OctoInformBeta

OpenAI preprint, September 2026 · §5. The marginal and exclusion of every Monge optimizer

§5 (proof of Theorem 1.1): an optimal coupling and the selection obstruction: Bounded certificates and disjoint branch images yield smooth Coulomb counterexamples with finite Monge approximation

Branch couplings and the obstruction to Monge optimality.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

The question is whether a common marginal on ℝ3\mathbb R^3 can have an optimal Coulomb coupling without an optimal transport given by two maps. A triple configuration is a point (x,y,z)(x,y,z) of (ℝ3)3(\mathbb R^3)^3✦; its coordinate projections are p1(x,y,z)=xp_1(x,y,z)=x✦, p2(x,y,z)=yp_2(x,y,z)=y✦ and p3(x,y,z)=zp_3(x,y,z)=z✦. The inverse distance is |x−y|−1|x-y|^{-1}, with value +∞+\infty at x=yx=y.✦ The extended Coulomb cost in (1) takes value +∞+\infty at every collision; away from collisions, the same reciprocal sum also defines a real scalar cost. c(x,y,z)=1|x−y|+1|x−z|+1|y−z|.c(x,y,z)=\frac1{|x-y|}+\frac1{|x-z|}+\frac1{|y-z|}.(1)✦

For a probability measure μ\mu on ℝ3\mathbb R^3, a three-marginal self-coupling π\pi of μ\mu is a probability measure on (ℝ3)3(\mathbb R^3)^3 whose three coordinate pushforwards equal μ\mu.✦ For a probability measure μ\mu on ℝ3\mathbb R^3, Γ(μ)\Gamma(\mu) is the set of its three-marginal self-couplings. A measure-preserving map for μ\mu is a measurable map T:ℝ3→ℝ3T:\mathbb R^3\to\mathbb R^3 with T#μ=μT_\#\mu=\mu.✦ The Kantorovich infimum E*(μ)E_*(\mu), the Monge transport cost GμG_\mu, and the Monge infimum EM(μ)E_{\mathrm M}(\mu) are defined in (2), where the last infimum is over measure-preserving pairs. E*(μ)=infπ∈Γ(μ)∫cdπ,Gμ(T2,T3)=∫c(x,T2(x),T3(x))dμ(x),EM(μ)=infT2,T3Gμ(T2,T3).\begin{aligned} E_*(\mu)&=\inf_{\pi\in\Gamma(\mu)}\int c\,d\pi,\\ G_\mu(T_2,T_3)&=\int c(x,T_2(x),T_3(x))\,d\mu(x),\\ E_{\mathrm M}(\mu)&=\inf_{T_2,T_3}G_\mu(T_2,T_3). \end{aligned}(2) Kantorovich attainment means that a coupling realizes E*(μ)E_*(\mu).✦ Monge nonattainment means that every measure-preserving pair has cost strictly above E*(μ)E_*(\mu).✦ Finite-cost Monge approximation means that, for each ε>0\varepsilon>0, a measure-preserving pair has finite cost at most E*(μ)+εE_*(\mu)+\varepsilon.✦

The density measure of a real function ρ\rho is [ρ]+dx[\rho]_+\,dx, where [a]+=max⁡{a,0}[a]_+=\max\{a,0\}; for nonnegative densities it is denoted μ=ρdx\mu=\rho\,dx.✦ A smooth compact probability density is a nonnegative function ρ\rho such that both ρ\rho and ρ\sqrt\rho are smooth and compactly supported, with ∫ρdx=1\int\rho\,dx=1.✦ A Coulomb counterexample density ρ\rho on ℝ3\mathbb R^3 is a smooth compact probability density whose measure μ=ρdx\mu=\rho\,dx is a probability measure with finite Kantorovich value, Kantorovich attainment and Monge nonattainment.✦ A density ρ\rho on ℝ3\mathbb R^3 satisfies the full Coulomb conclusion when it is a Coulomb counterexample density and its measure μ=ρdx\mu=\rho\,dx satisfies EM(μ)=E*(μ)E_{\mathrm M}(\mu)=E_*(\mu) and admits finite-cost Monge approximation.✦

