Finite boundary transport for real matching exponents

Fix a real exponent p ≥ 1. Each input is a finite family of birth–death points, x : I → ℝ² and y : J → ℝ², with birth < death. Deaths are finite and essential bars are absent. Labels carry unit mass; equal coordinates with different labels retain their multiplicity. The ground norm is the maximum norm, independently of p. Write c(i,j) = ‖x(i) − y(j)‖∞ and δ(z) = |death(z) − birth(z)|/2.

For a partial injection M between the occurrence labels, its powered cost is

Sₚ(M) = Σ matched(i,j) c(i,j)^p + Σ unmatched i δ(x(i))^p + Σ unmatched j δ(y(j))^p.

Both unmatched sums are necessary. Set mₚ = min_M Sₚ(M). The all-unmatched matching exists even for empty inputs, and the finite matching family makes this minimum attained. The rooted matching quantity is dₚ = mₚ^(1/p). Raising the first-order minimum to p is a different operation: the optimal matching can change.

Real plans and their optimum

A FiniteBoundaryPlan has real nonnegative entries a(i,j), with every row and column sum at most one. Its deletion and insertion masses are

u(i) = 1 − Σⱼ a(i,j), v(j) = 1 − Σᵢ a(i,j).

Its powered transport cost is

Cₚ(a) = Σᵢⱼ a(i,j)c(i,j)^p + Σᵢ u(i)δ(x(i))^p + Σⱼ v(j)δ(y(j))^p.

Distances are powered; mass coefficients stay linear. There is no cardinality or probability normalization. All fractional real plans satisfying these capacity conditions are admitted, not just incidence matrices of partial matchings.

The real plan infimum tₚ satisfies

tₚ = inf_a Cₚ(a) = mₚ, tₚ^(1/p) = dₚ.

Every plan admits a partial matching of no greater powered cost. Conversely, the incidence matrix of a mₚ-minimizing matching attains tₚ. Some optimal plans can still be fractional. Nonnegativity supplies lower bounds for both real infima; a zero mass matrix supplies a feasible plan.

At p ≥ 1 the reciprocal-power map is increasing on nonnegative numbers. The attaining matching and plan therefore also attain the rooted optima, and

dₚ = inf_M Sₚ(M)^(1/p), tₚ^(1/p) = inf_a Cₚ(a)^(1/p).

These identities use the proved minimizers, without an assumed optimizer or a claim that taking roots commutes with arbitrary real infima.

Shared assignment argument

The coefficient-dependent matching cost OccurrenceMatching.totalBoundaryCost and plan cost FiniteBoundaryPlan.linearCost take separate pair, deletion and insertion coefficients. The first-order APIs retain their definitions and types; their relabelling and assignment proofs reuse these helpers. The powered APIs supply c^p and δ^p as coefficients.

The augmentation has source-occurrence rows followed by target dummies, and target-occurrence columns followed by source dummies. Every source-row deletion coefficient is constant over its full dummy block, and every target-column insertion coefficient is constant over its full block. Dummy-to-dummy cost is zero. Restriction and completion preserve the three cost sums; completion is a section, not a two-sided inverse for arbitrary dummy permutations.

Reuse the existing padded mass matrix

D(a) = [[a, diag(u)], [diag(v), transpose(a)]].

After column relabelling it is doubly stochastic, including empty indices. Mathlib's Birkhoff theorem expresses it as a sum of permutation matrices with nonnegative coefficients summing to one. The assignment cost is linear in matrix mass. Each permutation restricts to a genuine matching and costs at least the finite matching minimum. Their convex average has the same lower bound. This is the fractional-to-matching argument for both first-order and powered coefficients; no metric property of the raw boundary blocks is used.

Empty inputs, labels and compatibility

With y empty, the powered optimum is Σᵢ δ(x(i))^p and its rooted quantity is the reciprocal-power root of that sum. The analogous target sum applies with x empty. Both-empty powered and rooted quantities are zero on p ≥ 1. Arbitrary equivalences of occurrence types preserve every optimum, including repeated coordinates.

At p = 1, candidate matching cost, plan cost, both powered minima and both rooted quantities agree with their existing first-order APIs. The supporting powered algebra also permits arbitrary real pairs and exponents; its PD interpretation is restricted to the domain at the start of this page. In particular, root-infimum monotonicity and rooted empty-input vanishing are stated with p ≥ 1.

