{"id":"f9a74869-4821-4697-984b-1662a6a4481e","arxiv_id":"2507.07200","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"In weak optimal transport, convex (or increasing convex) dual potentials are exactly characterized by the cost being decreasing in (increasing) convex order, with attainment under mild regularity.","lead":"A mathematics paper finds the exact condition on a weak optimal transport cost function that lets the dual optimization problem be solved over convex functions only: the cost must decrease when its second argument moves upward in convex order. This one condition explains and connects several known dual simplifications in martingale transport, mechanism design, and classical optimal transport.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the central Theorem 1.1 proof is sound; its most delicate imported ingredient, Lemma 2.1(i)/(vi), is a standard convex-analysis fact and does not appear to create a real gap.","rationale":"The reader correctly identifies Lemma 2.1(i)/(vi) as the most load-bearing step in the proof of Theorem 1.1: Lemma 2.3 converts this convex-hull representation into the equality psi^C = (conv psi)^C, which is the entire mechanism behind restricting the dual to convex potentials. I examined this step carefully rather than treating it as a black box. The representation is a classical consequence of the biconjugate theorem for lsc functions bounded below, and the monotone approximation lemma follows from the fact that the pointwise decreasing limit of finite convex functions is itself finite convex and hence continuous. I therefore do not see a correctness gap. I also checked the converse direction of Theorem 1.1, which uses only delta marginals and is valid. The unproved statements are Theorem 3.2 and the stable-cone extensions; both are explicitly described as modifications of earlier proofs, and neither is needed for the central equivalence in Theorem 1.1. Thus the reader's ACCEPT verdict remains appropriate, and I would not adjust it. The only recommendation is to add short proofs or precise references for Lemma 2.1(i)/(vi) and Theorem 3.2 to reduce reliance on 'obvious' or 'can be modified' assertions.","tokens_in":13130,"tokens_out":37346,"duration_ms":448323,"concrete_test":"Independently verify Lemma 2.1(i)/(vi) in the sharp class: take a non-convex, bounded-below, continuous psi with p-growth, for example psi(y) = -exp(-|y|^2) with p = 1, and compute conv psi(y) and inf_{mean rho = y} rho(psi) for several y including large |y|; if they match and conv(psi + |.|^p/n) decreases to conv psi, the mechanism is secure. A full analytic check via Fenchel-Moreau and Caratheodory would settle the identity for all lsc functions bounded below.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After checking the proof chain for the central claim, I find no load-bearing objection. Theorem 1.1 rests on Theorem 2.4, Lemma 2.3, and Lemma 2.1. The most delicate imported piece is Lemma 2.1(i)/(vi): the identity conv psi(y) = inf{rho(psi) : delta_y <=_c rho} and the monotone approximation conv(psi + |.|^p/n) down to conv psi. These are classical for lsc functions bounded below, and Lemma 2.1(vi) follows because the decreasing limit of finite convex functions is finite convex, hence continuous, and is majorized by psi, so it cannot exceed conv psi. I could not construct a Cb,p function for which either identity fails. The converse direction of Theorem 1.1 is a clean delta-marginal argument and is sound. The main unproved extensions, Theorem 3.2 and the stable-cone results in Section 6, are presented as modifications of the main proof; they support applications rather than the central equivalence, so they do not change the verdict.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the weak optimal transport duality and asks when the dual maximization can be restricted to convex (or increasing convex) potentials. The main result, Theorem 1.1, states that for a lower semicontinuous cost C that is convex in its second argument, equality between the primal weak transport problem and the dual restricted to convex potentials holds for all marginals if and only if C is nonincreasing in convex order in its second argument. The proof combines the general weak transport duality theorem with an approximation argument using the convex hull and a measurable selection. The paper also proves