Branches and Certificates

For maps f,g,h:ℝ3→ℝ3f,g,h:\mathbb R^3\to\mathbb R^3, the parametrized triple graph Ff,g,hF_{f,g,h} is the map from ℝ3\mathbb R^3 to (ℝ3)3(\mathbb R^3)^3 defined by Ff,g,h(x)=(f(x),g(x),h(x))F_{f,g,h}(x)=(f(x),g(x),h(x)).✦ Given a probability measure ν\nu and four measurable branches H1,…,H4H_1,\ldots,H_4, define the five-component marginal μ\mu and the cyclic branch coupling π0\pi_0 by (3) and (4). μ=13ν+16∑j=14(Hj)#ν.\mu=\frac13\nu+\frac16\sum_{j=1}^4(H_j)_\#\nu.(3)✦ π0=16((Fid,H1,H2)#ν+(FH1,H2,id)#ν+(FH2,id,H1)#ν+(Fid,H3,H4)#ν+(FH3,H4,id)#ν+(FH4,id,H3)#ν).\begin{aligned} \pi_0=\frac16\bigl(& (F_{\mathrm{id},H_1,H_2})_\#\nu +(F_{H_1,H_2,\mathrm{id}})_\#\nu +(F_{H_2,\mathrm{id},H_1})_\#\nu\\ &+(F_{\mathrm{id},H_3,H_4})_\#\nu +(F_{H_3,H_4,\mathrm{id}})_\#\nu +(F_{H_4,\mathrm{id},H_3})_\#\nu\bigr). \end{aligned}(4)✦

For local homeomorphisms Xk,ZkX_k,Z_k of a space EE, with k=1,2k=1,2, their common central parametrization consists of the inverse branch YkY_k and the reparametrized outer branch WkW_k, defined on the points where the successive maps in (5) are defined. Yk=Xk−1,Wk=Zk∘Xk−1,(H1,H2,H3,H4)=(Y1,W1,Y2,W2).Y_k=X_k^{-1},\qquad W_k=Z_k\circ X_k^{-1},\qquad (H_1,H_2,H_3,H_4)=(Y_1,W_1,Y_2,W_2).(5) For a central set BB and branches HjH_j, the component regions are B,H1(B),…,H4(B)B,H_1(B),\ldots,H_4(B).✦

A bounded real certificate on a measurable set U⊆ℝ3U\subseteq\mathbb R^3 is a measurable function u:ℝ3→ℝu:\mathbb R^3\to\mathbb R, bounded by |u|≤M|u|\leq M on UU for some M≥0M\geq0, satisfying (6). The contact relation 𝒞u\mathcal C_u of a bounded real certificate uu is the relation on triples (x,y,z)∈(ℝ3)3(x,y,z)\in(\mathbb R^3)^3 defined by the equality condition in (7); the positive part is regarded as a nonnegative extended real. [u(x)+u(y)+u(z)]+≤c(x,y,z),(x,y,z)∈U3.[u(x)+u(y)+u(z)]_+\leq c(x,y,z),\qquad (x,y,z)\in U^3.(6) 𝒞u(x,y,z)⇔[u(x)+u(y)+u(z)]+=c(x,y,z).\mathcal C_u(x,y,z)\quad\Longleftrightarrow\quad [u(x)+u(y)+u(z)]_+=c(x,y,z).(7)

The Claim

Theorem 1. For every probability measure ν\nu and measurable branches H1,…,H4H_1,\ldots,H_4, the marginal and cyclic coupling of (3) and (4) satisfy (8). π0∈Γ(μ).\pi_0\in\Gamma(\mu).(8)✦ Suppose ν\nu is concentrated on B⊆UB\subseteq U, each Hj(B)⊆UH_j(B)\subseteq U, and uu is a bounded real certificate on UU. Assume 𝒞u(x,H1(x),H2(x))\mathcal C_u(x,H_1(x),H_2(x)) and 𝒞u(x,H3(x),H4(x))\mathcal C_u(x,H_3(x),H_4(x)) for ν\nu-almost every xx. Then (9) holds. ∫cdπ0=E*(μ)<+∞.\int c\,d\pi_0=E_*(\mu)<+\infty.(9)✦

