. Rabin, G. Peyr´ e, and H. Pfister. Sliced and radon wasserstein barycen- ters of measures. Journal of Mathematical Imaging and Vision , 51(1):22–45, 2015

    2015

  7. [7]

    Y. Brenier. Polar factorization and monotone rearrange ment of vector-valued func- tions. Communications on pure and applied mathematics , 44(4):375–417, 1991

    1991

  8. [8]

    Calder and N

    J. Calder and N. G. Trillos. Improved spectral convergen ce rates for graph laplacians onε-graphs and k-nn graphs. Applied and Computational Harmonic Analysis , 60:123– 175, 2022. 26

    2022

  9. [9]

    Y. Chen, Y. Khoo, and L.-H. Lim. Convex relaxation for fok ker–planck equation. In Proceedings A, volume 481, page 20240001. The Royal Society, 2025

    2025

  10. [10]

    Y. Chen, Y. Khoo, and M. Lindsey. Multiscale semidefinit e programming approach to positioning problems with pairwise structure. Journal of Scientific Computing , 101(2):42, 2024

    2024

  11. [11]

    Cheng, J

    X. Cheng, J. Lu, Y. Tan, and Y. Xie. Convergence of flow-ba sed generative models via proximal gradient descent in wasserstein space. IEEE Transactions on Information Theory, 70(11):8087–8106, 2024

    2024

  12. [12]

    Cloninger, K

    A. Cloninger, K. Hamm, V. Khurana, and C. Moosm¨ uller. L inearized wasserstein dimensionality reduction with approximation guarantees. Applied and Computational Harmonic Analysis , 74:101718, 2025

    2025

  13. [13]

    M. Cuturi. Sinkhorn distances: Lightspeed computatio n of optimal transport. Ad- vances in Neural Information Processing Systems , 26, 2013

    2013

  14. [14]

    Demko, W

    S. Demko, W. F. Moss, and P. W. Smith. Decay rates for inve rses of band matrices. Mathematics of computation , 43(168):491–499, 1984

    1984

  15. [15]

    Dolgov, K

    S. Dolgov, K. Anaya-Izquierdo, C. Fox, and R. Scheichl. Approximation and sampling of multivariate probability distributions in the tensor tr ain decomposition. Statistics and Computing , 30(3):603–625, 2020

    2020

  16. [16]

    R. M. Dudley. The speed of mean glivenko-cantelli conve rgence. The Annals of Math- ematical Statistics , 40(1):40–50, 1969

    1969

  17. [17]

    Fournier and A

    N. Fournier and A. Guillin. On the rate of convergence in wasserstein distance of the empirical measure. Probability theory and related fields , 162(3):707–738, 2015

    2015

  18. [18]

    Friesecke

    G. Friesecke. Optimal Transport: A Comprehensive Introduction to Modeling , Analysis, Simulation, Applications. SIAM, 2024

    2024

  19. [19]

    Genevay, G

    A. Genevay, G. Peyr´ e, and M. Cuturi. Learning generati ve models with sinkhorn divergences. In International Conference on Artificial Intelligence and Stati stics, pages 1608–1617. PMLR, 2018

    2018

  20. [20]

    Grone, C

    R. Grone, C. R. Johnson, E. M. S´ a, and H. Wolkowicz. Posi tive definite completions of partial Hermitian matrices. Linear algebra and its applications , 58:109–124, 1984

    1984

  21. [21]

    D. Hou, T. Tang, and K.-C. Toh. A low-rank augmented lagr angian method for doubly nonnegative relaxations of mixed-binary quadratic progra ms. Operations Research, 2025

    2025

  22. [22]

    D. Hou, T. Tang, and K.-C. Toh. Rinnal+: a riemannian alm solver for sdp-rlt relaxations of mixed-binary quadratic programs. arXiv preprint arXiv:2507.13776 , 2025. 27

    Pith/arXiv arXiv 2025

  23. [23]

    H¨ utter and P

    J.-C. H¨ utter and P. Rigollet. Minimax estimation of sm ooth optimal transport maps. The Annals of Statistics , 49(2), 2021

    2021

  24. [24]

    L. V. Kantorovich. On the translocation of masses. In Dokl. Akad. Nauk. USSR (NS) , volume 37, pages 199–201, 1942

    1942

  25. [25]

    Y. Khoo, L. Lin, M. Lindsey, and L. Ying. Semidefinite rel axation of multimarginal op- timal transport for strictly correlated electrons in secon d quantization. SIAM Journal on Scientific Computing , 42(6):B1462–B1489, 2020

    2020

  26. [26]

    Khoo and L

    Y. Khoo and L. Ying. Convex relaxation approaches for st rictly correlated density functional theory. SIAM Journal on Scientific Computing , 41(4):B773–B795, 2019

    2019

  27. [27]

    S. Kim, M. Kojima, M. Mevissen, and M. Yamashita. Exploi ting sparsity in linear and nonlinear matrix inequalities via positive semidefinite ma trix completion. Mathematical programming, 129(1):33–68, 2011

    2011

  28. [28]

    Korotin, D

    A. Korotin, D. Selikhanovych, and E. Burnaev. Neural op timal transport. arXiv preprint arXiv:2201.12220, 2022

    Pith/arXiv arXiv 2022

  29. [29]

    Kullback and R

    S. Kullback and R. A. Leibler. On information and sufficie ncy. The Annals of Mathe- matical Statistics, 22(1):79–86, 1951

