Probability laws of sampled Rips diagrams

An iid tuple t : Fin n → X represents the finite metric subspace Set.range t. A point drawn several times appears once in this space. Rips.sampleDiagram k t q is its full degree-q Rips diagram over the field k. Rips time is diameter: an edge of length r enters at r.

For a probability mass function μ, Rips.sampleLaw k μ n q pushes the iid tuple law forward through this diagram map. Draws are independent and with replacement. The outer law has total mass one; the empty diagram is an ordinary possible outcome. Its mass cannot be discarded at the diagonal.

Couplings and distortion

Probability.Coupling μ ν is a probability mass function on X × Y with exact marginals μ and ν. It permits arbitrary splitting of mass. On finite input types this is a nonnegative matrix with prescribed row and column sums. Coupling.tuples draws iid pairs, and Rips.sampleLawCoupling pushes these pairs forward to a coupling of the diagram laws.

For a coupling π, the distortion moment is

\[ A_p(π)=\mathbb E_{a,b\sim π\text{ independently}} |d_X(a_X,b_X)-d_Y(a_Y,b_Y)|^p. \]

Probability.gromovWasserstein p μ ν is (1/2) inf_π A_p(π)^(1/p). The infinity definition uses the essential supremum on the support of π in each independent draw. Zero-mass input points therefore do not contribute.

Probability.wasserstein p bottleneck takes the infimum of the expected powered bottleneck cost over probability couplings, followed by the 1/p power. This is atomic transport between whole diagrams. It is distinct from transporting the counting measures of their bars.

Stability estimates

bottleneck_sampleDiagram_le proves that a bound on corresponding entries of two tuple distance matrices bounds their actual sample diagrams. The correspondence pairs values drawn at the same index, so it allows repetitions and different numbers of distinct sampled points.

For any input coupling π, transportCost_sampleLaw_le proves

\[ \inf_\gamma \mathbb E_\gamma[d_B(D,E)^p] \leq {n\choose2}\,A_p(π),\qquad p>0. \]

The proof bounds the power of the matrix maximum by the sum over unordered pairs of distinct positions. expectation_pairEnergy computes this sum's expectation exactly as binom(n,2) A_p(π). Taking infima gives wasserstein_sampleLaw_le_gromovWasserstein_allDegrees, for p ≥ 1, n ≥ 2, and every natural degree q, including degree zero:

\[ W_p(\operatorname{sampleLaw}(k,μ,n,q), \operatorname{sampleLaw}(k,ν,n,q)) \leq 2{n\choose2}^{1/p}\,GW_p(μ,ν). \]

The former wasserstein_sampleLaw_le_gromovWasserstein entry point remains the degree-q+1 specialization. At q = 0, the theorem compares full ordinary H₀ diagram laws; each nonempty sample retains its essential bar.

The separate theorem wassersteinInfinity_sampleLaw_le_gromovWassersteinInfinity_allDegrees proves W_∞ ≤ 2 GW_∞ for every sample size and every natural degree. It uses supported tuples and essential suprema, without a finite-p limit argument or an assumption that an optimal coupling exists. Neither atomic theorem assumes a compact ambient space, so both apply, for example, to the real-valued two-point law below on the noncompact space ℝ. The positive-degree wrappers wasserstein_sampleLaw_le_gromovWasserstein and wassersteinInfinity_sampleLaw_le_gromovWassersteinInfinity retain the specializations to degree q+1. At q = 0 the atomic laws are laws of full ordinary H₀ diagrams, not of reduced H₀ diagrams.

Finite-population total variation

For two probability mass functions on the same finite metric population, define δ(μ,ν) = ∑ₓ (μ(x) − min(μ(x),ν(x))). This is the mass absent from their common part. The finite half-L¹ formula gives the same value. It is symmetric, at most one, and zero exactly when the laws agree. The finite maximal coupling puts min(μ(x),ν(x)) on (x,x) and couples the two residual laws independently. Its probability of unequal coordinates is exactly δ. The zero-residual case uses the diagonal coupling directly; zero-mass atoms need no special hypothesis.

If all distances in the population are at most D, then distortion between two coupled pairs is at most D and vanishes when both pairs are diagonal. For p > 0, the resulting input bound is

\[ 2 GW_p(μ,ν) \le D(2δ(μ,ν))^{1/p}. \]

The atomic all-degree sampled-law estimate therefore gives, for p ≥ 1, resample size s ≥ 2, and every degree q,

\[ W_p(\operatorname{sampleLaw}(k,μ,s,q), \operatorname{sampleLaw}(k,ν,s,q)) \le D {s\choose2}^{1/p}(2δ(μ,ν))^{1/p}. \]

An empirical input law from m ≥ 1 training observations is the pushforward of the uniform law on their index set Fin m. Its mass at x counts every occurrence of x and divides by m; m and the resample size s are independent. Substituting this law for ν gives a deterministic error bound for the law of diagrams generated by future resampling. It does not bound the distance from one observed diagram to a population diagram, or supply a high-probability bound on the random training law. The full ordinary H₀ law retains its essential bar. This finite-TV deduction extends the existing sampled-law mechanism; Gómez–Mémoli does not state this empirical bound.

