Pith. sign in

REVIEW 3 major objections 6 minor 46 references

A Stiefel-manifold projection scheme approximates Wasserstein gradient flows without meshes, with an error bound that can be made arbitrarily small under geodesic convexity.

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

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

T0 review · deepseek-v4-flash

2026-08-01 18:37 UTC pith:O3TBLOYQ

load-bearing objection A genuinely new Stiefel-decoder framework for Wasserstein gradient flows with a clean a posteriori bound, but the adaptive convergence theorem rests on an equality that is false for nonlinear energies, and the numerics never exercise the adaptivity. the 3 major comments →

arxiv 2607.17235 v1 pith:O3TBLOYQ submitted 2026-07-19 math.NA cs.NA

A Dynamical Approximation Scheme on the Stiefel manifold for Wasserstein Gradient Flows

Isabella Carla Gonnella , Olga Mula , Federico Pichi , Gianluigi Rozza This is my paper
classification math.NA cs.NA MSC 49Q2265M9935Q8435K55
keywords Wasserstein gradient flowsStiefel manifoldtransport mapsDirac-Frenkel projectionTaylor-based background spaceadaptive error controlenergy dissipationmeshless methods
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper proposes a meshless way to compute Wasserstein gradient flows: instead of discretizing space or solving an optimal-transport problem at every step, it represents the evolving measure as the pushforward of the initial measure through a transport map, and approximates that map by a time-dependent linear combination of orthonormal basis functions constrained to a finite-dimensional background space. The parameters (coefficients and the orthonormal frame) evolve by projecting the exact flow velocity onto the decoder's tangent space, a Dirac–Frenkel dynamical principle on a Stiefel manifold. The authors prove that the resulting curve of measures is absolutely continuous in Wasserstein space, that energy dissipates along it up to the projection residual, and that for geodesically convex energies the residual—and hence the Wasserstein error at final time—can be made arbitrarily small by updating a Taylor-based background space at finitely many relinearization times. This matters because it offers an error-controlled alternative to particle and mesh-based Lagrangian methods that avoids both adaptive sampling instabilities and spatial grids.

Core claim

The central claim is that a Wasserstein gradient flow can be faithfully approximated by a low-dimensional transport map in L^2_mu0 of the form T_t = sum_i c_{t,i} v_{t,i}, where the orthonormal frame v_t moves on a Stiefel manifold constrained to a background space W_m, and the parameter dynamics are the least-squares projection of the pulled-back velocity v_{nu_t} composed with T_t onto the decoder tangent space. When the background space is built from the Taylor expansion of the Lagrangian velocity operator around a nearby linearization map—spanning L(Tbar) and DL(Tbar)[p_i]—the projection residual admits a computable bound. For geodesically convex energies satisfying the paper's assumptio

What carries the argument

The Stiefel decoder: a parameterization of transport maps as T_theta = sum_{i=1}^n c_i v_i, with v = (v_1, ..., v_n) an orthonormal frame in a finite-dimensional background space W_m contained in L^2_mu0; time integration is performed by exponentiating horizontal tangent directions on the Stiefel manifold. The background space W_m = span{v_1, ..., v_n, L(Tbar), DL(Tbar)[p_1], ..., DL(Tbar)[p_P]} is generated by the Lagrangian velocity operator and its derivative at a linearization map, so that the projected dynamics reproduce the first-order Taylor model of the velocity exactly. The error analysis splits the residual into a projection term (the part of DL(Tbar)[T - Tbar] not captured by the

Load-bearing premise

The central guarantee rests on the unproven practical assumption that at every scheduled relinearization time one can construct a linearization map close enough to the current transport map in H^2_mu0, with uniformly bounded Taylor constants, and that the approximate transport map remains a diffeomorphism throughout.

What would settle it

Take a lambda-convex energy (for example quadratic Fokker-Planck with Gaussian initial data), implement the adaptive relinearization loop with scheduled background updates, and measure the residual integral integral_0^{t_f} ||r_s||_{L^2_mu0} e^{-lambda(t_f-s)} ds as the partition is refined. If this quantity cannot be driven below a prescribed eta—because a sufficiently accurate linearization map cannot be produced or the constants C_{t_j,2} grow—then the claimed arbitrary-accuracy guarantee fails empirically.

