{"id":"f70fa712-d8e1-4477-9482-e1fa674f5047","arxiv_id":"2607.26671","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"On CAT(0) spaces, every probability measure has a unique Wasserstein projection onto the set of measures dominated by ν in convex order, and the optimal transport is a 1-Lipschitz map.","lead":"This paper proves that probability measures on CAT(0) spaces—spaces of non-positive curvature—have a unique nearest measure in convex order, transported by a 1-Lipschitz map. It also gives a local Hölder version on positively curved CAT(κ) spaces and a Strassen-type martingale characterization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the CAT(0) main theorem is internally consistent; the κ>0 restriction is an explicit scope condition, not a hidden flaw.","rationale":"The Reader's verdict ACCEPT with moderate confidence is reasonable. The Reader's weakest assumption—the localization condition (2.1) for κ>0—is real but is an explicitly stated hypothesis, not a hidden gap; the central CAT(0) theorem is unaffected. I agree with the Reader that the heaviest external dependencies are the conditions from [8] and the delegated measurability proof, but neither rises to a load-bearing objection: the weak-OT fundamental theorem is cited precisely, and the measurability gap is standard. A non-finding is therefore appropriate: the central claim is internally consistent, and the review should not be adjusted. The 'partial' agreement reflects that I accept the Reader's identification of the κ>0 condition as the least secure part of the statement, but I do not regard it as a correctness risk to the main theorem.","tokens_in":23790,"tokens_out":33490,"duration_ms":336422,"concrete_test":"Worth running: supply the omitted measurability argument—show that the set {(x,y): y∈C(p_x)} is Borel by expressing C(p_x) as the intersection over a countable determining family of convex Lipschitz functions of {y: f(y)≤∫f dp_x}, and then use Lusin-Novikov to obtain a Borel selector T(x)=P_{C(p_x)}x. If this cannot be done, the optimal plan (id,T)_#μ is not well-defined and Theorem 2.1's construction fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I traced the central chain: Proposition 2.2 supplies continuity and convexity of the weak cost; the verification of conditions (B)/(C) of [8] is plausible and the cited fundamental theorem supplies primal existence and duality; the construction T(x)=P_{C(p_x)}(x) yields a feasible \\bar\\mu≤_cvxν; k-convexity of d^2(x,·) forces uniqueness of both \\bar\\mu and the optimal plan; Proposition 3.1 together with Reshetnyak's quadruple comparison yields the 1-Lipschitz property. I found no internal inconsistency in the CAT(0) case. The κ>0 part is explicitly restricted by condition (2.1); failure of that condition limits the theorem's scope rather than invalidating the argument. The genuine soft spots are the reliance on the external fundamental theorem of weak OT [8] and the assertion that measurability of T is 'easy to see' and left to the reader. Both are standard and non-central for the main CAT(0) theorem; the latter is fillable via Lusin-Novikov once the graph of C(p_x) is shown to be Borel. I do not see a load-bearing flaw.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper establishes a non-Euclidean analogue of the Brenier–Strassen theorem. For a complete separable CAT(0) space E and μ,ν∈P2(E), it proves existence and uniqueness of the W2-projection μbar of μ onto {η≤cvxν}, and shows that the unique optimal plan from μ to μbar is induced by a 1-Lipschitz map, with no absolute-continuity assumption. The proof recasts the projection as a weak optimal transport problem with cost c(x,p)=d²(x,C(p)), establishes weak-OT duality, and then uses CAT(κ) comparison geometry to prove uniqueness and regularity. For κ>0, a localized version under condition (2.1) yields a 1/2-Hölder optimal map. A Strassen-type characterization of one-step barycentric martingales on proper CAT(0) spaces is also given.","tokens_in":24001,"tokens_out":19855,"duration_ms":185618,"significance":"If correct, this is a substantial advance: it transfers the Gozlan–Juillet Euclidean theory to spaces with curvature bounded above, with clean statements