Compact Borel laws

DiagramBottleneckTopology equips Diagram ℝ with the extended metric induced by bottleneck distance. Distances may be infinite for diagrams with different essential-bar multiplicities. Ordered finite-coordinate presentations prove separability and second countability, so the measurable space is the Borel space of this topology.

For a compact metric space X with Borel probability measure μ, Rips.sampleLawMeasure k μ n q is the pushforward of the iid product measure on X^n through the actual finite-sample diagram map. lipschitzWith_sampleDiagram proves its Lipschitz constant is 2, hence it is continuous and measurable. Its range is compact, and the sampled law assigns that range probability one. In positive degree, compact diameter bounds the bottleneck distance from every sample diagram to the empty diagram by twice the diameter of X. Consequently every real power p > 0 has finite moment from that base diagram.

An isometric embedding preserves the law after pushing forward the input measure. On a finite input type, sampleLawMeasure_eq_toMeasure_sampleLaw identifies the Borel law of a finite PMF with the toMeasure of the existing atomic sample law. For compact-law stability, write U_{n,q}(μ) for sampleLawMeasure k μ n q, with one fixed coefficient field k and any q : ℕ. A probability coupling π of μ and ν is a measure on X × Y with those exact marginals. In the diameter-time convention used here, GW_p(μ,ν) is one half the infimum over such couplings of the p-th root of the distortion moment integrated against π ⊗ π. For compact Borel probability inputs, the all-degree theorem Rips.wasserstein_sampleLawMeasure_le_measureGromovWasserstein_allDegrees proves the finite-p bound for the actual probability laws on Diagram ℝ, using bottleneck as the ground cost:

\[ W_p(U_{n,q}(μ),U_{n,q}(ν)) \le 2{n\choose2}^{1/p}\,GW_p(μ,ν), \qquad p\ge1,\ n\ge2,\ q\in\mathbb N. \]

The proof pushes each input coupling through iid pair sampling and the Rips diagram maps, bounds its cost, and then takes the infimum; it does not assume an optimal coupling exists. The positive-degree wrapper Rips.wasserstein_sampleLawMeasure_le_measureGromovWasserstein retains the specialization to degree q+1.

The compact-law infinity theorem Rips.wassersteinInfinity_sampleLawMeasure_le_measureGromovWassersteinInfinity_allDegrees proves, for every sample size n and natural degree q,

\[ W_\infty(U_{n,q}(μ),U_{n,q}(ν)) \le 2\,GW_\infty(μ,ν). \]

Here both Wasserstein and Gromov–Wasserstein infinity use essential suprema; the diagram cost is bottleneck, and the input cost is the distortion under two independent draws from one arbitrary probability coupling. The proof pushes iid tuples through the sample-law coupling and bounds the diagram essential supremum by that coupling's distortion essential supremum before taking the infimum. The factor 2 comes from the one-half GW normalization. This result assumes no optimal coupling and uses no limit in finite p. The positive-degree wrapper Rips.wassersteinInfinity_sampleLawMeasure_le_measureGromovWassersteinInfinity retains the specialization to degree q+1. At q = 0, both compact-law bounds concern the actual full ordinary H₀ diagrams; for a nonempty sample they retain the essential (0,∞) occurrence. The bounds do not assert stability for a probability law of reduced H₀ diagrams.

Scope and remaining bridges

The atomic theorems use probability mass functions and the library's full finite Rips diagrams. The finite-range projection is surjective and preserves all distances; sampleProjection_eq_iff characterizes its fibers by zero pseudodistance, and sampleDiagram_eq_of_range_eq establishes independence from repeated labels. tupleFilter puts the closed-diameter filtration directly on the tuple indices, including repeated indices at zero pseudodistance. mem_tupleFilter_iff identifies its faces with the projected Rips faces on the finite metric sample. For any chosen linear order on that finite sample, tupleHomologyIso gives the natural persistence-module isomorphism in every degree and over every coefficient ring. The identification of reduced and unreduced positive-degree persistence remains separate work.