Rooted distance laws

For a third finite family z : K → ℝ² and every real p ≥ 1, the rooted matching optimum satisfies

dₚ(x,x) = 0, dₚ(x,y) = dₚ(y,x), dₚ(x,z) ≤ dₚ(x,y) + dₚ(y,z).

The same laws hold for the rooted finite-real-plan optimum, by its proved equality with dₚ. On positive-lifetime inputs, zero cost has the precise characterization

dₚ(x,y) = 0 if and only if there is an equivalence e : I ≃ J with x(i) = y(e(i)) for all i.

This preserves values and their multiplicities while allowing different labels. Unequal raw functions related by a permutation can have zero cost. A diagonal occurrence has zero deletion cost, so positive lifetimes are needed for this separation statement. No metric-space instance on raw presentations is supplied. The self, symmetry and triangle algebra also holds for arbitrary real pairs.

Reversal exchanges the two unmatched parts and preserves the powered paired sum. For the triangle inequality, follow M from x to y by N from y to z. Let s_M(i) be M's pair-or-deletion charge at source i, and let t_M(j) be zero at a selected target j and δ(y(j)) at an unmatched target. Define s_N and t_N analogously. Use the disjoint union of source, middle and target labels, written I ⊕ J ⊕ K, as a common index for three nonnegative charge families A, B and C:

Index A B C
Source i s_M(i) s_N(j) if M(i)=j, otherwise 0 s_(N∘M)(i)
Middle j t_M(j) s_N(j) if M does not select j, otherwise 0 t_(N∘M)(k) if N(j)=k, otherwise 0
Target k 0 t_N(k) t_N(k)

Partial injectivity and the selected/complementary partition of J give exact sums

Σ A^p = Sₚ(M), Σ B^p = Sₚ(N), Σ C^p = Sₚ(N∘M).

Each input charge occurs once in its own sum. Pointwise C ≤ A+B follows from the ground triangle inequality and δ(a) ≤ ‖a−b‖∞+δ(b). These cover a middle occurrence paired on both sides, selected then deleted, and unselected then paired. A middle occurrence unmatched by both has no endpoint; its two nonnegative input charges remain unused. N's unmatched targets are also unmatched by the composite.

Monotonicity of positive real powers and finite-sum Minkowski now give

Sₚ(N∘M)^(1/p) ≤ (Σ (A+B)^p)^(1/p) ≤ Sₚ(M)^(1/p) + Sₚ(N)^(1/p).

Choose minimizing matchings for the two input costs to obtain the optimum triangle law. For separation, a sum of nonnegative powered charges is zero exactly when every raw charge is zero. The existing first-order zero-cost characterization then supplies the value-preserving equivalence. Conversely, reindexing a self-matching proves zero cost for such an equivalence.

Comparison with first-order distance

Let N = card I + card J count both occurrence families. For every real p ≥ 1, the unnormalized matching distances satisfy

dₚ(x,y) ≤ d₁(x,y) ≤ N^(1−1/p) dₚ(x,y).

The same comparison holds for the rooted finite-real-plan optima, by their equality with the matching optima. Empty families are included: if N=0, all distances are zero. At p=1 the factor is one and both inequalities are equalities.

For a matching M, form a nonnegative charge vector on I ⊕ J. At a source label, use its pair cost if selected and its deletion cost otherwise. At a target label, use zero if selected and its insertion cost otherwise. Its sum is S₁(M), and the sum of its p-th powers is Sₚ(M). Each paired cost occurs once; both unmatched parts remain. Finite-sum power inequalities give

Sₚ(M)^(1/p) ≤ S₁(M) ≤ N^(1−1/p) Sₚ(M)^(1/p).

For the first optimum inequality choose a first-order minimizing matching. For the second choose a p-minimizing matching. Both exist, including for empty inputs; one matching need not minimize both exponents. The algebra also applies to raw real pairs without positive lifetimes.

Two copies of (0,2) versus empty have d₁=2 and dₚ=2^(1/p), attaining the upper factor. At p=3/2 these are 2 and 2^(2/3), with factor 2^(1/3). The comparison is about optimized plans. A fractional candidate sending mass 3/4 between identical bars (0,2) has C₁=C₂=1/2, so its rooted second-order cost exceeds its first-order cost even though their common optimum is zero.

