Stochastic Euler Schemes and Dissipative Evolutions in the Space of Probability Measures

Giacomo Enrico Sodini; Giulia Cavagnari; Giuseppe Savar\'e

REVIEW 1 major objections 3 minor 1 cited by

Stochastic Euler Schemes and Dissipative Evolutions in the Space of Probability Measures

T0 review · 1 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper proves that stochastic Euler trajectories converge, in strong path-space Wasserstein distance, to the deterministic dissipative evolution driven by the averaged field.

desk verdict A solid, genuinely new path-space convergence theorem for stochastic Euler schemes in Wasserstein space, held together by an explicit but load-bearing uniform stability assumption. read the letter →

arxiv 2505.20801 v1 pith:DKQ7R27O submitted 2025-05-27 math.FA math.PR

Giulia Cavagnari , Giuseppe Savaré , Giacomo Enrico Sodini This is my paper · ORCID

classification math.FAmath.PR MSC 34A0647B4449Q2234A1234A60

keywords stochasticEulerschememultivaluedprobabilityvectorfieldsWassersteinspacedissipativeevolutionsgradientdescentinteractingparticlesystemsEVIsolutionsmeasuredifferentialequations

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

This paper proves a convergence theorem for stochastic time discretizations of dissipative evolutions: when a random velocity field is averaged over its randomness, the laws of the trajectories produced by the explicit Euler method converge, as the step size tends to zero, to the law of the characteristics of the deterministic evolution driven by the averaged field. The result covers stochastic gradient descent, interacting particle systems, and nonlocal variants in a single framework based on multivalued probability vector fields. The convergence is strong: it holds in the L2-Wasserstein distance on the space of continuous paths, not just at discrete times. The paper also shows that the random trajectories themselves converge in L2 to the deterministic solution when the random initialization converges almost surely.

What carries the argument

The central object is the multivalued probability vector field (MPVF), a set of probability measures on the tangent bundle TX whose spatial marginal is the current measure; each section F[µ] is a distribution of velocities, and its barycenter b_F(x,µ) is the averaged deterministic field. The argument proceeds by lifting the Explicit Euler scheme to a probability measure η_τ on piecewise-linear paths, proving a stability estimate in path-space Wasserstein distance between two such schemes, and then identifying all subsequential limits via the unique probabilistic representation of the contraction semigroup generated by any maximal totally λ-dissipative extension of bar(F). The sticky-particles lemma supplies uniqueness of that representation when the support is finite and non-increasing. The additional approximate-solvability condition provides the uniform stability bound needed to pass from finite-support initial measures to general ones.

What would settle it

Compute the path-space Wasserstein distance between stochastic Euler interpolants and the deterministic characteristics for a simple one-dimensional λ-dissipative field with controlled initial error, and compare the observed rate with the $τ^{{1/4}}$ bound of Theorem 5.7; a rate systematically worse than $τ^{{1/4}}$ with fixed initial error would contradict the theorem.

Watch

Extended reading notes

Core claim

Let F be a multivalued probability vector field on a separable Hilbert space, and suppose its barycenter bar(F) is totally λ-dissipative. If the Explicit Euler scheme for F is approximately solvable at the initial measure and the initial measures converge in W2, then the interpolated stochastic Euler schemes, viewed as probability measures on continuous curves, converge in W2,∞ to η, the probabilistic representation of the semigroup generated by any maximal totally λ-dissipative extension of bar(F). Equivalently, if the initial random variables converge almost surely, the stochastic Euler curves converge in L2(Ω; C([0,T]; X)) to the unique solution of the deterministic differential inclusion driven by the barycentric field. The proof combines a stability estimate for Euler schemes with respect to initial data, a sticky-particles uniqueness lemma for probabilistic representations with finite non-increasing support, and the semigroup stability of the limit.

Load-bearing premise

The convergence proof relies on the Explicit Euler scheme being approximately solvable at the initial measure: a single uniform stability bound must hold for finitely supported approximations across all small step sizes, and if that bound must blow up as the approximations approach the initial measure, the argument collapses.