the increasing-convex analogue, establishes dual attainment under additional boundedness and continuity conditions, and derives several known results as applications: barycentric transport, Strassen's theorem, the martingale Benamou–Brenier formulation, the multiple-good monopolist problem, and classical quadratic transport. A general stable-cone framework is proposed in the final section.","tokens_in":13226,"tokens_out":11037,"duration_ms":118499,"significance":"If the results are correct, Theorem 1.1 provides a sharp and unifying explanation for a number of disparate dual-restriction results in weak optimal transport, martingale transport, mechanism design, and classical Brenier duality. The proof of the main equivalence is clean and self-contained apart from two imported ingredients: the general weak transport duality theorem of Backhoff-Veraguas, Beiglböck, and Pammer, and standard convex-hull representations. The paper is honest about the conditional nature of the stable-cone framework and about the delicacy of dual attainment in unbounded settings, for example in Remark 5.1. No free parameters are introduced and the converse direction of Theorem 1.1 gives a falsifiable characterization, which strengthens confidence in the result.","major_comments":[{"comment":"Theorem 3.2 is stated without proof: the text says only that the proof of Theorem 1.1 'can be modified' to accommodate costs bounded below by -(a_l(x) + rho(b_l)). This theorem is then used in §5.1.2 and §5.1.3 to recover the martingale Benamou-Brenier duality and classical quadratic transport duality, so the missing argument is load-bearing for the applications claimed in the abstract. In particular, when the cost is not nonnegative and the potential class is enlarged to Cb,p + b_l, the C-conjugate may take the value -infinity and the finiteness arguments in Lemma 2.3 need to be reworked. Please supply a full proof or a precise reference for this extension.","section":"§3, Theorem 3.2"},{"comment":"Theorem 6.1 is stated as a theorem but its proof is not given; the text says only that the proof of Theorem 1.1 'naturally extends to the setting of stable cones.' The result depends crucially on the dual-representation assumption (5), and Theorem 6.4 relies on Theorem 6.1. Since the authors themselves note that the validity of such representations for semi-stable cones on noncompact spaces is subtle (with [11] flawed and [10] giving a counterexample), the proof should be written out to make clear exactly which properties (1)-(5) are used and where compactness or additional assumptions are needed.","section":"§6, Theorem 6.1"},{"comment":"In Step 1 of the proof of Theorem 4.1, the representation conv psi(y) = inf{xi(psi) : delta_y <=_c xi, |supp xi| <= d+1} is asserted without proof. This identity is used to control conv_R psi and to prove its nu-integrability, and is therefore central to the attainment result. Please add a proof or a precise reference for this Caratheodory-type hull formula.","section":"§4, Eq. (14)"}],"minor_comments":[{"comment":"Please add a reference or proof for the convex-hull representation conv psi(y) = inf{rho(psi) : delta_y <=_c rho}; the current proof is only written for the increasing convex version (ii).","section":"§2, Lemma 2.1(i)"},{"comment":"Egorov's theorem is applied to conv_R psi, which is only known to be nu-integrable and may take the value +infinity; please clarify that one first restricts to a set of full rho-measure on which these functions are finite, so that the uniform-convergence argument is justified.","section":"§4, Step 2"},{"comment":"The lower bound used for the martingale Benamou-Brenier example appears to be off by a dimension constant: for rho = gamma one has MCov(rho,gamma) = 1 in d = 1 while (1/2)rho(|y|^2) = 1/2. The bound should read -MCov(rho,gamma) >= -1/2 rho(|y|^2) - d/2, and b_l should be chosen accordingly.","section":"§5.1.2"},{"comment":"There are several typos and typesetting issues: the running title contains 'TRANSPOR T' with a stray space, the notation chi_{x=mean(pi_x)} in Eq. (3) should be defined, and the measure e_rho is introduced with the unusual notation 'de_rho' in the proof of Lemma 2.3.