Two-Parameter Flows for Learning Population Dynamics of Physical Systems

Benjamin Peherstorfer; Paul Schwerdtner; Tobias Blickhan

arxiv: 2605.26285 · v1 · pith:W5R4TTENnew · submitted 2026-05-25 · 💻 cs.LG · cs.NA· math.NA

Two-Parameter Flows for Learning Population Dynamics of Physical Systems

Paul Schwerdtner , Tobias Blickhan , Benjamin Peherstorfer This is my paper

Pith reviewed 2026-06-29 22:39 UTC · model grok-4.3

classification 💻 cs.LG cs.NAmath.NA

keywords two-parameter flowspopulation dynamicsconditional flow matchingprobability densitiesphysics-informed machine learningoptimal transport

0 comments

The pith

Two-parameter flows recover unique physics-time velocity fields for population dynamics by regressing on synthetic trajectories built from base-to-marginal transports.

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

The paper introduces two-parameter flows to learn how high-dimensional probability densities evolve over time when only unlabeled samples from each time marginal are available. It first learns independent transports from a fixed base distribution to each marginal using standard conditional flow matching, then builds coupled synthetic trajectories and regresses to extract a continuous physics-time velocity field. The authors establish that this velocity is unique and inherits regularity from the learned transports. A reader would care because the construction scales to high dimensions, skips the need for optimal-transport solves at every step, and permits non-gradient flows that can capture rotational or circulating behavior in physical systems.

Core claim

The paper claims that two-parameter flows, which learn sampling-time transports from a base distribution to successive marginals and then regress a velocity on the resulting synthetic trajectories, produce a unique physics-time dynamics that recovers the underlying population evolution and inherits regularity properties from the sampling-time maps.

What carries the argument

Two-parameter flows that first learn base-to-marginal transports at discrete sampling times via conditional flow matching and then regress a physics-time velocity field on coupled synthetic trajectories.

If this is right

The extracted physics-time dynamics are unique.
They inherit regularity from the sampling-time transports.
The method scales to high dimensions without per-step optimal-transport couplings.
It permits admissible non-gradient dynamics that can explain rotational or circulating phenomena.

Where Pith is reading between the lines

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

The construction could be applied to any domain where only snapshot distributions are observed, such as certain experimental measurements in fluid or particle systems.
Because it re-uses existing conditional flow matching implementations, the overhead of adding the regression step may be modest in practice.
One could check whether the learned velocities satisfy additional physical constraints, such as divergence-free conditions in incompressible flow, on benchmark problems where ground-truth velocities are known.

Load-bearing premise

That regressing on coupled synthetic trajectories constructed from independently learned base-to-marginal transports yields a well-defined, unique physics-time velocity field that correctly recovers the underlying population dynamics.

What would settle it

A concrete case in which two distinct physics-time velocity fields generate identical sequences of marginal distributions, yet the regression procedure returns only one of them, or where the extracted velocity fails to reproduce the observed marginal evolution when integrated forward.

Figures

Figures reproduced from arXiv: 2605.26285 by Benjamin Peherstorfer, Paul Schwerdtner, Tobias Blickhan.

**Figure 2.** Figure 2: Snapshots from a curve t 7→ ρ(t) (left to right: t ∈ [0, 1]). The first row displays piecewise optimal transport: samples retain their color for all t and move between marginals following piecewise OT as in (Terpin et al., 2024). The second row shows that TPF dynamics are similar, but more regular: Especially for later times, the flow defined by successive OT mappings develops sharp gradients in dxt,1/da. … view at source ↗

**Figure 3.** Figure 3: Barotropic flow: Our two-parameter flow approach scales population-dynamics inference to turbulent systems with states in d > 104 dimensions. It learns a fast-to-evaluate population-dynamics model that bypasses fine-scale advection motion (see lower kinetic energy of TPF model, which is emphasized in the trajectory videos found in the supplementary material) while preserving the self-organization/vortex-me… view at source ↗

