Distribution learning via neural differential equations: minimal energy regularization and approximation theory

Jakob Zech; Youssef Marzouk; Zhi Ren

Distribution learning via neural differential equations: minimal energy regularization and approximation theory

T0 review · 0 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Neural ODEs with straight-line velocity fields can approximate any bounded-below $C^k$ target density to accuracy $\epsilon$ with network size $O(\epsilon^{-(d+1)/(k-1)})$.

desk verdict A rigorous approximation-theory paper for neural ODEs with explicit density-dependent rates; worth a serious referee despite a statement-proof mismatch and some overclaiming in the intro. read the letter →

arxiv 2502.03795 v1 pith:37MAZ3KL submitted 2025-02-06 cs.LG math.CAstat.MEstat.ML

Youssef Marzouk , Zhi Ren , Jakob Zech This is my paper · ORCID

classification cs.LGmath.CAstat.MEstat.ML MSC 68T0749Q2265D15

keywords neuralODEsmeasuretransportKnothe–RosenblattmapsdisplacementinterpolationkineticenergyregularizationReLU^2networksapproximationratesdensityestimation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper establishes that neural ODEs can learn distributions at a provable, explicit rate: for $C^k$ source and target densities on a cube that are bounded above and below, there is a ReLU$^2$ network of width and sparsity $O(\epsilon^{-(d+1)/(k-1)})$ whose flow map pushes the source to within $\epsilon$ of the target in Wasserstein distance or KL divergence, and that also makes the regularized training objective $O(\epsilon^2)$. The key structural fact is that among all velocity fields producing a given transport map $T$, the unique minimizer of average kinetic energy is the straight-line interpolation field $f((1-t)x + tT(x),t) = T(x)-x$. The paper bounds the $C^k$ norm of that field polynomially in the $C^k$ norms of the two densities, with constants depending only on dimension, smoothness, and the density bounds. These bounds feed into Grönwall-type stability estimates and existing ReLU$^2$ approximation theorems, closing the loop from density smoothness to network size. A sympathetic reading is that this gives a fully explicit end-to-end guarantee for regularized neural ODE distribution learning, including the value of the loss being minimized.

What carries the argument

The load-bearing object is the displacement interpolation $T_t(x)=(1-t)x + tT(x)$, which draws a straight line from $x$ to its transported image. The corresponding velocity field $f$ is defined implicitly by $f(T_t(x),t)=T(x)-x$, and the paper proves it is the unique minimizer of the kinetic-energy regularization that penalizes Lagrangian acceleration. The proof machinery is a multivariate Faà di Bruno formula, a chain rule for high derivatives of composed and inverse maps in Banach spaces, used to bound derivatives of the inverse of $G(x,t)=(T_t(x),t)$. A second load-bearing ingredient is the Knothe–Rosenblatt construction, which produces a triangular $T$ whose Jacobian has positive diagonal entries and therefore satisfies the spectral condition automatically. A Grönwall stability estimate converts velocity-field error into distribution error, and ReLU$^2$ neural network approximation results convert the $C^k$ smoothness of the field into explicit width, depth, and sparsity bounds.

What would settle it

One concrete test: take a pair of $C^2$ densities on $[0,1]^2$ satisfying the boundedness assumption, compute the straight-line velocity field of their Knothe–Rosenblatt map, and measure the smallest ReLU$^2$ network width needed to reach $W_2$ error $10^{-3}$. If the required width grows faster than $\epsilon^{-(d+1)/(k-1)}$ as $\epsilon$ is decreased, the paper's central rate is wrong.

Watch

Extended reading notes

Core claim

The paper's central claim is that distribution learning with neural ODEs reduces to approximating one straight-line velocity field. For any transport map $T$ whose Jacobian has no non-positive real eigenvalue, the field defined by $f(T_t(x),t) = T(x)-x$ with $T_t(x) = (1-t)x + tT(x)$ is the unique velocity field whose time-one flow realizes $T$ with every trajectory a straight line of constant speed. The paper proves this field minimizes the average kinetic energy among all fields realizing $T$, which justifies using it as the target of training with an acceleration penalty. For triangular Knothe–Rosenblatt maps on $[0,1]^d$, the $C^k$ norm of this velocity field is bounded by $C_{k,d}(L_1/L_2)^{\beta_d k^{d+3}}$, where $L_1$ and $L_2$ are the upper and lower bounds on the densities. Combining this with a Grönwall-type stability estimate and deep ReLU$^2$ approximation results yields the network-size bounds and the $O(\epsilon^2)$ guarantee on the regularized objective.