Comparison at two real exponents

For real 1 ≤ r ≤ s and the same occurrence count N,

dₛ(x,y) ≤ dᵣ(x,y) ≤ N^(1/r−1/s) dₛ(x,y).

Both inequalities also hold for the rooted real-plan optima. Thus each optimized distance is antitone on [1,∞), with the maximum ground norm fixed. The families may be empty and distances may vanish. At r=s the factor is one, including 0^0=1 when N=0.

For the matching charge vector a, apply the preceding finite-sum inequalities to the vector with entries a_k^r at exponent s/r, then raise both sides to 1/r. Nonnegative-base power identities give the displayed factor without dividing by N. To pass to optima, choose an r-minimizer for the lower comparison and an s-minimizer for the upper comparison. Equality of matching and real-plan optima transfers the result to all feasible real reservoir matrices after optimization.

Two labelled copies of (0,2) versus empty attain the upper bound: their rooted costs at r=3/2 and s=3 are 2^(2/3) and 2^(1/3), with factor 2^(1/3). For the singletons (0,2) and (3/2,7/2), the optimum changes from pairing at exponent one, with cost 3/2, to deleting both at exponent two, with rooted cost √2. The comparison does not require a common optimizer.

Candidate plans need not be antitone. Sending half a unit from (0,2) to (6,8) and leaving half a unit at each boundary has powered costs 4 and 19 at exponents one and two. Its rooted second cost √19 exceeds 4, while the optimized distances are 2 and √2 and satisfy the comparison.

Continuity at finite exponents

Fix the two occurrence families. The four functions p ↦ mₚ, p ↦ dₚ, p ↦ tₚ and p ↦ tₚ^(1/p) are continuous on [1,∞). Empty families, nonempty zero-distance inputs and repeated coordinates are included. The statement concerns the optimized values even when their minimizing matchings change with p.

For a fixed matching, each branch of the source and target sums is fixed as p varies. Powers of its nonnegative geometric coefficients are continuous at positive exponents, including zero coefficients. A finite nonempty minimum of these continuous candidate functions is continuous. The proof identifies this finite minimum with the existing real infimum; it does not interchange a limit with an arbitrary infimum. The reciprocal exponent is positive, so rooting is continuous even when mₚ=0. The proved assignment equalities transfer the result to the infimum over all feasible real plans.

Identical singleton bars have zero optimized cost throughout the domain, while two labelled copies of (0,2) versus empty have powered cost 2 and rooted cost 2^(1/p), in either orientation. For (0,2) versus (3/2,7/2), the powered curve is min((3/2)^p,2). It remains continuous as the optimizer changes from pairing to deleting both.

Exponent zero lies outside this domain. Under the existing real-power conventions, the empty rooted cost is 1 at zero and 0 at every positive exponent. For identical singleton bars (0,2), the powered optimum is also 1 at zero and 0 at positive exponents. These positive-side limits differ from the values at zero. The continuity API keeps the definitions and the PH exponent domain unchanged.

Joint continuity with moving coordinates

Keep the finite occurrence types I and J fixed, and give their coordinate families the ordinary function-space topology. On ℝ × ((I → ℝ²) × (J → ℝ²)), restricted to p ≥ 1, the four optimized values mₚ(x,y), dₚ(x,y), tₚ(x,y) and tₚ(x,y)^(1/p) are jointly continuous. Consequently, a continuous parameter family a ↦ (p(a),x(a),y(a)) with p(a) ≥ 1 has continuous optimized costs. Both families and the exponent may move together.

For a fixed matching, pair coefficients are maxima of absolute coordinate differences, and boundary coefficients are half the absolute lifetime. These are continuous real functions. Joint continuity of real powers at positive exponents includes zero coefficients. Both unmatched sums remain, and their branch choices depend only on the fixed matching. Taking the finite nonempty minimum, then its positive reciprocal power, preserves continuity even at a zero minimum. The assignment equalities again transfer the result to all feasible real reservoir matrices.