**Figure 4.** Figure 4: Particle instabilities: Our TPF method learns population dynamics that accurately generalize across unseen Debye-length parameters in these Vlasov-Poisson instability benchmarks and match fine phase-space structure. means and covariances undergo a random walk over t. We train a two-parameter flow model as well as a flow with optimal-transport couplings between successive time steps following Terpin et al. … view at source ↗

read the original abstract

This work addresses the problem of learning the dynamics of high-dimensional probability densities over time using unlabeled samples, without assuming access to trajectory information. We introduce two-parameter flows that learn only sampling-time transports from a base distribution to each marginal and then extract a physics-time velocity by regressing on coupled synthetic trajectories. We prove that the resulting physics-time dynamics are unique and inherit regularity from the sampling-time transports. Because we can build on standard, well-developed conditional flow matching techniques for learning the base-to-marginal transports, our approach scales to high dimensions and avoids per-step optimal-transport couplings, while allowing admissible non-gradient dynamics that can naturally explain rotational or circulating physics phenomena.

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.

Desk Editor's Note private letter to a colleague

The two-parameter construction decouples transport learning from velocity regression in a workable way, but the uniqueness claim for the physics-time field rests on a regression step whose robustness to independent fitting errors is not yet obvious from the abstract.

read the letter

The paper's main move is to learn independent base-to-marginal transport maps at each time via conditional flow matching, push fixed base samples through those maps to create synthetic trajectories, and then regress a single velocity field on the resulting pairs to recover the physics-time dynamics. This splits the problem so that standard high-dimensional flow matching tools can be reused without needing paired trajectory data or repeated optimal transport at each step. The approach also keeps the door open for non-gradient vector fields, which is relevant for rotational or circulating behavior in fluids and similar systems.

What the work does cleanly is leverage existing conditional flow matching machinery for the sampling-time part, which immediately gives scalability in high dimensions and avoids some of the computational bottlenecks in trajectory-based or per-step coupling methods. The claimed uniqueness of the resulting physics-time velocity and the inheritance of regularity from the transport maps are presented as theorems, which at least makes the central guarantee explicit rather than purely empirical.

The soft spot is the regression step itself. Because each transport map is learned separately, the synthetic trajectories are not guaranteed to be consistent with any single continuous vector field; approximation errors or non-commuting maps across times could make the regression either ill-posed or produce a field whose induced continuity equation deviates from the observed marginals. The abstract asserts a uniqueness proof, but the stress-test concern about hidden inconsistencies in the coupling construction is the part that needs the most direct verification in the full derivation and experiments. If the proof rules this out under realistic assumptions and the numerical results hold up, the method is on solid ground; otherwise it functions more as a practical heuristic.

This is aimed at people working on learning continuous dynamics from unlabeled snapshots in scientific machine learning, particularly in fluid mechanics or molecular simulation. A reader already comfortable with flow matching will see the extension quickly and can judge whether the regression step adds enough reliability for their use case. It deserves a serious referee because the construction is specific, the claims are checkable, and the practical advantages are clear even if the uniqueness result requires tightening.

Referee Report

1 major / 0 minor

Summary. The paper introduces two-parameter flows to learn the dynamics of high-dimensional probability densities over time from unlabeled samples. It learns sampling-time transports from a fixed base distribution to each time marginal via conditional flow matching, constructs coupled synthetic trajectories by pushing base samples through these maps, and regresses a physics-time velocity field on the resulting trajectories. The authors claim to prove that the resulting physics-time dynamics are unique and inherit regularity from the sampling-time transports; the method is positioned as scalable to high dimensions, free of per-step optimal-transport couplings, and permissive of admissible non-gradient dynamics.

