Pith. sign in

REVIEW 3 major objections 6 minor 3 cited by

Totally convex functions, $L^2$-Optimal transport for laws of random measures, and solution to the Monge problem

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · 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 optimal transport between laws of random measures has a unique Monge solution for super-regular laws.

desk verdict Real new result and clean d=1 theory, but Theorem 6.15 rests on an unpublished companion paper ([LS25]) for a Borel measurability fact; referee it, but require that gap be closed. read the letter →

arxiv 2509.01768 v1 pith:VWEKOWEL submitted 2025-09-01 math.FA math.OCmath.PR

Alessandro Pinzi , Giuseppe Savaré This is my paper · ORCID
classification math.FAmath.OCmath.PR MSC 49Q2246N1060B05
keywords optimaltransportrandommeasuresWassersteinspacetotallyconvexfunctionalsMongeproblemGaussianLagrangianliftingsubdifferentials
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 that optimal transport between laws of random measures—probability measures on the Wasserstein space P2(H)—has the same convex-analysis structure as classical quadratic transport on a Hilbert space. The route is to define 'totally convex' functionals, whose conjugate is taken with respect to the maximal-correlation pairing, and to show their total subdifferentials are exactly the optimal random couplings. For a class of 'super-regular' first laws, the paper proves that the optimal random coupling is unique and fully deterministic, and that the transport is governed by a single nonlocal field f(x,μ) that moves each point x of each sampled measure μ. This class is non-empty and natural: in dimension one it contains every law obtained by pushing a nondegenerate Gaussian measure through the law map, so the result applies to occupation measures of Gaussian processes such as Brownian motion.

What carries the argument

Total subdifferentials and their minimal sections. A totally convex functional φ on P2(H) has a subdifferential ∂tφ consisting of optimal couplings γ∈P2(H×H); its minimal section ∂°tφ is a single deterministic coupling (i×∇Wφ(·,μ))♯μ. The argument lifts φ through the law map ι:H→P2(H), so total convexity becomes ordinary convexity of φ̂=φ∘ι on the Hilbert space H; the classical subdifferential of φ̂ is the Lagrangian representation of ∂tφ. This reduces single-valuedness of total subdifferentials to Gateaux differentiability of convex Lipschitz functions, which then is turned into a measure-theoretic negligibility statement by super-regularity.

What would settle it

Take H=R, Q=[0,1] with Lebesgue measure, and M=ι♯g for a nondegenerate Gaussian g on L2(Q). The theorem predicts that for every N∈PP2(R), Γo(M,N) has exactly one element and it is a graph. In a discretized or analytic example, produce two distinct couplings Π1,Π2∈Γo(M,N) with equal W2² value; that directly contradicts Theorem 6.15. The easiest target is N=M: the identity map must be the only optimal coupling, so any other optimal coupling of M to itself with the same W2 value refutes the theorem.

Watch

Extended reading notes

Core claim

At the level of PP2(H), the paper recovers the Euclidean trichotomy: optimal couplings are supported on total subdifferentials of totally convex Kantorovich potentials; total cyclical monotonicity characterizes optimal random couplings; and, whenever the first marginal is super-regular, the total subdifferential is single-valued almost everywhere. The conclusion is Theorem 6.15: for M super-regular and N arbitrary, both RΓo(M,N) and Γo(M,N) are singletons, the unique random coupling is concentrated on deterministic couplings, and there is a unique Borel map f∈L2(M-bar; H) solving the strict Monge problem. In the one-dimensional case, Theorem 6.19 shows every law G=ι♯g generated by a nondegen

Load-bearing premise

The paper's key conclusions rest on imported results—law-invariant Lipschitz maps admit pointwise nonlocal representations, deterministic couplings form a Gδ set, and convex Lipschitz functions are differentiable off a σ-d.c. hypersurface—so if any of those fails, the uniqueness and determinism claims lose their support.

Editorial extensions

If this is right

  • Theorem 6.15: for every super-regular law M and every target N, optimal random couplings and optimal couplings are both singletons; the unique random coupling is fully deterministic and the unique map solves the strict Monge problem.
  • Theorem 6.19: in one dimension, every law obtained by pushing a nondegenerate Gaussian measure through the law map is super-G-regular; this includes occupation measures of Gaussian processes such as Brownian motion.
  • The strict and usual Monge formulations are equivalent for an initial law that works for every target, and every Lipschitz totally convex potential is W-differentiable almost everywhere under super-regular M.
  • In finite dimension, super-G-regular measures are dense in PP2(H), so unique strict Monge solutions exist for a dense set of initial laws.
  • Minimal geodesics from a super-regular endpoint are uniquely determined: each intermediate point and each sub-geodesic is pinned down by a continuous nonlocal transport field.

