pith. machine review for the scientific record. sign in

arxiv: 2605.11944 · v1 · submitted 2026-05-12 · 🧮 math.OC

Recognition: 2 theorem links

· Lean Theorem

Generative Transfer for Entropic Optimal Transport with Unknown Costs

Antoine Debouchage, Francois Buet-Golfouse, Xiaozhen Wang, Zhenjie Ren

Pith reviewed 2026-05-13 05:06 UTC · model grok-4.3

classification 🧮 math.OC
keywords entropic optimal transportgenerative transferunknown costspath-wise tiltingconditional flow matchingWasserstein distanceconvergence rate
0
0 comments X

The pith

A generative framework transfers entropic optimal transport couplings to new marginals using samples from a reference coupling under an unknown shared cost.

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

The paper shows how to recover entropic optimal transport plans for new data distributions when the underlying cost function is not observed directly. It uses samples from one known optimal coupling as reference and develops an iterative generative process to adapt the coupling to the new marginals. This approach allows mass transport to extend beyond the original data support, unlike simpler reweighting methods. The method integrates flow matching techniques to produce practical samplers and includes proofs of convergence at rate O(δ) in Wasserstein distance.

Core claim

Given samples from a reference optimal coupling under a latent cost, an iterative path-wise tilting algorithm can be used to evolve the coupling to new marginals, yielding covariance-type equations for transport fields that, when integrated with conditional flow matching, generate the target EOT plan with global convergence rate O(δ) in W1 distance.

What carries the argument

The iterative path-wise tilting algorithm that evolves the coupling jointly with a marginal transport path, producing infinitesimal updates via covariance-type evolution equations for transport vector fields.

If this is right

  • The generated coupling converges to the target EOT plan in W1 distance at rate O(δ).
  • The joint evolution of coupling and transport path permits mass to move outside the support of the reference samples.
  • Sample-level learning rules derived from the tilting dynamics integrate with conditional flow matching to yield a practical paired-data sampler.
  • Global convergence guarantees apply to the full iterative process.

Where Pith is reading between the lines

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

  • Reference couplings can act as data-driven proxies for inferring latent costs when direct cost access is unavailable.
  • The covariance evolution equations may extend to other generative transport settings where plans must adapt to shifted marginals.
  • The approach enables sampling of paired observations consistent with the transferred EOT plan without recomputing costs from scratch.

Load-bearing premise

A single latent cost function is shared between the reference coupling and the new marginals and can be recovered sufficiently well through the iterative path-wise tilting process to allow mass to move beyond the reference support.

What would settle it

Running the tilting algorithm on synthetic data with known ground-truth cost, then measuring whether the W1 distance between the generated coupling and the exact target EOT plan for new marginals decreases proportionally to the step-size parameter δ.

Figures

Figures reproduced from arXiv: 2605.11944 by Antoine Debouchage, Francois Buet-Golfouse, Xiaozhen Wang, Zhenjie Ren.

