{"id":"d549ce2d-1c82-4849-a4ce-4a1a01d41f35","arxiv_id":"2509.04935","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The quadratic barycentric transport cost equals the infimum of an expected kinetic-energy integral over semimartingales, and the optimal processes are geodesics with Markovian dynamics.","lead":"This math paper proves a new dynamical formula for the barycentric transport cost between two probability distributions, analogous to the Benamou-Brenier formula for the classical Wasserstein distance. The result also shows that optimal random paths are geodesics with a Markov property, and it gives explicit Gaussian formulas.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Achievability in Theorem 9 is not proved as written: the stretched-Brownian kernel is constructed for the wrong marginal (μ instead of μ̄), so the asserted martingale M with M1∼ν may not exist.","rationale":"The reader's weakest_assumption concerned the external theorem [6, Theorem 2.2] and possible extra conditions. My stress-test identifies a more concrete and more damaging issue in the way that theorem is invoked: the proof applies it to the wrong marginal. The sentence 'Since μ ≤c ν' is false for arbitrary μ,ν, and the construction only works if the weak transport problem is posed with μ̄=∇ϕ#μ. This is a load-bearing gap in the achievability half of Theorem 9, the paper's central result. However, the fix is natural and likely correct: μ̄≤cν by construction, all second moments are finite, and [6, Theorem 2.2] should apply to μ̄. Thus I do not regard the central claim as false; I regard the manuscript as needing a corrected proof of this step before full acceptance. This is precisely a CONDITIONAL verdict rather than REJECT or UNCHANGED.","tokens_in":22347,"tokens_out":7090,"duration_ms":77000,"concrete_test":"Re-derive the construction in §4 with the corrected marginal: let q be a minimizer of inf_q ∫ W2²(q_z,γ) dμ̄(z) subject to ∫ y dq_z(y)=z for μ̄-a.e. z and μ̄ q=ν, and define M_t=g_t(∇ϕ(X0),B_t) using ∇F_z#γ=q_z. Verify three conditions: (a) M0=∇ϕ(X0) a.s.; (b) M1∼ν; (c) M is an F-martingale. If all hold, then X_t in (22) is admissible and E∫||v_t||²dt=T2(ν|μ), so Theorem 9 stands modulo the typo. If any fails, the central equality is unproved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is the equality (20). The lower bound is fine, but the achievability half relies on exhibiting the process (22) with a martingale M such that M0=∇ϕ(X0) and M1∼ν. In the proof, this M is imported from [6, Theorem 2.2] via the weak transport problem inf_p ∫ W2²(p_x,γ)dμ(x) with constraints ∫y dp_x(y)=x for μ-a.e. x and μp=ν. The text then says 'Since μ ≤c ν this set of kernels is non-empty.' But for arbitrary μ,ν∈P2(Rd), μ≤cν is false, so this set is generally empty. What is needed is a martingale starting from μ̄=∇ϕ#μ, not from μ. The correct weak transport problem should have first marginal μ̄, which does satisfy μ̄≤cν. As written, however, p* has first marginal μ. Setting M_t=g_t(∇ϕ(X0),B_t) gives Law(M1)=∫ p*_{∇ϕ(x)} dμ(x)=μ̄ p*, which need not equal ν; moreover p*_z is only defined μ-a.e., not necessarily μ̄-a.e. Thus the proof of existence of an admissible process achieving equality in (20) has a real gap. The gap is likely typographical and repairable by replacing μ with μ̄ throughout the weak transport construction, but the submitted argument does not do this.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the quadratic barycentric transport cost T_2(ν|μ) between probability measures on R^d. Its main result, Theorem 9, asserts a Benamou–Brenier type representation: T_2(ν|μ) equals the infimum of E∫_0^1 ||v_t||^2 dt over semimartingales dX_t = v_t dt + dM_t with X_0∼μ, X_1∼ν, and M an F-martingale. It further claims that equality is attained by processes X_t = X_0 + t(∇φ(X_0)-X_0) + M_t - ∇φ(X_0), where ∇φ#μ is the backward projection μ̄ of μ onto {η≤_c ν}, M_0=∇φ(X_0), M_1∼ν; these optimal processes satisfy the geodesic identity T_2^{1/2}(μ_t|μ_s)=(t-s)T_2^{1/2}(ν|μ) and are Markov when M is Markovian. The proof is built on a