Reading between the lines

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

  • Editorial inference: the same Lagrangian-lifting route should extend the strict Monge theory to Lp-Wasserstein costs built from smooth norms, with super-regularity defined by the corresponding null sets; the paper mentions the possibility but does not carry it out.
  • Editorial inference: reading f(x,μ) as a conditional map—where to send x once the whole measure μ is known—suggests a computational proxy: approximate a law-invariant convex potential from samples and compare its gradient field to the unique Monge map predicted by Theorem 6.15.
  • Editorial inference: the Gaussian-generated measures appear to be the right 'uniform' reference class on P2(H); if so, they offer a starting point for Sobolev and Dirichlet-form calculus on the space of laws of random measures, though such calculus is not developed in this paper.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper develops a convex-analytic framework for the L2 optimal transport problem on P P2(H), the Wasserstein space of laws of random measures on a Hilbert space H. It introduces totally convex functionals on P2(H), a Kantorovich–Legendre–Fenchel transform based on the maximal correlation pairing [µ,ν], total subdifferentials, and totally cyclically monotone multivalued probability vector fields. These tools are used to characterize optimal random couplings and minimal geodesics, and to define super-regular measures on P2(H) through the Hilbertian Lagrangian lifting. The central result (Theorem 6.15) states that if M is super-regular and N is arbitrary, then both the set of optimal random couplings and the set of optimal couplings are singletons, the unique random coupling is fully deterministic, and the strict Monge problem has a unique Borel solution f. The paper also shows that in dimension one every nondegenerate Gaussian pushforward is super-G-regular (Theorem 6.19), and provides further super-regular examples in higher dimensions via Berman-type conditions.

Significance. If the results are correct, the paper represents a substantial extension of the Brenier–Knott–Smith–Rachev–Rüschendorf convex-analytic theory to a genuinely non-Hilbertian, positively curved Wasserstein space. The strict Monge theorem for laws of random measures, including full-support examples generated by Gaussian measures, is new and likely useful for applications involving random measures. The paper contains many self-contained and carefully written arguments, including the biconjugation theorem, the Rockafellar-type characterization of totally cyclically monotone MPVFs, and the deterministic representation of optimal random couplings under regularity. The main limitation is the heavy reliance on external machinery, especially an in-preparation citation, and a demonstrable error in a stated measurability proposition; these issues are repairable but currently prevent full verification.

major comments (3)
  1. [§2.4, Eq. (2.4); §5.3–6.2] The claim that Pdet(X×Y) is a Gδ subset (Eq. (2.4)) is attributed to [LS25], which is marked 'In preparation'. This measurability fact is used in the deterministic-coupling/selection arguments leading to Proposition 6.5 and Theorem 6.15. In addition, Theorem 5.7(2) invokes [CSS25, Thm 4.8] as a black box to convert m.p.i.-invariant Lipschitz maps into nonlocal fields f_{t,i}; the non-branching structure and hence Theorem 6.15 depend on this. Please state precisely which external results are used, verify their hypotheses, and include a proof of the [LS25] fact or replace it with a published, citable reference.
  2. [Proposition 2.6] This proposition is not correct as stated. Claim 1 contains a sign error: the displayed characterization should be Pr_2(R)=⋂_k {µ×µ(D)<1/k}; as written, the intersection over the sets {µ×µ(D)>1/k} is empty. Claim 2 states J(C)=Pr_2(R^d), but C parametrizes densities with respect to Lebesgue measure, so J(C)=Pgr_2(R^d), not Pr_2(R^d). This is a false statement as printed. The Borel/Gδ measurability of Pr_2(H) needed in Proposition 6.5 may be recoverable from a corrected Claim 1, but the proposition must be rewritten and the two classes Pr and Pgr kept distinct.
  3. [§6.2, proof of Theorem 6.15] The truncation step for unbounded target measures asserts 'Since optimality is preserved by restriction' (after Eq. (6.26)) without proof. This is not immediate for arbitrary restrictions of optimal plans; it should be justified, for example by restricting a pair of Kantorovich potentials or by using cyclical monotonicity of supp Π. This step is load-bearing for extending the result from bounded targets to arbitrary N, so a precise justification should be supplied.
minor comments (6)
  1. [Theorem 4.11] The inequality (4.30) uses w2(µ,ν) without squares, but the proof and the standard cyclical monotonicity for the quadratic cost require w2^2(µ,ν). Please add the squares in the displayed definition.
  2. [Introduction, after (R1)–(R2)] The text promises 'We will show that those conditions are nearly optimal', but no theorem in Sections 5–6 proves this statement. Either supply a precise statement and proof, or rephrase this as an open question/remark.
  3. [Theorem 5.5, proof] The sentence 'Clearly 2⇒3' appears to be a typo for '2⇒4': the justification refers to the second part of Theorem 4.11, which concerns total cyclical monotonicity of the lifted MPVF. Please correct the implication labels.
  4. [Lemma 3.7, proof] The supremum is taken 'w.r.t. g∈G(H)' but should be over the measure-preserving isomorphisms g∈G(Q).
  5. [Definition 6.6] The notation 'P Pr^rr_2(H)' and 'P P^grr_2(H)' is difficult to parse. Consider clearer notation for the classes of regular/super-regular laws of random measures.
  6. [Theorem 6.15] The property 'essentially totally cyclically monotone' is not defined in the paper. Please define it or remove the qualifier 'essentially'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: the central uniqueness theorem reduces to external published lifting/subdifferential theory and to an independently verified regularity class, not to its own inputs.