Figure 1
Figure 1. Figure 1: Cost transfer via tilting along a marginal path. We observe a reference optimal coupling π ′ for (µ ′ , ν′ ) and seek the target optimal coupling π ⋆ for (µ, ν) under the same unknown cost. We bridge the two tasks through intermediate marginals (µt , νt) and couplings πt . EOT couplings along the path and infinitesimal tilts. Let πt be the entropic OT coupling between µt and νt for the same unknown cost c … view at source ↗
Figure 2
Figure 2. Figure 2: Transport map and coupling evolution on 2D-Moon. The crescent geometry has no analytical transport map, making evaluation purely metric-based. Our method (SW2 = 0.297) substantially outperforms all independent baselines (best: CFM† = 0.792), confirming that the cost-transfer approach generalises beyond Gaussian geometries. The coupling evolution shows the joint mass adapting from the reference crescent ali… view at source ↗
Figure 3
Figure 3. Figure 3: ODE trajectory evolution across training iterations (Colour MNIST-RGB simple experiment). Columns (left → right): forward model at different training iterations (Initial → Fi￾nal). Rows (top → bottom): ODE time t from source (t=0) to transported image (t=1), integrated with 200 Euler steps. The same PURPLE digit-7 source is used in every column. Adaptation progressively tilts the transport from the wrong c… view at source ↗
Figure 4
Figure 4. Figure 4: Proof dependency graph. Directed edges indicate logical dependencies; thick rose arrows (proves) lead into the final theorem nodes, while thin grey arrows denote depends on. B Assumptions In this section, we provide the theoretical guarantees for the discrete-time Weighted CFM algorithm. Our primary goal is to show that the accumulated error between the ideal continuous-time regression target β t s and the… view at source ↗
Figure 5
Figure 5. Figure 5: Transport map and coupling evolution on 2D-Simple. The pretrained model applies the reference anticorrelation map at the wrong location (near origin instead of (±10, ±10)), yielding SW2 = 14.59. After 50 outer iterations the model correctly recovers T(x)=−x at the new marginals, reaching SW2 = 0.025 — competitive with baselines that observe the ground-truth coupling. The coupling evolution (bottom) shows t… view at source ↗
Figure 6
Figure 6. Figure 6: Transport map and coupling evolution on 2D-Medium. The reference coupling encodes a 60◦ rotation on 2-blob Gaussians. The new task requires transferring this rotation to a 4-blob square arrangement, which the pretrained model handles poorly (SW2 ≈ 2.19). After 100 outer iterations, the adapted model (SW2 ≈0.756) is competitive with cost-aware baselines (best: OT-CFM† = 0.623). The coupling evolution shows … view at source ↗
Figure 7
Figure 7. Figure 7: Transport map and coupling evolution on 2D-Complex. The source is a five￾blob cross that is scaled 6.7× relative to the reference, making one-shot reweighting infeasible (the target is outside the reference support). The implicit-loss variant achieves the best SW2 = 0.544, improving 79% over the pretrained baseline (2.64) and outperforming cost-aware baselines on this geometry. N Perturbed Experiments In t… view at source ↗
Figure 8
Figure 8. Figure 8: Non-linear 2D-Circle. The transport maps a full circle to two spatially separated semi￾circles. The pretrained model spreads mass roughly evenly but the coupling is wrong (Sε = 1.958); adaptation correctly splits the mass and routes upper/lower particles to the respective semi-circle targets (Sε = 0.574). Q.2 CFM loss variants and the covariance update The coupling sampler update (Algorithm 1) trains the f… view at source ↗
Figure 9
Figure 9. Figure 9: Non-linear 2D-Radial warp. Ground-truth map: T(x) = x · (1 + 0.5∥x∥ 2 ). The SIREN-based model learns the non-linear radial expansion from the reference coupling, reducing Sinkhorn divergence from 1.557 to 0.384 without access to the explicit map formula. Explicit. The exponential weight is linearised via exp(h) ≈ 1 + h, giving the approximate factor hr = 1 + log wk. The new model is trained to match a pre… view at source ↗
Figure 10
Figure 10. Figure 10: Non-linear 2D-Polar twist. Ground-truth map: polar vortex r ′ = r, θ ′ = θ + sin(r). The pretrained model achieves superficially low SW2 = 0.054 but the Sinkhorn divergence 3.769 reveals a severely wrong coupling structure. Adaptation corrects this dramatically (Sε : 3.769 → 0.043, 99% reduction). Q.3 Method variant comparison [PITH_FULL_IMAGE:figures/full_fig_p040_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Transport quality metrics (2D-Simple). Sinkhorn divergence, MMD, energy dis￾tance, forward and backward marginal map errors, and forward and backward coupling-level map errors p Eπt [∥T(X) − Y ∥ 2], all per iteration. Shaded bands show mean ± 1 std across the best three seeds. All metrics decay steadily, confirming stable convergence of both the generator and the stored coupling πt . python -m dualwcfm.ex… view at source ↗
