{"id":"22cc293e-fc75-4d2b-bd9c-bca16bb5a55a","arxiv_id":"2501.16316","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Weak optimal transport has a fundamental theorem: strong duality, primal and dual attainment, and complementary slackness, with applications to martingale and entropic transport.","lead":"This paper proves a unified duality and attainment theorem for weak optimal transport with convex costs, extending the classical fundamental theorem of optimal transport. It recovers key known results and yields new structure theorems for martingale and entropic transport problems.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.5 delegates the central duality proof to [11] without showing the extension from jointly lsc to merely Borel-in-x, lsc-in-ρ costs; this is the load-bearing step behind Theorem 1.2(ii).","rationale":"The reader's CONDITIONAL verdict is appropriate. I do not see an internal contradiction in the main theorem, but the proof's most exposed point is the unverified extension of [11] to non-jointly-lsc costs. This is more basic than the B/(C) conditions: those are clearly stated hypotheses for dual attainment, and Example 2.10 shows that (C) is genuinely needed. By contrast, the duality equality in Theorem 2.2(ii) is asserted with only a pointer to prior work, and the manuscript's own remarks flag that the current setting is 'slightly stronger' than the cited one. This makes the central claim conditionally supported rather than fully demonstrated. Secondary issues in Theorem 2.7's proof, such as the separate Komlós reindexings and the exact level-set choice in Lemma 2.8, appear repairable and do not change the verdict. If the concrete check confirms the adaptation, the paper should be accepted; if not, Theorem 1.2's generality must be narrowed or the missing minimax argument must be supplied.","tokens_in":42144,"tokens_out":21337,"duration_ms":225699,"concrete_test":"Obtain the proof of [11, Theorem 3.1] and check every place where joint lower semicontinuity or joint continuity of C is used. Then attempt to re-run the proof under Assumption 2.1 without joint lsc. As a stress case, take a non-closed Borel set A ⊂ X and set C(x,ρ) = 1_A(x) H(ρ|ν), or use the entropic cost (4.2) with c(x,y) = 1_A(x), which is Borel in x and convex/lsc in ρ but not jointly lsc. Test whether WTC = DC and primal attainment still hold. If they fail, Theorem 2.2(ii) needs an additional joint-lsc hypothesis or a new proof; if they hold, the delegation is vindicated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Under Assumption 2.1, C is only required to be Borel in x and convex/lsc in ρ; it need not be jointly lsc. Theorem 2.5, which provides the central duality WTC = DC and primal attainment, does not prove this claim. After changing the topology on Y, the proof states only that 'we can follow line by line [11, Proof of Theorem 3.1].' The related-literature section describes the cited setting as lsc costs on Polish spaces and explicitly says Theorem 1.2 is 'slightly stronger' in order to include entropic transport. Section 2.5 and Remark 2.16 again indicate that the Borel-in-x, lsc-in-ρ case is not identical to the joint-lsc setting; notably, Theorem 2.15, the relaxed non-convex variant, is stated without proof. The condition that must hold for Theorem 1.2(ii) is therefore that the minimax/duality argument in [11] survives without any joint-lsc assumption. The manuscript does not supply that argument. If it does not survive, costs such as the entropic cost (4.2) with only Borel c can fail the claimed duality, and the applications in Sections 3–5 lose their foundation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fundamental theorem for weak optimal transport with costs C(x,ρ) that are measurable in x and convex and lower semicontinuous in ρ for the p-weak topology. The main theorem claims primal attainment, strong duality with a C-transform dual, dual attainment under conditions (B) and (C), and a complementary slackness criterion. The paper then derives applications to barycentric costs (a Gangbo–McCann–Strassen theorem and convex Kantorovich–Rubinstein formulae), to entropic and convexly regularized optimal transport, and to relaxed martingale optimal transport, including a martingale Benamou–Brenier interpolation and entropic martingale transport.","tokens_in":42401,"tokens_out":6123,"duration_ms":61944,"significance":"If the central theorem is fully proved, this is a substantial unification: it extends the classical Kantorovich duality package to weak transport at a high level of generality and it yields concise derivations of several known results plus new structural results for barycentric and martingale-type problems. The paper is careful in separating the roles of conditions (B) and (C), and Example 2.10 gives a useful demonstration that condition (C) cannot simply be