Significance. If the uniqueness result and the correctness of the regression step hold, the work offers a practical route to continuous-time population dynamics that builds directly on mature conditional flow matching techniques, thereby inheriting their high-dimensional scalability. The explicit allowance for non-gradient flows is a genuine strength for modeling rotational or circulatory phenomena that gradient-based methods cannot capture. The avoidance of repeated optimal-transport solves per time step is a clear computational advantage.

major comments (1)

[abstract, paragraph 2] Abstract, paragraph 2: the uniqueness claim for the physics-time velocity rests on regressing a single vector field onto synthetic trajectories that are assembled from independently learned base-to-marginal transport maps. Because each transport is trained separately, the induced trajectories need not lie on the integral curves of any single continuous velocity; the manuscript must supply the precise argument (presumably in the uniqueness proof) showing that regression nevertheless recovers a unique field whose continuity equation reproduces the observed marginal evolution. This step is load-bearing for the central claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying this load-bearing aspect of the uniqueness claim. We address the comment below and will revise the manuscript accordingly.

read point-by-point responses

Referee: [abstract, paragraph 2] Abstract, paragraph 2: the uniqueness claim for the physics-time velocity rests on regressing a single vector field onto synthetic trajectories that are assembled from independently learned base-to-marginal transport maps. Because each transport is trained separately, the induced trajectories need not lie on the integral curves of any single continuous velocity; the manuscript must supply the precise argument (presumably in the uniqueness proof) showing that regression nevertheless recovers a unique field whose continuity equation reproduces the observed marginal evolution. This step is load-bearing for the central claim.

Authors: We agree that the current presentation of the uniqueness argument is insufficiently explicit and requires expansion. The proof (Theorem 4.1 and its supporting lemmas in Section 4) proceeds by first noting that the synthetic trajectories are generated by transporting identical base samples through the family of learned maps; this common origin induces a consistent family of paths even though the maps are trained independently. The regression step then minimizes the expected squared residual between the finite-difference velocities along these paths and the candidate vector field v. Under the standing assumption that a Lipschitz velocity field exists whose continuity equation reproduces the observed marginals, we show that any other candidate field w that also satisfies the marginal evolution must coincide with v almost everywhere; the argument relies on integrating the difference (v - w) against the empirical measure of the synthetic trajectories and using the fact that the base-to-marginal maps converge (in the large-sample limit) to the true transport maps of the underlying dynamics. We will add a self-contained, step-by-step derivation of this regression-to-uniqueness implication in the revised main text, together with a short remark clarifying why independent training of the maps does not break consistency of the induced trajectories. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation self-contained with internal uniqueness proof

full rationale

The paper defines two-parameter flows by first learning independent sampling-time transports from base to each marginal (via standard conditional flow matching), constructs synthetic trajectories, regresses a physics-time velocity field, and states that it proves uniqueness plus regularity inheritance. No quoted step reduces the claimed uniqueness or velocity field to a self-definition, a fitted parameter renamed as prediction, or a load-bearing self-citation chain. The proof is presented as an internal contribution rather than imported from prior author work, and the method explicitly builds on external flow-matching techniques. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the uniqueness claim is presented as proved but its supporting assumptions are not enumerated.

pith-pipeline@v0.9.1-grok · 5642 in / 1238 out tokens · 17486 ms · 2026-06-29T22:39:54.527892+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 8 canonical work pages · 1 internal anchor

[1]

Alvarez-Melis, D., Schiff, Y ., and Mroueh, Y

URL https://openreview.net/forum? id=li7qeBbCR1t. Alvarez-Melis, D., Schiff, Y ., and Mroueh, Y . Optimizing functionals on the space of probabilities with input convex neural networks.Transactions on Machine Learning Research, 2022. ISSN 2835-8856. URL https:// openreview.net/forum?id=dpOYN7o8Jm. Ambrosio, L., Gigli, N., and Savar ´e, G.Gradient Flows. L...

2022
[2]

URL http://link.springer.com/10.1007/ s10208-019-09425-z

doi: 10.1007/s10208-019-09425-z. URL http://link.springer.com/10.1007/ s10208-019-09425-z. Berman, J., Blickhan, T., and Peherstorfer, B. Parametric model reduction of mean-field and stochastic systems via higher-order action matching. InProceedings of the 38th International Conference on Neural Information Processing Systems, NeurIPS ’24, Red Hook, NY , USA,

work page doi:10.1007/s10208-019-09425-z
[3]