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

If this is right

  • The discrete curve nu_t = T_t#mu_0 is an absolutely continuous curve in Wasserstein space whose velocity is the projection of the exact velocity, so the approximate measure needs no mollification or regularization.
  • The approximate flow dissipates energy: d/dt F(nu_t) equals minus the squared projected velocity norm, so the method preserves the monotone decay structure of the true gradient flow up to the projection residual.
  • For geodesically convex energies, the residual bound serves as a computable a posteriori error estimator, and the Taylor relinearization strategy guarantees arbitrarily small final-time Wasserstein error.
  • Taylor-based background spaces reach a given accuracy with fewer basis functions than fixed polynomial spaces in the tested Fokker-Planck, nonlinear diffusion, and interaction problems.
  • Because the method is meshless, the dimensional bottleneck shifts from spatial discretization to the size of the approximation space and the accuracy of Monte Carlo integration.

Where Pith is reading between the lines

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

  • The adaptive error-control theorem is never exercised numerically: the experiments build the background space once at t=0 and keep it fixed, so the finite-partition relinearization loop remains an untested algorithmic component.
  • The theory requires lambda > 0 and T_t to remain a diffeomorphism; the numerical rollback when the Jacobian approaches zero (as in the Gaussian-mixture merging test) falls outside the proven guarantees, so treating near-singular maps rigorously is a natural extension.
  • The same decoder and projection machinery could be applied to parametric or conditional gradient flows by treating parameters as extra variables, since nothing in the construction depends on the specific form of the time interval.
  • A practical test of the adaptive strategy on a two-dimensional porous-medium or double-well example, with scheduled background updates and measured residual integrals, would show whether the constants in the theorem remain bounded as predicted.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a meshless Lagrangian method for approximating Wasserstein gradient flows. The evolving measure is represented by pushforward of the initial measure through a transport map, which is approximated in a time-dependent linear subspace of L^2_{μ0}. The subspace is parametrized by a Stiefel decoder constrained to a finite-dimensional background space built from Taylor approximations of the velocity operator. The authors prove that the approximate transport map yields an absolutely continuous curve of measures with a projected velocity field, derive a Gronwall-type a posteriori Wasserstein error bound in terms of the projection residual, and prove an energy-dissipation identity up to a projection error. For geodesically convex energies, they state a convergence theorem (Theorem 5.1) claiming that adaptive relinearization of the background space can make the final-time Wasserstein error arbitrarily small. Numerical experiments on linear and nonlinear Fokker–Planck equations, porous-medium diffusion, and interaction energies illustrate the method's behavior.

Significance. If the main theorem is valid, the paper provides a novel and attractive bridge between Dirac–Frenkel dynamical approximation and Wasserstein gradient flows: it avoids meshes and optimal-transport solves at each step, produces regular measure approximations rather than empirical measures, and gives a rigorous a posteriori control of the projection error for geodesically convex energies. The proof of the a posteriori bound (Theorem 4.2) and the energy-dissipation identity (Proposition 4.2) are clear and appear correct under the stated regularity assumptions. The Taylor-based construction of the background space is a useful idea, and the fixed-background numerical experiments show competitive accuracy in low-dimensional benchmarks. However, the central convergence theorem is not proved as written because of an incorrect identity in the local residual analysis, and the adaptive relinearization mechanism it depends on is never implemented or tested. These issues are repairable in principle, but the manuscript currently overstates what is established.