    1951

  30. [30]

    J. B. Lasserre. Global optimization with polynomials a nd the problem of moments. SIAM Journal on optimization , 11(3):796–817, 2001

    2001

  31. [31]

    J. B. Lasserre. A semidefinite programming approach to t he generalized problem of moments. Mathematical Programming, 112(1):65–92, 2008

    2008

  32. [32]

    J. B. Lasserre. Moments, positive polynomials and their applications , volume 1. World Scientific, 2009

    2009

  33. [33]

    T. Lin, C. Fan, N. Ho, M. Cuturi, and M. Jordan. Projectio n robust wasserstein distance and riemannian optimization. Advances in neural information processing sys- tems, 33:9383–9397, 2020

    2020

  34. [34]

    Makkuva, A

    A. Makkuva, A. Taghvaei, S. Oh, and J. Lee. Optimal trans port mapping via input convex neural networks. In International Conference on Machine Learning , pages 6672–6681. PMLR, 2020

    2020

  35. [35]

    Mokrov, A

    P. Mokrov, A. Korotin, L. Li, A. Genevay, J. M. Solomon, a nd E. Burnaev. Large- scale wasserstein gradient flows. Advances in Neural Information Processing Systems , 34:15243–15256, 2021

    2021

  36. [36]

    G. Monge. M´ emoire sur la th´ eorie des d´ eblais et des re mblais. Mem. Math. Phys. Acad. Royale Sci. , pages 666–704, 1781

  37. [37]

    Mula and A

    O. Mula and A. Nouy. Moment-sos methods for optimal tran sport problems. Nu- merische Mathematik , 156(4):1541–1578, 2024. 28

    2024

  38. [38]

    Papamakarios, T

    G. Papamakarios, T. Pavlakou, and I. Murray. Masked aut oregressive flow for density estimation. Advances in neural information processing systems , 30, 2017

    2017

  39. [39]

    Paty and M

    F.-P. Paty and M. Cuturi. Subspace robust wasserstein d istances. In International conference on machine learning , pages 5072–5081. PMLR, 2019

    2019

  40. [40]

    Pelizzola

    A. Pelizzola. Cluster variation method in statistical physics and probabilistic graphical models. Journal of Physics A: Mathematical and General , 38(33):R309, 2005

    2005

  41. [41]

    J. Peng, T. Hazan, N. Srebro, and J. Xu. Approximate infe rence by intersecting semidefinite bound and local polytope. In Artificial Intelligence and Statistics , pages 868–876. PMLR, 2012

    2012

  42. [42]

    Y. Peng, S. Yang, Y. Khoo, and D. Wang. Tensor density est imator by convolution- deconvolution. arXiv preprint arXiv:2412.18964 , 2024

    arXiv 2024

  43. [43]

    Peyr´ e, M

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

    2019

  44. [44]

    Rabin, G

    J. Rabin, G. Peyr´ e, J. Delon, and M. Bernot. Wasserstei n barycenter and its appli- cation to texture mixing. In International conference on scale space and variational methods in computer vision , pages 435–446. Springer, 2011

    2011

  45. [45]

    Scarvelis and J

    C. Scarvelis and J. Solomon. Riemannian metric learnin g via optimal transport. In International Conference on Learning Representations . OpenReview, 2023

    2023

  46. [46]

    Sun, K.-C

    D. Sun, K.-C. Toh, Y. Yuan, and X.-Y. Zhao. Sdpnal+: A mat lab software for semidef- inite programming with bound constraints (version 1.0). Optimization Methods and Software, 35(1):87–115, 2020

    2020

  47. [47]

    Tang and K.-C

    T. Tang and K.-C. Toh. Exploring chordal sparsity in sem idefinite programming with sparse plus low-rank data matrices. arXiv preprint arXiv:2410.23849 , 2024

    Pith/arXiv arXiv 2024

  48. [48]

    Tolstikhin, O

    I. Tolstikhin, O. Bousquet, S. Gelly, and B. Schoelkopf . Wasserstein auto-encoders. arXiv preprint arXiv:1711.01558 , 2017

    Pith/arXiv arXiv 2017

  49. [49]

    L. N. Trefethen. Approximation theory and approximation practice, extende d edition . SIAM, 2019

    2019

  50. [50]

    Vacher, B

    A. Vacher, B. Muzellec, A. Rudi, F. Bach, and F.-X. Viala rd. A dimension-free com- putational upper-bound for smooth optimal transport estim ation. In Conference on Learning Theory, pages 4143–4173. PMLR, 2021

    2021

  51. [51]

    C. Villani. Topics in optimal transportation , volume 58. American Mathematical Soc., 2021

    2021

  52. [52]

    Villani et al

    C. Villani et al. Optimal transport: old and new , volume 338. Springer, 2008

    2008

  53. [53]

    M. J. Wainwright, M. I. Jordan, et al. Graphical models, exponential families, and variational inference. Foundations and Trends ® in Machine Learning , 1(1–2):1–305, 2008. 29

    2008

  54. [54]

    Weed and F

    J. Weed and F. Bach. Sharp asymptotic and finite-sample r ates of convergence of empirical measures in wasserstein distance. Bernoulli, 25(4A):2620–2648, 2019

    2019

  55. [55]

    Weed and Q

    J. Weed and Q. Berthet. Estimation of smooth densities i n wasserstein distance. In conference on Learning Theory , pages 3118–3119. PMLR, 2019. 30

    2019