{"id":"94d0ebfd-9cff-446a-ba5f-0d204a573a89","arxiv_id":"2505.20801","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The law of the interpolated stochastic Euler trajectory converges in L2-Wasserstein distance on path space to the deterministic dissipative evolution driven by the barycentric field.","lead":"Stochastic Euler schemes for dissipative random vector fields are shown to converge, in the Wasserstein space of paths, to the deterministic mean-field evolution generated by the averaged field. The result covers stochastic gradient descent and interacting particle systems, giving a rigorous continuous-time limit for their laws.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.9's convergence holds only under Definition 5.8's uniform stability bound; total dissipativity of the barycenter alone does not imply it, and the proof's double-limit argument collapses if L must blow up as the approximating finite measures approach the initial law.","rationale":"The reader's weakest_assumption (approximate solvability, Definition 5.8) is also the most load-bearing point in my reading. The central claim is valid under that hypothesis: Theorem 5.7 provides the uniform-in-τ stability estimate, Proposition 4.3 gives convergence as the finite-support approximations approach µ̄, and Theorem 5.6 supplies the exact limit for finite-support initial data. The triangle inequality then closes by a standard double-limit argument. The only place where the chain could break is the uniformity of L across the approximating sequence, because the constant in Theorem 5.7 depends on L and would blow up if the minimal stable L grew without bound. The paper's examples verify this uniformity through explicit growth bounds, so the theorem is internally consistent. My concrete superlinear example shows that total dissipativity of the barycenter alone does not guarantee approximate solvability, so the hypothesis is genuinely load-bearing and not redundant. This does not change the reader's CONDITIONAL verdict: the proof is sound under its stated assumptions, but the theorem's scope is narrower than the abstract's phrase 'suitable dissipativity and boundedness conditions' suggests. I agree with the reader that the only additional items needed for full acceptance are minor: providing the omitted proof of Claim 2 in Proposition 6.4 and restating the Gronwall estimate cited from [11] in Theorem 5.7.","tokens_in":38624,"tokens_out":23680,"duration_ms":246318,"concrete_test":"For the SDF example in Section 6.1, take g(x,u)=u x^2 with u∈{±1} equiprobable, so that bar(F)=0 is totally λ-dissipative with λ=0. Under the Explicit Euler scheme (EE), show that the second moment m_{n+1}=∫|x|² dM^{n+1}_τ satisfies m_{n+1} ≥ m_n + τ²∫x^4 dM^n_τ ≥ m_n + τ² m_n², by Jensen's inequality. Starting from a Gaussian initial law, this recurrence diverges in O(1/τ²) steps, so sup_{n,τ} |F[M^n_τ]|₂ = +∞. This verifies that Definition 5.8 fails for a totally dissipative barycenter, confirming that the uniform L bound is an additional, essential hypothesis and not a consequence of dissipativity alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 5.9 uses the triangle inequality W2,∞(η_h,η) ≤ C(√W2(µ̄_h,µ̄_j)+τ(h)^{1/4}) + W2,∞(η^j_h,η^j) + C̃√W2(µ̄,µ̄_j), with C from Theorem 5.7 and C̃ from Proposition 4.3. Both constants are controlled by the single stability bound L in Definition 5.8. If the minimal L for which the Explicit Euler scheme is solvable from the approximating finite measures µ̄_j grows as j→∞, then C → ∞ and the double limit (first h→∞, then j→∞) does not close. Total λ-dissipativity of bar(F) gives unconditional dissipativity of F, but it does not control the second moments of F[µ̄_j]; the paper invokes growth conditions (6.3), (6.12), (F1)-(F2), and Lemma 5.13 of [11] to obtain uniform L. A field such as g(x,u)=u|x|^{1+δ} (zero mean, so bar(F)=0 is trivially dissipative) has superlinear velocities; the Euler second moments can blow up as τ→0, so no finite L exists and the theorem does not apply. This is not an internal inconsistency, but it identifies approximate solvability as the true load-bearing hypothesis: the paper's scope is exactly the class of MPVFs satisfying a uniform moment bound, and whether convergence persists without it is left open.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":38921,"tokens_out":8511,"duration_ms":86287,"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":[{"comment":"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.","section":"§5, Definition 5.8 and Theorem 5.9"}],"minor_comments":[{"comment":"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.","section":"§6.1, Proposition 6.4, Claim 2"},{"comment":"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.","section":"§5, Theorem 5.7"},{"comment":"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.","section":"§2, Definition 2.5"}],"recommendation":"minor_revision","confidential_remarks":"This is a natural continuation of the authors' prior works [11, 12, 13], and the new result is the strong path-space convergence theorem. The refereeing process would benefit from a clearer separation of new versus recalled results, and from a more prominent discussion of the uniform stability bound as the key hypothesis governing the theorem's applicability. The paper is technically sound as far as I can verify; the main revision requests concern exposition and the omission of a proof in an example."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.","headline":"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.","tokens_in":674,"tokens_out":746,"would_cite":true,"duration_ms":32408,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["34A06","47B44","49Q22","34A12","34A60"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that stochastic Euler trajectories converge, in strong path-space Wasserstein distance, to the deterministic dissipative evolution driven by the averaged field.","keywords":["stochastic Euler scheme","multivalued probability vector fields","Wasserstein space","dissipative evolutions","stochastic gradient descent","interacting particle systems","EVI solutions","measure differential equations"],"falsifier":"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.","tokens_in":38399,"feed_emoji":"🎲","tokens_out":4354,"duration_ms":42551,"temperature":0.7,"pith_summary":"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.","feed_headline":"Stochastic Euler curves converge to a deterministic dissipative limit","feed_subtitle":"Proof that random Euler trajectories converge, in path-space Wasserstein distance, to the averaged field's deterministic evolution.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the Explicit Euler scheme for dissipative probability vector fields and the EVI well-posedness used as the starting point of the convergence argument.","marker":"[11]"},{"why":"Supplies the Lagrangian representation, contraction semigroup, and minimal selection for maximal totally dissipative MPVFs, which define the limit η.","marker":"[12]"},{"why":"Supplies the Lipschitz-stability estimate for semigroups that gives the rate in Proposition 4.3, used to control the limit as initial measures vary.","marker":"[13]"},{"why":"Provides the Wasserstein-space background, coupling lemmas, and path-space results used throughout the proof.","marker":"[2]"},{"why":"Provides the Hilbert-space maximal monotone operator theory that underpins dissipativity and semigroup generation for the barycentric field.","marker":"[8]"}],"fun_headline_variants":["Stochastic Euler curves converge to a single dissipative limit","Random step schemes yield deterministic averaged dynamics","Measure-space proof: stochastic Euler → deterministic flow","From random velocity fields to a unique limiting evolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Stochastic Euler curves converge to a single dissipative limit","Random step schemes yield deterministic averaged dynamics","Measure-space proof: stochastic Euler → deterministic flow","From random velocity fields to a unique limiting evolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1252,"prompt_tokens":820,"completion_tokens":432,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":373}},"tokens_in":436,"tokens_out":432,"duration_ms":4830,"temperature":1.0,"reasoning_tokens":373,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:48:12.158043+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Cavagnari, G","cited_arxiv_id":null,"evidence_quote":"Supplies the Explicit Euler scheme for dissipative probability vector fields and the EVI well-posedness used as the starting point of the convergence argument."},{"cited_title":"Br´ ezis","cited_arxiv_id":null,"evidence_quote":"Provides the Hilbert-space maximal monotone operator theory that underpins dissipativity and semigroup generation for the barycentric field."}],"review_version":1}