dual-potential construction: Section 3 associates to a dual optimizer f̄ a convex paving {D_{f̄}(z)} that is invariant for all martingales from μ̄ to ν (Lemma 4, Proposition 5, Corollary 6). Section 5 connects martingales from μ to a forward projection ν̃ with martingales from μ̄ to ν via ∇φ. Section 6 gives explicit Gaussian formulas: the backward projection is Gaussian with covariance solving a matrix optimization, and a closed form in the commuting case.","tokens_in":22701,"tokens_out":5800,"duration_ms":64515,"significance":"If the main theorem is correct, the paper gives a genuine semimartingale-transport formulation of the barycentric cost, a geodesic structure analogous to McCann interpolation, and a Markov property for optimal processes. This is a valuable bridge between weak optimal transport, martingale transport, and the de March–Touzi convex paving theory. The paper is written in a modular way, with detailed proofs of Lemma 4 and Proposition 5, and it contains explicit, falsifiable statements in the Gaussian case (Propositions 15 and 18). The authors also honestly acknowledge the concurrent Gaussian results of Alfonsi–Jourdain. However, the proof of the achievability half of Theorem 9 contains a genuine error in the construction of the martingale M, as detailed below; the error appears repairable by replacing μ with μ̄, but it is load-bearing for the central claim in its current form.","major_comments":[{"comment":"The weak optimal transport problem is formulated with first marginal μ: inf_p ∫ W_2^2(p_x,γ)dμ(x) subject to ∫ y dp_x(y)=x for μ-a.e. x and μp=ν. The text then says this set of kernels is non-empty because ‘μ ≤_c ν’. For arbitrary μ,ν∈P_2(R^d), μ≤_cν is false and the set is generally empty. What is needed is a martingale starting from μ̄=∇φ#μ, not from μ. The subsequent construction M_t=g_t(∇φ(X_0),B_t) uses the kernel p* at the point ∇φ(X_0); if p* is defined only μ-a.e., this is not defined μ̄-a.e., and even if extended, M_1 would have law μ̄p*, not necessarily ν. The correct fix is to replace μ by μ̄ throughout the weak transport formulation: inf_p ∫ W_2^2(p_z,γ)dμ̄(z) with ∫y dp_z(y)=z for μ̄-a.e. z and μ̄p=ν. Since μ̄≤_cν, the set is non-empty, and the same argument yields M_0=∇φ(X_0), M_1∼ν and a martingale M. This correction is necessary for the existence of the process (22) and h","section":"Section 4, proof of Theorem 9, paragraph ‘In order to construct the martingale part…’"},{"comment":"The proof of Proposition 14 is omitted with the statement ‘proof is identical and thus omitted’. Since Proposition 14 is used in the Gaussian section to identify C_{f̄}(z) as an affine subspace, the omission is not merely cosmetic. If the proof is truly identical to that of Propositions 11–12, a short indication of the dictionary between f̄ and ḡ should be supplied. In addition, Proposition 26(ii) says details are left to the reader; given that it is an adaptation of an external result, this is acceptable but should be made precise.","section":"Section 5, Proposition 14 and Section 6"}],"minor_comments":[{"comment":"In the bullet list defining the optimal process, the notation ‘µ := argmin_{η≤_cν} W_2^2(µ,η)’ reuses μ for the backward projection. It should read μ̄ = argmin…, and later W_2^2(μ,μ̄) rather than W_2^2(μ,μ).","section":"Section 4, after (22)"},{"comment":"In the second bullet, ‘if g is an optimizer of (8), then f=P_2g is an optimizer of (8)’ should refer to problem (7).","section":"Section 2, Lemma 1"},{"comment":"In the displayed chain, the term ‘(t−s)^2 T_2^{1/2}(ν|μ)’ should be ‘(t−s)^2 T_2(ν|μ)’, since the expectation involving E||∇φ(X_0)−X_0||^2 equals T_2(ν|μ), not its square root.","section":"Remark 10"},{"comment":"The phrase ‘let P be an orthonormal matrix P’ is redundant; also the expression D_μ := PΣ_μP^T is used before P is fully specified. Minor rewording would help.","section":"Section 6, Proposition 18"},{"comment":"The paper contains a number of small typographical slips (e.g. ‘Item 2.’ for ‘Item (ii)’, the use of ar µ versus µ in the backward projection, and the notation µ_s for the backward projection of µ_s in Remark 10). These do not affect the mathematics but should be cleaned up.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The gap in the proof of Theorem 9 appears to be a genuine but localized typographical error: the weak transport problem must be posed with first marginal μ̄, not μ. Once this correction is made, the argument lines up with the known theory (backward projection, Strassen's theorem, stretched Brownian motion from [6]). I therefore view this as a repairable load-bearing issue rather than a fatal flaw, and recommend major revision. The paper otherwise fits the journal well and the Gaussian overlap with [2] is properly acknowledged."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does two real things: it proves a Benamou-Brenier type formula for T2 and shows optimal processes are geodesics with a Markov property. That is new and worth having. The main proof is modular and mostly rigorous. But the achievability half of Theorem 9 has a genuine gap: the stretched-Brownian kernel is built from the wrong first marginal.\n\nThe lower bound in (20) is a clean Jensen argument. The convex paving in Section 3 is a nice constructive structural result, and the Markov property (iv) follows from it. The paper also honestly flags that the Gaussian results duplicate Alfonsi-Jourdain, which is the right call.\n\nThe gap: in the proof of Theorem 9, they introduce a weak optimal transport problem with first marginal μ and constraint ∫y dp_x(y)=x, claiming \"Since μ ≤c ν this set of kernels is non-empty.\" That is false in general; μ ≤c ν would be a martingale coupling, which is not assumed. What is true is μ̄ ≤c ν, where μ̄ = ∇ϕ# μ. The kernel should have first marginal μ̄, and then M1 ~ ν follows. As written, p* has first marginal μ, so M1 would have law μ̄p*, not necessarily ν, and p*_z is only defined for μ-a.e. z. This is repairable by replacing μ with μ̄ throughout the construction; the rest of the proof stands. It's a typo in a load-bearing place, but the fix is clear.\n\nMinor issues: Proposition 14 omits a proof that is indeed identical; Proposition 26(ii) leaves details; the Gaussian section overlaps with [2] but they say so. No circularity—the main formula is proved independently.\n\nWho it's for: people working in weak optimal transport, semimartingale transport, and martingale transport. A good referee can handle this. I'd send it to peer review; the gap is real but likely fixable. If the authors replace μ by μ̄ in that paragraph, the theorem goes through.","headline":"Genuinely new Benamou-Brenier formula for quadratic barycentric transport, with a load-bearing typo in the achievability proof that is easy to fix.","tokens_in":23185,"tokens_out":3273,"would_cite":true,"duration_ms":32867,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60A10","49J55","60G42"],"pacs":[],"model":"deepseek-v4-flash","headline":"The quadratic barycentric transport cost between two probability measures is equal to the minimum expected squared drift of a semimartingale with those endpoint laws.","keywords":["weak optimal transport","quadratic barycentric transport","Benamou-Brenier formula","convex order","martingale transport","convex paving","geodesics","Gaussian measures"],"falsifier":"Test the theorem on a pair μ,ν in P₂(R^d) with finite variance for which ν is purely atomic and d≥2, e.g. μ uniform on the vertices of a cube and ν uniform on a larger cube's vertices. Solve the finite-dimensional linear program for T2 from the decomposition (11), and independently solve a finely time-discretized version of the semimartingale problem in (20) by minimizing over v and martingale increments. If the two minima differ, Theorem 9(i) fails.","tokens_in":22292,"feed_emoji":"📐","tokens_out":15538,"duration_ms":150220,"temperature":0.7,"pith_summary":"The paper studies the quadratic barycentric transport cost between two probability measures—the smallest possible value of the expected squared conditional mean difference E‖E(Y−X|X)‖² over couplings (X,Y). It proves a Benamou–Brenier-type theorem for