{"id":"a2450edf-ed35-4924-9f2c-70fd8924cc91","arxiv_id":"2509.01768","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Using totally convex functionals and their Lagrangian liftings, the authors prove existence and uniqueness of Monge solutions for optimal transport between super-regular laws of random measures.","lead":"This paper proves that optimal transport between laws of random measures has a unique deterministic solution for a wide class of super-regular measures. This gives the Monge formulation, not just existence of couplings, for optimal transport in the Wasserstein space built on another Wasserstein space.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central uniqueness proof depends on unproven companion results [CSS25, Thm 4.8] and [LS25] (in preparation); without them, Theorem 6.15 lacks a verifiable proof.","rationale":"The reader's conditional verdict correctly identifies the dependence on [CSS25] and, especially, on the in-preparation [LS25]. The paper's own text locates the two uses: the nonlocal-field representation in Theorem 5.7 and the Gδ-measurability of deterministic couplings in (2.4). Both are load-bearing for the uniqueness and Monge-map conclusion in Theorem 6.15. I did not find a clear internal contradiction in the main strategy; the overall approach is coherent and the examples appear to support the super-regularity claims. I also noticed a misstated inequality in (5.29), but the conclusion it is used for can be recovered from the true lower bound |X̃−X̃′|² ≥ (1−t)²|Δ0|² + t²|Δ1|², so I do not treat that as the deciding issue. The decisive issue remains the unresolved companion support. Thus the reader's CONDITIONAL verdict should stand unchanged: the result is plausible but cannot be accepted as fully verified until [LS25] is available and the [CSS25] hypotheses are confirmed.","tokens_in":51569,"tokens_out":29115,"duration_ms":337175,"concrete_test":"Obtain the proof of [LS25] (or give an independent verification) that Pdet(X×Y) is a Gδ subset of P(X×Y), using the universal-disintegration characterization (2.4). This single check settles whether the measurability/selection step in Proposition 6.5 and Theorem 6.15 is justified. Separately, verify that [CSS25, Thm 4.8] applies to the set \\hat F_t of Theorem 5.7 under the correct Lipschitz bounds from monotonicity; if both hold, the central proof goes through, otherwise Theorem 6.15 must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 6.15 is the paper's central claim, and its proof is not self-contained. Two imported results carry the load. First, Theorem 5.7 invokes [CSS25, Thm 4.8] to convert m.p.i.-invariant Lipschitz maps F_{t,i} on the lifted set \\hat F_t into continuous nonlocal fields f_{t,i}; this inversion (5.23) is what produces the geodesic non-branching structure used in Theorem 6.4 and hence in the strict-Monge equivalence. Second, the identification of Pdet(X×Y) as a Gδ set, stated just before (2.4), is attributed to [LS25], a paper marked 'In preparation'. That measurability fact is used in Proposition 6.5 and the selection step of Theorem 6.15 to assert that the deterministic-coupling restrictions and the final Monge map F are Borel. If either companion result has additional hypotheses, or if [LS25] is not released with a correct proof, then the uniqueness conclusion in Theorem 6.15 is unsupported as written. This is not an internal inconsistency in the strategy, but a missing support pillar in a chain where every link is needed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":51926,"tokens_out":13824,"duration_ms":157578,"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":[{"comment":"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.","section":"§2.4, Eq. (2.4); §5.3–6.2"},{"comment":"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.","section":"Proposition 2.6"},{"comment":"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.","section":"§6.2, proof of Theorem 6.15"}],"minor_comments":[{"comment":"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.","section":"Theorem 4.11"},{"comment":"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.","section":"Introduction, after (R1)–(R2)"},{"comment":"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.","section":"Theorem 5.5, proof"},{"comment":"The supremum is taken 'w.r.t. g∈G(H)' but should be over the measure-preserving isomorphisms g∈G(Q).","section":"Lemma 3.7, proof"},{"comment":"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.","section":"Definition 6.6"},{"comment":"The property 'essentially totally cyclically monotone' is not defined in the paper. Please define it or remove the qualifier 'essentially'.","section":"Theorem 6.15"}],"recommendation":"major_revision","confidential_remarks":"To the editor: this paper is part of a systematic research program and relies heavily on companion papers ([CSS25], [PS25], [LS25]). Reliance on published work is acceptable, but the central theorem currently depends on an in-preparation citation ([LS25]) and on a measurability proposition (Proposition 2.6) that is demonstrably wrong as written. The overall strategy appears sound and the issues are local and repairable, so I do not recommend rejection. I would urge the editor to ensure that the companion-paper results used as black boxes are available to referees in their final form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a real new result — uniqueness of Monge solutions for super-regular laws of random measures — and the Hilbertian lifting strategy is genuinely elegant. The one thing that keeps me from endorsing it as-is is that Theorem 6.15 leans on [LS25], a paper marked 'In preparation,' for a measurability fact (Pdet being Gδ) that is used in the Borel selection steps. If that companion doesn't come out with a correct proof, the central uniqueness claim is unsupported as written. I don't see an internal contradiction in the strategy; it's a missing support pillar, not a broken beam.