ISBN 979-8-3313-1438-5

Curran Associates Inc. ISBN 979-8-3313-1438-5. Blickhan, T., Berman, J., Stuart, A., and Peherstorfer, B. DICE: Discrete inverse continuity equation for learning population dynamics.arXiv, 2507.05107, 2025. URL https://arxiv.org/abs/2507.05107. Bunne, C., Papaxanthos, L., Krause, A., and Cuturi, M. Proximal optimal transport modeling of popula- tion dynam...

work page arXiv 2025
[4]

Lipschitz-Guided Design of Interpolation Schedules in Generative Models

URL https://proceedings.mlr.press/ v151/bunne22a.html. Chen, T., Liu, G.-H., Tao, M., and Theodorou, E. Deep momentum multi-marginal schr¨odinger bridge. InThirty- seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/ forum?id=ykvvv0gc4R. Chen, Y ., Conforti, G., Georgiou, T. T., and Ripani, L. Multi-marginal Schr¨...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/978-3-030-26980-7_75 2023
[5]

Daniels, M

URL https://proceedings.mlr.press/ v130/chewi21a.html. Daniels, M. On the Contractivity of Stochastic Interpola- tion Flow, 2025. URL https://arxiv.org/abs/ 2504.10653. Version Number: 2. 10 Two-Parameter Flows Gigli, N.Second order analysis on (P2(M), W2). Number 1018 in Memoirs of the American Mathematical Society. American Mathematical Society, Provide...

work page doi:10.1137/1 2025
[6]

cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper. pdf. Hong, W., Shi, Y ., and Niles-Weed, J. Trajectory in- ference with smooth schr ¨odinger bridges. InForty- second International Conference on Machine Learning,

2020
[7]

Hoogeboom, E., Gritsenko, A

URL https://openreview.net/forum? id=GHyvvWu1XC. Hoogeboom, E., Gritsenko, A. A., Bastings, J., Poole, B., van den Berg, R., and Salimans, T. Autoregressive dif- fusion models. InInternational Conference on Learning Representations, 2022. URL https://openreview. net/forum?id=Lm8T39vLDTE. Hyv¨arinen, A. Estimation of non-normalized statistical models by sc...

2022
[8]

2406783121

ISSN 0036-1410, 1095-7154. doi: 10.1137/ S0036141096303359. Kochkov, D., Smith, J. A., Alieva, A., Wang, Q., Brenner, M. P., and Hoyer, S. Machine learn- ing–accelerated computational fluid dynamics.Pro- ceedings of the National Academy of Sciences, 118(21):e2101784118, 2021. doi: 10.1073/pnas. 2101784118. URL https://www.pnas.org/doi/ abs/10.1073/pnas.21...

work page doi:10.1073/pnas 2021
[9]

doi: https://doi.org/10.1016/bs.hna.2024.05

work page doi:10.1016/bs.hna.2024.05 2024
[10]

Lavenant, H., Zhang, S., Kim, Y .-H., and Schiebinger, G

URL https://www.sciencedirect.com/ science/article/pii/S1570865924000097. Lavenant, H., Zhang, S., Kim, Y .-H., and Schiebinger, G. Toward a mathematical theory of trajectory inference.The Annals of Applied Probability, 34(1A):428–500, February
[11]

URL https://doi

doi: 10.1214/23-AAP1969. URL https://doi. org/10.1214/23-AAP1969. Li, L., Hurault, S., and Solomon, J. Self-consistent velocity matching of probability flows. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.),Advances in Neural Information Processing Sys- tems, volume 36, pp. 57038–57057. Curran Associates, Inc., 2023. Li...

work page doi:10.1214/23-aap1969 2023
[12]

Lipman, Y ., Chen, R

OpenReview.net, 2021. Lipman, Y ., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Represen- tations, 2023. URL https://openreview.net/ forum?id=PqvMRDCJT9t. Liu, X., Gong, C., and Liu, Q. Flow Straight and Fast: Learn- ing to Generate and Transfer Data with Re...