For example, let p(t)=1+t² and take singleton families x(t)=(t,t+2) and y(t)=(-t,2-t). Both bars have lifetime 2, and their powered optimum is min((2|t|)^(1+t²),2). All four optimized path costs are continuous. At t=0 they vanish; at t=1/2 pairing costs 1 while deleting both costs 2; at t=1 pairing costs 4 while deleting both still costs 2. Thus the result includes a change of minimizing matching with both coordinates and exponent moving.

The examples also vary the lifetime of two repeated labels in both one-empty orientations. A raw singleton (t,-t) versus empty has rooted cost |t|, continuous through its diagonal collision at zero. It is a positive bar for t<0 and lies below the diagonal for t>0. The diagonal and below-diagonal cases exercise the raw cost algebra. Fixed label types and cardinalities are part of the joint domain. This finite library derivation introduces no topology on changing presentations or general persistence measures. The exponent-zero discontinuities above still apply: identical singleton bars (q,q+2) have powered optimum 1 at exponent q=0 and optimum 0 for q>0, even as their coordinates move.

Diagonal translation and nonnegative scaling

Fix p ≥ 1 and use mₚ and dₚ on raw real coordinate families, which permit diagonal outputs. This makes the zero-scale case meaningful as a cost identity. For a real t, let Tₜ(b,d) = (b+t,d+t). For a ≥ 0, let Sₐ(b,d) = (ab,ad). Apply the same map to both input families. Their powered minima and rooted distances satisfy

mₚ(Tₜx,Tₜy) = mₚ(x,y), dₚ(Tₜx,Tₜy) = dₚ(x,y),

mₚ(Sₐx,Sₐy) = a^p mₚ(x,y), dₚ(Sₐx,Sₐy) = a dₚ(x,y).

The finite real-plan optima satisfy the same laws by their equality with the matching minima. A diagonal shift cancels in every pair and boundary distance. Scaling multiplies each powered coefficient by a^p, so it multiplies every matching's sum by that factor. Attained minimizers give both inequalities for the minimum, including when a=0; no division by a is needed. Taking the nonnegative reciprocal-power root gives the rooted scaling law.

Every real diagonal shift preserves birth < death, including shifts to negative births. Applications imposing nonnegative births must separately preserve that extra condition. Positive scaling preserves positive lifetimes. At scale zero, the raw coordinate families consist of diagonal points and all costs vanish. This case belongs to the raw-cost API; Diagram.scale still requires a positive factor. Repeated coordinates retain their occurrence labels throughout.

The shift must be along the diagonal. For example, the first-order distance from the singleton (0,2) to the empty family is 1; translating vertically to (0,4) changes it to 2, although both lifetimes are positive. These laws are direct library derivations from the cost definitions at fixed maximum ground norm.

Finite-part perturbation bounds

Let D and E be library diagrams and let N be the sum of their finite occurrence counts. Write B for their full bottleneck distance, mₚᶠⁱⁿ for the powered matching minimum on their finite occurrence coordinate functions, and dₚᶠⁱⁿ for its root. This root is also the rooted finite-real-plan optimum by the equality above. Essential occurrences are excluded from this transport cost and from N, even if D and E contain them. Assume B is finite. For real p≥1,

mₚᶠⁱⁿ ≤ N B^p, dₚᶠⁱⁿ ≤ N^(1/p) B.

Choose a full matching attaining B and retain its finite partial injection. Each finite source contributes a pair or deletion charge at most B. Each target contributes zero when selected and a boundary charge at most B otherwise. Summing their p-th powers gives at most N B^p; the finite powered minimum is no larger than this candidate. This uses a bottleneck optimizer as an upper-bound witness, without asserting that it also minimizes the p-cost. Nonnegativity, p≠0 and real-power factorization give the rooted estimate, including N=0 and B=0. Equality with the real-plan optimum transfers this to the infimum over all feasible matrices.

More generally, if N≤m for a natural cardinality budget m and 0≤r with B≤r, then

dₚᶠⁱⁿ ≤ m^(1/p) r.

A uniform family estimate requires this budget explicitly. Cardinality need not remain constant under perturbations: bars can reach the diagonal and disappear. Finiteness of B is also material. In the Lean API an extended infinite value converts to real zero; essential-cardinality mismatch can make B infinite while the finite-part transport cost is positive.