dropped. However, the main duality proof delegates the decisive minimax step to a prior paper, and the non-convex relaxed theorem is stated without proof; both points are load-bearing for the advertised scope and for the Section 5 applications.","major_comments":[{"comment":"The proof of the central duality WTC(μ,ν)=DC(μ,ν) is not self-contained. After establishing lower semicontinuity of ν↦WTC(μ,ν), the text states that 'we can follow line by line [11, Proof of Theorem 3.1]' and obtain the dual representation. The cited result is presented in the related-literature section as covering lsc costs on Polish spaces, whereas Theorem 2.5 only assumes C is Borel in x and lsc in ρ. The paper itself notes in §1.3 that Theorem 1.2 is 'slightly stronger' precisely in order to include entropic optimal transport in its usual generality. Since this extension is the load-bearing step, the manuscript needs to supply the actual argument or a precise statement from the literature that covers Borel-in-x, lsc-in-ρ costs; otherwise the duality for costs such as (4.2) with merely Borel c is not established.","section":"§2.1, Theorem 2.5"},{"comment":"Theorem 2.15, the fundamental theorem for relaxed WOT without convexity, is stated without proof. The surrounding text and Remark 2.16 only say that the generalization follows 'line by line' as in Theorem 2.2 and refer to [17] for the equivalence of formulations. This theorem is subsequently used in an essential way in Section 5: Theorem 5.1 relies on it for duality and dual attainment of the non-convex cost Cϑ, and Theorem 5.4 and Theorem 5.8 inherit that reliance. A proof, or a reference whose assumptions match exactly, must be provided before the Section 5 results can be considered established.","section":"§2.5, Theorem 2.15"},{"comment":"In the proof of Lemma 5.9 it is asserted that 'the reasoning in Corollary 2.13 also works for P∈Λ(μ,ν)' that are optimal for the relaxed non-convex problem. Corollary 2.13 is proved in the convex setting and relies on Theorem 2.2, complementary slackness, and C-monotonicity; no analogue is proved for the relaxed, non-convex setting of Theorem 2.15. Since Lemma 5.9 is used to prove the Gibbs-type structure in Theorem 5.8, this transfer from the convex theory to the relaxed setting needs to be justified explicitly.","section":"§5.2, Lemma 5.9"},{"comment":"The complementary slackness criterion is stated as an 'if and only if' for a pair (π,(f,g)) of candidates, under Assumption 2.1. The forward implication uses that both are optimal, and the reverse implication uses weak duality from Lemma 2.4 together with D≤WTC. This is correct given Theorem 2.5. However, the statement of Proposition 2.6 itself does not mention conditions (B) and (C) for dual attainment, which is fine, but it would help the reader to clarify explicitly that the equivalence is between simultaneous primal/dual optimality and the pointwise equality, not between individual optimality and the equality alone.","section":"§2.2, Proposition 2.6"}],"minor_comments":[{"comment":"There are several typographical issues: 'FUNDAMENT AL' in the title, 'Tentali' for 'Tetali' in the introduction, 'Propsition 4.1' in the proof of Theorem 4.2, 'Benaumou–Brenier' in Section 5.1, and 'vaild' in Remark 5.3. These should be corrected in a revision.","section":"Global"},{"comment":"The dual formula (3.3) is written as a supremum over 'ψ convex, lsc' without explicitly stating the integrability condition ψ∈L1(ν). Since ν(ψ) can be infinite, please add the domain convention or state that the supremum is over convex lsc ψ with ψ∈L1(ν) and ϑ□ψ∈L1(μ).","section":"§3, Theorem 3.1(i)"},{"comment":"In the converse direction of Theorem 4.2, the proof assumes that the functions f,g in the representation (4.5) belong to L1(μ)×L1(ν), while the theorem statement only says 'measurable'. If the representation can hold with non-integrable f,g, the converse needs a brief justification; if integrability is intended, the statement should say so.","section":"§4.1, Theorem 4.2"},{"comment":"The corollary states that every optimal π is C-monotone, but the proof uses dual attainment and hence conditions (B) and (C). This is clear from the proof, but the statement of the corollary only says 'Suppose that Assumption 2.1, (B) and (C) are satisfied', so the dependence is explicit. No change needed beyond ensuring the same conditions are cited in later uses of C-monotonicity.","section":"§2.4, Corollary 2.13"}],"recommendation":"major_revision","confidential_remarks":"The paper is ambitious and, if the missing proofs are supplied, would be a strong contribution to weak optimal transport. The main concern is not circularity or internal inconsistency but rather that the central duality theorem delegates the decisive extension to [11] without proof, and the non-convex relaxed theorem used in Section 5 is unproven. These issues are