Load-bearing premise

The load-bearing premise is Assumption 2.2: both densities are $C^k$ and bounded below by a positive constant $L_2$ on their compact supports, so they never vanish; if either density has thin tails or touches zero, the polynomial bounds on the velocity field blow up and the stated rates no longer apply, a restriction the paper itself acknowledges.

Editorial extensions

If this is right

Smoothness buys a concrete improvement: with $C^k$ densities the required network width and sparsity scale as $\epsilon^{-(d+1)/(k-1)}$, so every additional order of differentiability reduces the exponent by one.
Wasserstein approximation requires only uniform ($C^0$) control of the velocity field, giving the rate $\epsilon^{-(d+1)/k}$, while KL divergence requires $C^1$ control and therefore uses the $\epsilon^{-(d+1)/(k-1)}$ rate; because the estimates are $L^\infty$, the same argument yields guarantees for Hellinger, chi-square, and total variation.
The same network class makes the full training objective, divergence plus kinetic-energy penalty, smaller than $C\epsilon^2$, so optimizing the regularized loss is provably finding a near-minimal-energy straight-line flow.
The quadratic-cost optimal transport map has a Jacobian with nonnegative eigenvalues, so the zero-loss minimizer of the regularized objective is always of the straight-line form, connecting the regularizer to the fluid-dynamics formulation of optimal transport.
Exchanging source and target covers the Bayesian setting where the target is known only up to a normalizing constant.

Reading between the lines

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

The explicit polynomial dependence on $L_1/L_2$ suggests a natural stress test: shrink the lower density bound toward zero and the theory predicts polynomial growth in required network size; experiments with tail or boundary-vanishing densities could show whether this degradation is fundamental or an artifact of the proof.
Because the straight-line velocity field is the same object used by rectified flow and flow-matching constructions, the $C^k$ bounds here likely transfer to those settings, giving them network-size guarantees without new regularity analysis.
Extending the stability analysis to unbounded domains with Gaussian tails would require a different interpolation than straight lines, since the lower-bound assumption fails there; the paper's restriction to compact, lower-bounded densities is the main barrier to such an extension.
The paper notes its rates are not minimax optimal; whether a modified architecture or a different interpolation can close the gap to the classical $n^{-2k/(d+2k)}$ rate remains an open question that this work makes precise.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The paper you'll want to know about: Marzouk, Ren, Zech prove explicit, density-dependent network size bounds for neural ODE distribution approximation. That's new. They show that for C^k densities bounded away from zero on the cube, a ReLU^2 velocity field with width/sparsity O(ε^{-(d+1)/(k-1)}) drives the source to within ε in Wasserstein or KL, and makes their regularized objective O(ε^2). The rates are not in the prior literature, and the proof structure is sound.

What's actually new: the C^k bound on the straight-line velocity field for Knothe–Rosenblatt maps, polynomial in the density norms, and the derived KR map C^k bound that goes with it. The minimal-energy characterization of straight-line interpolation is a rigorous statement of what's known in OT/flow matching, but they state it cleanly and connect it to their objective. The stability lemmas are standard Gronwall-type estimates, but they are needed for the whole chain and are done carefully.

Soft spots, in order. First, Theorem 3.12's proof uses the KL-based training objective (A.1) with a weighted regularization parameter, while the theorem statement is about the general objective (OP) with unweighted R(f). That's a statement-proof mismatch that should be fixed; the claim is true for the weighted version, but the general claim needs a careful statement. Second, the intro's 'first unified attempt' is a bit strong given the existing flow-matching and OT literature; it's the first with explicit density-dependent rates, not the first unified treatment. Third, the concrete network rate theorems (6.4, 6.6) import results from the companion paper Marzouk et al. 2024; that's acceptable if the companion is available, but a referee should verify the hypotheses match. The uniform lower bound L2 on densities is a real restriction—Gaussian tails are out—but the authors flag it in Section 7, so it's a stated limitation, not a hidden flaw.

