Same-mass finite weighted approximation
Let H be the strict birth–death half-plane, with diagonal distance
δ(x) = (death − birth)/2. For real p ≥ 1, every finite Borel measure μ
on H satisfying
∫ δ(x)^p dμ(x) < ∞
can be approximated arbitrarily closely in native boundary transport by
atomicMeasure D, for a native FiniteWeightedDiagram D, a finitely
supported nonnegative weight function on H, of exactly the same total
mass. Here BT_p is the reciprocal-p root of the native all-boundary-plan
powered infimum. The approximants also form a sequence converging in BT_p.
The input can be nonatomic or have noncompact support; no Euclidean moment
along the diagonal is required. The zero measure is included.
The public theorem exists_atomicMeasure_approx_same_mass gives the
arbitrary-accuracy diagram. quantizedDiagram μ n is an actual native
approximation sequence; quantizedDiagram_mass preserves mass at every
index, and boundaryTransportPowerValue_quantizedDiagram_tendsto and
boundaryTransportDistance_quantizedDiagram_tendsto prove powered and
rooted convergence. quantizedDiagram_zero identifies every zero-input
approximant with the zero diagram.
Quantizers and native coefficients
A measurable finite-valued map T : H → H partitions the input into its
fibres. Give each value y the coefficient μ(T⁻¹ {y}). These coefficients
are finite because the original measure is finite. The existing indexed
coalescing constructor forms a native diagram from this finite family,
including zero coefficients. Its canonical atomic measure is exactly
μ.map T. In particular, its total mass is μ(H), rather than a normalized
probability mass or the mass of an augmented reservoir.
The quantizers are chosen using the existing collapsed space X = H ⊔ {*}
with metric
ρ(x,y) = min(‖x−y‖∞, δ(x)+δ(y)), ρ(x,*) = δ(x).
The interior image H → X is dense. Pinned mathlib's metric
SimpleFunc.approxOn can take all its values in that image, with one fixed
interior basepoint of depth one. Pulling those values back to H gives
finite-valued measurable T_n : H → H, with
ρ(x,T_n(x)) → 0 pointwise. Each target is an interior point from the
outset: the finite approximants carry no boundary atom whose mass would
have to be discarded or added to the original measure.
Cost convergence and capacity
The nearest-point bound supplied by approxOn gives
ρ(x,T_n(x)) ≤ 1 + δ(x).
For p ≥ 1, the powered errors are therefore bounded by
2^p (1 + δ(x)^p). This function is integrable using both finite original
mass and finite diagonal p-persistence. Dominated convergence gives
∫ ρ(x,T_n(x))^p dμ(x) → 0.
Let m = μ(H) and fix one capacity R = 2m+1. Since every target measure
has mass m, this satisfies the full sum-of-masses condition for every
pair, including μ = 0. Push μ along x ↦ (x,T_n(x)) into X × X
and add (R−m)δ_(*,*). This is a genuine competitor with the two required
reservoir marginals, and its powered cost is the displayed error integral.
The existing all-competitor reservoir/boundary comparison bounds the native
boundary infimum by this cost. Taking reciprocal-p roots yields convergence
in BT_p and hence the arbitrary-accuracy approximant with unchanged mass.
Neither an approximating measure nor an optimizer is a caller premise.
Finite original mass matters here: it controls the constant term in the dominating function. A locally finite input of infinite original mass is outside this theorem even when its diagonal persistence moment is finite.
Sources and remaining scope
The construction consumes pinned mathlib's
SimpleFuncDense
and dominated convergence,
together with the proved native coalescing, atomic-measure and reservoir
comparison APIs. It requires no additional OT package or toolchain change.
Mathematical correspondence is to the repository contract in
#505; the corrected article's
separate AI-input/reuse condition remains unresolved, and no article-body
comparison is claimed.
This is same-original-mass approximation by the canonical native atomic
constructor. Ordinary Wasserstein API identification remains the other open endpoint in the
dependency map; the sharp native triangle inequality and
separation are proved in the metric laws
(boundaryTransportDistance_triangle, boundaryTransportDistance_eq_zero_iff).