Editorial extensions

If this is right

For stochastic gradient descent with a λ-dissipative averaged field and stable initializations, the interpolated SGD trajectories converge in L2 to the gradient-flow ODE solution, giving a continuous-time justification of SGD in the vanishing step-size regime.
For interacting particle systems whose pairwise interaction field satisfies the two-sided dissipativity condition, the stochastic particle scheme converges to the deterministic McKean-Vlasov-type ODE.
The convergence is strong in path space, so functionals of the whole trajectory, not just time marginals, converge to their deterministic limits.
Theorem 5.7 yields an explicit path-space Wasserstein bound of order W2(initial)^{1/2} + τ^{1/4}, quantifying how fast the stochastic scheme approaches the deterministic evolution as the step size vanishes.

Reading between the lines

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

The strong path-space convergence suggests that finite-dimensional statistics of stochastic Euler trajectories, such as exit times or hitting distributions, converge to those of the deterministic flow under dissipativity; the paper does not state this directly.
The framework likely extends to time-dependent fields and projected or stochastic proximal variants, where the barycenter is the averaged operator; the approximate-solvability condition would need to be verified in each such setting.
The τ^{1/4} rate in path-space Wasserstein distance is probably not sharp; a natural numerical test would measure the actual convergence order in smooth dissipative examples to see whether τ^{1/2} holds.
If approximate solvability fails, weak convergence of marginals may still hold while strong path-space convergence fails, suggesting the uniform stability bound is the threshold separating genuine stochastic approximation from a deterministic limit.

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 main result here is real: Theorem 5.9 proves strong W2,∞ convergence of the laws of interpolated stochastic Euler trajectories to the deterministic dissipative evolution driven by the barycentric field. That is a genuine step beyond your earlier marginal convergence results in [11] and [12], and the proof technique—sticky-particles uniqueness plus finite-support approximation—is a smart way to close the path-space problem. The finite-support case (Theorem 5.6) is clean, and the stability estimate (Theorem 5.7) is intricate but coherent. The paper is honest about what it needs: Definition 5.8 states the uniform stability bound L explicitly, and the examples verify it through growth conditions. I give credit for that. The stress-test concern about L blowing up is worth taking seriously, but it is not a load-bearing flaw in the proof. It correctly identifies that total dissipativity of the barycenter alone does not give you a uniform L, and that the triangle inequality in Theorem 5.9 collapses if the minimal L for the approximating finite measures diverges. That means the theorem's scope is exactly the class of MPVFs with a uniform second-moment bound. The paper does not overclaim—it lists approximate solvability as a hypothesis—but it would help if a remark stated plainly that this uniformity is essential and that the counterexample g(x,u)=u|x|^{1+δ} shows the theorem stops at the boundary. That is a scope clarification, not a correction. Minor issues: Claim 2 in Proposition 6.4 is stated without proof. It is straightforward, but a two-line proof would make the paper self-contained. The Gronwall lemma from [11] is cited without restatement; again, that is a self-containedness issue rather than a correctness one. The paper leans heavily on the authors' previous work for the semigroup machinery, which is legitimate but means a referee should check those citations carefully. This is a paper for the optimal transport and Wasserstein gradient flow community, and for anyone working on continuous-time limits of stochastic approximation. It deserves a serious referee. I would send it to review, with requests for the omitted proof, a remark on the role of L, and a brief restatement of the Gronwall lemma. The central argument holds up; these are polish, not overhaul.

Referee Report

1 major / 3 minor

Summary. The paper develops a measure-theoretic framework for the convergence of stochastic explicit Euler schemes to deterministic dissipative evolutions in Wasserstein space. It constructs probabilistic representations of the affine interpolants of the Euler scheme and proves, under a total dissipativity condition on the barycentric field together with an approximate solvability condition involving a uniform stability bound, that the laws of the interpolated trajectories converge strongly in the path-space Wasserstein distance W2,∞ to the semigroup generated by any maximal totally dissipative extension of the barycenter. The main result is Theorem 5.9, with applications to stochastic gradient-type flows and interacting particle systems.