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The central equivalence in Theorem 1.1 is sound in my reading, and the stress-test concern about Lemma 2.1 does not land: the hull representation and monotone approximation are standard convex-analysis facts. My recommendation of major_revision is driven by the unproved Theorem 3.2, which is used for several headline applications, and by the unproved stable-cone theorem in Section 6. Both are fixable, but they need to be written out before the paper appears in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean, genuinely unifying result. The main theorem — convex-order monotonicity of the cost is equivalent to restricting the WOT dual to convex potentials — is stated in that generality for the first time, and the proof is largely sound. The converse direction via delta marginals is a nice argument. I agree with the reader's ACCEPT and soundness 8.\n\nWhat the paper does well: it takes a set of known dual restrictions (barycentric, martingale Benamou–Brenier, mechanism design, Strassen) and shows they all check the same simple condition. Lemma 2.3 is the heart: under monotonicity, psi^C = (conv psi)^C. The use of Lemma 2.1(i)/(vi) is legitimate; the stress-test concern about that imported identity does not land on reading the proof. I could not construct a counterexample either. The application section is a useful checklist for the condition.\n\nSoft spots, in order of softness: Theorem 3.2 (allowing negative costs) is asserted as a modification of earlier proofs without being written out. Same for the stable-cone Theorem 6.1, which relies on semi-stable cone condition (5) — the paper itself calls it 'somewhat artificial'. These are not flaws in the central argument, but they mean the extensions are more sketch than proof. The attainment section is dense and relies on the external [7, Theorem 2.2] plus conditions (B)–(C); it is coherent but technical. A referee should ask for Theorem 3.2 to be proved, not just described.\n\nI'd send this to a serious referee. The core result is important enough and the proof detailed enough. It will be a well-cited unification paper. My only wish is that the sketched parts were fuller, but that's a revision, not a rejection.","headline":"A genuinely unifying, sharp condition for convex dual potentials in weak optimal transport; the core proof is sound and it deserves a serious referee, with some extensions sketched too briefly.","tokens_in":13853,"tokens_out":2008,"would_cite":true,"duration_ms":21748,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","60E15"],"pacs":[],"model":"deepseek-v4-flash","headline":"One order condition decides when weak transport duals can be convex","keywords":["weak optimal transport","convex dual potentials","convex order","increasing convex order","Strassen's theorem","martingale Benamou-Brenier","mechanism design","stable cones"],"falsifier":"Compute the primal weak transport value and the convex-restricted dual value for $C(x,\\rho)=\\int y^4\\,d\\rho(y)$ with $\\mu=\\delta_0$ and $\\nu=(\\delta_{-1}+\\delta_1)/2$; they are $1$ and $0$, a gap that the converse direction demands for a cost that is not convex-order decreasing, so an analogous computation producing a gap for a cost that is convex-order decreasing would refute the forward direction.","tokens_in":12814,"feed_emoji":"📐","tokens_out":15255,"duration_ms":162098,"temperature":0.7,"pith_summary":"This paper identifies the exact condition under which the dual of a weak optimal transport problem can be maximized over convex potentials: the cost function must be nonincreasing in convex order in its second argument. The main theorem, Theorem 1.1, proves that this monotonicity is necessary and sufficient for the restricted duality to hold for all marginals, and a parallel theorem does the same for increasing convex potentials with the increasing convex order. The paper then shows that, under a boundedness condition and a mild continuity condition, the restricted supremum is attained by a $\\nu$-integrable convex function. Applications recover several results that previously looked separate: Strassen's martingale theorem, barycentric transport, the martingale Benamou--Brenier problem, the multiple-good monopolist problem, and classical quadratic-cost Brenier duality. If the criterion is correct, checking one monotonicity property of the cost tells you whether the dual simplifies to convex potentials and whether an optimal convex potential exists.","feed_headline":"One cost condition decides when dual potentials stay