RipsReducedH0 defines reduced sublevel H₀ as the kernel of the natural augmentation induced by collapsing each Rips sublevel to a point. On a finite nonempty metric space, reducedSublevelH0Diagram_points_eq_finitePart identifies the reduced diagram points with the finite part of the ordinary H₀ diagram; reducedSublevelH0Diagram_points_eq_eraseEssential proves this removes exactly one (0, ⊤) occurrence while preserving all other multiplicities.

The compact Borel theorems use the independent MeasureCoupling transport definition directly on whole diagrams. They do not yet identify that transport infimum with the existing finite-PMF transport cost on finite support, and the finite-space reduced-H₀ bridge does not supply a reduced-diagram probability-law bound. The compact bounds above apply to full diagrams, including full H₀ at degree zero.

Why the whole support diagram is different

Consider the real-valued two-point law on 0 and 1 with masses 1−ε and ε. Against a singleton, its distortion moment is 2ε(1−ε), so the finite-p distortion GW distance is (1/2)(2ε(1−ε))^(1/p). The Lean regression proves that both real-law atoms are positive for 0 < ε < 1. It also uses the finite Bool metric chart induced by the same coordinates false ↦ 0, true ↦ 1: epsilonTwoPoint_fullH0_points gives its full H₀ diagram, with an essential component and a finite bar (0,1), and epsilonTwoPoint_bottleneck_singleton proves the chart's diagram is exactly bottleneck distance 1/2 from the one-point diagram. Thus the whole-support diagram stays nontrivial even as the rare atom's mass tends to zero.

The exact sample-law regression is first stated on that Bool metric chart with twoPointWeights. For n = 2, epsilonTwoPoint_sampleLaw_nontrivial_mass proves that the atom at the diagram with the finite H₀ bar has mass exactly 2ε(1−ε), and epsilonTwoPoint_sampleLaw_nontrivial_mass_tendsto_zero proves this mass tends to zero as ε → 0. The reusable sampleLaw_map_isometry theorem transports the full diagram law through an isometric embedding; its specialization epsilonTwoPoint_realSampleLaw_nontrivial_mass proves the same exact atom for the actual real-valued twoPointLaw. This transports the entire law for any sample size, but the explicit rare-atom formula here is for n = 2; no arbitrary-n atom formula is asserted. The full-support diagram therefore cannot replace the sampled diagram law.

Two degree-zero fixtures apply the atomic all-degree theorem to this input. weighted_twoPoint_sampleLaw_bound_degreeZero states the finite-p bound between the H₀ sample law of twoPointLaw and that of a one-point input, for every n ≥ 2. For n = 2 the source law has the rare atom of exact mass 2ε(1−ε) above; for 0 < ε < 1 this mass lies strictly between zero and one, so the source law is not a Dirac mass. twoPoint_pair_sampleLaw_bound_degreeZero compares the n = 2 H₀ sample laws of twoPointLaw at two weights ε and ε'. It proves W_p ≤ 2 GW_p between them, the exact rare-atom mass 2ε(1−ε) and 2ε'(1−ε') in each law, and that the two laws differ whenever these masses differ. It does not compute the GW distance between the two inputs. The positive-degree fixture weighted_twoPoint_sampleLaw_bound only instantiates the q+1 wrapper.

The singleton distortion identity is exposed as gromovWasserstein_of_subsingleton. A separate uniform two-point fixture at distance two proves that distinct draws have one essential and one finite full H₀ occurrence, while repeated draws retain only the essential occurrence. Each of these distinct diagrams receives at least half of the two-draw law's probability. uniformTwoPoints_sampleLaw_nonconstant therefore witnesses a nonconstant pushforward law with positive mass at both outputs. The ε-dependent full-support diagram, bottleneck obstruction, exact two-draw atom and vanishing-mass limit are formalized on the Bool metric chart described above.

Sources and reuse

The source is Gómez and Mémoli, Curvature Sets Over Persistence Diagrams, arXiv:2103.04470v5, Definitions 2.9–2.10 and 3.18, Theorems 3.8 and 3.19. The finite-p estimate combines their distance-matrix bound with Rips' diameter-scale Lipschitz constant two. The atomic and compact Borel probability-law bounds use full finite sample diagrams. Both the atomic and the compact Borel bounds apply in every degree; at degree zero they are laws of full ordinary H₀ diagrams. The separate finite Rips reduced-H₀ diagram correspondence is pointwise and does not assert compact reduced-law stability.

Pinned mathlib supplies probability mass functions, their monad laws, and extended-real sums and powers. TauCeti's OptimalTransport/Coupling and GromovWasserstein/Finite provide related measure-coupling and matrix-objective APIs. Their inspected revision uses a different Lean toolchain and a GW convention without the factor one half; no TauCeti code or dependency is imported here.