Bottom line: this is a serious theory paper. The central approximation chain holds together under its assumptions. A careful referee will want the Theorem 3.12 mismatch clarified and the intro toned down, but I'd send it out rather than desk reject. Worth a spot in your reading group if you follow generative modeling theory.

Referee Report

0 major / 6 minor

Summary. The paper develops an approximation theory for distribution learning with neural ODEs. It proves that, under a spectrum condition on a transport map T, the straight-line interpolation (1-t)x + tT(x) is realized by a unique C^k velocity field on the space-time domain; that this velocity field minimizes a kinetic-energy regularizer among all fields inducing T; that the C^k norm of the field is bounded polynomially in the C^k norm of T and, for Knothe-Rosenblatt maps, polynomially in the C^k norms of the source and target densities; and that Wasserstein/KL approximation error and the regularized objective can be made arbitrarily small with ReLU^2 neural networks whose width and sparsity scale as epsilon^{-(d+1)/(k-1)} (with a slightly different exponent in the pure Wasserstein case). The paper is explicitly restricted to compact domains and densities that are uniformly bounded above and below, and it transparently relies on companion-paper results for the concrete network-size and KL-stability statements.

Significance. These are valuable explicit bias bounds for a practically important model class. The minimal-energy characterization is self-contained and elegant, the C^k bounds via Faà di Bruno and the inverse function theorem are explicit, and the paper is honest about its main restrictions (compact support, uniform lower bounds on densities) and about the fact that the final network-rate results import theorems from a companion paper. If the imported hypotheses are verified, the results go substantially beyond universal approximation statements by providing concrete rates and objective-value guarantees. The acknowledgment in Section 7 that the resulting statistical rates may be suboptimal relative to minimax density estimation is a useful and candid limitation.

minor comments (6)

[Sections 5.2, 6.4, 6.6] Theorems 5.3, 6.4, and 6.6 depend on specific results from Marzouk et al. (2024) (Theorems 16 and 20, Lemmas 6-8), but the exact hypotheses and statements of these results are not reproduced in the manuscript. Since these carry the final network-size and KL-stability claims, please state the precise companion theorems in an appendix or otherwise verify that their hypotheses match the ansatz space FNN and the C^k extension norms used in the present paper.
[Sections 3.2, 4.2, 6] Theorem 3.4 establishes f in C^k on the interior of Omega_{[0,1]}, while the neural-network approximation results use f in C^k([0,1]^{d+1}). Please state explicitly how f is extended to the closed cube, for example by extending T to a C^k function on an open neighborhood and applying the inverse function theorem there; the Stein extension cited in Section 3.2 gives W^{k,infinity}, not C^k, so a short clarification would remove ambiguity.
[Theorems 4.11 and 4.12] The condition '0 < L1 <= L2 < infinity' appears to be a typographical error; it should read '0 < L2 <= L1 < infinity' to be consistent with Assumption 2.2 and with the repeated use of L1/L2 >= 1.
[Proposition 6.3] Proposition 6.3 states k >= 1 while Assumption 2.2 is stated with k >= 2; please reconcile these assumptions or add a note explaining that the relevant parts of Assumption 2.2 are used with the adjusted smoothness order.
[Theorem 3.10 proof] The equality condition in the Jensen step is stated as 'equality holds iff d_t X(x,t) = X(x,1)-X(x,0)', but the first Jensen inequality also requires that |d_t X(x,t)| is constant in t for almost every t; the conclusion is correct, but the statement of the equality condition should be more precise.
[Appendix C] There are several small typographical issues in the proofs: in the proof of Theorem C.4, some denominators read 'n!' where 'alpha_r!' is intended, and Lemma C.9 writes 'det dT(x)' where 'det(∇T(x))' is meant. These do not affect the arguments but should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is self-contained, and the companion-paper approximation results used are independent parameter-free theorems.