major comments (3)
  1. [§5.2, Eq. (31)] The identity \|\dot T_t - L^app_{\bar t,\bar\varepsilon}(T_t)\|=0 is false in general. By Eq. (24), \dot T_t = P_{W_m} L(T_t) in the non-degenerate regime, while L^app is P_{W_m} applied to the first-order Taylor expansion of L at \bar T. Hence \dot T_t - L^app = P_{W_m} R_2(T_t), where R_2 is the Taylor remainder. For nonlinear L (e.g., porous-medium diffusion in §6.3) R_2 is generically nonzero and not contained in W_m^\perp. Thus the 'first term vanishes' step leading to (32) is not justified. The proof can likely be repaired by replacing (31) with the inequality \|P_{W_m}R_2(T_t)\| ≤ C_{\bar t,2}\|T_t-\bar T\|_{H^2}^2 and absorbing the extra term into the existing quadratic bound, but as written Lemma 5.1 is not proved.
  2. [§5.2, Theorem 5.1 and H3] The convergence guarantee is conditional on the existence, at every relinearization time, of an H^4_{\mu0} linearization map \bar T within H^2 tolerance \bar\varepsilon, with uniformly bounded constants C_{\bar t_j,1}, C_{\bar t_j,2}. No constructive procedure is given for producing such a map, and Section 6 explicitly states that 'the background space is built only once, at the initial time, and then kept fixed'. The adaptive error-control theorem is therefore never exercised, and the numerical experiments do not demonstrate the claimed advantages of the adaptive construction. This is a significant gap between the theoretical claim and the computational evidence.
  3. [§4.2, §5.2, §6.1] The theory assumes throughout that T_t remains a diffeomorphism so that ν_t has a density and the subdifferential ∂F(ν_t) is defined. The numerical scheme in Section 6.1 monitors the Jacobian determinant and, when invertibility is lost, rolls back to an earlier time and restarts; this mechanism is not part of the analytical framework and is not covered by the convergence guarantees. The same applies to the composition of maps after rollback. This should be stated explicitly as a limitation rather than left implicit.
minor comments (6)
  1. [Section 3.2] Typo: 'Dirac-Fenkel' should be 'Dirac-Frenkel'.
  2. [Theorem 4.2 heading] 'Gronwald-type' should be 'Gronwall-type'.
  3. [Section 5.1 / Eq. (26)] The weighted Sobolev spaces H^k_{\mu0} are used throughout the error analysis but never defined. Please specify the norm and the regularity assumptions on μ0.
  4. [Eq. (35)] The term \|T_{\bar t} - T_{\bar t}\|_{L^2} appears twice in a way that is notationally confusing; one of the two occurrences should refer to the exact trajectory and the other to the approximate one. Please clarify.
  5. [Section 6.1 / Figure 3] The Jacobian diagnostic and rollback procedure are described only informally. A precise algorithmic description (how the threshold is chosen, how the composition is defined) would help reproducibility.
  6. [Figure 5 caption] There is a spacing artifact in 'T aylor-based' in the caption.

Circularity Check

0 steps flagged

No circular step: the scheme is a projection of the exact velocity with a posteriori residual bounds; Eq. (31) is a proof gap, not a circularity.

full rationale

After walking the derivation chain, I find no circular step. The approximate curve is defined by the projected dynamics (24), \dot T_t = P_{W_m}(v_{\nu_t}\circ T_t) in the nondegenerate regime, and the residual r_t is definitionally the orthogonal projection error of the exact velocity onto the current subspace. Theorem 4.2's Wasserstein bound and Proposition 4.2's dissipation identity follow from the chain rule, the two-curve estimate (72), and orthogonality of the projection, not from fitting any target quantity. The Taylor background space (29)-(30) is built from L(\bar T) and DL(\bar T)[p_i], and the residual bound (32) is an a posteriori estimate whose constants C_{\bar t_j,i} are hypotheses (H3), not fitted data; no fitted parameter is renamed as a prediction. The only author-overlap citation is [33] (Mula) in the literature survey on dynamical sampling, and it is not load-bearing. Two flagged issues are correctness/limitation concerns rather than circularity: (i) Eq. (31) asserts \|\dot T_t - L^{app}_{\bar t_j,\bar\varepsilon_j}(T_t)\|_{L^2_{\mu_0}}=0 'by construction,' but from (24) \dot T_t - L^{app}_{\bar t_j,\bar\varepsilon_j}(T_t) = P_{W_m} R_2(T_t), and for nonlinear L the Taylor remainder need not lie in W_m^\perp, so Lemma 5.1's first-term vanishing is not proved as written; (ii) Section 6 states 'the background space is built only once, at the initial time, and then kept fixed,' so the adaptive relinearization of Theorem 5.1 is never exercised. These do not make the derivation circular.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The method introduces no new physical entities. The free parameters are numerical and algorithmic choices (basis size, perturbation directions, update schedule) rather than constants fitted to make the theory match data. The axioms are mostly standard WGF theory plus the stated hypotheses (H1)-(H3); the least supported axiom is the existence/availability of accurate linearization maps in the adaptive strategy, which the paper does not construct or test.