For two face filters f,g on the same finite complex, let Δ be the maximum absolute change of their face values, with Δ=0 for an empty face set. Over the coefficient field and ordered-vertex assumptions of the degree-n sublevel diagrams, put D=dgm(f,n) and E=dgm(g,n), and let N(f,g,n) count their finite occurrences on both sides. The proved full-diagram bottleneck stability gives B≤Δ and hence

dₚᶠⁱⁿ(dgm(f,n),dgm(g,n)) ≤ N(f,g,n)^(1/p) Δ.

The cardinality-budget version replaces N(f,g,n) by any explicit upper bound m. The existing bottleneck argument allows the barcode pairing and compatible face order to change; no fixed pairing or order is required here. Essential bars in these diagrams still belong to the full bottleneck comparison and remain outside the finite-part transport quantity.

Limit as the matching exponent tends to infinity

Fix two finite diagrams D,E with empty essential parts. Let B be their full bottleneck distance and let N count both finite occurrence families. Empty essential parts imply B<∞, so its real value is meaningful. For every real p≥1,

B ≤ dₚ(D,E) ≤ N^(1/p) B.

Here dₚ uses the canonical finite occurrence labels of D,E. Each charge of a matching is at most its rooted powered sum. Choose a minimizing finite matching and extend it by the unique matching of empty essential families. Its maximum bounds B from above, giving the lower comparison. The upper comparison is the occurrence-count bound above. Both comparisons concern optimized quantities; the minimizing matching may change with p.

If N>0, then N^(1/p) tends to one as real p tends to infinity. Squeezing gives

lim_(p→∞) dₚ(D,E) = B.

If N=0, both finite families are empty and both sides are zero; no limit claim about 0^(1/p) is used. The exact matching/plan identity gives the same limit for the rooted real boundary-reservoir optimum. The API provides real limits and extended nonnegative real limits, with B's finiteness proved before conversion.

The comparison can fail for unequal essential births. Essential-only diagrams with births 0 and 1 have zero finite-part transport at every p≥1, but full bottleneck one. Finiteness of B alone therefore does not suffice. The common essential-birth multiset condition below gives another sufficient case.

Uniform recovery under count and radius bounds

Let m be a natural number and R≥0. Consider all pairs D,E with empty essential parts, total finite occurrence count N≤m, and bottleneck B≤R. Set M=max(1,m). For either optimized rooted cost and every real p≥1,

|dₚ(D,E) − B| ≤ (M^(1/p) − 1) R.

Indeed, B≤dₚ≤N^(1/p)B≤M^(1/p)B. Subtract B before using the radius bound: the error is at most (M^(1/p)−1)B, and this is at most (M^(1/p)−1)R. Since M≥1, its reciprocal-power factor tends to one. The common error envelope therefore tends to zero. The API gives TendstoUniformlyOn in real p along atTop: one exponent threshold works for every pair in this budget class, even when its occurrence types and minimizing matchings vary. This needs no topology or compactness of the set of diagrams. Empty inputs satisfy m=0; identical nonempty positive diagrams also lie in a zero-radius class.

For R>0, two copies of (0,2R) versus empty have B=R and rooted cost 2^(1/p)R. They attain the budget-two error bound, including p=3/2. The singleton pair (0,2), (3/2,7/2) belongs to the same count-two, radius-one class while its optimizer changes from pairing at one to deleting both at two.

Both budgets matter. At exponent n≥1, take 2^n copies of (0,2) versus empty. Their bottleneck is one and their rooted costs are two, so the error stays one along cofinal real exponents without a count cap. With count two fixed, choose a(p)=(2^(1/p)−1)⁻¹>0 and two copies of (0,2a(p)) versus empty. Their error is one for every p≥1 without a common radius bound. These examples obstruct a uniform epsilon-one-half tail on the corresponding larger classes; each fixed pair still has the pointwise limit above. The essential-birth example also excludes removing the empty-essential hypotheses.

This uniform estimate is a finite library derivation using the DL matching convention and the existing optimized bounds. The probability-measure theorem TauCeti.tendsto_wassersteinEDist_atTop is a related pointwise result: its normalized costs increase to their supremum, using compactness of couplings. These unit-occurrence costs approach bottleneck from above. The count/radius uniformity here concerns finite essential-free diagrams; it gives no uniform all-Radon or powered-cost convergence theorem.