full rationale

The paper's central claims do not reduce to their inputs by construction. Theorem 3.10 proves that among all velocity fields realizing a fixed transport map T, the straight-line interpolation minimizes average kinetic energy; this follows from Jensen's inequality applied to the trajectory integral, not from the definition of the regularization R(f). The existence of a zero-loss velocity field in Theorem 3.12 is established via the Brenier optimal transport map and the Monge-Ampere equation, with regularity imported from an external source (Panaretos and Zemel, 2020); the spectrum condition (3.2) is derived from strict convexity of the potential and the uniform lower bounds on the densities. The C^k regularity bounds on the velocity field (Theorems 4.5, 4.11, 4.12) are proven in the paper using the Banach-space Faà di Bruno formula, the inverse function theorem, and explicit estimates on Knothe-Rosenblatt components; these are self-contained analytic derivations rather than fitted or assumed quantities. The neural-network approximation rates and KL stability results in Sections 5.2 and 6 are imported from the authors' companion paper (Marzouk et al., 2024), which is a self-citation; however, the cited theorems are parameter-free mathematical statements with explicit assumptions about C^k functions and ansatz spaces, and they do not assume the present paper's conclusions. Under the review rules, such cited results count as independent support and do not constitute circularity. The paper explicitly acknowledges the restriction to bounded, lower-bounded densities in Section 7, and this limitation is stated rather than hidden. No fitted parameter is renamed as a prediction, no uniqueness theorem is invoked to forbid alternatives, and no known result is merely relabeled. The only external dependency is the companion paper's approximation theorems, which are load-bearing but not circular. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The paper introduces no free parameters fitted to data and no new physical or mathematical entities. It relies on standard mathematical tools (Faà di Bruno, inverse function theorem, Stein's extension theorem) and on domain assumptions about the probability distributions. The main external dependency is the companion paper's neural network approximation results, which are mathematical theorems rather than fitted quantities.

assumptions (5)

domain assumption Assumption 2.1: both source and target are supported on compact convex sets and are absolutely continuous with respect to Lebesgue measure.
Used throughout; enables the use of Lipschitz domain theory and extension theorems. Stated in Section 2.2.
domain assumption Assumption 2.2: densities are in C^k with uniform upper bound L1 and uniform lower bound L2 > 0, for k >= 2.
Crucial for the C^k bound on the Knothe-Rosenblatt map (Theorem 4.11) and for the Jacobian determinant lower bound in Lemma 4.10. Stated in Section 2.2.
domain assumption Assumption 3.2: the transport map T satisfies the spectrum condition that no eigenvalue of its Jacobian lies on the negative real axis.
Necessary for the straight-line flow to be well-defined (Lemma 3.3, Theorem 3.4). Satisfied by optimal transport maps and monotone triangular maps, as shown in Theorem 3.12 and Lemma 4.9.
standard math External regularity theorem for optimal transport maps (Theorem C.3, citing Panaretos-Zemel 2020).
Used in Theorem 3.12 to assert that the optimal transport map is the gradient of a strictly convex C^{k+2} potential with positive Jacobian determinant, under the stated density assumptions.
standard math Neural network approximation rates for ReLU^2 networks, taken from the companion paper Marzouk et al. (2024), Theorems 16 and 20.
Used as a black box in Section 6 to convert the C^k norm of the velocity field into network size bounds. The companion paper's results are stated as published theorems with explicit assumptions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Distribution learning via neural differential equations: minimal energy regularization and approximation theory." pith.science (2026). https://pith.science/paper/37MAZ3KL

@misc{pith2026250203795,
  author       = {Pith},
  title        = {Pith review of: Distribution learning via neural differential equations: minimal energy regularization and approximation theory},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/37MAZ3KL}},
  note         = {Machine review of arXiv:2502.03795}
}

abstract

Neural ordinary differential equations (ODEs) provide expressive representations of invertible transport maps that can be used to approximate complex probability distributions, e.g., for generative modeling, density estimation, and Bayesian inference. We show that for a large class of transport maps $T$, there exists a time-dependent ODE velocity field realizing a straight-line interpolation $(1-t)x + tT(x)$, $t \in [0,1]$, of the displacement induced by the map. Moreover, we show that such velocity fields are minimizers of a training objective containing a specific minimum-energy regularization. We then derive explicit upper bounds for the $C^k$ norm of the velocity field that are polynomial in the $C^k$ norm of the corresponding transport map $T$; in the case of triangular (Knothe--Rosenblatt) maps, we also show that these bounds are polynomial in the $C^k$ norms of the associated source and target densities. Combining these results with stability arguments for distribution approximation via ODEs, we show that Wasserstein or Kullback--Leibler approximation of the target distribution to any desired accuracy $\epsilon > 0$ can be achieved by a deep neural network representation of the velocity field whose size is bounded explicitly in terms of $\epsilon$, the dimension, and the smoothness of the source and target densities. The same neural network ansatz yields guarantees on the value of the regularized training objective.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 13 canonical work pages