Figure 12
Figure 12. Figure 12: Wasserstein distances (2D-Simple). W1 and W2 between the transported distribu￾tion and the target (forward steps only), and the squared Bures–Wasserstein distance BW2 over all iterations. Shaded bands show mean ± 1 std across the best three seeds. All three distances con￾verge to near-zero, indicating that the adapted generator accurately matches the target marginal. 0 10 20 30 40 50 Iteration 10 1 SW1(^º… view at source ↗
Figure 13
Figure 13. Figure 13: Multi-seed stability. Mean ± 1 std over 5 independent seeds for 2D-Simple (blue), 2D-Complex (pink), and 2D-Moon (teal). Each seed uses independent pretraining to capture full pipeline variance. The shaded envelopes remain narrow throughout, confirming that TACO is robust to random initialisation. Experiment. In order to assess the veracity of the theorem on the convergence rate we delib￾erately design th… view at source ↗
Figure 14
Figure 14. Figure 14: Extended ablation results on 2D-Simple. (a) N = 50 is sufficient; gains plateau beyond N = 100 and the algorithm converges stably without overfitting. (b) Any α >0 is strongly preferred; α = 0.2 is the sweet spot balancing diversity and coupling quality. Large α = 1.0 (pure particles) performs well but is more expensive since the Euler bias dominates and lead to worse Sinkhorn divergence (see 15 below). (… view at source ↗
Figure 15
Figure 15. Figure 15: Final map error heatmap over (σ, α) on 2D-Simple. The heatmap confirms that the two hyperparameters interact mildly. A broad region around (σ= 0.1, α= 0.2) gives low map error; the main failure mode is α= 0 (any σ), which activates the circularity collapse. 8 16 32 64 128 Dimension d 0.06 0.08 0.10 0.12 0.14 W2(^º; º) 0:060 0:066 0:063 0:100 0:092 Scalability anti-correlation benchmark T(x) = ¡ x W2(º^; º… view at source ↗
Figure 16
Figure 16. Figure 16: Scalability. W2(ˆν, ν) (blue, left axis) remains near-constant across dimensions while BW2 (orange, right axis) grows as O( √ d). 56 [PITH_FULL_IMAGE:figures/full_fig_p056_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Empirical O(δ) convergence on 2D-Far. Map error (RMSE) at the final outer iteration vs. the number of iterations N (log–log scale). Filled circles: measured values; dashed line: least-squares C/N fit with C ≈ 5641. The grey right-triangle in the lower right marks the reference slope −1, with each leg spanning log10 3 ≈ 0.5 log-decades. The linear decay in δ = 1/N is consistent with the bound in Theorem 6.… view at source ↗
Figure 18
Figure 18. Figure 18: Single-cell transport (TACO). PCA visualisation of transported control cells (source) onto the treated distribution (target). Left: source control cells (µ). Centre: ground￾truth treated cells (ν). Right: TACO-transported cells (T♯µ). 58 [PITH_FULL_IMAGE:figures/full_fig_p058_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Convergence on single-cell transfer. Target-marginal W1 vs. outer iteration k for the standard (K562) and perturbed (MCF7 0.1 µM) settings. Both converge to low error within ≈30 iterations. 10 5 0 5 10 PC 1 6 4 2 0 2 4 6 8 PC 2 Pretrained K562 true K562 control Predicted 10 5 0 5 10 PC 1 6 4 2 0 2 4 6 8 PC 2 DualWCFM K562 true K562 control Predicted Coupling Arrows K562: Givinostat (K562) [PITH_FULL_IMAG… view at source ↗
Figure 20
Figure 20. Figure 20: Entropic coupling arrows. Each arrow connects a source control cell to its weighted centre of mass under the learned coupling π. Colours indicate transport distance. 59 [PITH_FULL_IMAGE:figures/full_fig_p059_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Evolution of transported distribution across TACO iterations. Each panel shows a PCA projection of the transported cells at a given outer iteration k. The distribution converges toward the ground-truth target within ≈30 iterations. SW MMD E-dist Sink. 0 50 100 150 200 250 300 350 400 Metric value (lower is better) 2.377 0.672 12.537 382.929 0.470 0.169 0.470 46.556 TACO final metric comparison Pretrained … view at source ↗
Figure 22
Figure 22. Figure 22: Single-cell transport metrics (bar chart). Grouped bars compare OT-CFM (scratch), pretrained baseline, and TACO across W1, W2, MMD, Energy, and Sinkhorn diver￾gence for the standard and perturbed settings. Lower is better for all metrics. 60 [PITH_FULL_IMAGE:figures/full_fig_p060_22.png] view at source ↗
read the original abstract