and explicit constants. The central CAT(0) theorem is internally consistent and the proof is mostly self-contained, relying on recent weak-OT duality [8] and on known comparison tools (Reshetnyak quadruple comparison, Kuwae/Yokota barycenter results). The paper is honest about its scope: the κ>0 results require the support restriction (2.1). It also includes a useful Strassen-type result for barycentric martingales. The main technical gaps are local and fixable: a terse verification of the hypotheses of [8] and an omitted measurability proof for the transport map T.","major_comments":[{"comment":"The argument invokes [8, Theorem 1.2] and states that conditions (B)/(C) are satisfied, but the conditions are not stated. Please state the hypotheses of [8] and verify them explicitly. In particular, explain why continuity and convexity of c from Proposition 2.2 imply condition (C), and confirm that the displayed growth bound on p. 13 is exactly condition (B). Since this is the foundation for existence of the optimal kernel and the duality formula, the verification should not be left to the reader.","section":"§2.4, proof of Theorem 2.1"},{"comment":"Measurability of the map T is asserted with 'details are left to the reader.' This is load-bearing because \\bar μ=T#μ and the optimal plan are defined through T. Please add a short proof: by measurability of the kernel x↦p_x and Proposition 2.4, the map x↦C(p_x) is measurable into closed convex sets with the Hausdorff topology; by Lemma A.1 the metric projection (K,x)↦P_Kx is jointly continuous, so T is measurable. Alternatively, use the Lusin–Novikov theorem on the graph of C(p_x).","section":"§2.4, definition of T after (2.7)"}],"minor_comments":[{"comment":"The symbol D appears in the definition of Qf(x):=inf_{y∈D}... but D has not been defined. For κ=0 it should be D=E; please correct.","section":"Theorem 2 (Introduction)"},{"comment":"'A is dense inside the support of μ' deserves a one-sentence justification (any open set meeting supp μ has positive μ-measure, so a full-measure subset must intersect it).","section":"§3.2, proof of Theorem 1"},{"comment":"The gluing argument used to construct (Z,X0,X1) is omitted; please include a sentence explaining the standard construction.","section":"§2.4, uniqueness of the optimal plan"},{"comment":"The use of [38, Theorem A] on the product space requires a brief justification that the product B×B falls under the hypotheses of [38].","section":"Lemma 1.3(2)"}],"recommendation":"minor_revision","confidential_remarks":"The paper is mathematically sound in its main lines; I found no fatal flaw. The AI-use disclosure is transparent and does not affect my assessment. The main request is to make the verification of the external weak-OT theorem explicit and to fill the small measurability gap. Both are local and easily addressed. If the authors cannot fully verify [8]'s conditions in the κ>0 case, the CAT(0) claims remain significant and could be published separately."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real extension of the Euclidean Brenier–Strassen theorem to CAT(0) and CAT(κ) spaces, and the central argument looks sound. The main theorem gives a unique W2-projection of a measure to the convex-order sublevel set of a reference measure, with a deterministic 1-Lipschitz optimal map in the CAT(0) case and a 1/2-Hölder version under a local radius condition in the CAT(κ) case.\n\nWhat's actually new: the projection theorem itself, the identification of the projection problem with a weak optimal transport problem whose cost is squared distance to the set of convex means, and the Lipschitz/Hölder regularity of the optimal map without any absolute continuity assumption. I also think Proposition 2.4 is a real contribution: it gives Lipschitz (or Hölder, for κ>0) stability of the set of convex means under W1, extending the known barycenter contraction property. The Strassen-type barycentric martingale characterization in Section 4 is a nice extra, though not the heart of the paper.