repairable within the manuscript's scope by adding complete arguments. The heavy reliance on the authors' own prior work is acceptable because those references are published and independent, but the new manuscript should make the incremental proof step explicit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers what the title promises: a fundamental theorem for weak optimal transport that extends the classical duality, attainment, and complementary slackness story to costs that are Borel in x and convex–lsc in the measure argument. The main theorem is plausible and the applications are the real payoff. The barycentric-cost results (Section 3) recover the convex Kantorovich–Rubinstein formula and the Brenier–Strassen theorem in a clean way, and the relaxed martingale transport results in Section 5—especially the martingale Benamou–Brenier interpolation with the explicit dual (5.16) and the Gibbs-type structure (5.24)—are new and worth knowing. The paper also gives a genuine necessity example (Example 2.10) for condition (C), which is more than many papers do.\n\nNow the soft spots. The proof of Theorem 2.5, which is the load-bearing duality step underlying Theorem 1.2(ii), does not actually prove the extension from jointly lsc costs to Borel-in-x, lsc-in-ρ costs. After a change of topology, it asserts that one can follow [11, Proof of Theorem 3.1] line by line. The cited result is for jointly lsc costs, and the manuscript explicitly states that Theorem 1.2 is 'slightly stronger' to include EOT. That stronger regime may well be true—the lsc of the functional is established via [22, Prop. A.12(d)]—but the minimax argument in [11] might rely on joint lsc in a way that is not captured by the one-sentence delegation. This is not a demonstrated flaw; it is an omitted proof. Similarly, Theorem 2.15, on relaxed WOT for non-convex costs, is stated without proof, and the Section 5 applications rely on it. A serious referee will need to see those arguments or explicit references to where they appear.\n\nThe main theorem itself is well-supported by the included arguments for dual attainment under (B) and (C) and complementary slackness; those parts are detailed and convincing. The self-citations to [11] and [22] are fine—they are published results with independent grounding, not circular restatements. The paper is honest about where assumptions are needed and where dual attainment fails (e.g., pure martingale costs).\n\nWho is this for? Anyone working on weak transport, martingale optimal transport, or entropic regularisation. It deserves a serious referee: the claims are important and the gaps look fillable, but the manuscript as posted is not ready as is. I would send it to peer review and ask for a real proof (or precise citation) for the Borel-in-x duality step and for Theorem 2.15.","headline":"A genuine and largely convincing generalization of the fundamental theorem to weak costs, with real new applications, but the key duality step is delegated to prior work with a 'line by line' claim that may not cover the Borel-in-x setting, and the relaxed non-convex theorem is stated without proof.","tokens_in":42958,"tokens_out":1828,"would_cite":true,"duration_ms":19916,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","49N15","60G42","28A33"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves a fundamental theorem for weak optimal transport: convex-in-measure costs admit optimal plans, strong duality, and, under two regularity conditions, dual attainment and complementary slackness.","keywords":["weak optimal transport","strong duality","dual attainment","complementary slackness","barycentric transport","entropic optimal transport","martingale optimal transport","C-transform"],"falsifier":"Take a two-point version of the problem, $X=Y=\\{0,1\\}$, with $\\mu=\\nu$ uniform, and a convex lsc cost satisfying (B) and (C), for example $C(x,\\rho)=|x-\\operatorname{mean}(\\rho)|^2+\\varepsilon(\\rho(1)\\log \\rho(1)+\\rho(0)\\log \\rho(0))$. Compute $WT_C$ by enumerating the one-dimensional coupling polytope and $D_C$ by a convex one-dimensional search over $g(0),g(1)$, and check equality. A single finite example where the min differs from the sup would refute the theorem; independently, searching for an optimal $(\\pi,(f,g))$ pair that violates $C(x,\\pi_x)=f(x)+\\pi_x(g)$ would refute the complementarity criterion.","tokens_in":41964,"feed_emoji":"🚚","tokens_out":11118,"duration_ms":95521,"temperature":0.7,"pith_summary":"Classical optimal transport has a central structure theorem: the cheapest way to move one probability measure to another equals the best dual certificate, and optimal plans are exactly the ones satisfying a pointwise equality. This paper extends that theorem to weak optimal transport, where the cost of sending a point $x$ is a nonlinear function $C(x,\\rho)$ of the entire conditional