free parameters (3)
  • n (number of basis functions / frame size) = 2 to 40 depending on the experiment
    Chosen by hand per test; larger n lowers W2 error (Figure 5) but is not determined by the method.
  • P / perturbation directions p_i = all monomials up to a prescribed degree, e.g. degree 5 in 2D when n=40
    Ad hoc choice; the paper notes there is no selection criterion and that this causes dimension-dependent growth of the background space.
  • Relinearization times and tolerances = not instantiated
    Theorem 5.1 asserts existence of a finite partition and tolerances but gives no concrete rule, and the experiments do not use relinearization.
axioms (6)
  • standard math Standard Wasserstein-gradient-flow theory: AC curves, chain rule, metric derivative identity, and lambda-convexity monotonicity estimates from Ambrosio-Gigli-Savare.
    Used throughout Sections 2, A, and D; standard background results are invoked without proof.
  • domain assumption The energy F has the form (8) and is sufficiently regular that the Wasserstein subdifferential lies in the tangent space (H1).
    Restricts the method to densities with first variation depending on F, rho, and grad rho, excluding singular measures and some nonlocal cases without diffusion.
  • domain assumption F is lambda-geodesically convex with lambda > 0 (H2).
    Needed for the exponential/Gronwall estimates in Theorem 4.2 and Theorem 5.1; it excludes nonconvex energies like the double-well examples from the convergence guarantee.
  • domain assumption The Lagrangian velocity operator L is C^2 on the affine spaces W_m + T-bar, with uniformly bounded constants C_{t_j,1} and C_{t_j,2} (H3).
    Needed for the Taylor expansion and local residual control; not verified algorithmically.
  • ad hoc to paper At each relinearization time an H^4_{mu0} linearization map T-bar within H^2_{mu0} tolerance exists and is available, with fixed perturbation directions p_i.
    No construction or algorithm is supplied; Section 6 states the background is built only once, so this hypothesis is the main non-constructive step in Theorem 5.1.
  • domain assumption The approximate transport maps T_t are diffeomorphisms on [0,t_f].
    Needed to define nu_t = T_t#mu_0 with a Lebesgue density and to evaluate dF(nu_t); the implementation instead detects loss of invertibility via the Jacobian and rolls back (Figure 3).

pith-pipeline@v1.3.0-alltime-deepseek · 30852 in / 15287 out tokens · 154091 ms · 2026-08-01T18:37:04.157981+00:00 · methodology

0 comments
read the original abstract

We propose a meshless Lagrangian dynamical method for approximating Wasserstein gradient flows (WGFs). The evolving measure is represented as the pushforward of the initial measure $\mu_0$ through a transport map in the weighted Hilbert space $L^2_{\mu_0}$. We approximate this map in time-dependent linear subspaces of $L^2_{\mu_0}$, whose orthonormal frames are evolved by a Dirac--Frenkel dynamical principle on a Stiefel manifold constrained to a finite-dimensional background space, adaptively constructed via local approximations of the WGF velocity field. We prove that the resulting transport map induces an absolutely continuous curve of probability measures in Wasserstein space, whose velocity is obtained by projecting the exact WGF velocity onto the background space, and we show that the approximation preserves the energy dissipation structure up to the projection error of the velocity. Moreover, for geodesically convex energies, we derive an a posteriori estimate controlling such projection error through the adaptive construction of the background space, yielding as well a bound on the approximation error of the pushforward measure in the Wasserstein metric. Numerical experiments on linear and nonlinear Fokker--Planck equations, porous-medium diffusion, and interaction energies demonstrate the accuracy of the method, its energy-dissipation properties, and the advantages of the adaptive construction.

Figures

Figures reproduced from arXiv: 2607.17235 by Federico Pichi, Gianluigi Rozza, Isabella Carla Gonnella, Olga Mula.