[1]

Albergo and Eric V anden-Eijnden

Michael S. Albergo and Eric V anden-Eijnden. Building norma lizing ﬂows with stochastic interpolants. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1- 5,

work page 2023
[3]

URL https://doi.org/10.48550/arXiv.2303.08797

doi: 10.48550/ARXIV .2303.08797. URL https://doi.org/10.48550/arXiv.2303.08797. Antonio Álvarez-López, Borjan Geshkovski, and Domènec Rui z-Balet. Constructive approximate transport maps with normalizing ﬂows. arXiv preprint arXiv:2412.19366,

arXiv doi:10.48550/arxiv.2303.08797
[4]

doi: 10.1007/s10208-023-09630-x

ISSN 1615-3383. doi: 10.1007/s10208-023-09630-x. URL https://doi.org/10.1007/s10208-023-09630-x . Ricardo Baptista, Bamdad Hosseini, Nikola B Kovachki, Y ous sef M Marzouk, and Amir Sagiv. An ap- proximation theory framework for measure-transport sampl ing algorithms. Mathematics of Computation,

doi:10.1007/s10208-023-09630-x
[10]

doi: 10.1007/978-3-662-00547-7

ISBN 3-540-13928-1. doi: 10.1007/978-3-662-00547-7. URL https://doi.org/10.1007/978-3-662-00547-7 . Laurent Dinh, David Krueger, and Y oshua Bengio. NICE: non-l inear independent components estima- tion. In Y oshua Bengio and Y ann LeCun, editors, 3rd International Conference on Learning Represen- tations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Wor...

doi:10.1007/978-3-662-00547-7
[13]

URL https://doi.org/10.48550/arXiv.2404.00551

doi: 10.48550/ARXIV .2404.00551. URL https://doi.org/10.48550/arXiv.2404.00551. Amir Gholaminejad, Kurt Keutzer, and George Biros. ANODE: u nconditionally accurate memory- efﬁcient gradients for neural odes. In Sarit Kraus, editor, Proceedings of the Twenty- Eighth International Joint Conference on Artiﬁcial Intell igence, IJCAI 2019, Macao, China, Au- gu...

arXiv doi:10.48550/arxiv.2404.00551
[15]

Courville

Chin-Wei Huang, David Krueger, Alexandre Lacoste, and Aaro n C. Courville. Neural autoregressive ﬂows. In Jennifer G. Dy and Andreas Krause, editors, Proceedings of the 35th International Con- ference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden, July 10-15, 2018 , 41 volume 80 of Proceedings of Machine Learning Research , pages 208...

work page 2018
[17]

doi: 10.48550/ARXIV .2404. 09730. URL https://doi.org/10.48550/arXiv.2404.09730. Diederik P . Kingma, Tim Salimans, Rafal Józefowicz, Xi Chen , Ilya Sutskever, and Max Welling. Improving variational autoencoders with inverse autoregr essive ﬂow. In Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg, Isabelle Guyon, and Roman Gar nett, editors, Advances i...

arXiv doi:10.48550/arxiv.2404.09730
[19]

Normalizi ng ﬂows: An introduction and review of current methods

URL https://proceedings.neurips.cc/paper_files/paper/2018/file/d139db6a236200b21cc7f7 Ivan Kobyzev, Simon Prince, and Marcus Brubaker. Normalizi ng ﬂows: An introduction and review of current methods. IEEE Transactions on Pattern Analysis and Machine In- telligence, page 1–1,

work page 2018

Show all 34 references

[20]

doi: 10.1109/tpami.2020

ISSN 1939-3539. doi: 10.1109/tpami.2020. 2992934. URL http://dx.doi.org/10.1109/TPAMI.2020.2992934. Zhifeng Kong and Kamalika Chaudhuri. The expressive power o f a class of normalizing ﬂow mod- els. In Silvia Chiappa and Roberto Calandra, editors, The 23rd International Confer...