2021
[13]

doi: 10.3934/dcds.2014.34.1533

ISSN 1553-5231. doi: 10.3934/dcds.2014.34.1533. URL http://aimsciences.org//article/ doi/10.3934/dcds.2014.34.1533. McWilliams, J. C. The emergence of isolated coherent vortices in turbulent flow.Journal of Fluid Mechanics, 146:21–43, 1984. doi: 10.1017/S0022112084001750. Neklyudov, K., Brekelmans, R., Severo, D., and Makhzani, A. Action Matching: Learnin...

work page doi:10.3934/dcds.2014.34.1533 2014
[14]

Peherstorfer, B., Willcox, K., and Gunzburger, M

URL https://proceedings.mlr.press/ v202/neklyudov23a.html. Peherstorfer, B., Willcox, K., and Gunzburger, M. Survey of multifidelity methods in uncertainty propagation, in- ference, and optimization.SIAM Review, 60(3):550–591, 2018. Persiianov, M., Chen, J., Mokrov, P., Tyurin, A., Burnaev, E., and Korotin, A. Learning of population dynamics: In- verse op...

2018
[15]

Petrovi´c, K., Atanackovic, L., Moro, V ., Kapusniak, K., Ceylan, I

URL https://openreview.net/forum? id=tVJIKd6CLF. Petrovi´c, K., Atanackovic, L., Moro, V ., Kapusniak, K., Ceylan, I. I., Bronstein, M., Bose, J., and Tong, A. Curly flow matching for learning non-gradient field dynamics. In Belgrave, D., Zhang, C., Lin, H., Pascanu, R., Ko- niusz, P., Ghassemi, M., and Chen, N. (eds.),Advances in Neural Information Proce...

2025
[16]

Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S

URL https://proceedings.mlr.press/ v258/shen25b.html. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequi- librium thermodynamics. In Bach, F. and Blei, D. (eds.), Proceedings of the 32nd International Conference on Ma- chine Learning, volume 37 ofProceedings of Machine Learning Research, pp. 2256–2...

2015
[17]

cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93-Paper. pdf. Song, Y ., Garg, S., Shi, J., and Ermon, S. Sliced score matching: A scalable approach to density and score estimation. InProceedings of the Thirty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI 2019, Tel Aviv, Israel, July 22-25,...

2019
[18]

Terpin, A., Lanzetti, N., Gadea, M., and Dorfler, F

URL https://openreview.net/forum? id=PxTIG12RRHS. Terpin, A., Lanzetti, N., Gadea, M., and Dorfler, F. Learn- ing diffusion at lightspeed. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems,
[19]

Tong, A., Huang, J., Wolf, G., Van Dijk, D., and Krishnaswamy, S

URL https://openreview.net/forum? id=y10avdRFNK. Tong, A., Huang, J., Wolf, G., Van Dijk, D., and Krishnaswamy, S. TrajectoryNet: A dynamic opti- mal transport network for modeling cellular dynam- ics. In III, H. D. and Singh, A. (eds.),Proceed- ings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Res...
[20]

field difference

URL https://proceedings.mlr.press/ v119/tong20a.html. Villani, C.Optimal transport: old and new. Number 338 in Grundlehren der mathematischen Wissenschaften. Springer, Berlin, 2009. ISBN 978-3-540-71049-3. 12 Two-Parameter Flows A. Details about examples A.1. Evolving mixture of Gaussians A conditional flow matching model learns velocity fields to map a s...

2009

[1] [1]

Alvarez-Melis, D., Schiff, Y ., and Mroueh, Y

URL https://openreview.net/forum? id=li7qeBbCR1t. Alvarez-Melis, D., Schiff, Y ., and Mroueh, Y . Optimizing functionals on the space of probabilities with input convex neural networks.Transactions on Machine Learning Research, 2022. ISSN 2835-8856. URL https:// openreview.net/forum?id=dpOYN7o8Jm. Ambrosio, L., Gigli, N., and Savar ´e, G.Gradient Flows. L...