Figure 1
Figure 1. Figure 1: One-dimensional linear Fokker–Planck equation with n = 2, initialized with a centered Gaussian µ0 ∼ N (0.0, 0.5). The density νt and the transport map Tt are compared with the analytic solutions in 1a, and the Wasserstein error is reported together with the energy decay and its theoretical upper bound in 1b. The results are obtained with a time-step size ∆t = 2−6 , tf = 3 and Nsamples = 215 . In this conte… view at source ↗
Figure 2
Figure 2. Figure 2: Two-dimensional linear Fokker–Planck equation with n = 6, initialized with a non-centered Gaussian µ0 ∼ N (2.0, 2.0), diag((0.3, 0.3)) . The density evolution is displayed in 2a, and the Wasserstein error is reported together with the energy decay and its theoretical upper bound in 2b. The results are obtained with a time-step size ∆t = 2−6 and Nsamples = 215. The reference solution is computed with a fin… view at source ↗
Figure 3
Figure 3. Figure 3: One-dimensional linear Fokker–Planck equation with n = 8, initialized with µ0 ∼ 1 2 N (−3, 0.3) + 1 2 N (3, 0.5). The density evolution is shown in 3a: the upper panel is computed using the exact pushforward formula, while the lower one from the histogram of the transported samples. The Wasserstein error is reported together with the energy decay and its theoretical upper bound in 3b, as well as the tempor… view at source ↗
Figure 4
Figure 4. Figure 4: One-dimensional double-well Fokker–Planck equation for kL = 0.001 with n = 12, initialized with centered Gaussian µ0 ∼ N (0, 0.5). The density evolution is displayed in 4a, and the Wasserstein error is reported together with the energy decay in 4b. The results are obtained with a time-step size ∆t = 2−5 and Nsamples = 217 . We next consider a non-convex confinement potential, namely the double-well potenti… view at source ↗
Figure 5
Figure 5. Figure 5: Error comparison for the one-dimensional double-well Fokker–Planck equation initialized both with a centered Gaussian µ0 ∼ N (0, 0.5) (CG) and multi-Gaussian µ0 ∼ 1 2 N (−3, 0.3) + 1 2 N (3, 0.5) (MG). The plots report the Wasserstein error with respect to a finite element reference solution, displayed for different values of the number of basis functions, the number of samples, and the time-step size used… view at source ↗
Figure 6
Figure 6. Figure 6: Two-dimensional double-well Fokker–Planck equation for kL = 0.01 with n = 40, initialized with a centered Gaussian µ0 ∼ N (0, 0), diag(0.5, 0.5) . The density evolution is displayed in 6a, and the sliced Wasserstein error is reported together with the energy decay in 6b. The results are obtained with a time-step size ∆t = 2−8 and Nsamples = 215. The reference solution is computed with a finite element met… view at source ↗
Figure 7
Figure 7. Figure 7: One-dimensional nonlinear diffusion equation with m = 2 and n = 13, initialized with a centered Gaussian µ0 ∼ N (0.0, 0.5). The density evolution is displayed in 7a, and the Wasserstein error is reported together with the energy decay in 7b. The results are obtained with a time-step size ∆t = 2−6 and Nsamples = 215 . We finally consider the nonlinear diffusion equation, corresponding to the Wasserstein gra… view at source ↗
Figure 8
Figure 8. Figure 8: Energy decay (on the left) and spreading rate (on the right) for the one-dimensional nonlinear diffusion equation for different values of m with n = 13, ∆t = 2−6 and Nsamples = 215, initialized with a centered Gaussian µ0 ∼ N (0.0, 0.5). The standard deviation of the numerical solution is compared with the expected self-similar scaling law. and the energy decreases along the whole time interval. The associ… view at source ↗
Figure 9
Figure 9. Figure 9: Two-dimensional interaction equation with n = 30, initialized with a centered Gaussian µ0 ∼ N [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

46 extracted references · 11 linked inside Pith

  1. [1]

    The geometry of dissipative evolution equations: The porous medium equation.Comm

    Felix Otto. The geometry of dissipative evolution equations: The porous medium equation.Comm. Partial Differential Equations, 26(1–2):101–174, January 2001

    2001

  2. [2]

    Carrillo, Robert J

    Jos´ e A. Carrillo, Robert J. McCann, and C´ edric Villani. Kinetic equilibration rates for granular media and related equations: entropy dissipation and mass transportation estimates.Rev. Mat. Iberoameri- cana, 19(3):971–1018, December 2003

    2003

  3. [3]

    Wasserstein quantum Monte Carlo: A novel approach for solving the quantum many-body Schr¨ odinger equation.arXiv preprint arXiv:2307.07050, 2023

    Kirill Neklyudov, Jannes Nys, Luca Thiede, Juan Carrasquilla, Qiang Liu, Max Welling, and Alireza Makhzani. Wasserstein quantum Monte Carlo: A novel approach for solving the quantum many-body Schr¨ odinger equation.arXiv preprint arXiv:2307.07050, 2023

    Pith/arXiv arXiv 2023

  4. [4]

    Lauren Conger, Franca Hoffmann, Eric Mazumdar, and Lillian J. Ratliff. Coupled Wasserstein gradient flows for min-max and cooperative games.arXiv preprint arXiv:2411.07403, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Dataset dynamics via gradient flows in probability space.arXiv preprint arXiv:2010.12760, 2020