1939
[21]

Gen Li, Y uting Wei, Y uejie Chi, and Y uxin Chen

URL http://proceedings.mlr.press/v108/kong20a.html. Gen Li, Y uting Wei, Y uejie Chi, and Y uxin Chen. A sharp conver gence theory for the probability ﬂow odes of diffusion models. CoRR, abs/2408.02320,

arXiv
[22]

URL https://doi.org/10.48550/arXiv.2408.02320

doi: 10.48550/ARXIV .2408.02320. URL https://doi.org/10.48550/arXiv.2408.02320. Qianxiao Li, Ting Lin, and Zuowei Shen. Deep learning via dyn amical systems: An approximation perspec- tive. CoRR, abs/1912.10382,

1912 doi arXiv
[23]

Y aron Lipman, Ricky T

URL http://arxiv.org/abs/1912.10382. Y aron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Ni ckel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Rep- resentations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023 . Op...

1912 arXiv
[24]

42 Xingchao Liu, Chengyue Gong, and Qiang Liu

URL https://openreview.net/forum?id=PqvMRDCJT9t. 42 Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight an d fast: Learning to gener- ate and transfer data with rectiﬁed ﬂow. In The Eleventh International Conference on Learn- ing Representations, ICLR 2023, Kigali, Rwand...

2023
[25]

Yiping Lu, Aoxiao Zhong, Quanzheng Li, and Bin Dong

URL https://openreview.net/forum?id=XVjTT1nw5z. Yiping Lu, Aoxiao Zhong, Quanzheng Li, and Bin Dong. Beyond ﬁ nite layer neural networks: Bridging deep architectures and numerical differential equations. In 6th International Conference on Learning Rep- resentations, ICLR 2018,...

2018
[26]

U RL http://dx.doi.org/10.1007/978-3-319-11259-6_23-1

doi: 10.1007/978-3-319-11259-6_23-1. U RL http://dx.doi.org/10.1007/978-3-319-11259-6_23-1 . Y oussef Marzouk, Zhi Ren, Sven Wang, and Jakob Zech. Distrib ution learning via neural differential equa- tions: a nonparametric statistical perspective. Journal of Machine Learning R...

doi
[27]

arXiv:2309.01043. Tarek A. Moselhy and Y oussef M. Marzouk. Bayesian inference with op- timal maps. Journal of Computational Physics , 231(23):7815–7850,

arXiv
[29]

URL https://link.springer.com/content/pdf/10.1007%2F978-3-030-38438-8.pdf

ISBN 978-3-030-38437-1. URL https://link.springer.com/content/pdf/10.1007%2F978-3-030-38438-8.pdf . Danilo Jimenez Rezende and Shakir Mohamed. V ariational inf erence with normalizing ﬂows. In Francis R. Bach and David M. Blei, editors, Proceedings of the 32nd International Co...
[30]

URL https://doi.org/10.1137/21m1411433

doi: 10.1137/21M1411433. URL https://doi.org/10.1137/21m1411433. Lars Ruthotto and Eldad Haber. Deep neural networks motivat ed by partial differential equa- tions. J. Math. Imaging Vis. , 62(3):352–364,

doi
[31]

URL https://doi.org/10.1007/s10851-019-00903-1

doi: 10.1007/s10851-019-00903-1. URL https://doi.org/10.1007/s10851-019-00903-1 . 43 Filippo Santambrogio. Optimal transport for applied mathe maticians. calculus of variations, pdes and mod- eling

doi
[32]

Y ang Song, Jascha Sohl-Dickstein, Diederik P

URL https://www.math.u-psud.fr/~filippo/OTAM-cvgmt.pdf. Y ang Song, Jascha Sohl-Dickstein, Diederik P . Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic diffe rential equations. In 9th International Confer- ence on Learni...

2021
[35]

Sparse approximation of tri angular transports, part i: The ﬁnite-dimensional case

URL https://proceedings.neurips.cc/paper/2020/hash/979a3f14bae523dc5101c52120c535e9-A Jakob Zech and Y oussef Marzouk. Sparse approximation of tri angular transports, part i: The ﬁnite-dimensional case. Constructive Approximation , 2022a. doi: https://doi.org/10.1007/ s00365-0...