2022

[2] [2]

URL http://link.springer.com/10.1007/ s10208-019-09425-z

doi: 10.1007/s10208-019-09425-z. URL http://link.springer.com/10.1007/ s10208-019-09425-z. Berman, J., Blickhan, T., and Peherstorfer, B. Parametric model reduction of mean-field and stochastic systems via higher-order action matching. InProceedings of the 38th International Conference on Neural Information Processing Systems, NeurIPS ’24, Red Hook, NY , USA,

work page doi:10.1007/s10208-019-09425-z

[3] [3]

ISBN 979-8-3313-1438-5

Curran Associates Inc. ISBN 979-8-3313-1438-5. Blickhan, T., Berman, J., Stuart, A., and Peherstorfer, B. DICE: Discrete inverse continuity equation for learning population dynamics.arXiv, 2507.05107, 2025. URL https://arxiv.org/abs/2507.05107. Bunne, C., Papaxanthos, L., Krause, A., and Cuturi, M. Proximal optimal transport modeling of popula- tion dynam...

work page arXiv 2025

[4] [4]

Lipschitz-Guided Design of Interpolation Schedules in Generative Models

URL https://proceedings.mlr.press/ v151/bunne22a.html. Chen, T., Liu, G.-H., Tao, M., and Theodorou, E. Deep momentum multi-marginal schr¨odinger bridge. InThirty- seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/ forum?id=ykvvv0gc4R. Chen, Y ., Conforti, G., Georgiou, T. T., and Ripani, L. Multi-marginal Schr¨...

work page internal anchor Pith review Pith/arXiv arXiv doi:10.1007/978-3-030-26980-7_75 2023

[5] [5]

Daniels, M

URL https://proceedings.mlr.press/ v130/chewi21a.html. Daniels, M. On the Contractivity of Stochastic Interpola- tion Flow, 2025. URL https://arxiv.org/abs/ 2504.10653. Version Number: 2. 10 Two-Parameter Flows Gigli, N.Second order analysis on (P2(M), W2). Number 1018 in Memoirs of the American Mathematical Society. American Mathematical Society, Provide...

work page doi:10.1137/1 2025

[6] [6]

cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2020/file/ 4c5bcfec8584af0d967f1ab10179ca4b-Paper. pdf. Hong, W., Shi, Y ., and Niles-Weed, J. Trajectory in- ference with smooth schr ¨odinger bridges. InForty- second International Conference on Machine Learning,

2020

[7] [7]

Hoogeboom, E., Gritsenko, A

URL https://openreview.net/forum? id=GHyvvWu1XC. Hoogeboom, E., Gritsenko, A. A., Bastings, J., Poole, B., van den Berg, R., and Salimans, T. Autoregressive dif- fusion models. InInternational Conference on Learning Representations, 2022. URL https://openreview. net/forum?id=Lm8T39vLDTE. Hyv¨arinen, A. Estimation of non-normalized statistical models by sc...

2022

[8] [8]

2406783121

ISSN 0036-1410, 1095-7154. doi: 10.1137/ S0036141096303359. Kochkov, D., Smith, J. A., Alieva, A., Wang, Q., Brenner, M. P., and Hoyer, S. Machine learn- ing–accelerated computational fluid dynamics.Pro- ceedings of the National Academy of Sciences, 118(21):e2101784118, 2021. doi: 10.1073/pnas. 2101784118. URL https://www.pnas.org/doi/ abs/10.1073/pnas.21...

work page doi:10.1073/pnas 2021

[9] [9]

doi: https://doi.org/10.1016/bs.hna.2024.05

work page doi:10.1016/bs.hna.2024.05 2024

[10] [10]

Lavenant, H., Zhang, S., Kim, Y .-H., and Schiebinger, G

URL https://www.sciencedirect.com/ science/article/pii/S1570865924000097. Lavenant, H., Zhang, S., Kim, Y .-H., and Schiebinger, G. Toward a mathematical theory of trajectory inference.The Annals of Applied Probability, 34(1A):428–500, February

[11] [11]

URL https://doi

doi: 10.1214/23-AAP1969. URL https://doi. org/10.1214/23-AAP1969. Li, L., Hurault, S., and Solomon, J. Self-consistent velocity matching of probability flows. In Oh, A., Naumann, T., Globerson, A., Saenko, K., Hardt, M., and Levine, S. (eds.),Advances in Neural Information Processing Sys- tems, volume 36, pp. 57038–57057. Curran Associates, Inc., 2023. Li...

work page doi:10.1214/23-aap1969 2023

[12] [12]

Lipman, Y ., Chen, R

OpenReview.net, 2021. Lipman, Y ., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Represen- tations, 2023. URL https://openreview.net/ forum?id=PqvMRDCJT9t. Liu, X., Gong, C., and Liu, Q. Flow Straight and Fast: Learn- ing to Generate and Transfer Data with Re...