This paper addresses the practical challenge in Entropic Optimal Transport (EOT) where the underlying ground cost function is typically latent and unobserved. Rather than assuming a fixed geometric cost, we adopt a data-driven approach where a shared cost is revealed only through samples from a reference optimal coupling. The question is then: given samples from a reference optimal coupling, can we recover the optimal coupling for new marginals under the same latent cost? We introduce a generative transfer framework that recovers the optimal coupling for new marginals by utilizing an iterative path-wise tilting algorithm. Unlike static importance reweighting, this method evolves the coupling jointly with a marginal transport path, allowing mass to move beyond the reference support. We derive sample-level learning rules for these infinitesimal updates, which yield covariance-type evolution equations for the associated transport vector fields. By integrating this dynamics with Conditional Flow Matching (CFM), we produce a practical sampler for paired data. Finally, we provide theoretical guarantees establishing a global convergence rate of \mathcal{O}(\delta), ensuring the generated coupling converges to the target EOT plan in W_1 distance.

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 / 1 minor

Summary. The paper addresses Entropic Optimal Transport (EOT) with unknown latent costs by proposing a generative transfer framework. Given samples from a reference optimal coupling, an iterative path-wise tilting algorithm evolves the coupling jointly with a marginal transport path to recover the optimal coupling for new marginals under the shared cost. Sample-level learning rules are derived to yield covariance-type evolution equations for the transport vector fields; these are integrated with Conditional Flow Matching (CFM) to produce a practical paired-data sampler. Theoretical guarantees are provided for an O(δ) global convergence rate in W1 distance of the generated coupling to the target EOT plan.

Significance. If the O(δ) rate holds for the full generative pipeline (including CFM approximation), the work would provide a practical, data-driven method for transferring EOT plans across marginals without explicit cost specification. The path-wise tilting combined with covariance evolution equations and CFM integration offers a novel sampler construction, and the explicit convergence guarantee is a strength if the error terms are fully controlled.

major comments (2)
  1. [theoretical guarantees (abstract and convergence theorem)] The abstract and theoretical guarantees claim an O(δ) W1 convergence rate for the generated coupling produced by integrating the tilting dynamics with CFM. However, the analysis appears to control only the ideal continuous tilting path (assuming exact vector fields), without propagating the CFM training error, discretization error, or extrapolation error beyond the reference support into the final W1 bound. This is load-bearing for the central claim that the practical sampler converges at the stated rate.
  2. [path-wise tilting algorithm and learning rules] The weakest assumption—that a single latent cost function is shared and can be recovered sufficiently well via the iterative path-wise tilting to permit mass transport beyond the reference support—is stated but lacks explicit quantitative bounds or verification in the derivation of the evolution equations. This directly affects whether the sample-level learning rules remain valid for the new marginals.
minor comments (1)
  1. [abstract] The abstract could specify the dependence of δ on dimension, sample size, and training budget to clarify the practical scope of the O(δ) rate.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thorough review and constructive feedback on our manuscript. We value the positive remarks on the novelty of the path-wise tilting approach combined with CFM and the potential practical impact. We address each major comment below, clarifying the scope of our current analysis and outlining targeted revisions to strengthen the theoretical claims and assumptions.

read point-by-point responses

  1. Referee: [theoretical guarantees (abstract and convergence theorem)] The abstract and theoretical guarantees claim an O(δ) W1 convergence rate for the generated coupling produced by integrating the tilting dynamics with CFM. However, the analysis appears to control only the ideal continuous tilting path (assuming exact vector fields), without propagating the CFM training error, discretization error, or extrapolation error beyond the reference support into the final W1 bound. This is load-bearing for the central claim that the practical sampler converges at the stated rate.

    Authors: We appreciate the referee identifying this gap in the error propagation. Theorem 4.1 and the associated analysis establish the O(δ) rate specifically for the continuous-time tilting dynamics under exact vector fields obtained from the covariance evolution equations. The CFM component is introduced in Section 3.3 as a practical neural approximation to these fields, with the sampler in Algorithm 1 relying on it for implementation. We agree that the abstract and main claim would benefit from explicit control of approximation errors. In the revision, we will add a new corollary (Corollary 4.2) that decomposes the total W1 error into the ideal tilting term O(δ) plus additive terms for CFM training error (bounded via existing CFM generalization results by increasing training samples or network width), Euler discretization error (standard O(h) bound for step size h), and a support extrapolation term controlled by the tilting parameter δ. This preserves the core O(δ) rate for the dynamics while making the practical pipeline's convergence rigorous up to controllable terms. The abstract will be updated to reflect this decomposition. revision: yes

  2. Referee: [path-wise tilting algorithm and learning rules] The weakest assumption—that a single latent cost function is shared and can be recovered sufficiently well via the iterative path-wise tilting to permit mass transport beyond the reference support—is stated but lacks explicit quantitative bounds or verification in the derivation of the evolution equations. This directly affects whether the sample-level learning rules remain valid for the new marginals.