Significance. If the result holds, it provides a general strong-convergence theorem for stochastic approximation schemes at the level of probability measures on path space, extending earlier marginal-convergence results. The paper's main contribution is the identification of the limit as the deterministic semigroup and the strong W2,∞ convergence, not merely convergence of the time-marginals. The sticky particles representation (Theorem 5.5 and Appendix B) is a useful independent result. The proofs are detailed and largely self-contained up to the authors' previous works [11, 12, 13], which supply the well-posedness and semigroup machinery.

major comments (1)

[§5, Definition 5.8 and Theorem 5.9] The approximate solvability hypothesis is the true load-bearing condition for the main theorem. In the proof of Theorem 5.9, the triangle inequality combines the constants from Theorem 5.7 and Proposition 4.3, both of which depend on the single uniform stability bound L. If the minimal L for the approximating finite-support measures µ̄_j diverges as j→∞, then the constants blow up and the double-limit argument (first h→∞, then j→∞) does not close. Total λ-dissipativity of bar(F) alone does not control the second moments of F[µ̄_j]; the uniform bound in Definition 5.8 is genuinely additional. The paper should state this explicitly and discuss the scope of the theorem, for instance by giving a concrete example such as g(x,u)=u|x|^{1+δ} with zero-mean u, where the barycenter is trivially dissipative but the Euler second moments can diverge as τ→0, so the theorem does not apply.

minor comments (3)

[§6.1, Proposition 6.4, Claim 2] Claim 2 in Proposition 6.4 is stated without proof: 'This proof, being straightforward, is omitted.' Since this claim is used to identify the abstract path-space measure η_τ with the law of the Stochastic Dissipative Flow, a short proof or a precise reference should be provided.
[§5, Theorem 5.7] The proof of Theorem 5.7 invokes the Gronwall estimate from [11, Lemma B.1 and equation (6.8)] without restating it. To make the paper more self-contained, the exact form of the lemma used should be stated or the relevant equations reproduced.
[§2, Definition 2.5] The notation bar(F) is used both for the map from MPVFs to MPVFs and for the barycenter projection of a single Φ. The distinction is clear from context, but a parenthetical clarification at first use would help the reader.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction: the path-space convergence theorem is proved from prior well-posedness and stability results whose assumptions do not include the target result.

full rationale

Theorem 5.9 establishes a new W2,∞ convergence statement for laws of stochastic Euler interpolants to the semigroup of a maximal totally dissipative extension of bar(F). The proof combines Theorem 5.6 (finite-support case), Theorem 5.7 (stability in initial data), and Proposition 4.3 (semigroup stability). Each ingredient is either proved in the paper (sticky-particles uniqueness in Appendix B, stability estimates in Section 5) or is a prior theorem from [11], [12], [13] with stated assumptions that do not include the path-space convergence result. The approximate-solvability condition (Definition 5.8) is an explicit extra hypothesis requiring a uniform stability bound L and solvability from finite approximations; it is not a restatement of the conclusion. The skeptical concern that L might blow up as the finite approximations approach the initial measure identifies a genuine limitation of scope, but the paper states this as an assumption in Definition 5.8 and verifies it in examples via growth conditions such as (6.3), (6.12), (F1), and (F2). Heavy self-citation is present, but the cited results are independent, parameter-free theorems with proofs, so they do not make the derivation circular.

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

No free parameters are fitted to data. The constants λ and L in the theorems are structural hypotheses on the dynamics, not tuned to make the result pass. The axioms are the standard background of optimal transport, the cited results from the authors' previous papers, and the technical assumptions of the explicit Euler scheme. No new physical or mathematical entities (in the sense of new particles, forces, or dimensions) are introduced.

assumptions (5)

standard math Standard properties of Wasserstein spaces, the superposition principle, and tightness criteria for probability measures on metric spaces (as in [2]).
Used throughout Sections 2, 3, and Appendix A; treated as known background.
domain assumption The semigroup S_t and the probabilistic representation η for a maximal totally λ-dissipative MPVF exist and are unique, with the stated Lipschitz and sticky particle properties (Theorem 4.1, Theorem 4.2 from [12]).
The paper relies on this for the limit object; it is cited, not reproven.
domain assumption The explicit Euler scheme for a metrically λ-dissipative MPVF yields, up to subsequences, the unique λ-EVI solution for the marginal measures (Theorem 3.5 and Proposition 3.7 from [11]).
Used in Proposition 3.7 and Theorem 5.6 to identify the marginals of any limit point.
domain assumption The Gronwall comparison lemma [11, Lemma B.1] applies to the differential inequality for σ2 in the proof of Theorem 5.7 and yields the stated exponential estimate.
This lemma is cited without statement; the stability rate in Theorem 5.7 depends on it.
domain assumption The probability space (Ω,B,P) is standard Borel and non-atomic, and the relevant random variables are in L2, as required for the Lagrangian representation.
Stated at the beginning of Section 4 and used in Theorem 5.10.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Stochastic Euler Schemes and Dissipative Evolutions in the Space of Probability Measures." pith.science (2026). https://pith.science/paper/DKQ7R27O

