Ordinary optimal transport and the finite-Radon boundary problem
The native boundary-plan infimum already equals the collapsed-space reservoir infimum for arbitrary finite inputs. Identifying this value with an existing ordinary Wasserstein API still needs a compatible optimal-transport package. The library proves native finite-measure attainment, sharp triangle and separation, and a literal mathlib coupling-norm formulation using the current Lean/mathlib pin. The independently maintained ordinary Wasserstein interface remains a separate dependency.
Existing boundary-transport results
Let H be the strict birth–death half-plane, δ(x) = (death − birth)/2, and
X = H ⊔ {*} with collapsed metric
ρ(x,y) = min(‖x−y‖∞, δ(x)+δ(y)), ρ(x,*) = δ(x).
For finite Borel measures μ,ν on H, choose finite R ≥ μ(H)+ν(H) and put
μ_R = μ + (R−μ(H))δ_*, with the interior measure pushed into X.
Both reservoir measures have total mass R. The topology on X is induced by
ρ; it is Polish and has the proved interior Borel identification.
For real p ≥ 1, the library proves equality between the infimum of full
powered boundary-plan costs and the infimum of ρ^p over all measures with
reservoir marginals μ_R,ν_R. This includes infinite competitor costs and
does not require atomic inputs. The capacity is independent of the chosen
admissible R; positive capacities admit probability normalization with the
factor R^(1/p). The basic plan-fiber lemmas interiorIncl_mem_image and
diagonalProjection_not_mem_image expose membership for the interior inclusion
and exclude diagonal projections. Self, symmetry, finite-distance bounds from
finite p-persistence, and the exact zero-target moment with its
[IsFiniteMeasure μ] variants (valid even when the moment is infinite) are
proved, along with the native finite atomic and indexed/coalesced bridges. The
metric-law bridge gives sharp triangle and
separation for every finite original measure, without a moment premise.
Finite moments make the extended distance finite.
The reservoir infimum has a native definition. The attainment theorem below constructs general finite-input minimizers for that value and the boundary value. Identification with a separately implemented ordinary Wasserstein distance remains open. See the exported contracts.
Imports and compatibility
BoundaryTransportGeometryCore and BoundaryTransportReverseCore are the canonical
imports for reusable geometry and the general reverse construction. The former keeps
the density and small-positive-boundary-distance results; the latter keeps the
arbitrary-point capacity-two competitor and exports Reverse.boundaryPlan_add and
Reverse.boundaryPlan_dirac_points_cost. The Dirac formula gives the cost of the
selected reverse plan for every real exponent, using the direct leg when distance is
less than or equal to the sum of boundary depths. It does not assert optimality.
The fixed far-point witnesses are in BoundaryTransportGeometryExamples and
BoundaryTransportReverseExamples. The old BoundaryTransportGeometry and
BoundaryTransportReverse imports remain available with their old declarations and
types; the Reverse entrance also retains transitive Geometry access. Ordinary
TDA.Persistence imports the Core modules only. Clients that relied on aggregate
imports for fixed far-point names should explicitly import the corresponding
Examples module or compatibility entrance. TDA.Persistence.Examples retains both
example modules and both old entrances.
Compatible API boundary
The library uses Lean 4.35.0 and mathlib
56a005855ea17e0b0958f73ec801ae140b0a1679.
That source has no IsCoupling, transportCost or wassersteinEDist
declarations. It does provide the following relevant ingredients:
| Existing mathlib API | Contribution to the missing argument |
|---|---|
ProbabilityMeasure.continuous_map |
Prescribed marginal equations define a weakly closed set of probability plans. |
isTightMeasureSet_singleton and IsTightMeasureSet.prodMk |
All plans with fixed finite Polish marginals form a tight family. |
isCompact_closure_of_isTightMeasureSet |
Tight probability-plan families have compact closure. Closedness then gives compactness of the feasible set. |
FiniteMeasure.continuous_lintegral_boundedContinuousFunction |
The integral of each bounded continuous cost truncation is continuous under weak convergence. |
lintegral_iSup, lowerSemicontinuous_iSup and LowerSemicontinuousOn.exists_isMinOn |
Increasing truncations give a lower semicontinuous full cost and an attained minimum on a nonempty compact feasible set. |
Pinned sources: probability maps, tightness, Prokhorov and weak integral continuity. These ingredients are not a packaged coupling minimizer or a Wasserstein triangle/separation theorem.
TauCeti candidates
The licensed OT source at 575dbdf
and the inspected main snapshot 6a7e27a
both pin Lean 4.36.0-rc1 and mathlib
2cabef97b63d3275776b8b060d802de6a6fa1055. They are not directly compatible
with this project's pinned toolchain and mathlib. The first main-history
Wasserstein/Basic snapshot
already pins Lean 4.34.0-rc2 and mathlib 03616a12cdf1e3f499349f69e80eee82b28745bc.
The distinct releases/v4.33.0 snapshot
pins Lean 4.33.0 and mathlib db584cd6d46c92f209a44c0f1c829460d327499d.
It contains gluing and multimarginal constructions, but no coupling/cost,
existence or Wasserstein modules. Its version name therefore does not supply
the required OT spine. No claim is made about every uninspected side branch
or independently maintained fork.
