Pith. sign in

REVIEW 5 minor 1 cited by

Linear convergence of proximal descent schemes on the Wasserstein space

T0 review · 0 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read This paper proves that three proximal descent schemes on the Wasserstein space converge linearly to the minimizer of an entropy-regularized objective, under flat convexity rather than geodesic convexity.

desk verdict The first discrete-time linear convergence rates for JKO-type schemes under flat convexity; the uniform boundedness assumption on the flat derivative is the real price of admission. read the letter →

arxiv 2411.15067 v2 pith:24RRUGGY submitted 2024-11-22 math.OC cs.LGmath.PR

Razvan-Andrei Lascu , Mateusz B. Majka , David Šiška , {L}ukasz Szpruch This is my paper · ORCID
classification math.OCcs.LGmath.PR MSC 46N1049Q2249K3058E30
keywords entropyregularizationproximalJKOschemesWassersteinspacemean-fieldoptimizationlogarithmicSobolevinequalityflatconvexityoptimaltransportlinearconvergence
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 linear convergence for three discrete-time proximal methods on the space of probability measures: a proximal point scheme, a prox-linear scheme, and a proximal gradient scheme. Each method targets the minimizer of an entropy-regularized objective F_sigma(mu) = F(mu) + sigma KL(mu|pi), and the authors show the objective gap contracts by a fixed factor every step. This relaxes the common requirement that F be geodesically convex in the Wasserstein space, replacing it with flat convexity, a bounded flat derivative, and a strongly convex reference potential. The proof avoids discrete-time Evolution Variational Inequalities and instead uses a uniform logarithmic Sobolev inequality for the proximal Gibbs measures together with an entropy sandwich lemma. The main technical challenge is proving that the relative entropy is Wasserstein subdifferentiable along the iterates, which is handled by showing the iterates stay in a Sobolev regularity class.

What carries the argument

The load-bearing object is the proximal Gibbs measure Phi[mu] proportional to exp(-$sigma^{{-1}}$ delta F/delta mu (mu, .) - U), which for every mu satisfies a logarithmic Sobolev inequality with uniform constant alpha_U $e^{{-4 C_F / sigma}}$, obtained by the Holley–Stroock criterion applied to the strongly log-concave reference pi proportional to $e^{{-U}}$. The entropy sandwich lemma (Lemma 2.7) bounds the objective gap F_sigma(mu) - F_sigma(mu*_sigma) between $\sigma$ KL(mu|mu*_sigma) and $\sigma$ KL(mu|Phi[mu]). The proof then shows each scheme's first-order optimality condition identifies the relative Fisher information I(mu_{n+1}|Phi[mu]) with a Wasserstein displacement, and the log-Sobolev inequality converts that Fisher information into KL, producing a contraction in the objective gap. A Sobolev regularity class C of densities is introduced so that the relative entropy KL(·|pi) has a unique Wasserstein subgradient, given by the log-density gradient, at every iterate, which makes the Fisher information finite and the optimality conditions valid.

What would settle it

In the exactly solvable Gaussian case with U quadratic and F = 0, the JKO iterates are explicit; compute the objective gap and compare the contraction ratio with the predicted kappa = 1 + tau sigma alpha_U. A single asymptotic contraction factor smaller than the predicted one would refute the claimed Q-linear rate of Theorem 3.1(i).

Watch

Extended reading notes

Core claim

The central claim is Theorem 3.1: for each of the schemes (1.3), (1.4), and (1.5), under Assumptions 2.1–2.3 and 2.8 (plus Assumption 2.9 for the latter two, with small step-size bounds), there exists kappa > 1 such that 0 <= F_sigma(mu_n) - F_sigma(mu*_sigma) <= $kappa^{{-n}}$(F_sigma(mu_0) - F_sigma(mu*_sigma)) for all n. Corollary 3.2 transfers this exponential contraction to the relative entropy KL(mu_n|mu*_sigma) and the squared Wasserstein distance $W_2^{2}$(mu_n, mu*_sigma). The purpose is to show that geodesic convexity is not needed for linear rates in mean-field optimization: flat convexity plus a uniformly bounded flat derivative and a strongly log-concave reference measure suffice. The proof works by identifying, through first-order optimality conditions, the relative Fisher information with the squared Wasserstein displacement at each step, then using the uniform log-Sobolev inequality and the sandwich lemma to convert Fisher information into a one-step objective decrease.