Common essential-birth multisets

The same error bound and uniform limit hold when the entire essential-birth multisets agree. Equality includes values and multiplicities; the common multiset may vary across pairs, with no bound on its count or birth values. Keep the total finite count N≤m and full bottleneck B≤R≥0. For both existing rooted optima and every real p≥1,

|dₚ−B| ≤ ((max(1,m))^(1/p)−1)R.

The uniform domain is the literal set of pairs satisfying those three conditions. One exponent tail works for every pair in it, including m=0 and R=0. There is no topology on diagrams or chosen common optimizer.

The canonical essential occurrence equivalence preserves each birth value, including repeated births. Extend a minimizing finite matching by that equivalence. Its essential charges vanish, so its full maximum equals the cost of the same finite matching with empty essential types. The existing rooted-sum bound gives the lower comparison; the arbitrary-essential finite upper bound gives the same count factor. Equality of essential cardinalities first establishes full bottleneck finiteness before conversion to real. optimized real plans use the proved matching equality after optimization.

For example, reordered essential births [-3,-3,7] have full bottleneck and both finite costs zero. Any number of identical common births, at any real value, lies in the same zero-budget class. Adding two copies of (0,2R), R>0, on one side gives B=R and dₚ=2^(1/p)R, attaining the error envelope while retaining the essential block. Adding that block to (0,2) and (3/2,7/2) retains the change from pairing at p=1 to deletion at p=2. Essential births 0 and 1 have equal counts but full B=1 and finite costs zero, so cardinality alone cannot replace multiset equality. Unequal counts give B=∞ and exclude every finite-radius class.

This is an extension of the existing library model, with no essential p-cost. DL corrected v4 §1.1, Remark 1.2, describes a sum of independent finite and essential terms; this library's full bottleneck uses a maximum. Common essential values give zero essential displacement in either convention. No identification of general essential metrics or source uniform theorem is claimed. The essential-free class is a checked restriction of this domain.

Čech sublevel diagrams under a displacement budget

Fix a finite linearly ordered label type V, a coefficient field k in the same universe, and a pseudometric ambient space X in an independent universe. A cloud is an arbitrary function a : V → X; repeated points and empty labels are allowed. Write D(a) for the diagram dgm k (Cech.filter a) n. The filter uses the infimal enclosing radius over all ambient centers. Its sublevels contain the closed-ball Čech nerves, and equal them under the existing proper-space specialization. Attainment is not assumed for the results here.

For n≥1, each D(a) has empty essential part. In degree zero, nonempty labels give exactly the essential-birth multiset {0}; empty labels give the empty diagram without choosing an ambient point. Thus the entire essential multisets of D(a) and D(b) agree in every natural degree. Each diagram has at most choose(card V,n+2) finite occurrences, and same-label displacement δ(a,b) bounds the full bottleneck B(a,b). If δ(a,b)≤R with R≥0, put M=max(1,2 choose(card V,n+2)). For each optimized rooted cost and every real p≥1,

|dₚ(D(a),D(b)) − B(a,b)| ≤ (M^(1/p) − 1) R.

Both rooted costs converge uniformly to B on the literal cloud-pair class δ(a,b)≤R as real p tends to infinity. The fixed data are V,k,X,n,R; the exponent tail precedes every cloud pair in that class. The proof derives the diagram budgets, precomposes the generic uniform result with (a,b)↦(D(a),D(b)), and restricts to the displacement class. No topology on clouds or diagrams, fixed barcode length or common optimizing matching is required.

For labels (0,0),(2,0),(1,2) versus (0,0),(2,0),(2,2) in the Euclidean plane, the degree-one diagrams over ZMod 2 are the singleton (√5/2,5/4) and the empty diagram. Displacement is exactly one; B and both rooted optima are r=(5/4−√5/2)/2>0 for every p≥1. This tests nonzero quantities, including p=2 and p=3/2; its error is zero and does not attain the coarse budget-two rate. The acute triangle paired with itself tests radius zero with a retained bar. Empty and repeated-label examples also instantiate the two models.