For the local homeomorphisms in (5), suppose (x,y,z)(x,y,z) is a permutation of (Xk(s),s,Zk(s))(X_k(s),s,Z_k(s)), where x∈Bx\in B, ss belongs to the source of XkX_k, and both ss and Zk(s)Z_k(s) lie outside BB. Then (10) holds. y∈{Y1(x),W1(x),Y2(x),W2(x)}.y\in\{Y_1(x),W_1(x),Y_2(x),W_2(x)\}.(10)✦

Let μ\mu be a probability measure, ν≪μ\nu\ll\mu, and suppose x∈Bx\in B for ν\nu-almost every xx. Let 𝒞\mathcal C be a relation on triples and 𝒮\mathcal S a relation on pairs. Assume every optimal coupling is concentrated on 𝒞\mathcal C, and 𝒞(x,y,z)\mathcal C(x,y,z) with x∈Bx\in B implies 𝒮(x,y)\mathcal S(x,y). Suppose no measure-preserving map TT for μ\mu satisfies 𝒮(x,T(x))\mathcal S(x,T(x)) for ν\nu-almost every xx. Then (11) holds. Monge nonattainment holds for μ.\text{Monge nonattainment holds for }\mu.(11)✦

Suppose the hypotheses for (9) hold, BB is measurable, and μ=ρdx\mu=\rho\,dx for a smooth compact probability density ρ\rho. Assume the branches are injective on BB and send every measurable subset of BB to a measurable set. Suppose the five component regions are pairwise disjoint, and every contact triple (x,y,z)∈B×U×U(x,y,z)\in B\times U\times U satisfies y∈{H1(x),…,H4(x)}y\in\{H_1(x),\ldots,H_4(x)\}. Then (12) holds. ρ is a Coulomb counterexample density.\rho\text{ is a Coulomb counterexample density}.(12)✦

Let B=B(a,r)⊆ℝ3B=B(a,r)\subseteq\mathbb R^3 with r>0r>0, where aa is its center. Suppose four local homeomorphisms HjH_j have sources containing BB, are smooth on their sources, and have smooth inverses on their targets. Assume their five component regions are pairwise disjoint and contained in a measurable set UU. Let uu be a bounded real certificate on UU, with both branch triples in contact for every x∈Bx\in B. Suppose every contact triple in B×U×UB\times U\times U has its second entry among the four branches at its first entry. Then (13) holds. ∃ρ:ℝ3→ℝsuch that the full Coulomb conclusion holds for ρ.\exists\rho:\mathbb R^3\to\mathbb R \quad\text{such that the full Coulomb conclusion holds for }\rho.(13)✦

In the paper, the result reads: choosing either branch uniformly and permuting the entries of (x,Yk(x),Wk(x))(x,Y_k(x),W_k(x)), with xx distributed as ν\nu, gives a coupling π0\pi_0 with all marginals μ\mu and ∫cdπ0=3∫Uudμ=E*(μ)<∞\int c\,d\pi_0=3\int_Uu\,d\mu=E_*(\mu)<\infty by saturation of the certificate; every optimal coupling is concentrated on the contact set, and an optimal graph plan would force T2(x)∈{Y1(x),W1(x),Y2(x),W2(x)}T_2(x)\in\{Y_1(x),W_1(x),Y_2(x),W_2(x)\} for ν\nu-almost every central xx, contradicting Lemma 2.1, which excludes a measure-preserving map with this selection property.

What Is Checked

The assertions are stated over probability measures and measurable branches, arbitrary contact and selection relations, and smooth local homeomorphisms on a positive-radius ball, with the certificate, disjointness and selection hypotheses specified in the theorem.