    David Alvarez-Melis and Nicol` o Fusi. Dataset dynamics via gradient flows in probability space.arXiv preprint arXiv:2010.12760, 2020

    Pith/arXiv arXiv 2010

  6. [6]

    A unifying view of variational generative Wasser- stein flows.arXiv preprint arXiv:2605.31369, 2026

    Paul Caucheteux, Cl´ ement Bonet, and Anna Korba. A unifying view of variational generative Wasser- stein flows.arXiv preprint arXiv:2605.31369, 2026

    Pith/arXiv arXiv 2026

  7. [7]

    Flowing with confidence.arXiv preprint arXiv:2605.18472, 2026

    Friso de Kruiff, Dario Coscia, Max Welling, and Erik Bekkers. Flowing with confidence.arXiv preprint arXiv:2605.18472, 2026

    Pith/arXiv arXiv 2026

  8. [8]

    The variational formulation of the fokker–planck equation.SIAM Journal on Mathematical Analysis, 29(1):1–17, 1998

    Richard Jordan, David Kinderlehrer, and Felix Otto. The variational formulation of the fokker–planck equation.SIAM Journal on Mathematical Analysis, 29(1):1–17, 1998

    1998

  9. [9]

    Birkh¨ auser Basel, 2008

    Luigi Ambrosio, Nicola Gigli, and Giuseppe Savar´ e.Gradient flows: in metric spaces and in the space of probability measures. Birkh¨ auser Basel, 2008

    2008

  10. [10]

    A computational fluid mechanics solution to the Monge- Kantorovich mass transfer problem.Numer

    Jean-David Benamou and Yann Brenier. A computational fluid mechanics solution to the Monge- Kantorovich mass transfer problem.Numer. Math., 84(3):375–393, January 2000

    2000

  11. [11]

    An augmented lagrangian approach to wasserstein gradient flows and applications.ESAIM: Proceedings and surveys, 54:1–17, 2016

    Jean-David Benamou, Guillaume Carlier, and Maxime Laborde. An augmented lagrangian approach to wasserstein gradient flows and applications.ESAIM: Proceedings and surveys, 54:1–17, 2016

    2016

  12. [12]

    Primal dual methods for wasserstein gradient flows.Foundations of Computational Mathematics, 22(2):389–443, 2022

    Jos´ e A Carrillo, Katy Craig, Li Wang, and Chaozhen Wei. Primal dual methods for wasserstein gradient flows.Foundations of Computational Mathematics, 22(2):389–443, 2022

    2022

  13. [13]

    Entropic approximation of wasserstein gradient flows.SIAM Journal on Imaging Sciences, 8(4):2323–2351, 2015

    Gabriel Peyr´ e. Entropic approximation of wasserstein gradient flows.SIAM Journal on Imaging Sciences, 8(4):2323–2351, 2015. 28

    2015

  14. [14]

    Convergence of en- tropic schemes for optimal transport and gradient flows.SIAM Journal on Mathematical Analysis, 49(2):1385–1418, 2017

    Guillaume Carlier, Vincent Duval, Gabriel Peyr´ e, and Bernhard Schmitzer. Convergence of en- tropic schemes for optimal transport and gradient flows.SIAM Journal on Mathematical Analysis, 49(2):1385–1418, 2017

    2017

  15. [15]

    Large-scale wasserstein gradient flows.Advances in Neural Information Processing Systems, 34:15243– 15256, 2021

    Petr Mokrov, Alexander Korotin, Lingxiao Li, Aude Genevay, Justin M Solomon, and Evgeny Burnaev. Large-scale wasserstein gradient flows.Advances in Neural Information Processing Systems, 34:15243– 15256, 2021

    2021

  16. [16]

    Deep jko: Time-implicit particle methods for general nonlinear gradient flows.Journal of Computational Physics, 514:113187, 2024

    Wonjun Lee, Li Wang, and Wuchen Li. Deep jko: Time-implicit particle methods for general nonlinear gradient flows.Journal of Computational Physics, 514:113187, 2024

    2024

  17. [17]

    A blob method for the aggregation equation.Mathematics of computation, 85(300):1681–1717, 2016

    Katy Craig and Andrea Bertozzi. A blob method for the aggregation equation.Mathematics of computation, 85(300):1681–1717, 2016