distribution $\\rho$ it is sent to, not just of the destination point. The main result states that for costs convex and lower semicontinuous in the measure argument, an optimal plan always exists and the primal value equals the dual value; under two mild boundedness and continuity conditions the dual optimum is attained, and a plan with a dual pair is jointly optimal exactly when $C(x,\\pi_x)=f(x)+\\pi_x(g)$ holds almost surely. Because entropic regularization, barycentric transport, and martingale-type problems are all weak transport problems in disguise, the theorem gives a single route to results that previously needed separate arguments. The consequence is that the classical duality toolkit now applies to a broad class of nonlinear transportation problems.","feed_headline":"Nonlinear transport costs now have a fundamental theorem","feed_subtitle":"Convex-in-measure costs get existence, strong duality, dual attainment, and a slackness test.","key_machinery":"The load-bearing object is the $C$-transform, $g^C(x)=\\inf_{\\rho\\in\\mathcal P_p(Y)}\\{C(x,\\rho)-\\rho(g)\\}$, the nonlinear analogue of the classical $c$-transform. Lemma 2.3 shows any admissible dual pair $(f,g)$ can be replaced by $(g^C,g)$, so the dual is a one-function maximization $\\sup_g \\mu(g^C)+\\nu(g)$. Primal attainment and strong duality come from lower semicontinuity of $\\pi\\mapsto\\int C(x,\\pi_x)\\,\\mu(dx)$ under the adapted-weak topology; dual attainment is proved by Komlós-style convex combinations together with uniform integrability supplied by condition (B), while condition (C) passes admissibility to the limit. Complementary slackness then characterizes joint optimality. In the applications, the same transform is computed explicitly: it becomes the infimal convolution $\\vartheta\\square\\psi$ for barycentric costs, the relative-entropy potential for entropic costs, and $\\vartheta\\square g^C$ or $(\\psi^*\\star\\check\\gamma)^*$ for relaxed martingale costs.","core_discovery":"The paper's central claim is Theorem 1.2: for a measurable cost $C:X\\times \\mathcal P_p(Y)\\to[0,\\infty]$ that is convex and lower semicontinuous in the measure argument, the weak transport value $WT_C(\\mu,\\nu)=\\inf_{\\pi}\\int C(x,\\pi_x)\\,\\mu(dx)$ is attained and equals the dual value $D_C(\\mu,\\nu)=\\sup\\{\\mu(f)+\\nu(g): f(x)+\\rho(g)\\le C(x,\\rho)\\}$. Under conditions (B) and (C) the dual supremum is attained, and a coupling $\\pi$ together with an admissible pair $(f,g)$ is optimal if and only if $C(x,\\pi_x)=f(x)+\\pi_x(g)$ holds $\\mu$-almost surely. The proof reduces duality to the $C$-transform $g^C(x)=\\inf_\\rho (C(x,\\rho)-\\rho(g))$, which replaces the two-variable dual constraint by a single-function formula. The paper applies this theorem to barycentric costs $\\vartheta(x-\\operatorname{mean}(\\pi_x))$, to entropic transport, and to relaxed martingale transport, deriving uniqueness of optimal barycenters, the Gibbs form of entropic optimizers, and dual attainment for problems where classical martingale duality fails.","pith_inferences":["Beyond the paper's claims, the $C$-transform computation is likely to be the first move in any future weak transport application: the theorem reduces the whole dual to $\\sup_g \\mu(g^C)+\\nu(g)$, so tractability of a problem is essentially the tractability of one infimum over measures.","A natural testable extension is to weaken condition (B) to polynomial growth without the entropy term; the proof's uniform-integrability step would then need a different super-coercivity argument, and the theorem's boundary might move.","The relaxed lifted formulation suggests a quantitative measure of non-convexity: the gap $WT_C(\\mu,\\nu)-WT_{\\bar C}(\\mu,\\nu)$ between a non-convex cost and its convex hull could be studied as a function of the atom sizes of $\\mu$, with the paper's equality cases marking when the gap vanishes.","In the financial reading of the convex Kantorovich–Rubinstein corollary, the maximum locked-in arbitrage under trading restrictions is exactly the weak transport value; one could extend the formula to multi-step strategies, where the barycentric cost would involve conditional expectations at intermediate times."],"forward_implications":["Every weak transport problem with convex lower semicontinuous cost has an optimal plan and satisfies strong duality, so existence and dual certificates are available without compactness of the state space.","When (B) and (C) hold, the dual problem is attained, meaning optimality of a plan can be certified by a pair of potentials and checked through the pointwise equality $C(x,\\pi_x)=f(x)+\\pi_x(g)$.","For barycentric costs of the form $\\vartheta(x-\\operatorname{mean}(\\pi_x))$, strictly convex $\\vartheta$ yields a unique optimal barycenter and a Monge-type transport map, extending Strassen's theorem to a quantitative projection