Metric laws for finite boundary measures

Boundary transport separates finite positive measures on the strict birth–death half-plane and satisfies the sharp triangle inequality. Finite diagonal p-persistence makes its extended distance finite. Its reservoir comparison also has the literal coupling L^p-norm form, including probability scaling.

Measures, distance and common capacity

Write H = {(b,d) ∈ ℝ² : b < d} and δ(b,d) = (d−b)/2. The collapsed space X = H ⊔ {*} has metric

ρ(x,y) = min(‖x−y‖∞, δ(x)+δ(y)), ρ(x,*) = δ(x).

The native boundary distance BT_p(μ,ν) is the reciprocal-p root of the infimum of full ground costs over all boundary plans with interior marginals μ,ν. The diagonal marginals are free. For finite original measures and finite R ≥ μ(H)+ν(H), the reservoir measures

μ_R = μ + (R−μ(H))δ_*, ν_R = ν + (R−ν(H))δ_*

have common mass R; interior measures are pushed into X. The proved all-competitor comparison identifies the native powered infimum with the infimum of ∫ρ^p dγ over every measure γ with these reservoir marginals. Original masses may be unequal or zero. No Euclidean moment along the diagonal or compact-support condition is imposed.

Literal coupling norm and normalization

For real p > 0, eLpNorm_dist_eq_collapsedPowerCost identifies mathlib's eLpNorm (fun z => dist z.1 z.2) (ENNReal.ofReal p) γ with (∫ρ^p dγ)^(1/p). Positive reciprocal powers preserve arbitrary infima, including empty competitor families and infinite values. eLpNorm_edist_eq_dist identifies the extended-metric norm used in the coupling formulas with the real-distance norm. Consequently reservoirDistance_eq_iInf_eLpNorm expresses the native reservoir value as the infimum of those actual norms over the two marginal equations.

For p ≥ 1 and the full finite capacity above, boundaryTransportDistance_eq_iInf_eLpNorm gives the same expression for BT_p. If also R > 0, boundaryTransportDistance_eq_normalized_iInf_eLpNorm proves

BT_p(μ,ν) = R^(1/p) inf { ‖ρ‖_(L^p(γ)) : γ₁=μ_R/R, γ₂=ν_R/R }.

The two normalized marginals are probability measures by normalizedReservoirPair_isProbability. The factor uses the reservoir's total mass, not the original mass of either input. These statements use mathlib's existing norm and literal measure-marginal equations; they do not install an independently maintained Wasserstein package.

Sharp triangle and separation

For every real p ≥ 1 and arbitrary finite original Borel measures, boundaryTransportDistance_triangle proves

BT_p(μ,κ) ≤ BT_p(μ,ν) + BT_p(ν,κ).

One capacity R = μ(H)+ν(H)+κ(H)+1 is valid for all three pairs. Actual reservoir minimizing plans exist by BoundaryTransportAttainment. A private construction on this fixed collapsed space glues their common middle marginal using mathlib's finite-measure conditional kernel and measure-kernel composition. The resulting endpoint plan is a genuine competitor. Applying the pointwise metric triangle and mathlib's eLpNorm_add_le to the glued measure gives coefficient one. A gluing law or minimizing plan is never a caller premise, and infinite optimal or competitor costs remain allowed.

boundaryTransportDistance_eq_zero_iff proves

BT_p(μ,ν) = 0 ↔ μ = ν.

An actual zero-cost reservoir optimizer has equal endpoints almost everywhere, so its marginals agree. Restricting those marginals to the interior gives equality of the original measures. Conversely an identity boundary plan has zero cost. Neither this result nor the extended triangle requires finite persistence moments. Finite p-persistence supplies finiteness through boundaryTransportDistance_lt_top; together with the existing self and symmetry laws these are the finite-distance metric laws.

Native consumers and capacity countercontrol

Two far unit atoms, each of diagonal distance one, have native powered cost two at p=2. Reservoir capacity two recovers that value. At capacity one, every feasible plan is concentrated on the far point pair, and its powered cost is four. Thus a capacity bound using only the maximum original mass does not give the boundary value. The fixture computes the whole wrong capacity infimum, rather than merely one expensive competitor.

Translated unit bars realize equality in the triangle inequality. Combining the metric laws with same-original-mass finite atomic approximation bounds the error in a distance by the sum of the two approximation errors. These consumers use the existing atomicMeasure constructor and its proved native finite matching compatibility.

Correspondence boundary

This argument follows the mathematical contract in #505, existing native comparison and attainment, and the current pinned mathlib sources. It is a fixed-space bridge, with no new general optimal-transport definitions. Identification with a separately maintained ordinary Wasserstein API and the authorized corrected-article comparison remain unresolved in the dependency and source-use map.