full rationale

The paper's central derivation is a functional-analytic chain: total convexity is defined via lifted convexity on L2(Q,M;H), and the lifting/subdifferential correspondence is imported from [CSS23a]/[CSS25], which are published external works with proofs not including the present Monge-uniqueness claim. Theorem 6.15's assumption M in PPrr_2(H) is not defined in terms of the existence of the optimal Monge map; instead Theorem 6.13 proves, using Zajicek's theorem, that the singular set of a Lipschitz totally convex potential is a random exceptional set, so the super-regularity hypothesis kills exactly the singular set. The Gaussian examples in Section 6.3 are verified independently (occupation measures, Berman condition, nondegeneracy of the Karhunen-Loeve increments), not fitted. The one fragile input is the measurability assertion at (2.4): "It is possible to prove [LS25] that Pdet(X×Y) is a Gδ", attributed to a paper marked 'In preparation' and authored by a co-author. This fact is used in auxiliary lifting/selection lemmas, but it is a missing-support/correctness issue rather than a reduction of the central claim to its own inputs. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported to forbid alternatives, and the super-regular condition is not defined in terms of the desired Monge solution. Hence there is no circular step under the rubric; the central content is independent of its citations.

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

No free parameters are fitted: eigenvalues, Hurst parameters, and coefficient sequences in the examples are construction data, not fit parameters. The paper imports standard theorems of convex analysis and Gaussian analysis as well as results from the authors' own prior program. No physical or probabilistic entity is invented.

assumptions (5)
  • standard math Zajicek's theorem: the set of Gateaux non-differentiability points of a convex Lipschitz function on a separable Hilbert space is contained in a sigma-d.c. hypersurface (cited [Zaj79], [BL00, Thm 4.20]).
    Used in Theorem 2.5 (Brenier theorem) and in Theorem 6.13 to show the singular set Singr(phi) is exceptional. This is the main external analytic fact converting differentiability into negligibility.
  • standard math Csornyei's theorem: Gaussian-null, Aronszajn-null, and cube-null sets coincide in separable Banach spaces (cited [Cso99]).
    Justifies that sigma-d.c. hypersurfaces are Gaussian-null, which is used for the super-regularity and density arguments in Sections 2.2 and 6.3.
  • ad hoc to paper Representation and extension theorems for law-invariant Lipschitz or monotone sets from [CSS25, Thm 4.8], and the G_delta measurability of deterministic couplings from [LS25] (In preparation).
    These companion results are the foundation for the deterministic-coupling structure in Lemma 5.10, Theorem 5.7, and Theorem 6.15. They are not proved in this preprint.
  • domain assumption (Q, FQ, M) is a standard Borel space with non-atomic probability measure, and the law map iota from H = L2(Q;H) to P2(H) is surjective and 1-Lipschitz.
    This is the standing setting for the Lagrangian lifting of Section 2.3 and is used throughout.
  • domain assumption For the examples, nondegenerate Gaussian measures satisfy the stated analytic conditions: occupation measures are absolutely continuous via Berman / Geman-Horowitz conditions (6.48)-(6.51), or C1 maps have nondegenerate Jacobians a.e.
    These external probabilistic results verify super-regularity of LGGRM in Section 6.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Totally convex functions, $L^2$-Optimal transport for laws of random measures, and solution to the Monge problem." pith.science (2026). https://pith.science/paper/VWEKOWEL