    Authors: The shared latent cost is the modeling foundation that allows the reference coupling samples to inform transport for new marginals; the path-wise tilting evolves the joint while enforcing the fixed-cost optimality condition at each step, and the covariance-type evolution equations (Eq. 3.4) are derived by differentiating the entropic OT dual along the marginal path without explicit cost estimation. The sample-level rules (Eq. 3.5) arise as Monte Carlo estimators of the resulting expectations. We concur that quantitative verification of implicit cost recovery would improve rigor, especially for mass transport outside the reference support. In the revised version, we will insert a new supporting lemma (Lemma 3.3) that bounds the deviation between the learned vector fields and the ideal fields in terms of reference sample size n and tilting increment δ, with high-probability guarantees under standard concentration assumptions on the reference samples. This lemma will be invoked to justify validity of the rules for new marginals and to tighten the extrapolation error term in the updated convergence analysis. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation builds from tilting dynamics to sampler and rate independently.

full rationale

The paper derives sample-level rules from the iterative path-wise tilting process to obtain covariance-type evolution equations for transport vector fields, then integrates those dynamics with Conditional Flow Matching to produce the generative sampler, and separately states an O(δ) W1 convergence guarantee for the generated coupling to the target EOT plan. No equations or steps are shown that reduce the claimed convergence rate, the sampler, or the vector fields to a fitted parameter or input quantity defined inside the paper by construction. The central guarantee is presented as a derived bound on the tilting path rather than a tautology, and no self-citation load-bearing, ansatz smuggling, or renaming of known results is evident in the provided derivation outline. The chain remains self-contained against external benchmarks for the tilting and CFM components.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on standard optimal transport existence results plus the domain assumption that a latent cost is consistent across reference and target marginals; no explicit free parameters or new invented entities with independent evidence are stated in the abstract.

axioms (2)
  • domain assumption Existence and uniqueness properties of entropic optimal couplings for given marginals
    Invoked implicitly to define both the reference coupling and the target EOT plan.
  • domain assumption The latent ground cost is identical for the reference and new marginals
    Core premise enabling transfer from reference samples to new marginals.
invented entities (1)
  • Path-wise tilting algorithm no independent evidence
    purpose: To evolve the coupling jointly with a marginal transport path
    New algorithmic component introduced to allow mass movement beyond reference support.

pith-pipeline@v0.9.0 · 5497 in / 1518 out tokens · 72610 ms · 2026-05-13T05:06:34.421418+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

Reference graph

Works this paper leans on

73 extracted references · 73 canonical work pages · 3 internal anchors

  1. [1]

    Topics in Optimal Transportation , series =

    Villani, C. Topics in Optimal Transportation , series =

    work page

  2. [2]

    Optimal Transport: Old and New , series =

    Villani, C. Optimal Transport: Old and New , series =. 2009 , doi =

    work page 2009

  3. [3]

    2019 , publisher=

    Computational optimal transport: With applications to data science , author=. 2019 , publisher=

    work page 2019

  4. [4]

    Santambrogio, Filippo , title =

    work page

  5. [5]

    Advances in Neural Information Processing Systems , volume =

    Cuturi, Marco , title =. Advances in Neural Information Processing Systems , volume =. 2013 , publisher =

    work page 2013

  6. [6]

    Lecture Notes, Columbia University , year =

    Nutz, Marcel , title =. Lecture Notes, Columbia University , year =

    work page

  7. [7]

    SIAM Journal on Mathematical Analysis , volume=

    Convergence of entropic schemes for optimal transport and gradient flows , author=. SIAM Journal on Mathematical Analysis , volume=. 2017 , publisher=

    work page 2017

  8. [8]

    Probability Theory and Related Fields , volume =

    Conforti, Giovanni , title =. Probability Theory and Related Fields , volume =

    work page

  9. [9]

    Identifiability of the Optimal Transport Cost on Finite Spaces , journal =

    Gonz. Identifiability of the Optimal Transport Cost on Finite Spaces , journal =

    work page

  10. [10]

    arXiv preprint arXiv:2109.12004 , year =

    Pooladian, Aram-Alexandre and Niles-Weed, Jonathan , title =. arXiv preprint arXiv:2109.12004 , year =

    work page arXiv

  11. [11]

    Large-Scale Optimal Transport and Mapping Estimation , booktitle =

    Seguy, Vivien and Damodaran, Bharath Bhushan and Flamary, R. Large-Scale Optimal Transport and Mapping Estimation , booktitle =

    work page

  12. [12]