convex","feed_subtitle":"Convex-order monotonicity is necessary and sufficient for the weak transport dual to reduce to convex potentials.","key_machinery":"The machine that carries the argument is the convex-hull representation $\\operatorname{conv}\\psi(y)=\\inf\\{\\rho(\\psi):\\delta_y\\preceq_c\\rho,\\ \\rho\\in P_p(\\mathbb{R}^d)\\}$ from Lemma 2.1(i), together with its increasing-convex analogue and the monotone approximation $\\operatorname{conv}(\\psi+|\\cdot|^p/n)\\downarrow\\operatorname{conv}\\psi$. This identity converts convex-order monotonicity of the cost into an equality of $C$-conjugates, $\\psi^C=(\\operatorname{conv}\\psi)^C$: if replacing $\\rho$ by a measure above it in convex order only lowers the cost, then the cheapest measure for $\\psi$ is no cheaper for its convex hull. That equality is precisely what lets the unrestricted dual over all test functions be restricted to convex ones, and the converse half of the theorem runs the same machinery backwards to force monotonicity of $C$. Section 6 generalizes the hull representation into a stability condition on cones, so the same argument works for any order generated by a stable cone.","core_discovery":"On the paper's own terms, the central discovery is that convexity of dual potentials in weak optimal transport is governed by convex-order monotonicity of the cost. For a lower semicontinuous cost $C:X\\times P_p(\\mathbb{R}^d)\\to[0,\\infty]$ that is convex in its second argument, the paper proves the identity $\\inf_{\\pi\\in\\Pi(\\mu,\\nu)}\\int C(x,\\pi_x)\\,d\\mu(x)=\\sup_{\\psi\\in C_{b,p}(\\mathbb{R}^d),\\ \\psi\\text{ convex}}(\\mu(\\psi^C)-\\nu(\\psi))$ holds for every pair of marginals if and only if $C$ is $\\preceq_c$-decreasing in its second argument. The same statement with increasing convex functions and the increasing convex order $\\preceq_{icx}$ is Theorem 3.1. Under boundedness condition (B) and continuity condition (C), Theorem 4.1 upgrades the equality to attainment: the supremum is reached by a $\\nu$-integrable convex function $\\psi_{\\mathrm{opt}}:\\mathbb{R}^d\\to(-\\infty,\\infty]$. Section 6 abstracts the mechanism to stable cones of functions, yielding the same restriction for any order whose cone satisfies a hull-representation condition.","pith_inferences":["A screening rule follows that the paper leaves implicit: one can certify the convex-potential simplification by checking $C(x,\\cdot)$ against convex-order comparisons on a finite set of test measures, without solving the transport problem itself.","The stable-cone framework suggests the same reduction should hold for other orders, such as directionally convex or $k$-convex functions, whenever the associated hull admits an infimum-over-measures representation; the paper only instantiates convex and increasing convex.","In the martingale Benamou--Brenier example the boundedness condition (B) fails, so attainment is delicate; an extension to that setting would need a problem-specific replacement for (B) rather than a direct application of Theorem 4.1.","The converse direction could be used diagnostically: if a transport-type model provably has convex optimal dual potentials for all marginals, then its cost must be convex-order decreasing, which may reveal hidden monotonicity in models not currently formulated as weak transport."],"forward_implications":["Applying the theorem to the order-indicator cost $C(x,\\rho)=0$ if $\\delta_x\\preceq_c\\rho$ and $\\infty$ otherwise gives Strassen's theorem: $\\mu\\preceq_c\\nu$ exactly when a martingale coupling exists.","For barycentric costs $C(x,\\rho)=\\theta(x-\\operatorname{mean}\\rho)$ with convex $\\theta$, the cost is automatically $\\preceq_c$-decreasing, reproducing the known restriction of the dual to convex functions and, for quadratic $\\theta$, the Brenier--Strassen mixture result.","The increasing-convex version recovers the multiple-good monopolist duality and the increasing-convex Kantorovich--Rubinstein formula, with dual attainment by increasing convex potentials under the regularity conditions.","For classical transport costs $C(x,\\rho)=\\int c(x,y)\\,d\\rho(y)$, convex-order monotonicity is equivalent to concavity of $c$ in its second argument, so the quadratic-cost duality in the paper is a direct instance of the same criterion.","Under conditions (B) and (C), the convex-restricted dual supremum is