2021

[13] [13]

doi: 10.3934/dcds.2014.34.1533

ISSN 1553-5231. doi: 10.3934/dcds.2014.34.1533. URL http://aimsciences.org//article/ doi/10.3934/dcds.2014.34.1533. McWilliams, J. C. The emergence of isolated coherent vortices in turbulent flow.Journal of Fluid Mechanics, 146:21–43, 1984. doi: 10.1017/S0022112084001750. Neklyudov, K., Brekelmans, R., Severo, D., and Makhzani, A. Action Matching: Learnin...

work page doi:10.3934/dcds.2014.34.1533 2014

[14] [14]

Peherstorfer, B., Willcox, K., and Gunzburger, M

URL https://proceedings.mlr.press/ v202/neklyudov23a.html. Peherstorfer, B., Willcox, K., and Gunzburger, M. Survey of multifidelity methods in uncertainty propagation, in- ference, and optimization.SIAM Review, 60(3):550–591, 2018. Persiianov, M., Chen, J., Mokrov, P., Tyurin, A., Burnaev, E., and Korotin, A. Learning of population dynamics: In- verse op...

2018

[15] [15]

Petrovi´c, K., Atanackovic, L., Moro, V ., Kapusniak, K., Ceylan, I

URL https://openreview.net/forum? id=tVJIKd6CLF. Petrovi´c, K., Atanackovic, L., Moro, V ., Kapusniak, K., Ceylan, I. I., Bronstein, M., Bose, J., and Tong, A. Curly flow matching for learning non-gradient field dynamics. In Belgrave, D., Zhang, C., Lin, H., Pascanu, R., Ko- niusz, P., Ghassemi, M., and Chen, N. (eds.),Advances in Neural Information Proce...

2025

[16] [16]

Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S

URL https://proceedings.mlr.press/ v258/shen25b.html. Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequi- librium thermodynamics. In Bach, F. and Blei, D. (eds.), Proceedings of the 32nd International Conference on Ma- chine Learning, volume 37 ofProceedings of Machine Learning Research, pp. 2256–2...

2015

[17] [17]

cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93-Paper

URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ 3001ef257407d5a371a96dcd947c7d93-Paper. pdf. Song, Y ., Garg, S., Shi, J., and Ermon, S. Sliced score matching: A scalable approach to density and score estimation. InProceedings of the Thirty-Fifth Conference on Uncertainty in Artificial Intelligence, UAI 2019, Tel Aviv, Israel, July 22-25,...

2019

[18] [18]

Terpin, A., Lanzetti, N., Gadea, M., and Dorfler, F

URL https://openreview.net/forum? id=PxTIG12RRHS. Terpin, A., Lanzetti, N., Gadea, M., and Dorfler, F. Learn- ing diffusion at lightspeed. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems,

[19] [19]

Tong, A., Huang, J., Wolf, G., Van Dijk, D., and Krishnaswamy, S

URL https://openreview.net/forum? id=y10avdRFNK. Tong, A., Huang, J., Wolf, G., Van Dijk, D., and Krishnaswamy, S. TrajectoryNet: A dynamic opti- mal transport network for modeling cellular dynam- ics. In III, H. D. and Singh, A. (eds.),Proceed- ings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Res...

[20] [20]

field difference

URL https://proceedings.mlr.press/ v119/tong20a.html. Villani, C.Optimal transport: old and new. Number 338 in Grundlehren der mathematischen Wissenschaften. Springer, Berlin, 2009. ISBN 978-3-540-71049-3. 12 Two-Parameter Flows A. Details about examples A.1. Evolving mixture of Gaussians A conditional flow matching model learns velocity fields to map a s...

2009