This is a finite library composition using the DL matching convention. The same-label Čech stability argument supplies its geometric input; its CDSO Hausdorff comparison and the radius-filter attainment distinction have separate scopes. No uniform p-limit is attributed to that source.

The birth-zero proof uses a vertex at time zero whose inclusion into a full late stage has nonzero homology range. The diagram-rank formula gives a point alive between those stages. Every finite death is a critical face value at most that late time, so this point is essential. Its birth is at most zero and is a nonnegative critical value, hence zero. Only then does the essential count one identify the full multiset. An eventual count alone would not determine the birth.

This ordinary H₀ calculation uses the radius filter's empty negative tail. CDSO retains vertices at nonpositive scales, so no whole real-indexed H₀ module identification is asserted. The proper-ambient closed-nerve specialization here is at nonnegative radius. Translating a singleton's filter values by one gives an essential birth at one, finite costs zero and full bottleneck one against the original diagram; equal essential counts alone remain insufficient. The H₀ log module still has no finite-grid zero left tail. The all-degree real-parameter consumers do not change that fact.

For the clouds (0,2) and (0,6), the H₀ finite bars are (0,1) and (0,3), with a common essential birth zero. Displacement is four, while full bottleneck is 3/2. Both rooted finite optima are two at exponent one and √(5/2) at exponent two: pairing is optimal at one, and deleting both is strictly better at two. The exponent-one error is 1/2. Comparing the acute triangle with a constant cloud gives two finite deletions with charges 1/2 and √5/4; repeated labels (0,0,2) give only the positive bar (0,1), since the simultaneous zero-length merge is filtered out.

Examples

Example Powered/rooted calculation
Both diagrams empty, every p≥1 Both optimized costs and bottleneck are zero; the limit includes N=0
One bar (0,2) versus empty, every p≥1 Both rooted optima and bottleneck are one
Empty target, diagonal charges 1 and 2, p=2 Powered optimum 5; rooted optimum √5
Half mass from (0,2) to (0,4), p=2 Plan cost 9/2; both boundary deficits are 1/2; optimum is 4
Two copies of (0,2) versus one copy, every p≥1 Rooted optimum and bottleneck are one, with a pair-one/delete-one matching
(0,2) versus (3/2,7/2) p=1 pairs, cost 3/2; p=2 deletes both, powered cost 2, strictly below the paired cost 9/4
(0,8) versus empty, p=3/2 Powered optimum 8; rooted optimum 4
Two copies of (0,2) versus empty, any p≥1 Powered optimum 2; rooted optimum 2^(1/p); bottleneck 1, attaining the occurrence-count factor
Essential-only births 0 versus 1, every p≥1 Finite-part cost zero, full bottleneck one; equal essential counts do not suffice for recovery
Two bars (0,2R) versus none, with common essential births [-3,-3,7], R>0 Full bottleneck R; both rooted optima 2^(1/p)R; the count-two error bound is attained
filled triangle versus its tied filter, H₁ The finite parts have one bar and zero bars; their transport is bounded by the face perturbation, including p=2 and p=3/2
Four middle roles with unmatched endpoints, p=2 Candidate powers 11 and 16; composite power 19; √19 ≤ √11+4
Singletons (0,10), (1,11), (2,12), every p≥1 Consecutive rooted optima 1 and 1; direct rooted optimum 2; sharp triangle at p=2 and p=3/2

The examples also calculate a triangle-inequality failure for the squared optimum: translating the singleton (0,10) by one twice gives successive squared costs 1 and 1, but the direct squared optimum is 4. This distinguishes powered sums from metric claims about rooted quantities.