attained by a $\\nu$-integrable convex function, meaning the simplification is not only an equality of values but an achieved optimum."],"supporting_citations":[{"why":"Supplies the general weak transport duality theorem (Theorem 2.4) from which the paper starts.","marker":"[3]"},{"why":"Introduced weak optimal transport and the barycentric example; the paper extends its observation to a criterion.","marker":"[14]"},{"why":"The barycentric transport restriction result that Theorem 1.1 recovers.","marker":"[12]"},{"why":"The martingale Benamou--Brenier dual formulation recovered via Theorem 3.2.","marker":"[5]"},{"why":"The multiple-good monopolist dual restriction recovered via Theorem 3.1.","marker":"[6]"},{"why":"Provides the duality and attainment framework used in Theorem 4.1 for existence of optimal convex potentials.","marker":"[7]"},{"why":"Strassen's martingale existence theorem proved as a corollary of the convex-order criterion.","marker":"[18]"},{"why":"Classical optimal transport duality recovered for the quadratic cost through the same monotonicity condition.","marker":"[19]"}],"fun_headline_variants":["Weak transport duals convex iff cost is convex-order monotone","One cost condition: convex dual potentials in weak transport","Convex-order monotonicity pins down weak dual convexity","Sharp condition for convex potentials in weak optimal transport","Weak transport: convex duals characterized by cost monotonicity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument relies on the classical representation of the convex hull of a test function as an infimum of expectations taken over probability measures that dominate the Dirac mass at that point in convex order, and on the monotone approximation that realizes that hull; if this representation failed for some admissible function, the equality of conjugates that powers the restriction theorem would break.","fun_headline_variants_meta":{"raw":{"variants":["Weak transport duals convex iff cost is convex-order monotone","One cost condition: convex dual potentials in weak transport","Convex-order monotonicity pins down weak dual convexity","Sharp condition for convex potentials in weak optimal transport","Weak transport: convex duals characterized by cost monotonicity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1432,"prompt_tokens":911,"completion_tokens":521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":440}},"tokens_in":527,"tokens_out":521,"duration_ms":56962,"temperature":1.0,"reasoning_tokens":440,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:47:15.704080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the primal weak transport value and the convex-restricted dual value for $C(x,\\rho)=\\int y^4\\,d\\rho(y)$ with $\\mu=\\delta_0$ and $\\nu=(\\delta_{-1}+\\delta_1)/2$; they are $1$ and $0$, a gap that the converse direction demands for a cost that is not convex-order decreasing, so an analogous computation producing a gap for a cost that is convex-order decreasing would refute the forward direction.","supporting_citations":[{"cited_title":"Backhoff-Veraguas, M","cited_arxiv_id":null,"evidence_quote":"Supplies the general weak transport duality theorem (Theorem 2.4) from which the paper starts."},{"cited_title":"Kantorovich duality for general transport costs and applications","cited_arxiv_id":null,"evidence_quote":"Introduced weak optimal transport and the barycentric example; the paper extends its observation to a criterion."},{"cited_title":"On a mixture of Brenier and Strassen theorems","cited_arxiv_id":null,"evidence_quote":"The barycentric transport restriction result that Theorem 1.1 recovers."},{"cited_title":"Existence of Bass martingales and the martingale Benamou −Brenier problem in Rd, 2025","cited_arxiv_id":null,"evidence_quote":"The martingale Benamou--Brenier dual formulation recovered via Theorem 3.2."},{"cited_title":"Applications of weak transport theory","cited_arxiv_id":null,"evidence_quote":"The multiple-good monopolist dual restriction recovered via Theorem 3.1."},{"cited_title":"Strassen","cited_arxiv_id":null,"evidence_quote":"Strassen's martingale existence theorem proved as a corollary of the convex-order criterion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Classical optimal transport duality recovered for the quadratic cost through the same monotonicity condition."}],"review_version":1}