2020
[2000]

URL https://doi.org/10.1007/s002110050002

doi: 10.1007/ s002110050002. URL https://doi.org/10.1007/s002110050002. Joe Benton, George Deligiannidis, and Arnaud Doucet. Error bounds for ﬂow matching methods. Trans. Mach. Learn. Res., 2024,

2024 doi
[2007]

40 Y ann Brenier

doi: 10.1070/SM2005v196n03ABEH000882 . 40 Y ann Brenier. Polar factorization and monotone rearrangement of vector-valued functions. Communications on Pure and Applied Mathematics , 44:375–417,

doi
[2010]

Coupling-based invertible neural networks are universal d iffeomorphism approximators

Takeshi Teshima, Isao Ishikawa, Koichi Tojo, Kenta Oono, Ma sahiro Ikeda, and Masashi Sugiyama. Coupling-based invertible neural networks are universal d iffeomorphism approximators. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia Hadsell, Maria-Fl orina Balcan, and Hsuan-Tien ...

2020 arXiv
[2012]

doi: https://doi.org/10.1016/j.jcp.201 2.07.022

ISSN 0021-9991. doi: https://doi.org/10.1016/j.jcp.201 2.07.022. URL https://www.sciencedirect.com/science/article/pii/S0021999112003956. Derek Onken, Samy Wu Fung, Xingjian Li, and Lars Ruthotto. Ot -ﬂow: Fast and accurate con- tinuous normalizing ﬂows via optimal transport. ...

doi
[2015]

Chris Finlay, Jörn-Henrik Jacobsen, Levon Nurbekyan, and A dam M

URL http://arxiv.org/abs/1410.8516. Chris Finlay, Jörn-Henrik Jacobsen, Levon Nurbekyan, and A dam M. Oberman. How to train your neural ODE: the world of jacobian and kinetic regularizatio n. In Proceedings of the 37th In- ternational Conference on Machine Learning, ICML 2020,...

arXiv
[2016]

Glow: Generative ﬂow wi th invertible 1x1 convolutions

URL https://proceedings.neurips.cc/paper/2016/hash/ddeebdeefdb7e7e7a697e1c3e3d8ef54-A Durk P Kingma and Prafulla Dhariwal. Glow: Generative ﬂow wi th invertible 1x1 convolutions. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bi anchi, and R. Garnett, editors, Ad...

2016
[2017]

U RL https://doi.org/10.1016/j.neunet.2017.07.002

doi: 10.1016/j.neunet.2017.07.002. U RL https://doi.org/10.1016/j.neunet.2017.07.002. 44 Dmitry Y arotsky and Anton Zhevnerchuk. The phase diagram of approximation rates for deep neural networks. In Hugo Larochelle, Marc’Aurelio Ranzato, Raia H adsell, Maria-Florina Balcan, an...

2017 doi
[2018]

URL http://arxiv.org/abs/1806.07366. G. Constantine and T. H. Savits. A multivariate faa di bruno f ormula with applications. Transactions of the American Mathematical Society, 348:503–520,

arXiv
[2019]

URL https://doi.org/10.24963/ijcai.2019/103

doi: 10.24963/ijcai.201 9/103. URL https://doi.org/10.24963/ijcai.2019/103. Will Grathwohl, Ricky T. Q. Chen, Jesse Bettencourt, Ilya Su tskever, and David Duvenaud. FFJORD: free- form continuous dynamics for scalable reversible generati ve models. In 7th International Confere...

2019 doi
[2020]

Y uan Gao, Jian Huang, Y uling Jiao, and Shurong Zheng

URL http://proceedings.mlr.press/v119/finlay20a.html. Y uan Gao, Jian Huang, Y uling Jiao, and Shurong Zheng. Conver gence of continuous normalizing ﬂows for learning probability distributions. CoRR, abs/2404.00551,

arXiv
[2021]

Daniel Zhengyu Huang, Jiaoyang Huang, and Zhengjiang Lin

URL https://openreview.net/forum?id=te7PVH1sPxJ. Daniel Zhengyu Huang, Jiaoyang Huang, and Zhengjiang Lin. C onvergence analysis of probability ﬂow ODE for score-based generative models. CoRR, abs/2404.09730,

arXiv
[2023]

Michael S

URL https://openreview.net/forum?id=li7qeBbCR1t. Michael S. Albergo, Nicholas M. Bofﬁ, and Eric V anden-Eijnd en. Stochastic interpolants: A unifying framework for ﬂows and diffusions. CoRR, abs/2303.08797,

arXiv
[2024]

Jens Behrmann, Will Grathwohl, Ricky T. Q. Chen, David Duven aud, and Jörn-Henrik Jacobsen. Invertible residual networks. In Kamalika Chaudhuri and Ruslan Salakh utdinov, editors, Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9 -15 June 2019,...