Load-bearing premise

The uniform bound |delta F/delta mu (mu, x)| <= C_F for all mu and x is what keeps the log-Sobolev constant of the proximal Gibbs measure from degenerating; if that bound fails, the proof has no positive constant to work with.

Editorial extensions

If this is right

  • The objective gap F_sigma(mu_n) - F_sigma(mu*_sigma) contracts by a fixed factor at every step for all three schemes, so the methods are Q-linearly convergent in objective value.
  • The same exponential contraction transfers to KL(mu_n|mu*_sigma) and W_2^2(mu_n, mu*_sigma), with explicit constants in Corollary 3.2.
  • Geodesic convexity of F is not needed; flat convexity, bounded flat derivative, and a strongly convex reference potential are enough for linear rates.
  • The small step-size restrictions are explicit: tau < 2/L'_F for the prox-linear scheme and tau < 1/L'_F for the proximal gradient scheme.
  • The two-layer mean-field neural network with a clipped activation function satisfies the assumptions, so the rates apply to that training objective.

Reading between the lines

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

  • The explicit contraction factors suggest the practical step-size constraints are real and quantitative: the rate degrades as the Wasserstein gradient becomes less smooth, and particle implementations would need to respect the same tau bounds.
  • Remark 2.6 points to a relaxation: if uniform Lipschitzness of x |-> delta F/delta mu (mu, x) replaces the boundedness of delta F/delta mu, the two-layer network example can drop its clipping function; this relaxation is left as a remark but is a testable route to broader applicability.
  • The linear rates are in the mean-field limit; the paper does not quantify how the contraction interacts with particle discretization error, which seems to be the natural next question for practical algorithms.
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.

Referee Report

0 major / 5 minor

Summary. The paper studies the entropy-regularized mean-field optimization problem min_μ F_σ(μ)=F(μ)+σ KL(μ|π) on the Wasserstein space and analyzes three discrete-time proximal/JKO-type schemes: the proximal point scheme (1.3), the prox-linear scheme (1.4), and the proximal gradient scheme (1.5). Under flat convexity of F (Assumption 2.1), Lipschitzness and uniform boundedness of the flat derivative (Assumption 2.2), strong convexity and growth conditions on the potential U of π (Assumption 2.3), and suitable Wasserstein smoothness assumptions (Assumptions 2.8–2.9), Theorem 3.1 establishes Q-linear convergence of the objective gap with explicit contraction rates; Corollary 3.2 transfers these rates to KL divergence and squared Wasserstein distance. The proof strategy combines a uniform logarithmic Sobolev inequality for the proximal Gibbs measures Φ[μ], obtained via Holley–Stroock, with the entropy sandwich lemma (Lemma 2.7), first-order optimality conditions, and a technical regularity result showing that the relative entropy admits a unique Wasserstein subgradient along the iterates.

Significance. If the result is correct, it is a useful discrete-time counterpart to the continuous-time LSI-based analysis of mean-field Langevin dynamics and extends JKO-type convergence theory beyond geodesically convex functionals. The paper's strengths are its rigor and completeness: existence and uniqueness of minimizers for each scheme are proved rather than assumed, the Wasserstein subdifferentiability of the relative entropy along the iterates is handled via a Sobolev-regularity class C, and all rates are given with explicit constants. The main caveat is that Assumption 2.2(2.3), the uniform bound on δF/δμ, is essential for the uniform LSI constant in (2.7) and also enters the proof of existence in Theorem 5.1; this condition excludes natural flat-convex objectives such as F(μ)=∫|x|^2 dμ(dx), so the advertised relaxation from geodesic convexity is narrower than the abstract suggests. This is a scope limitation rather than an internal inconsistency of the proof.