    2016

  18. [18]

    A blob method for diffusion.Calculus of Variations and Partial Differential Equations, 58(2):53, 2019

    Jos´ e Antonio Carrillo, Katy Craig, and Francesco S Patacchini. A blob method for diffusion.Calculus of Variations and Partial Differential Equations, 58(2):53, 2019

    2019

  19. [19]

    A lagrangian scheme for the solution of nonlinear diffusion equations using moving simplex meshes.Journal of Scientific Computing, 75(3):1463–1499, 2018

    Jos´ e A Carrillo, Bertram D¨ uring, Daniel Matthes, and David S McCormick. A lagrangian scheme for the solution of nonlinear diffusion equations using moving simplex meshes.Journal of Scientific Computing, 75(3):1463–1499, 2018

    2018

  20. [20]

    Jos´ e A Carrillo and J Salvador Moll. Numerical simulation of diffusive and aggregation phenomena in nonlinear continuity equations by evolving diffeomorphisms.SIAM Journal on Scientific Computing, 31(6):4305–4329, 2010

    2010

  21. [21]

    Lagrangian schemes for wasserstein gradient flows.Handbook of Numerical Analysis, 22:271–311, 2021

    Jose A Carrillo, Daniel Matthes, and Marie-Therese Wolfram. Lagrangian schemes for wasserstein gradient flows.Handbook of Numerical Analysis, 22:271–311, 2021

    2021

  22. [22]

    P. A. M. Dirac. Note on exchange phenomena in the thomas atom.Proceedings of the Cambridge Philosophical Society, 26(3):376–385, 1930

    1930

  23. [23]

    Natural gradient via optimal transport.Information Geometry, 1(2):181–214, 2018

    Wuchen Li and Guido Mont´ ufar. Natural gradient via optimal transport.Information Geometry, 1(2):181–214, 2018

    2018

  24. [24]

    Neural parametric Fokker–Planck equations

    Shu Liu, Wuchen Li, Hongyuan Zha, and Haomin Zhou. Neural parametric Fokker–Planck equations. arXiv preprint arXiv:2002.11309, 2020

    Pith/arXiv arXiv 2002

  25. [25]

    Numerical analysis on neural net- work projected schemes for approximating one dimensional Wasserstein gradient flows.arXiv preprint arXiv:2402.16821, 2024

    Xinzhe Zuo, Jiaxi Zhao, Shu Liu, Stanley Osher, and Wuchen Li. Numerical analysis on neural net- work projected schemes for approximating one dimensional Wasserstein gradient flows.arXiv preprint arXiv:2402.16821, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Natural gradient works efficiently in learning.Neural Computation, 10(2):251–276, 1998

    Shun-ichi Amari. Natural gradient works efficiently in learning.Neural Computation, 10(2):251–276, 1998

    1998

  27. [27]

    Shun-ichi Amari and Scott C. Douglas. Why natural gradient? InProceedings of the IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1213–1216, 1998

    1998

  28. [28]

    American Mathematical Society, Providence, RI, 2000

    Shun-ichi Amari and Hiroshi Nagaoka.Methods of Information Geometry, volume 191 ofTranslations of Mathematical Monographs. American Mathematical Society, Providence, RI, 2000

    2000

  29. [29]

    Absil, Robert Mahony, and Rodolphe Sepulchre.Optimization Algorithms on Matrix Manifolds

    P.-A. Absil, Robert Mahony, and Rodolphe Sepulchre.Optimization Algorithms on Matrix Manifolds. Princeton University Press, Princeton, NJ, 2008

    2008

  30. [31]

    Neural galerkin schemes with active learning for high-dimensional evolution equations.arXiv preprint arXiv:2203.01360, 2022

    Joan Bruna, Benjamin Peherstorfer, and Eric Vanden-Eijnden. Neural galerkin schemes with active learning for high-dimensional evolution equations.arXiv preprint arXiv:2203.01360, 2022

    Pith/arXiv arXiv 2022

  31. [32]

    Lubich and J

    C. Lubich and J. Nick. Regularized dynamical parametric approximation of stiff evolution problems, 2025. 29

    2025

  32. [33]

    Stable nonlinear dynamical approximation with dynamical sampling.arXiv preprint arXiv:2505.11938, 2025

    Daan Bon, Benjamin Caris, and Olga Mula. Stable nonlinear dynamical approximation with dynamical sampling.arXiv preprint arXiv:2505.11938, 2025