    Advances in Neural Information Processing Systems , volume =

    Kassraie, Parnian and Pooladian, Aram-Alexandre and Klein, Michal and Thornton, James and Niles-Weed, Jonathan and Cuturi, Marco , title =. Advances in Neural Information Processing Systems , volume =

    work page

  13. [13]

    Advances in Neural Information Processing Systems , volume =

    Gushchin, Nikita and Kolesov, Alexander and Korotin, Alexander and Vetrov, Dmitry and Burnaev, Evgeny , title =. Advances in Neural Information Processing Systems , volume =. 2023 , note =

    work page 2023

  14. [14]

    Advances in Neural Information Processing Systems (Datasets and Benchmarks Track) , volume =

    Gushchin, Nikita and Kolesov, Alexander and Mokrov, Petr and Karpikova, Polina and Spiridonov, Andrey and Burnaev, Evgeny and Korotin, Alexander , title =. Advances in Neural Information Processing Systems (Datasets and Benchmarks Track) , volume =

    work page

  15. [15]

    International Conference on Learning Representations , year =

    Mokrov, Petr and Korotin, Alexander and Kolesov, Alexander and Gushchin, Nikita and Burnaev, Evgeny , title =. International Conference on Learning Representations , year =

    work page

  16. [16]

    International Conference on Learning Representations , year =

    Korotin, Alexander and Selikhanovych, Daniil and Burnaev, Evgeny , title =. International Conference on Learning Representations , year =

    work page

  17. [17]

    , title =

    Klein, Dominik and Prigent, Hugues and Cuturi, Marco and Peyré, Gabriel and Theis, Fabian J. , title =. Advances in Neural Information Processing Systems , volume =. 2024 , note =

    work page 2024

  18. [18]

    Advances in Neural Information Processing Systems , volume =

    Diffusion. Advances in Neural Information Processing Systems , volume =

    work page

  19. [19]

    Diffusion

    Shi, Yuyang and. Diffusion. Advances in Neural Information Processing Systems , volume =

    work page

  20. [20]

    2025 , eprint=

    Iterative Tilting for Diffusion Fine-Tuning , author=. 2025 , eprint=

    work page 2025

  21. [21]

    International Conference on Machine Learning , year =

    Gushchin, Nikita and Kholkin, Sergei and Burnaev, Evgeny and Korotin, Alexander , title =. International Conference on Machine Learning , year =

    work page

  22. [22]

    Advances in Neural Information Processing Systems , volume =

    De Bortoli, Valentin and Korshunova, Iryna and Mnih, Andriy and Doucet, Arnaud , title =. Advances in Neural Information Processing Systems , volume =. 2024 , note =

    work page 2024

  23. [23]

    International Conference on Learning Representations , year =

    Kim, Beomsu and Kwon, Gihyun and Kim, Kwanyoung and Ye, Jong Chul , title =. International Conference on Learning Representations , year =

    work page

  24. [24]

    International Conference on Learning Representations , year =

    Chen, Tianrong and Liu, Guan-Horng and Theodorou, Evangelos , title =. International Conference on Learning Representations , year =

    work page

  25. [25]

    Lipman, Yaron and Chen, Ricky T. Q. and Ben-Hamu, Heli and Nickel, Maximilian and Le, Matt , title =. International Conference on Learning Representations , year =

    work page

  26. [26]

    Lipman, Yaron and Chen, Ricky T. Q. and Ben-Hamu, Heli and Nickel, Maximilian and Le, Matt , title =. arXiv preprint arXiv:2210.02747 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    Improving and generalizing flow-based generative models with minibatch optimal transport

    Improving and generalizing flow-based generative models with minibatch optimal transport , author=. arXiv preprint arXiv:2302.00482 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Transactions on Machine Learning Research , year =

    Tong, Alexander and Malkin, Nikolay and Huguet, Guillaume and Zhang, Yanlei and Rector-Brooks, Jarrid and Fatras, Kilian and Wolf, Guy and Bengio, Yoshua , title =. Transactions on Machine Learning Research , year =

    work page

  29. [29]

    International Conference on Learning Representations , year =

    Liu, Xingchao and Gong, Chengyue and Liu, Qiang , title =. International Conference on Learning Representations , year =

    work page

  30. [30]

    and Vanden-Eijnden, Eric , title =

    Albergo, Michael S. and Vanden-Eijnden, Eric , title =. International Conference on Learning Representations , year =

    work page

  31. [31]