API and source scope

FinitePMatchingCost supplies powered and rooted matching quantities, attainment, candidate bounds, root-infimum equality, empty formulas, singleton alternatives, reindexing and first-order compatibility. FinitePBoundaryTransport supplies the real-plan costs, matching/transport equalities, attaining incidence plans and the corresponding bounds and formulas. The example module uses these definitions. FinitePMatchingCostLaws supplies powered-cost reversal, rooted candidate composition, the rooted distance laws and positive-lifetime separation. FinitePBoundaryTransportLaws transfers those laws to the real-plan optimum. FinitePMatchingCostBottleneck supplies the powered and rooted occurrence-count bounds, finite radius and cardinality-budget forms. FinitePFilterStability applies them to the finite parts of sublevel diagrams. FinitePMatchingCostComparison supplies first-order and two-real-exponent comparisons for matching and real-plan optima, including empty occurrence families, and their antitonicity on [1,∞). FinitePCostContinuity supplies continuity in the real exponent on [1,∞) for the powered and rooted matching and real-plan optima. FinitePCostSimilarity supplies common diagonal-translation invariance and nonnegative scaling for powered and rooted matching and real-plan optima. FinitePBottleneckLimit supplies the lower bottleneck comparison and real-exponent limits for matching and real-plan optima on diagrams with empty essential parts. FinitePBottleneckUniformLimit supplies the common absolute-error bound and uniform convergence on essential-free count/radius budget classes for both optima. FinitePCommonEssentialBottleneck supplies the same error and uniform conclusions when the essential-birth multisets agree, without essential count or range bounds. CechFinitePBottleneckUniformLimit derives both bounds and uniform limits for positive-degree Čech sublevel diagrams from a same-label displacement budget. CechEssentialBirth proves the degree-zero essential multiset, and CechFinitePBottleneckAllDegrees derives both bounds and uniform limits for every natural degree from the same input budget.

Figalli–Gigli retains priority for boundary-reservoir transport; its original bounded-domain quadratic Euclidean setting is distinct from this model. Divol–Lacombe corrected v4, formula (1), separates p from the ground norm q. This finite occurrence model uses q=∞ and real p≥1. Definition 2.1 and Proposition 3.2 concern measure transport; the latter's finite case invokes Lacombe–Cuturi–Oudot, Proposition 1 / Supplement §7.1. Their displayed convention ties the ground norm to p. Reusing the assignment proof with maximum-norm powered coefficients is a checked library adaptation, not a verbatim specialization of that displayed theorem.

DL Proposition 3.1 and its Appendix B metric-law proof use boundary gluing and Minkowski; FG Lemma 2.1 and Theorem 2.2 give the original quadratic antecedent. The finite occurrence allocation above is checked library reasoning corresponding to that mechanism. It uses pinned mathlib's Real.Lp_add_le, not a measure-gluing assumption or a triangle law for the raw augmented blocks.

The exponent comparisons are finite library derivations from DL formula (1) and pinned mathlib's power inequalities, rather than theorems quoted from DL. They use Real.add_rpow_le_rpow_add and Real.rpow_sum_le_const_mul_sum_rpow_of_nonneg on the matching charge vector. Probability-normalized Wasserstein exponent comparisons concern a different normalization and do not supply these unnormalized occurrence bounds.

Finite-exponent continuity is also a finite library derivation using the same DL convention. Pinned mathlib's finite-infimum identity and continuity theorem apply to the finite matching family; its real-power continuity includes zero bases at positive exponents. No continuity theorem for infinite diagrams or general persistence measures is inferred from this finite-minimum proof.

The perturbation estimates are checked finite sum-to-maximum reasoning from DL formulas (1)/(2) and the existing bottleneck attainment and stability theorems. DL Remarks 1.1/1.2 distinguish unstable point counts and the separate essential term. The formula conventions are compared on finite diagrams. Allowing essential points on the right of the bound is a one-sided library statement about the finite projection, not an identification of full essential p-transport conventions. The arbitrary-real-pair supporting algebra is a library extension; no all-measure stability theorem is inferred from this count estimate.

The exponent limit is a checked finite-diagram specialization of DL §1.1 formulas (1)/(2), at fixed q=∞. Its finite charge-vector squeeze is proved independently in the library. Supporting algebra also permits off-diagonal pairs below the diagonal; the source interpretation uses positive lifetimes. No limit theorem for infinite diagrams or general persistence measures is claimed.

The proved equality here concerns finite real reservoir matrices; the comparison with all Borel boundary plans is proved separately for unit-mass atomic measures (boundaryTransportPowerValue_ofIndexed_unit, see finite weighted boundary transport). Essential or infinite bars, arbitrary weighted inputs, completeness and barycenters remain outside it. Quotient and metric-space instances are not supplied. Existing first-order distance laws remain in the first-order matching module. The raw augmented cost is not an ordinary metric obtained by merely collapsing the diagonal.