\n\nSoft spots, in proportion: the proof leans on the fundamental theorem of weak optimal transport from [8], a recent preprint, and conditions (B)/(C) of that theorem are asserted rather than checked in detail. I traced the surrounding argument and the assertions look plausible — Proposition 2.2 plus the quadratic growth bound appear to cover both conditions — but a referee should ask for a short explicit verification. Measurability of the map T is also delegated to \"easy to see\" using Lemma A.1; that is fillable but should be spelled out. The κ>0 localization condition (2.1) is load-bearing but honestly stated; failure of that condition just restricts the theorem's scope, not the proof's validity.\n\nOne thing I want to flag: the AI disclosure for Proposition 2.4 is transparent and the authors say they checked the proof. I read the proof and it is an ordinary mathematical argument, not a black box. No red flag.\n\nThis paper is for people working on optimal transport in NPC/CAT spaces, convex order, and metric martingales. If the weak-OT technical details hold up, it will be a standard reference. I would send it to a serious referee; my recommendation is accept after light-to-moderate revision asking for the (B)/(C) verification and a few more details on measurable selections.","headline":"A genuine CAT(0)/CAT(κ) extension of the Brenier–Strassen projection theorem, with a plausible proof chain; worth refereeing carefully.","tokens_in":24597,"tokens_out":1596,"would_cite":true,"duration_ms":17660,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","60E15","53C23","60G48"],"pacs":[],"model":"deepseek-v4-flash","headline":"On complete separable CAT(0) spaces, every probability measure has a unique Wasserstein projection onto the set of convex-order-dominated measures, and the optimal transport from μ to its projection is always given by a 1-Lipschitz map.","keywords":["CAT(κ) spaces","convex order","Wasserstein projection","optimal transport","barycenter","convex means","weak optimal transport","barycentric martingales"],"falsifier":"Take the round sphere of radius 1/√κ as a CAT(κ) space, let ν be uniform on the equator and μ a Dirac mass at the north pole, so the diameter bound in (2.1) is exactly at the boundary ε = π/(2√κ); compute the projection of μ onto {η ≤_cvx ν}. If the projection is not unique or the optimal map is not 1/2-Hölder, the localization condition is shown to be sharp.","tokens_in":23635,"feed_emoji":"📐","tokens_out":5921,"duration_ms":49650,"temperature":0.7,"pith_summary":"This paper extends the classical Brenier–Strassen theorem—about projecting a probability measure onto the set of measures dominated by another in convex order—from flat Euclidean space to spaces with curvature bounded above. On any complete separable CAT(0) space, the authors prove that the projection exists, is unique, and is realized by an optimal transport plan that is always deterministic: a single 1-Lipschitz transport map, with no absolute-continuity assumption on the source measure. The proof proceeds by identifying the projection problem with a weak optimal transport problem whose cost is the squared distance from a point to the set of convex means of a measure, and by establishing a strong duality with convex potentials. A localized version holds on CAT(κ) spaces with positive curvature, where the optimal map is 1/2-Hölder continuous. As a by-product, the paper gives a Strassen-type characterization of one-step barycentric martingales on proper CAT(0) spaces.","feed_headline":"Every measure has a unique convex-order projection on CAT(0) spaces","feed_subtitle":"The nearest dominated measure is reached by a deterministic 1-Lipschitz map, with no regularity assumption on μ.","key_machinery":"The key object is the set C(p) of convex means of a probability measure p—the points x such that δ_x ≤_cvx p. The central mechanism is the weak optimal transport cost c(x,p)=d²(x,C(p)), which is continuous and convex in p on CAT(κ) spaces (locally for κ>0), and the identification min_{η≤_cvx ν} W2²(μ,η) = min_{p: μp=ν} ∫ d²(x,C(p_x)) μ(dx) = max_f ∫ Qf dμ − ∫ f dν, where Qf(x)=inf_y {f(y)+d²(x,y)} is the infimal convolution with squared distance. This equivalence turns the projection problem into a dual convex optimization problem, and the k-convexity of the squared distance (k=2 for κ=0; k=2√κ ε/tan(√κ ε) locally for κ>0) yields the regularity