    Stochastic Interpolants: A Unifying Framework for Flows and Diffusions

    Albergo, Michael S. and Boffi, Nicholas M. and Vanden-Eijnden, Eric , title =. arXiv preprint arXiv:2303.08797 , year =

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    and Goldstein, Mark and Boffi, Nicholas M

    Albergo, Michael S. and Goldstein, Mark and Boffi, Nicholas M. and Ranganath, Rajesh and Vanden-Eijnden, Eric , title =. International Conference on Machine Learning , year =

    work page

  33. [33]

    International Conference on Artificial Intelligence and Statistics , year =

    Tong, Alexander and Malkin, Nikolay and Fatras, Kilian and Atanackovic, Lazar and Zhang, Yanlei and Huguet, Guillaume and Wolf, Guy and Bengio, Yoshua , title =. International Conference on Artificial Intelligence and Statistics , year =

    work page

  34. [34]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

    work page

  35. [35]

    and Kumar, Abhishek and Ermon, Stefano and Poole, Ben , title =

    Song, Yang and Sohl-Dickstein, Jascha and Kingma, Diederik P. and Kumar, Abhishek and Ermon, Stefano and Poole, Ben , title =. International Conference on Learning Representations , year =

    work page

  36. [36]

    and Mohamed, Shakir , title =

    Rezende, Danilo J. and Mohamed, Shakir , title =. International Conference on Machine Learning , volume =

    work page

  37. [37]

    International Conference on Machine Learning , volume =

    Amos, Brandon and Luise, Giulia and Cohen, Samuel and Redko, Ievgen , title =. International Conference on Machine Learning , volume =

    work page

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Supervised training of conditional monge maps , author=. Advances in Neural Information Processing Systems , volume=

    work page

  39. [39]

    arXiv preprint arXiv:2511.19741 , year=

    Efficient Transferable Optimal Transport via Min-Sliced Transport Plans , author=. arXiv preprint arXiv:2511.19741 , year=

    work page arXiv

  40. [40]

    Advances in Neural Information Processing Systems , volume=

    Mapping estimation for discrete optimal transport , author=. Advances in Neural Information Processing Systems , volume=

    work page

  41. [41]

    arXiv preprint arXiv:2512.21829 , year=

    Tilt Matching for Scalable Sampling and Fine-Tuning , author=. arXiv preprint arXiv:2512.21829 , year=

    work page arXiv

  42. [42]

    arXiv preprint arXiv:2410.02601 , year=

    Diffusion & adversarial schr " odinger bridges via iterative proportional markovian fitting , author=. arXiv preprint arXiv:2410.02601 , year=

    work page arXiv

  43. [43]

    Nature methods , volume=

    Learning single-cell perturbation responses using neural optimal transport , author=. Nature methods , volume=. 2023 , publisher=

    work page 2023

  44. [44]

    Cell , volume=

    Optimal-transport analysis of single-cell gene expression identifies developmental trajectories in reprogramming , author=. Cell , volume=. 2019 , publisher=

    work page 2019

  45. [45]

    International conference on machine learning , pages=

    Trajectorynet: A dynamic optimal transport network for modeling cellular dynamics , author=. International conference on machine learning , pages=. 2020 , organization=

    work page 2020

  46. [46]

    arXiv preprint arXiv:2504.08328 , year=

    Towards generalizable single-cell perturbation modeling via the Conditional Monge Gap , author=. arXiv preprint arXiv:2504.08328 , year=

    work page arXiv

  47. [47]

    Proceedings of the IEEE international conference on computer vision , pages=

    Unpaired image-to-image translation using cycle-consistent adversarial networks , author=. Proceedings of the IEEE international conference on computer vision , pages=

    work page

  48. [48]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Photo-realistic single image super-resolution using a generative adversarial network , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    work page

  49. [49]

    International Conference on Medical image computing and computer-assisted intervention , pages=

    U-net: Convolutional networks for biomedical image segmentation , author=. International Conference on Medical image computing and computer-assisted intervention , pages=. 2015 , organization=

    work page 2015

  50. [50]