At 575dbdf, the substantial ordinary APIs include:
| API | Exact role and relevant hypotheses |
|---|---|
IsCoupling |
The two raw measure marginals, expressed as Measure.fst and Measure.snd, agree with the prescribed measures. These match the native reservoir marginal equations. |
wassersteinEDist_rpow_eq_transportCost and wassersteinEDist_eq_transportCost_rpow |
Identify the eLpNorm infimum and the powered-integral infimum for a measurable ground extended distance and a finite nonzero ENNReal exponent. For real p ≥ 1, use exponent ENNReal.ofReal p. |
wassersteinEDist_smul |
Common finite nonzero measure scaling gives the factor a^((1/p).toReal). Together with the native reservoir normalization, this supplies the requested probability convention. |
exists_isCoupling_eLpNorm_eq_wassersteinEDist |
Produces an actual optimizer on a complete second-countable Borel metric space for finite source mass, a finite nonzero exponent and existence of a coupling. Equal finite reservoir masses provide the last condition; an optimizer is not a caller premise. |
wassersteinEDist_triangle and wassersteinEDist_eq_zero_iff |
Sharp triangle uses standard-Borel gluing and finite intermediate mass. Separation uses a genuine metric, Polish/Borel hypotheses, finite source mass and finite nonzero exponent. |
exists_ae_mem_finset_wassersteinEDist_le |
On a separable pseudo-metric space with measurable opens and measurable singletons, approximates a probability law of finite moment at finite exponent p ≥ 1 and nonzero ENNReal accuracy. The native same-mass theorem instead quantizes the original measure into interior values, preserving the common-reservoir mass constraint. |
Sources: Coupling,
Wasserstein/Basic
and FiniteSupport.
The full spine includes general coupling and cost laws, weak compactness,
lower semicontinuity, kernel gluing and Wasserstein norm/metric arguments.
It is not a single theorem that can be imported independently of its package.
Native finite-measure attainment
Attainment of the existing native reservoir value can be separated from that
full ordinary-OT dependency. Keep every auxiliary plan set on the already
proved space X × X. Normalize the two reservoirs to probability measures;
their product makes the feasible set nonempty. The pinned APIs above prove
its weak compactness. For the continuous cost ρ^p, the functions
min(ρ^p,n) are bounded continuous nonnegative functions increasing to the
full cost. Their weakly continuous integrals have the full integral as
supremum, so the full cost is lower semicontinuous, including at infinite
values. It therefore attains its minimum.
exists_minimizing_reservoirPlan produces an actual minimum for every
nonnegative real exponent and every finite positive full sum-capacity.
For p ≥ 1 and arbitrary finite Borel μ,ν,
exists_minimizing_boundaryPlan then produces an admissible boundary
minimizing measure by the existing reverse construction and the proved
equality of infima. Choosing R = μ(H)+ν(H)+1 handles zero inputs
without division by zero. Finite p-persistence is needed for finite cost,
not for existence of a minimizer whose value may be infinite.
exists_finite_cost_minimizing_boundaryPlan records this finite-cost
consequence. exists_minimizing_boundaryPlan_rooted gives one plan realizing
both powered and rooted values. Neither compact support nor a Euclidean
moment along the diagonal is required.
This is a fixed-space bridge to the native infimum. The boundary-transport lane adds no
generic coupling class, ordinary Wasserstein definition or gluing library. (The separate area
TDA.Probability defines MeasureCoupling, measureTransportCost and measureWasserstein on
ProbabilityMeasure with exact probability marginals for the sampled Rips laws; the
boundary-transport value is not identified with them.) The closed/tight-plan argument can be adapted from
TauCeti Compactness;
the specific continuous-cost truncation avoids importing its more general
lower-semicontinuous approximation layer. Compactness and a minimizing
coupling are proved internally, not requested as caller premises.
The consumers include a nonatomic uniform band of bars (t,t+2) and the
countable law with weights 2^(−n−1) at (2^(n+2),2^(n+2)+2). Both have
total mass one and diagonal moment one at every real exponent. The latter
has divergent first birth-coordinate moment and gives positive mass outside
every compact subset of H. Sending its mass to the fixed diagonal point
(0,0) is an admissible infinite-cost competitor at p=1, while an attained
optimum has cost one. Thus finite optimal cost is compatible with infinite
competitor costs.
The native metric laws use existing current-pin APIs directly.
Measure.condKernel, Measure.disintegrate and Measure.fst_compProd
glue finite minimizing plans on the fixed space; measurable kernel sections
identify the second prescribed pair marginal. eLpNorm_add_le gives the sharp
triangle estimate at a common full sum-capacity for all three pairs. A
zero-cost actual optimizer has equal endpoints almost everywhere, which
proves separation by equality of interior restrictions. The norm bridge
expresses both the unnormalized reservoir value and its probability scaling
as literal infima of mathlib eLpNorm over all marginal equations. This lane adds no generic
coupling class, Wasserstein definition or general gluing API, and no identification with
TDA.Probability.measureWasserstein is claimed.
Native same-mass finite atomic approximation
is proved directly using the current-pin metric SimpleFunc.approxOn and
dominated convergence. Every target is chosen in the interior image;
quantizing only the original measure preserves its mass and leaves the
common boundary reservoir padding unchanged. No ordinary Wasserstein
finite-support theorem is needed for this native endpoint.
For identification with an independently maintained ordinary Wasserstein API, the recommended dependency decision is a reviewed external backport of its required closure to this exact Lean/mathlib pin, with compilation and axiom validation. Merely selecting the older release branch is insufficient. A project-wide toolchain/mathlib upgrade is a different dependency decision; none is made here. The native attainment, metric laws and literal coupling-norm forms do not depend on that choice.
Source-use boundary
The inspected TauCeti code is Apache-2.0 licensed. A source adaptation must retain its applicable notices and identify the source revision and modifications. Code availability does not certify a compiled backport or authorize a dependency change.
The corrected Divol–Lacombe article has a separate, unresolved article-text AI-input/reuse condition in #505. Open-access metadata does not establish that permission. The native contract and the route above use the repository's mathematical digest and licensed Lean sources; they do not claim a new comparison with the article body.