of the optimal map.","core_discovery":"The central claim is Theorem 1: for any μ,ν in P2(E) on a complete separable CAT(0) space, there exists a unique ar μ ≤_cvx ν with W2²(μ,ar μ)=inf_{η≤_cvx ν} W2²(μ,η), and there is a unique W2-optimal transport plan from μ to ar μ, induced by a 1-Lipschitz transport map T defined on the support of μ. The structural Theorem 2 recasts the projection problem as a weak optimal transport problem with cost c(x,p)=d²(x,C(p)), where C(p) is the closed convex set of convex means of p, and shows that the optimal map is simultaneously the metric projection of x onto C(p_x), the proximal operator of an optimal convex potential f, and the unique optimal transport map from μ to ar μ. The same duality hold","pith_inferences":["The proof suggests that the convex-order projection can be computed by solving a weak optimal transport problem, which may lead to practical algorithms for sampling and dimension reduction in Hadamard spaces.","The sharpness of the localization condition (2.1) is worth testing: if supports are allowed to approach distance Dκ/2 on a sphere, the Hölder exponent may degrade or the optimal map may fail to exist, delineating the true boundary of the theorem.","The 1-Lipschitz regularity of the projection map on CAT(0) spaces gives a curved analogue of Caffarelli's contraction theorem, potentially transferring concentration inequalities from ν to μ.","The barycentric martingale characterization may open the way to martingale optimal transport on CAT(0) spaces, with applications in robust finance in non-Euclidean state spaces."],"forward_implications":["The set of measures dominated by ν in convex order is a Chebyshev set in the Wasserstein space P2(E): every μ has a unique nearest point in it.","The nearest-point map is deterministic and 1-Lipschitz even when μ is singular, extending Brenier's theorem beyond absolute continuity.","The strong duality gives a computable dual formulation: the squared projection distance equals sup over convex potentials f of ∫ Qf dμ − ∫ f dν.","A Pythagorean inequality W2²(μ,ν) ≥ W2²(μ,ar μ) + W2²(ar μ,ν) and non-expansiveness W2(ar μ1,ar μ2) ≤ W2(μ1,μ2) hold for the convex-order projection.","On proper CAT(0) spaces, existence of a barycentric martingale (in the Sturm sense) between μ and ν is characterized by the barycentric envelope inequality ∫ bar(f) dμ ≤ ∫ f dν for all admissible l.s.c. functions f."],"fun_headline_variants":["Unique convex-order projection on CAT(0) spaces","1-Lipschitz map for convex-order projection, no regularity","Brenier-Strassen on CAT(0): unique projection","No absolute continuity needed for CAT(0) convex-order projection","CAT(0) spaces: unique projection with 1-Lipschitz map"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"For positive curvature, the whole argument rests on the condition that the support of ν lies in a closed convex set D with all pairwise distances between supp μ and D strictly less than π/(2√κ); if that fails, the key k-convexity of the squared distance and the compactness arguments break down.","fun_headline_variants_meta":{"raw":{"variants":["Unique convex-order projection on CAT(0) spaces","1-Lipschitz map for convex-order projection, no regularity","Brenier-Strassen on CAT(0): unique projection","No absolute continuity needed for CAT(0) convex-order projection","CAT(0) spaces: unique projection with 1-Lipschitz map"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001327,"raw_usage":{"total_tokens":5245,"prompt_tokens":764,"completion_tokens":4481,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":4391}},"tokens_in":508,"tokens_out":4481,"duration_ms":32712,"temperature":1.0,"reasoning_tokens":4391,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T11:03:26.910113+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the round sphere of radius 1/√κ as a CAT(κ) space, let ν be uniform on the equator and μ a Dirac mass at the north pole, so the diameter bound in (2.1) is exactly at the boundary ε = π/(2√κ); compute the projection of μ onto {η ≤_cvx ν}. If the projection is not unique or the optimal map is not 1/2-Hölder, the localization condition is shown to be sharp.","supporting_citations":[],"review_version":1}