@misc{pith2026250520801,
  author       = {Pith},
  title        = {Pith review of: Stochastic Euler Schemes and Dissipative Evolutions in the Space of Probability Measures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DKQ7R27O}},
  note         = {Machine review of arXiv:2505.20801}
}

read the original abstract

We study the convergence of stochastic time-discretization schemes for evolution equations driven by random velocity fields, including examples like stochastic gradient descent and interacting particle systems. Using a unified framework based on Multivalued Probability Vector Fields, we analyze these dynamics at the level of probability measures in the Wasserstein space. Under suitable dissipativity and boundedness conditions, we prove that the laws of the interpolated trajectories converge to those of a limiting evolution governed by a maximal dissipative extension of the associated barycentric field. This provides a general measure-theoretic study for the convergence of stochastic schemes in continuous time.

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

Analysis on spaces of measures
math.AP 2026-07 conditional novelty 4.0 of 10

A synthesis monograph unifying vertical and horizontal differential calculus on spaces of measures, applied to well-posedness of monotone mean field game master equations and to viscosity solutions of mean field Hamil...

Reference graph

Works this paper leans on

22 extracted references · 17 canonical work pages · cited by 1 Pith paper

[11]

Cavagnari, G

G. Cavagnari, G. Savar´ e, and G. E. Sodini. Dissipative probability vector ﬁelds and generation of evolution semigroups in Wasserstein spaces. Probab. Theory Related Fields , 185(3-4):1087—-1182, 2023

work page 2023
[12]

Cavagnari, G

G. Cavagnari, G. Savar´ e, and G. E. Sodini. A Lagrangian approach to totally dissipative evolutions in W asser- stein spaces. arXiv:2305.05211, 2023

arXiv 2023
[1]

Ambrosio, E

L. Ambrosio, E. Bru´ e, and D. Semola. Lectures on optimal transport , volume 130 of Unitext. Springer, Cham,

work page
[2]

Ambrosio, N

L. Ambrosio, N. Gigli, and G. Savar´ e. Gradient Flows In Metric Spaces and in the Space of Probabili ty Measures. Lectures in Mathematics. ETH Z¨ urich. Birkh¨ auser Basel,2008

2008
[3]

H. H. Bauschke and P. L. Combettes. Convex analysis and monotone operator theory in Hilbert spa ces. CMS Books in Mathematics/Ouvrages de Math´ ematiques de la SMC. Springer, Cham, second edition, 2017. With a foreword by H´ edy Attouch

work page 2017
[4]

Bena ¨ ım

M. Bena ¨ ım. Dynamics of stochastic approximation algorithms. In J. Az´ ema, M.´Emery, M. Ledoux, and M. Yor, editors, S´ eminaire de Probabilit´ es XXXIII, pages 1–68, Berlin, Heidelberg, 1999. Springer Berlin Hei delberg

work page 1999
[5]

Billingsley

P. Billingsley. Probability and measure . Wiley Series in Probability and Statistics. John Wiley & So ns, Inc., Hoboken, NJ, anniversary edition, 2012. With a foreword by S teve Lalley and a brief biography of Billingsley by Steve Koppes

work page 2012
[6]

Bogachev

V. Bogachev. Measure Theory, volume II. Springer Berlin Heidelberg, 2006

work page 2006

Show all 22 references

[7]

L. Bottou. Large-scale machine learning with stochasti c gradient descent. In Y. Lechevallier and G. Saporta, editors, Proceedings of COMPSTAT’2010, pages 177–186. Physica-Verlag HD, 2010. Invited talk, COM PSTAT 2010, Paris, France

2010
[8]

Br´ ezis

H. Br´ ezis. Op´ erateurs maximaux monotones et semi-groupes de contrac tions dans les espaces de Hilbert . North-Holland Publishing Co., Amsterdam-London; America n Elsevier Publishing Co., Inc., New York, 1973. North-Holland Mathematics Studies, No. 5. Notas de Matem´ a tica (50)

1973
[9]

H. Brezis. Functional Analysis, Sobolev Spaces and Partial Diﬀerenti al Equations . Universitext. Springer New York, 2010

2010
[10]