minor comments (5)
  1. [Lemma 6.6 and §4.2] Lemma 6.6 is stated only for μ′, μ ∈ C, but in the proof of Theorem 3.1(ii) it is applied with μ = μ0, and μ0 is assumed only to lie in P^λ_2. The proof of the lemma itself only needs μ ∈ P^λ_2 and μ′ ∈ C, so the statement should be relaxed accordingly; the sentence just before (4.4) claiming that all iterates (μn)n∈N lie in C is also inaccurate at n = 0.
  2. [Remark 2.6 and Theorem 5.1 Step 3] The remark that condition (2.3) can be replaced by uniform Lipschitzness of x ↦ δF/δμ(μ,x) is not sufficient as written, because (2.3) is also used to obtain the dominated convergence bound in Step 3 of the proof of Theorem 5.1. Either prove the full statement under the weaker assumption or restrict the remark to the LSI constant.
  3. [Example 3.3] The displayed formula for δF/δμ appears to omit the factor 2 from the derivative of the squared L2 loss and the normalization constant required by Definition B.1; please verify the computation and report the normalized derivative.
  4. [Abstract and Introduction] The phrase 'relaxing the common reliance on geodesic convexity' should be qualified: Assumption 2.2(2.3), the uniform boundedness of the flat derivative, is essential for the uniform LSI constant in (2.7) and excludes basic flat-convex functionals such as F(μ)=∫|x|^2 dμ(dx). A sentence stating this scope would improve the paper.
  5. [Various] There are minor typographical issues ('auxilliary' in Section 7, 'if identical' in Proposition 6.3, and the header of Theorem 3.1 reading '1.5'); these do not affect the mathematics.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the linear-rate proof is derived from explicit assumptions and independent LSI/sandwich results; the only self-citation is [17] in Example 3.3 and is not load-bearing.

full rationale

The main claim (Theorem 3.1) is proved directly from Assumptions 2.1, 2.2, 2.3, 2.8 and 2.9. The rates are explicit functions of the assumed constants α_U, C_F, L'_F, τ and σ; no parameter is fitted to the target quantities. The two imported ingredients, Lemma 2.7 (entropy sandwich) and the uniform LSI (2.7), come from [27,12] and from the Bakry–Emery and Holley–Stroock criteria [3,16], all independent of the present authors. The self-citation is [17] (Hu, Ren, Šiška, Szpruch), used only in Example 3.3 to verify that the two-layer mean-field neural-network loss is flat-convex; the convergence theorems do not rely on this example. The boundedness condition (2.3) is a genuine hypothesis that drives the Holley–Stroock constant e^{-4C_F/σ} in the rates; dropping it would degenerate the constant, but this is a scope limitation, not a circular reduction. The proof does not use its conclusion as an input, and no equation of the form 'prediction equals fitted input' appears.

Assumptions & free parameters 0 free parameters · 8 assumptions · 0 invented entities

This is a pure mathematics paper with no fitted parameters, no numerical constants tuned to data, and no invented entities. It relies on standard tools from optimal transport, logarithmic Sobolev inequalities, and convex analysis, and imports the entropy sandwich lemma from prior independent work. The paper's own assumptions (strong convexity of U, boundedness and Lipschitzness of the flat derivative) are domain assumptions that provide the hypotheses under which the external theorems apply.