    Nature methods , volume=

    CryoDRGN: reconstruction of heterogeneous cryo-EM structures using neural networks , author=. Nature methods , volume=. 2021 , publisher=

    work page 2021

  51. [51]

    arXiv preprint arXiv:1904.04554 , year=

    From (Martingale) Schrodinger bridges to a new class of Stochastic Volatility Models , author=. arXiv preprint arXiv:1904.04554 , year=

    work page arXiv 1904

  52. [52]

    Mathematics of Operations Research , year=

    Convergence of Sinkhorn’s algorithm for entropic martingale optimal transport problem , author=. Mathematics of Operations Research , year=

    work page

  53. [53]

    arXiv preprint arXiv:2503.11843 , year=

    Entropic Optimal Transport Problem with Convex Functional Cost , author=. arXiv preprint arXiv:2503.11843 , year=

    work page arXiv

  54. [54]

    BIT Numerical Mathematics , volume=

    A dynamical systems framework for intermittent data assimilation , author=. BIT Numerical Mathematics , volume=. 2011 , publisher=

    work page 2011

  55. [55]

    Journal of Geophysical Research: Oceans , volume=

    Sequential data assimilation with a nonlinear quasi-geostrophic model using Monte Carlo methods to forecast error statistics , author=. Journal of Geophysical Research: Oceans , volume=. 1994 , publisher=

    work page 1994

  56. [56]

    Journal of the Operational Research Society , volume=

    Nonlinear programming , author=. Journal of the Operational Research Society , volume=. 1997 , publisher=

    work page 1997

  57. [57]

    Journal of optimization theory and applications , volume=

    Convergence of a block coordinate descent method for nondifferentiable minimization , author=. Journal of optimization theory and applications , volume=. 2001 , publisher=

    work page 2001

  58. [58]

    International Conference on Machine Learning , pages=

    Optimal transport mapping via input convex neural networks , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    work page 2020

  59. [59]

    International Conference on Artificial Intelligence and Statistics , pages=

    Learning generative models with sinkhorn divergences , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2018 , organization=

    work page 2018

  60. [60]

    2015 , edition =

    Measure Theory and Fine Properties of Functions , author =. 2015 , edition =

    work page 2015

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Light unbalanced optimal transport , author=. Advances in Neural Information Processing Systems , volume=

    work page

  62. [62]

    Advances in Neural Information Processing Systems , volume=

    Equivariant flow matching , author=. Advances in Neural Information Processing Systems , volume=

    work page

  63. [63]

    2004 , publisher=

    Feynman-Kac formulae: genealogical and interacting particle systems with applications , author=. 2004 , publisher=

    work page 2004

  64. [64]

    Advances in neural information processing systems , volume=

    Maximum likelihood training of implicit nonlinear diffusion model , author=. Advances in neural information processing systems , volume=

    work page

  65. [65]

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , url =

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , booktitle =. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , url =

    work page

  66. [66]

    2018 , eprint=

    Demystifying MMD GANs , author=. 2018 , eprint=

    work page 2018

  67. [67]

    Implicit Neural Representations with Periodic Activation Functions , url =

    Sitzmann, Vincent and Martel, Julien and Bergman, Alexander and Lindell, David and Wetzstein, Gordon , booktitle =. Implicit Neural Representations with Periodic Activation Functions , url =

    work page

  68. [68]

    International conference on machine learning , pages=

    On the spectral bias of neural networks , author=. International conference on machine learning , pages=. 2019 , organization=

    work page 2019

  69. [69]

    Science , volume=

    Massively multiplex chemical transcriptomics at single-cell resolution , author=. Science , volume=. 2020 , publisher=

    work page 2020

  70. [70]

    Sliced and Radon Wasserstein Barycenters of Measures , journal =

    Bonneel, Nicolas and Rabin, Julien and Peyr. Sliced and Radon Wasserstein Barycenters of Measures , journal =. 2015 , volume =. doi:10.1007/s10851-014-0506-3 , url =

    work page doi:10.1007/s10851-014-0506-3 2015

  71. [71]

    Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics , pages =

    Interpolating between Optimal Transport and MMD using Sinkhorn Divergences , author =. Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics , pages =. 2019 , editor =

    work page 2019

  72. [72]

    2014 , eprint=

    Adam: A Method for Stochastic Optimization , author=. 2014 , eprint=

    work page 2014

  73. [73]

    and Bottou, L

    Lecun, Y. and Bottou, L. and Bengio, Y. and Haffner, P. , journal=. Gradient-based learning applied to document recognition , year=

    work page