\n\nThe genuinely new content: Theorem 6.15 (unique optimal random coupling, deterministic Monge map for super-regular M and arbitrary N) and Theorem 6.19 (all nondegenerate Gaussian push-forwards are super-regular in d=1). The d=1 proof is clean — it uses the diagonal condition, Karhunen-Loève, and a neat Fourier/Walsh argument to show the singular set is Gaussian-null. The higher-dimensional examples (C1 maps, fractional Brownian motion, Walsh-type expansions) are concrete and tie into Berman's condition; they are conditional on an integral condition like (6.55), which is explicit and checkable. The paper also makes good use of the authors' prior framework ([CSS23a], [CSS25]) — [CSS25] is published (Canad. J. Math.) so quoting it is not a problem. The concept of super-regularity is carefully introduced and they show it's stable under absolute continuity and that the class is dense; the 'nearly optimal' remarks are honest.\n\nSoft spots in proportion: (1) the [LS25] dependence is real and load-bearing. They use Pdet ⊂ P(X×Y) is Gδ to build Borel sets in Prop 6.5 and to ensure the Monge map in Thm 6.15 is Borel. Without a public proof of that fact, the uniqueness proof doesn't close. This is the main issue. (2) Minor typos: Prop 2.6's one-dimensional proof writes '> 1/k' where it must be '< 1/k' (upper semicontinuity gives open sublevel sets), and the implication labels in Thm 5.5's proof are scrambled (2⇒4, not 2⇒3). Neither affects the mathematics. (3) The paper is long and the notation is heavy; the reader who isn't already inside the authors' machinery will need patience. That's a presentation cost, not a flaw.\n\nFor a reader: someone working on Wasserstein-over-Wasserstein, hierarchical OT, or random-measure Monge problems will get real value. It deserves a serious referee. I'd send it to review now, with an explicit request that the authors either prove the measurability fact or point to a publicly available version of [LS25]; that's a condition, not a desk-reject.","headline":"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.","tokens_in":52363,"tokens_out":4634,"would_cite":true,"duration_ms":49577,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","46N10","60B05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that optimal transport between laws of random measures has a unique Monge solution for super-regular laws.","keywords":["optimal transport","random measures","Wasserstein space","totally convex functionals","Monge problem","Gaussian measures","Lagrangian lifting","subdifferentials"],"falsifier":"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.","tokens_in":51480,"feed_emoji":"📐","tokens_out":10527,"duration_ms":121672,"temperature":0.7,"pith_summary":"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.","feed_headline":"Super-regular laws get unique Monge transport maps","feed_subtitle":"For these 'super-regular' random measures, which include Gaussian-built examples, every target has one optimal transport map.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the law-invariance and extension theorem for Lipschitz maps invariant under measure-preserving isomorphisms; gives the nonlocal field representation used for minimal sections and geodesic inversion.","marker":"[CSS25]"},{"why":"Introduces totally convex functionals, total subdifferentials, and their Lagrangian lifting; the paper's Kantorovich-Legendre transform and minimal-section results build on it.","marker":"[CSS23a]"},{"why":"Proves the set of non-differentiability points of a convex Lipschitz function is contained in a σ-d.c. hypersurface, turning total subdifferential singletons into almost-everywhere W-differentiability under super-regular measures.","marker":"[Zaj79]"},{"why":"Classical result giving unique Monge transport maps via convex subdifferentials; the paper reproduces this pattern at the random-measure level.","marker":"[Bre91]"},{"why":"Knott-Smith result characterizing optimal couplings for the quadratic cost; the paper's total-subdifferential criteria are its random-measure analogue.","marker":"[KS84]"},{"why":"Shows deterministic couplings form a Gδ (Borel) set; this measurability is needed to select fully deterministic optimal random couplings.","marker":"[LS25]"},{"why":"Provides the Wasserstein-space toolkit: P2 metric structure, glueing lemma, displacement interpolation, and geodesic and nonbranching background used throughout Sections 5 and 6.","marker":"[AGS08]"},{"why":"Gives the Hilbert-space Gaussian framework, including Karhunen-Loève expansions, used to construct and analyze Gaussian-generated random measures.","marker":"[Bog98]"},{"why":"Provides occupation-density and Berman criteria used to verify that Gaussian-generated random measures have atomless or absolutely continuous realizations.","marker":"[GH80]"}],"fun_headline_variants":["Unique Monge solutions for super-regular random measures","Total convexity yields unicity in Wasserstein space","Super-regular laws guarantee a single Monge map","Monge problem unique for Gaussian-generated random measures","Total subdifferentials give unique Monge solutions"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Unique Monge solutions for super-regular random measures","Total convexity yields unicity in Wasserstein space","Super-regular laws guarantee a single Monge map","Monge problem unique for Gaussian-generated random measures","Total subdifferentials give unique Monge solutions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001037,"raw_usage":{"total_tokens":4256,"prompt_tokens":857,"completion_tokens":3399,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":3324}},"tokens_in":601,"tokens_out":3399,"duration_ms":27665,"temperature":1.0,"reasoning_tokens":3324,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:11:44.158832+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}