2019

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Albergo and Eric V anden-Eijnden

Michael S. Albergo and Eric V anden-Eijnden. Building norma lizing ﬂows with stochastic interpolants. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1- 5,

work page 2023

[2] [3]

URL https://doi.org/10.48550/arXiv.2303.08797

doi: 10.48550/ARXIV .2303.08797. URL https://doi.org/10.48550/arXiv.2303.08797. Antonio Álvarez-López, Borjan Geshkovski, and Domènec Rui z-Balet. Constructive approximate transport maps with normalizing ﬂows. arXiv preprint arXiv:2412.19366,

arXiv doi:10.48550/arxiv.2303.08797

[3] [4]

doi: 10.1007/s10208-023-09630-x

ISSN 1615-3383. doi: 10.1007/s10208-023-09630-x. URL https://doi.org/10.1007/s10208-023-09630-x . Ricardo Baptista, Bamdad Hosseini, Nikola B Kovachki, Y ous sef M Marzouk, and Amir Sagiv. An ap- proximation theory framework for measure-transport sampl ing algorithms. Mathematics of Computation,

doi:10.1007/s10208-023-09630-x

[4] [10]

doi: 10.1007/978-3-662-00547-7

ISBN 3-540-13928-1. doi: 10.1007/978-3-662-00547-7. URL https://doi.org/10.1007/978-3-662-00547-7 . Laurent Dinh, David Krueger, and Y oshua Bengio. NICE: non-l inear independent components estima- tion. In Y oshua Bengio and Y ann LeCun, editors, 3rd International Conference on Learning Represen- tations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Wor...

doi:10.1007/978-3-662-00547-7

[5] [13]

URL https://doi.org/10.48550/arXiv.2404.00551

doi: 10.48550/ARXIV .2404.00551. URL https://doi.org/10.48550/arXiv.2404.00551. Amir Gholaminejad, Kurt Keutzer, and George Biros. ANODE: u nconditionally accurate memory- efﬁcient gradients for neural odes. In Sarit Kraus, editor, Proceedings of the Twenty- Eighth International Joint Conference on Artiﬁcial Intell igence, IJCAI 2019, Macao, China, Au- gu...

arXiv doi:10.48550/arxiv.2404.00551

[6] [15]

Courville

Chin-Wei Huang, David Krueger, Alexandre Lacoste, and Aaro n C. Courville. Neural autoregressive ﬂows. In Jennifer G. Dy and Andreas Krause, editors, Proceedings of the 35th International Con- ference on Machine Learning, ICML 2018, Stockholmsmässan, Stockholm, Sweden, July 10-15, 2018 , 41 volume 80 of Proceedings of Machine Learning Research , pages 208...

work page 2018

[7] [17]

doi: 10.48550/ARXIV .2404. 09730. URL https://doi.org/10.48550/arXiv.2404.09730. Diederik P . Kingma, Tim Salimans, Rafal Józefowicz, Xi Chen , Ilya Sutskever, and Max Welling. Improving variational autoencoders with inverse autoregr essive ﬂow. In Daniel D. Lee, Masashi Sugiyama, Ulrike von Luxburg, Isabelle Guyon, and Roman Gar nett, editors, Advances i...

arXiv doi:10.48550/arxiv.2404.09730

[8] [19]

Normalizi ng ﬂows: An introduction and review of current methods

URL https://proceedings.neurips.cc/paper_files/paper/2018/file/d139db6a236200b21cc7f7 Ivan Kobyzev, Simon Prince, and Marcus Brubaker. Normalizi ng ﬂows: An introduction and review of current methods. IEEE Transactions on Pattern Analysis and Machine In- telligence, page 1–1,

work page 2018