Camilli, G

F. Camilli, G. Cavagnari, R. De Maio, and B. Piccoli. Sup erposition principle and schemes for measure diﬀer- ential equations. Kinet. Relat. Models , 14(1):89–113, 2021

2021
[13]

Cavagnari, G

G. Cavagnari, G. Savar´ e, and G. E. Sodini. Extension of monotone operators and Lipschitz maps invariant for a group of isometries. Canad. J. Math. , 77(1):149–186, 2025

2025
[14]

Kiefer and J

J. Kiefer and J. W olfowitz. Stochastic estimation of th e maximum of a regression function. The Annals of Mathematical Statistics , 23(3):462–466, 1952. 34 GIULIA CA V AGNARI, GIUSEPPE SA V AR ´E, AND GIACOMO ENRICO SODINI

1952
[15]

H. J. Kushner and G. G. Yin. Stochastic approximation and recursive algorithms and app lications, volume 35 of Applications of Mathematics (New York) . Springer-Verlag, New York, second edition, 2003. Stochas tic Modelling and Applied Probability

2003
[16]

Moulines and F

˜A. Moulines and F. R. Bach. Non-asymptotic analysis of stoch astic approximation algorithms for machine learning. In Advances in Neural Information Processing Systems , volume 24, pages 451–459. Curran Associates, Inc., 2011

2011
[17]

B. Piccoli. Measure diﬀerential inclusions. 2018 IEEE Conference on Decision and Control (CDC) , pages 1323–1328, 2018

2018
[18]

B. Piccoli. Measure diﬀerential equations. Arch. Ration. Mech. Anal. , 233(3):1289–1317, 2019

2019
[19]

B. Piccoli. Control of multi-agent systems: results, o pen problems, and applications. Open Math. , 21(1):Paper No. 20220585, 26, 2023

2023
[20]

Robbins and S

H. Robbins and S. Monro. A stochastic approximation met hod. Ann. Math. Statistics , 22:400–407, 1951

1951
[21]

D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learni ng representations by back-propagating errors. Nature, 323(6088):533–536, 1986. Giulia Cavagnari: Politecnico di Milano, Dipartimento di M atematica, Piazza Leonardo Da Vinci 32, 20133 Milano (Italy) Email address : giu...

1986
[2021]

La Matematica per il 3+2

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [11]

Cavagnari, G

G. Cavagnari, G. Savar´ e, and G. E. Sodini. Dissipative probability vector ﬁelds and generation of evolution semigroups in Wasserstein spaces. Probab. Theory Related Fields , 185(3-4):1087—-1182, 2023

work page 2023

[2] [12]

Cavagnari, G

G. Cavagnari, G. Savar´ e, and G. E. Sodini. A Lagrangian approach to totally dissipative evolutions in W asser- stein spaces. arXiv:2305.05211, 2023

arXiv 2023

[3] [1]

Ambrosio, E

L. Ambrosio, E. Bru´ e, and D. Semola. Lectures on optimal transport , volume 130 of Unitext. Springer, Cham,

work page

[4] [2]

Ambrosio, N

L. Ambrosio, N. Gigli, and G. Savar´ e. Gradient Flows In Metric Spaces and in the Space of Probabili ty Measures. Lectures in Mathematics. ETH Z¨ urich. Birkh¨ auser Basel,2008

2008

[5] [3]

H. H. Bauschke and P. L. Combettes. Convex analysis and monotone operator theory in Hilbert spa ces. CMS Books in Mathematics/Ouvrages de Math´ ematiques de la SMC. Springer, Cham, second edition, 2017. With a foreword by H´ edy Attouch

work page 2017

[6] [4]

Bena ¨ ım

M. Bena ¨ ım. Dynamics of stochastic approximation algorithms. In J. Az´ ema, M.´Emery, M. Ledoux, and M. Yor, editors, S´ eminaire de Probabilit´ es XXXIII, pages 1–68, Berlin, Heidelberg, 1999. Springer Berlin Hei delberg

work page 1999

[7] [5]

Billingsley

P. Billingsley. Probability and measure . Wiley Series in Probability and Statistics. John Wiley & So ns, Inc., Hoboken, NJ, anniversary edition, 2012. With a foreword by S teve Lalley and a brief biography of Billingsley by Steve Koppes

work page 2012

[8] [6]

Bogachev

V. Bogachev. Measure Theory, volume II. Springer Berlin Heidelberg, 2006

work page 2006