assumptions (8)
  • standard math Bakry-Emery criterion: a log-concave measure with potential U that is alpha_U-strongly convex satisfies LSI with constant alpha_U.
    Invoked in Section 2 after Definition 2.5 to assert pi satisfies LSI with constant alpha_U.
  • standard math Holley-Stroock criterion: a bounded perturbation of a measure satisfying LSI preserves LSI, with the constant degraded by an exponential of the oscillation of the perturbation.
    Used in Section 2 to derive the uniform LSI (2.7) for Phi[mu] from boundedness (2.3) of the flat derivative.
  • standard math Otto-Villani theorem: LSI implies Talagrand's inequality.
    Used in (2.8) and Corollary 3.2 to pass from KL convergence to W2^2 convergence.
  • standard math Entropy sandwich lemma (Lemma 2.7) from [27,12]: under flat convexity of F, sigma KL(mu|mu*_sigma) <= F_sigma(mu)-F_sigma(mu*_sigma) <= sigma KL(mu|Phi[mu]).
    Used as a black box in the main proof to lower-bound KL(mu|Phi[mu]) by the objective gap.
  • standard math Characterization of the metric slope and subdifferential of KL(.|pi) (Theorem A.5 from [1]).
    Used in Propositions 5.3, 6.3, 7.4 to identify the Wasserstein subgradient of KL as grad log(dmu/dpi) when mu lies in the Sobolev class C.
  • standard math Geodesic convexity of KL(.|pi) for log-concave pi (Ambrosio-Gigli-Savare Theorem 9.4.10).
    Used in Lemma 6.6 to obtain the KL descent inequality via interpolation along generalized geodesics.
  • domain assumption Assumption 2.3: U is alpha_U-strongly convex with at least quadratic growth.
    Ensures pi has LSI and KL(.|pi) is geodesically convex; stated as a hypothesis of the paper.
  • domain assumption Assumption 2.2(2.3): the flat derivative delta F/delta mu is uniformly bounded by C_F.
    Needed for the Holley-Stroock perturbation argument to yield a uniform LSI constant for Phi[mu].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Linear convergence of proximal descent schemes on the Wasserstein space." pith.science (2026). https://pith.science/paper/24RRUGGY

@misc{pith2026241115067,
  author       = {Pith},
  title        = {Pith review of: Linear convergence of proximal descent schemes on the Wasserstein space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24RRUGGY}},
  note         = {Machine review of arXiv:2411.15067}
}
read the original abstract

We investigate proximal descent methods, inspired by the minimizing movement scheme introduced by Jordan, Kinderlehrer and Otto, for optimizing entropy-regularized functionals on the Wasserstein space. We establish linear convergence under flat convexity assumptions, thereby relaxing the common reliance on geodesic convexity. Our analysis circumvents the need for discrete-time adaptations of the Evolution Variational Inequality (EVI). Instead, we leverage a uniform logarithmic Sobolev inequality (LSI) and the entropy ``sandwich" lemma, extending the analysis from arXiv:2201.10469 and arXiv:2202.01009. The major challenge in the proof via LSI is to show that the relative Fisher information is well-defined at every step of the scheme. Since the relative entropy is not Wasserstein differentiable, we prove that along the scheme the iterates belong to a certain class of Sobolev regularity, and hence the relative entropy has a unique Wasserstein sub-gradient, and that the relative Fisher information is indeed finite.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Non-asymptotic entropic bounds for non-linear kinetic Langevin sampler with second-order splitting scheme

    math.PR 2024-12 conditional novelty 6.0 of 10

    A non-asymptotic relative entropy bound for unadjusted kinetic Langevin Monte Carlo with a second-order splitting scheme, under defective log-Sobolev and Lyapunov conditions more general than prior work.

Reference graph

Works this paper leans on

40 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    Ambrosio, N

    L. Ambrosio, N. Gigli, and G. Savare. Gradient Flows: In Metric Spaces and in the Space of Probability Measures. Lectures in Mathematics. ETH Z¨ urich. Birkh¨ auser Basel, 2008

    work page 2008

  2. [2]

    Arbel, A

    M. Arbel, A. Korba, A. Salim, and A. Gretton. Maximum mean discre pancy gradient flow. In Advances in Neural Information Processing Systems , volume 32. Curran Associates, Inc., 2019. 26

    work page 2019

  3. [3]

    Bakry and M

    D. Bakry and M. ´Emery. Diffusions hypercontractives. S´ eminaire de probabilit´ es de Strasbourg, 19:177–206, 1985

    work page 1985

  4. [4]

    Benamou, G

    J.-D. Benamou, G. Carlier, Q. M´ erigot, and ´E. Oudet. Discretization of functionals involving the Monge–Amp` ere operator.Numerische Mathematik , 134:611–636, 2014

    work page 2014

  5. [5]