of $\\mu$ onto the convex-order sublevel set of $\\nu$.","For entropic optimal transport, the theorem recovers the Gibbs structure $d\\pi/d(\\mu\\otimes\\nu)=\\exp((f+g-c)/\\varepsilon)$ directly from complementary slackness, and the same route works for general convex regularizers.","For martingale-type costs, where classical dual attainment can fail, the paper's relaxed formulation yields dual attainment and uniqueness, with optimizers built from a Bass-martingale kernel or a Gibbs-type density."],"supporting_citations":[{"why":"Introduces weak transport costs and proves Kantorovich duality under additional regularity; this paper extends that duality.","marker":"[52]"},{"why":"Provides the existence–duality–cyclical-monotonicity framework for weak transport that the proof follows line by line.","marker":"[11]"},{"why":"Extends weak transport existence and duality to Polish spaces via the adapted weak topology, the starting level of generality here.","marker":"[12]"},{"why":"States the classical fundamental theorem of optimal transport that Theorem 1.2 generalizes.","marker":"[6]"},{"why":"Derives the convex Kantorovich–Rubinstein formula on compact spaces, recovered here as an application.","marker":"[5]"},{"why":"Establishes the Brenier–Strassen theorem for squared barycentric costs, recovered and extended by Theorem 3.1.","marker":"[49]"},{"why":"Strassen's theorem on martingale couplings is recovered as the zero-cost case of the barycentric theorem.","marker":"[79]"},{"why":"Supplies the lifted transport plans and non-convex relaxation used in Theorem 2.15 and in the relaxed martingale sections.","marker":"[1]"},{"why":"Its structure theorem for costs regularized by general convex functions is recovered in Section 4.2.","marker":"[67]"}],"fun_headline_variants":["Weak optimal transport gets its fundamental theorem","Fundamental theorem for nonlinear transport costs","Convex-in-measure costs: duality and slackness","Master duality for weak optimal transport","Weak transport duality: fundamental theorem"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole structure rests on assuming the cost is convex and lower semicontinuous in the measure argument; for dual attainment, it also assumes the cost is bounded above by an integrable envelope plus an entropy term and is continuous under truncations, assumptions that are not consequences of convexity.","fun_headline_variants_meta":{"raw":{"variants":["Weak optimal transport gets its fundamental theorem","Fundamental theorem for nonlinear transport costs","Convex-in-measure costs: duality and slackness","Master duality for weak optimal transport","Weak transport duality: fundamental theorem"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000419,"raw_usage":{"total_tokens":2165,"prompt_tokens":959,"completion_tokens":1206,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":1143}},"tokens_in":575,"tokens_out":1206,"duration_ms":8570,"temperature":1.0,"reasoning_tokens":1143,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:32:39.644630+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a two-point version of the problem, $X=Y=\\{0,1\\}$, with $\\mu=\\nu$ uniform, and a convex lsc cost satisfying (B) and (C), for example $C(x,\\rho)=|x-\\operatorname{mean}(\\rho)|^2+\\varepsilon(\\rho(1)\\log \\rho(1)+\\rho(0)\\log \\rho(0))$. Compute $WT_C$ by enumerating the one-dimensional coupling polytope and $D_C$ by a convex one-dimensional search over $g(0),g(1)$, and check equality. A single finite example where the min differs from the sup would refute the theorem; independently, searching for an optimal $(\\pi,(f,g))$ pair that violates $C(x,\\pi_x)=f(x)+\\pi_x(g)$ would refute the complementarity criterion.","supporting_citations":[{"cited_title":"Gozlan, C","cited_arxiv_id":null,"evidence_quote":"Introduces weak transport costs and proves Kantorovich duality under additional regularity; this paper extends that duality."},{"cited_title":"Backhoff-Veraguas, M","cited_arxiv_id":null,"evidence_quote":"Provides the existence–duality–cyclical-monotonicity framework for weak transport that the proof follows line by line."},{"cited_title":"Backhoff-Veraguas, M","cited_arxiv_id":null,"evidence_quote":"Extends weak transport existence and duality to Polish spaces via the adapted weak topology, the starting level of generality here."},{"cited_title":"Gozlan and N","cited_arxiv_id":null,"evidence_quote":"Establishes the Brenier–Strassen theorem for squared barycentric costs, recovered and extended by Theorem 3.1."},{"cited_title":"Strassen","cited_arxiv_id":null,"evidence_quote":"Strassen's theorem on martingale couplings is recovered as the zero-cost case of the barycentric theorem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Its structure theorem for costs regularized by general convex functions is recovered in Section 4.2."}],"review_version":1}