this cost: T2 equals the minimum expected integral of squared drift over processes X_t that start with law μ, end with law ν, and evolve as a drift plus a martingale. It also shows the minimizers are geodesics—the cost between any two times along an optimal path scales as the squared time difference—and that the initial drift is readable from the current position, which makes the optimal dynamics Markov whenever the martingale part is Markov. This turns a static optimization over couplings into a stochastic control problem and connects barycentric transport to semimartingale and martingale transport.","feed_headline":"Barycentric transport cost equals a semimartingale control value","feed_subtitle":"Static barycentric transport now has a dynamic Benamou–Brenier formula with martingale noise.","key_machinery":"The paper's load-bearing construction is the convex paving {D_f(z)} attached to a dual optimizer f of the barycentric problem. Each D_f(z) is the set of points y such that f is affine on the segment from z to y (and just past y); equivalently, it is the projection of the relative face of the epigraph of f. Lemma 4 and Proposition 5 show the paving partitions the domain into convex cells that are invariant under every martingale connecting μ̄ and ν: if M_0∼μ̄ and M_1∼ν, then M_t∈D_f(M_0) almost surely for all t. This invariance makes the drift u=X_0−∇ϕ(X_0) recoverable from X_t=(1−t)u+M_t: on each cell u lies in ∂f(m) for every m in the cell, and the map (m,u)↦(1−t)u+m is injective on the sub","core_discovery":"The central claim is Theorem 9: for all μ,ν∈P₂(R^d), T2(ν|μ)=inf E∫₀¹‖v_t‖²dt, where the infimum ranges over progressively measurable drifts v and martingales M for which X_t=X_0+∫₀^t v_s ds+M_t−M_0, X_0∼μ, and X_1∼ν. Equality is attained by the explicit family X_t=(1−t)X_0+t∇ϕ(X_0)+M_t−∇ϕ(X_0), where ∇ϕ is the 1-Lipschitz gradient of a convex function pushing μ forward to μ̄, the backward projection of μ onto the set of measures dominated by ν in convex order, and M is any martingale with M_0=∇ϕ(X_0) and M_1∼ν. Along these processes the marginal laws satisfy T2(μ_t|μ_s)=(t−s)²T2(ν|μ), and if M is Markovian, X is Markovian.","pith_inferences":["Beyond the paper, the dynamic formula suggests a numerical route to T2 that does not require dual potentials: discretize time and optimize over v and M with the martingale constraint imposed on conditional increments; the equality with T2 provides a built-in check on any discretization.","The martingale-invariant convex paving may transfer to entropic or regularized versions of barycentric transport, where the same cells could localize the effect of regularization on martingale constraints.","The correspondence between martingales from μ to a forward projection and martingales from μ̄ to ν hints at a full dictionary between backward and forward Wasserstein projections: optimizers of one problem may be transported to optimizers of the other by ∇ϕ and ∇ϕ*, beyond the Gaussian examples worked out here."],"forward_implications":["Barycentric optimal transport becomes a semimartingale transport problem: costs, couplings, and geodesics can be described by drift processes and martingale components rather than by two-point couplings.","The optimal interpolations are constant-speed geodesics for √T2: T2(μ_t|μ_s)=(t−s)²T2(ν|μ) for every pair of times along a process of the form (22).","The initial drift X_0−∇ϕ(X_0) is σ(X_t)-measurable for each t<1, so the optimal process is Markov whenever the martingale part is Markov; this is a stochastic analogue of non-crossing in McCann interpolation.","Every optimal T2 plan can be factored as the deterministic map ∇ϕ to the backward projection μ̄ followed by an arbitrary martingale coupling from μ̄ to ν; in the commuting Gaussian case μ̄ is Gaussian with covariance min(Σ_μ,Σ_ν) in a common eigenbasis and T2 has the closed form |m_ν−m_μ|²+Σ_i[σ_i(μ)−σ_i(ν)]²₊."],"supporting_citations":[{"why":"Supplies the backward projection μ̄, the transport map ∇ϕ, the identity T2=inf_{η≤_cν}W₂², and the martingale-kernel structure of optimal plans used