    Bogachev

    V. Bogachev. Measure Theory: Volume 1 . Springer Berlin Heidelberg, 2007

    work page 2007

  6. [6]

    Cardaliaguet, F

    P. Cardaliaguet, F. Delarue, J. Lasry, and P. Lions. The Master Equation and the Convergence Problem in Mean Field Games . Annals of Mathematics Studies. Princeton University Press, 2019

    work page 2019

  7. [7]

    R. A. Carmona and F. Delarue. Probabilistic Theory of Mean Field Games with Applications I: Mean Field FBSDEs, Control, and Games . Springer International Publishing, 2018

    2018

  8. [8]

    Cattiaux and A

    P. Cattiaux and A. Guillin. Functional inequalities for perturbed me asures with applications to log-concave measures and to some Bayesian problems. Bernoulli, 28(4):2294 – 2321, 2022

    work page 2022

Show all 40 references
  1. [9]

    F. Chen, Z. Ren, and S. Wang. Entropic fictitious play for mean field optimization problem. Journal of Machine Learning Research , 24(211):1–36, 2023

    2023

  2. [10]

    F. Chen, Z. Ren, and S. Wang. Uniform-in-time propagation of c haos for mean field Langevin dynamics, 2023. arXiv:2212.03050

    2023

  3. [11]

    Chewi, A

    S. Chewi, A. Nitanda, and M. S. Zhang. Uniform-in- n log-Sobolev inequality for the mean-field Langevin dynamics with convex energy, 2024. arXiv:2409.10440

    2024 arXiv

  4. [12]

    L. Chizat. Mean-field Langevin dynamics : Exponential converg ence and annealing. Transactions on Machine Learning Research , 2022

    2022

  5. [13]

    Chizat and F

    L. Chizat and F. R. Bach. On the global convergence of gradien t descent for over-parameterized models using optimal transport. In NeurIPS, 2018

    2018

  6. [14]

    Drusvyatskiy and C

    D. Drusvyatskiy and C. Paquette. Efficiency of minimizing compos itions of convex functions and smooth maps. Mathematical Programming, pages 1–56, 2016

    2016

  7. [15]

    Gangbo and R

    W. Gangbo and R. J. McCann. The geometry of optimal transpo rtation. Acta Mathematica , 177(2):113 – 161, 1996

    1996

  8. [16]

    Holley and D

    R. Holley and D. W. Stroock. Logarithmic Sobolev inequalities and s tochastic Ising models. Journal of Statistical Physics , 46:1159–1194, 1987

    1987

  9. [17]

    K. Hu, Z. Ren, D. ˇSiˇ ska, and /suppress L. Szpruch. Mean-field Langevin dynamics and energy landscape of neural networks. Annales de l’Institut Henri Poincar´ e, Probabilit´ es et Statistiques, 57(4):2043 – 2065, 2021

    2021

  10. [18]

    Huang and Z

    Y.-J. Huang and Z. Malik. Generative modeling by minimizing the Wass erstein-2 loss, 2024. arXiv:2406.13619

    2024 arXiv

  11. [19]

    M. L. Jean-David Benamou, Guillaume Carlier. An augmented Lagr angian approach to Wasserstein gradient flows and applications. ESAIM: Proceedings and Surveys , 54(1):1–17, 2016

    2016

  12. [20]

    Jordan, D

    R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker–Planck equation. SIAM Journal on Mathematical Analysis , 29(1):1–17, 1998

    1998

  13. [21]

    Jourdain and A

    B. Jourdain and A. Tse. Central limit theorem over non-linear functionals of empirical measures with applications to the mean-field fluctuation of interacting diffusions. Electronic Journal of Probability, 26:1 – 34, 2021

    2021

  14. [22]

    Leahy, B

    J.-M. Leahy, B. Kerimkulov, D. ˇSiˇ ska, and /suppress L. Szpruch. Convergence of policy gradient for entropy regularized MDPs with neural network approximation in the mean-fie ld regime. In Proceedings of the 39th International Conference on Machine Learning , volume 162 of Pr...