    Pith/arXiv arXiv 2025

  33. [34]

    Regularized dynamical parametric approximation.Numerische Mathematik, pages 1–45, 2026

    Michael Feischl, Caroline Lasser, Christian Lubich, and J¨ org Nick. Regularized dynamical parametric approximation.Numerische Mathematik, pages 1–45, 2026

    2026

  34. [35]

    Learning Monge maps by lifting and constraining Wasserstein gradient flows.arXiv preprint arXiv:2603.25182, 2026

    Th´ eo Dumont, Th´ eo Lacombe, and Fran¸ cois-Xavier Vialard. Learning Monge maps by lifting and constraining Wasserstein gradient flows.arXiv preprint arXiv:2603.25182, 2026

    Pith/arXiv arXiv 2026

  35. [36]

    Polar factorization and monotone rearrangement of vector-valued functions.Comm

    Yann Brenier. Polar factorization and monotone rearrangement of vector-valued functions.Comm. Pure Appl. Math., 44(4):375–417, June 1991

    1991

  36. [37]

    An introduction to the analysis of gradients systems.arXiv preprint arXiv:2306.05026, 2023

    Alexander Mielke. An introduction to the analysis of gradients systems.arXiv preprint arXiv:2306.05026, 2023

    Pith/arXiv arXiv 2023

  37. [38]

    Birkhauser Verlag, 2015

    Filippo Santambrogio.Optimal Transport for Applied Mathematicians Calculus of Variations, PDEs, and Modeling. Birkhauser Verlag, 2015

    2015

  38. [39]

    Frenkel.Wave Mechanics: Advanced General Theory

    J. Frenkel.Wave Mechanics: Advanced General Theory. Clarendon Press, Oxford, 1934

    1934

  39. [40]

    A. D. McLachlan. A variational solution of the time-dependent schrodinger equation.Molecular Physics, 8(1):39–44, 1964

    1964

  40. [41]

    European Mathematical Society, 2008

    Christian Lubich.From Quantum to Classical Molecular Dynamics: Reduced Models and Numerical Analysis. European Mathematical Society, 2008

    2008

  41. [42]

    Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computational Physics, 404:108973, 2020

    Kookjin Lee and Kevin Carlberg. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders.Journal of Computational Physics, 404:108973, 2020

    2020

  42. [43]

    Model reduction for transport-dominated problems via online adaptive bases and adaptive sampling.SIAM Journal on Scientific Computing, 42(5):A2803–A2836, 2020

    Benjamin Peherstorfer. Model reduction for transport-dominated problems via online adaptive bases and adaptive sampling.SIAM Journal on Scientific Computing, 42(5):A2803–A2836, 2020

    2020

  43. [44]

    Dynamically orthogonal field equations for contin- uous stochastic dynamical systems.Physica D: Nonlinear Phenomena, 238(23-24):2347–2360, 2009

    Themistoklis P Sapsis and Pierre FJ Lermusiaux. Dynamically orthogonal field equations for contin- uous stochastic dynamical systems.Physica D: Nonlinear Phenomena, 238(23-24):2347–2360, 2009

    2009

  44. [45]

    Dynamical low-rank approximation.SIAM Journal on Matrix Analysis and Applications, 29(2):434–454, 2007

    Othmar Koch and Christian Lubich. Dynamical low-rank approximation.SIAM Journal on Matrix Analysis and Applications, 29(2):434–454, 2007

    2007

  45. [46]

    Efficient natural gradient descent methods for large- scale pde-based optimization problems.SIAM Journal on Scientific Computing, 45(4):A1621–A1655, 2023

    Levon Nurbekyan, Wanzhou Lei, and Yunan Yang. Efficient natural gradient descent methods for large- scale pde-based optimization problems.SIAM Journal on Scientific Computing, 45(4):A1621–A1655, 2023

    2023

  46. [47]

    Geodesics in infinite dimensional stiefel and grassmann manifolds.Comptes Rendus

    Philipp Harms and Andrea CG Mennucci. Geodesics in infinite dimensional stiefel and grassmann manifolds.Comptes Rendus. Math´ ematique, 350(15-16):773–776, 2012. 30 A Convexity in the Wasserstein space In this appendix, we first recall the definition of Wasserstein geodesics in Definition A.1, and then use this geodesic structure to state the notion ofλ-c...

    2012