throughout Theorem 9.","marker":"[16]"},{"why":"Establishes the Kantorovich-type duality (7) for T2 from which the dual optimizers and the paving are built.","marker":"[19]"},{"why":"Imported existence theorem for a stretched Brownian motion; it constructs the martingale M with M_0=∇ϕ(X_0) and M_1∼ν that achieves equality in (20).","marker":"[6, Theorem 2.2]"},{"why":"Provides the convex-paving technique for decomposition of multidimensional martingale transport plans that the paper adapts to dual potentials.","marker":"[13]"},{"why":"Source for the P2-dual formulation (8) and the theory of backward and forward Wasserstein projections used in Lemmas 1 and 3.","marker":"[21]"},{"why":"Strassen's convex-order characterization ensures the existence of the martingale kernels from μ̄ to ν underlying optimal plans and the martingale construction.","marker":"[26]"}],"fun_headline_variants":["Semimartingale formula for barycentric transport cost","Barycentric transport as martingale-controlled path","Dynamic Benamou-Brenier for barycentric transport","Quadratic barycentric cost from martingale paths","Martingale paths reveal barycentric cost"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Everything rests on the imported result that, on a sufficiently rich filtered probability space, a continuous martingale with prescribed marginals M_0=∇ϕ(X_0) and M_1∼ν can always be constructed from a Brownian motion; if that theorem needs extra moment or support conditions beyond finite variance, the equality can fail for some pairs.","fun_headline_variants_meta":{"raw":{"variants":["Semimartingale formula for barycentric transport cost","Barycentric transport as martingale-controlled path","Dynamic Benamou-Brenier for barycentric transport","Quadratic barycentric cost from martingale paths","Martingale paths reveal barycentric cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001021,"raw_usage":{"total_tokens":4074,"prompt_tokens":605,"completion_tokens":3469,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":349,"completion_tokens_details":{"reasoning_tokens":3411}},"tokens_in":349,"tokens_out":3469,"duration_ms":24531,"temperature":1.0,"reasoning_tokens":3411,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:46:27.463181+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Test the theorem on a pair μ,ν in P₂(R^d) with finite variance for which ν is purely atomic and d≥2, e.g. μ uniform on the vertices of a cube and ν uniform on a larger cube's vertices. Solve the finite-dimensional linear program for T2 from the decomposition (11), and independently solve a finely time-discretized version of the semimartingale problem in (20) by minimizing over v and martingale increments. If the two minima differ, Theorem 9(i) fails.","supporting_citations":[{"cited_title":"On a mixture of bren ier and strassen theorems","cited_arxiv_id":null,"evidence_quote":"Supplies the backward projection μ̄, the transport map ∇ϕ, the identity T2=inf_{η≤_cν}W₂², and the martingale-kernel structure of optimal plans used throughout Theorem 9."},{"cited_title":"Kantorovich duality for general transport c osts and applications","cited_arxiv_id":null,"evidence_quote":"Establishes the Kantorovich-type duality (7) for T2 from which the dual optimizers and the paving are built."},{"cited_title":"Irreducible convex pa ving for decomposition of multidimensional martingale tra nsport plans","cited_arxiv_id":null,"evidence_quote":"Provides the convex-paving technique for decomposition of multidimensional martingale transport plans that the paper adapts to dual potentials."},{"cited_title":"Backward and forward W asserstein projections in stochastic order","cited_arxiv_id":null,"evidence_quote":"Source for the P2-dual formulation (8) and the theory of backward and forward Wasserstein projections used in Lemmas 1 and 3."},{"cited_title":"Strassen","cited_arxiv_id":null,"evidence_quote":"Strassen's convex-order characterization ensures the existence of the martingale kernels from μ̄ to ν underlying optimal plans and the martingale construction."}],"review_version":1}