    2022

  15. [23]

    H. P. H. Luu, H. Yu, B. Williams, P. Mikkola, M. Hartmann, K. Puolam ¨ aki, and A. Klami. Non- geodesically-convex optimization in the Wasserstein space. In The 38th Annual Conference on Neural Information Processing Systems , 2024

    2024

  16. [24]

    S. Mei, A. Montanari, and P.-M. Nguyen. A mean field view of the lan dscape of two-layer neu- ral networks. Proceedings of the National Academy of Sciences of the Unite d States of America , 115:E7665 – E7671, 2018

    2018

  17. [25]

    Monmarch´ e, Z

    P. Monmarch´ e, Z. Ren, and S. Wang. Time-uniform log-Sobolev inequalities and applications to propagation of chaos. Electronic Journal of Probability , 29(none):1 – 38, 2024

    2024

  18. [26]

    Nitanda and T

    A. Nitanda and T. Suzuki. Stochastic particle gradient descent for infinite ensembles, 2017. arXiv:1712.05438

    2017 arXiv

  19. [27]

    Nitanda, D

    A. Nitanda, D. Wu, and T. Suzuki. Convex analysis of the mean fie ld Langevin dynamics. In Proceedings of The 25th International Conference on Artific ial Intelligence and Statistics , 2022

    2022

  20. [28]

    Otto and C

    F. Otto and C. Villani. Generalization of an inequality by Talagrand a nd links with the logarithmic Sobolev inequality. Journal of Functional Analysis , 173(2):361–400, 2000. 27

    2000

  21. [29]

    Parikh and S

    N. Parikh and S. Boyd. Proximal algorithms. Found. Trends Optim., 1(3):127–239, Jan. 2014

    2014

  22. [30]

    G. M. Rotskoff and E. Vanden-Eijnden. Trainability and accurac y of artificial neural networks: An interacting particle system approach. Communications on Pure and Applied Mathematics , 75, 2018

    2018

  23. [31]

    Salim, A

    A. Salim, A. Korba, and G. Luise. The Wasserstein proximal grad ient algorithm. In Advances in Neural Information Processing Systems , volume 33, pages 12356–12366. Curran Associates, Inc., 2020

    2020

  24. [32]

    Santambrogio

    F. Santambrogio. Optimal Transport for Applied Mathematicians: Calculus of Variations, PDEs, and Modeling. Progress in Nonlinear Differential Equations and Their Applications. Springer Inter- national Publishing, 2015

    2015

  25. [33]

    Santambrogio

    F. Santambrogio. Euclidean, Metric, and Wasserstein gradient flows: an overview. Bulletin of Mathematical Sciences, 7, 09 2016

    2016

  26. [34]

    Sirignano and K

    J. Sirignano and K. Spiliopoulos. Mean field analysis of neural netw orks: A law of large numbers. SIAM Journal on Applied Mathematics , 80(2):725–752, 2020

    2020

  27. [35]

    Teter, I

    A. Teter, I. Nodozi, and A. Halder. Proximal mean field learning in shallow neural networks. Transactions on Machine Learning Research , 2024

    2024

  28. [36]

    C. Villani. Optimal Transport: Old and New . Grundlehren der mathematischen Wissenschaften. Springer Berlin Heidelberg, 2008

    2008

  29. [37]

    S. Wang. Uniform log-Sobolev inequalities for mean field particles w ith flat-convex energy, 2024. arXiv:2408.03283

    2024 arXiv

  30. [38]

    Wibisono

    A. Wibisono. Sampling as optimization in the space of measures: Th e Langevin dynamics as a composite optimization problem. In Conference on Learning Theory , 2018

    2018

  31. [39]

    Xu and Q

    Y. Xu and Q. Li. Forward-Euler time-discretization for Wassers tein gradient flows can be wrong,

  32. [40]

    Zhang, C

    R. Zhang, C. Chen, C. Li, and L. Carin. Policy optimization as Wass erstein gradient flows. In J. Dy and A. Krause, editors, Proceedings of the 35th International Conference on Machin e Learning , volume 80 of Proceedings of Machine Learning Research, pages 5737–5746. PMLR, 10–...

    2018

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.