@misc{pith2026250901768,
  author       = {Pith},
  title        = {Pith review of: Totally convex functions, $L^2$-Optimal transport for laws of random measures, and solution to the Monge problem},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VWEKOWEL}},
  note         = {Machine review of arXiv:2509.01768}
}
abstract

We study the Optimal Transport problem for laws of random measures in the Kantorovich-Wasserstein space $\mathcal{P}_2(\mathcal{P}_2(\mathrm{H}))$, associated with a Hilbert space $\mathrm{H}$ (with finite or infinite dimension) and for the corresponding quadratic cost induced by the squared Wasserstein metric in $\\mathcal{P}_2(\mathrm{H}).$ Despite the lack of smoothness of the cost, the fact that the space $\mathcal{P}_2(\mathrm{H})$ is not Hilbertian, and the curvature distortion induced by the underlying Wasserstein metric, we will show how to recover at the level of random measures in $\mathcal{P}_2(\mathcal{P}_2(\mathrm{H}))$ the same deep and powerful results linking Euclidean Optimal Transport problems in $\mathcal{P}_2(\mathrm{H})$ and convex analysis. Our approach relies on the notion of totally convex functionals, on their total subdifferentials, and their Lagrangian liftings in the space square integrable $\mathrm{H}$-valued maps $L^2(\mathrm{Q},\mathbb{M};\mathrm{H}).$ With these tools, we identify a natural class of regular measures in $\mathcal{P}_2(\mathcal{P}_2(\mathrm{H}))$ for which the Monge formulation of the OT problem has a unique solution and we will show that this class includes relevant examples of measures with full support in $\mathcal{P}_2(\mathrm{H})$ arising from the push-forward transformation of nondegenerate Gaussian measures in $L^2(\mathrm{Q},\mathbb{M};\mathrm{H}).$

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

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

  1. On the stability of proximal operators in Wasserstein spaces under different notions of convexity

    math.OC 2026-07 accept novelty 7.0 of 10

    Wasserstein proximal operators are non-expansive under total or 2-base generalized geodesic convexity and locally 1/2-Hölder under ordinary generalized geodesic convexity.

  2. Mean-Field Model for Two-Layer Neural Networks Trained with Consensus-Based Optimization

    cs.LG 2025-11 conditional novelty 5.0 of 10

    CBO can train small two-layer networks, a hybrid CBO-Adam method improves convergence and stability, and a Wasserstein mean-field model of CBO has monotonically decreasing variance.

  3. 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

3 extracted references · 3 canonical work pages · cited by 3 Pith papers

  1. [60]

    On the optimal mapping of distributions

    issn: 0022-1236,1096-0783. doi: 10.1016/j.jfa.2024.110342. url: https://doi. org/10.1016/j.jfa.2024.110342. 50 [KS84] M. Knott and C. S. Smith. “On the optimal mapping of distributions”. In: J. Optim. The- ory Appl. 43.1 (1984), pp. 39–49. issn: 0022-3239,1573-2878. doi: 10.1007/BF00934745. url: https://doi.org/10.1007/BF00934745. [LS25] Hugo Lavenant and...

    work page arXiv 2024

  2. [299]

    Flowing Datasets with Wasserstein over Wasserstein Gradient Flows

    issn: 0002-9947. doi: 10.1090/s0002-9947-1969-0239652-5 . [Bia+08] Francesca Biagini et al. Stochastic calculus for fractional Brownian motion and applica- tions. Probability and its Applications (New York). Springer-Verlag London, Ltd., Lon- don, 2008, pp. xii+329. isbn: 978-1-85233-996-8. doi: 10.1007/978-1-84628-797-8 . url: https://doi.org/10.1007/978...

    work page Pith review arXiv doi:10.48550/arxiv.2506.07534 1969

  3. [7243]

    Entropic measure and Wasserstein dif- fusion

    doi: 10.1016/0047-259X(90)90070-X . url: https://doi.org/10.1016/0047- 259X(90)90070-X. [RR98] Svetlozar T. Rachev and Ludger R¨ uschendorf. Mass transportation problems. Vol. I, II . Probability and its Applications (New York). Theory. Springer-Verlag, New York, 1998. [RS09] Max-K. von Renesse and Karl-Theodor Sturm. “Entropic measure and Wasserstein dif...

    work page doi:10.1016/0047-259x(90)90070-x 1998

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.