Distance-matrix stability of Rips persistence images

Fix two metrics d,e on the same finite labelled vertex set V, a coefficient field k and a homological degree n. Use the diameter-time Rips filtration: vertices enter at zero, thresholds are closed, and stages below zero are empty. Let D_d,D_e be the diagrams of sublevel homology, and write

δ = max { |d(v,w) − e(v,w)| : v,w ∈ V },

with empty maximum zero. The existing sublevel-diagram theorem bounds their full bottleneck distance by δ, including their essential bars.

Finite image and bound

For a positive diagram D, let D_fin be its finite birth–death occurrences and N_D their number. Each pair satisfies birth < death. Equal pairs keep separate occurrences, so N_D and every image sum count multiplicity. Define PI_fin(D) by applying the existing Gaussian image to D_fin. Essential bars are omitted from this feature; their deaths are never replaced by finite cutoffs.

Fix the same bandwidth σ > 0, weight w and pixels for both inputs. On the birth–persistence plane, require 0 ≤ w ≤ B, maximum-norm Lipschitz coefficient A, and w(b,0)=0, with A,B ≥ 0. Pixels form a finite measurable pairwise disjoint family; the family may be empty. Set C = 2A + 3B sqrt(2/π)/σ. If N_d,N_e are the finite counts of D_d,D_e, then

‖PI_fin(D_d) − PI_fin(D_e)‖₁ ≤ C (N_d + N_e) δ.

This explicit nonsharp pixel-vector bound allows different finite-bar counts, changed pairings, coincident critical values and empty V. Counts are measured on the resulting diagrams; no bound in terms of the vertex count is assumed. The coefficient is independent of the number of pixels. No uniform cardinality-free bottleneck estimate is asserted.

From the matching maximum to the finite sum

More generally, let D,E have finite full bottleneck β. The existing occurrence matching has a finite partial injection and a separate essential-occurrence bijection. Its type already forbids finite/essential cross-matches. Keeping the finite map and replacing the essential types by the empty type preserves every finite charge: one pair or diagonal charge per source, and one diagonal charge per unmatched target.

Every charge is at most the full matching maximum. There are at most N_D+N_E charges, so the restricted total cost is at most (N_D+N_E) times that maximum. Empty sums and maxima are zero. Each individual matching has finite cost because all its occurrence types are finite and all coordinates, including essential births, are real. This justifies monotonicity when converting its costs from ℝ≥0∞ to ℝ.

Finite β implies equal essential cardinalities and supplies an attained optimal full matching through the existing API. Use its restriction as a candidate for the finite first-order minimum m₁. Thus

m₁(D_fin,E_fin) ≤ (N_D+N_E) β.

The finite-bottleneck hypothesis is necessary. A diagram with finite pair (0,2) and one essential bar, compared with the empty diagram, has full bottleneck ⊤ and finite first-order cost 1. In Lean, ENNReal.toReal ⊤ = 0; an unrestricted real version of the bound would therefore be false. The Rips bound is below ENNReal.ofReal δ, hence finite, so its consumer needs no extra essential-cardinality hypothesis.

Apply the existing bound ‖PI(x)−PI(y)‖₁ ≤ C m₁(x,y), then the finite-sum comparison and the Rips bottleneck bound. Positivity of diagrams supplies every finite positive-length hypothesis. Pairing changes require no fixed template.

Multiplicity explains the count factor: m+1 copies of (0,2) against the empty input have first-order cost m+1 and bottleneck 1. Essential-only diagrams have zero finite image. These cases and a nonzero two-point metric perturbation exercise the interfaces.

Interfaces and source scope

FiniteMatchingCostBottleneck supplies OccurrenceMatching.finiteRestriction, its sum-to-maximum estimate and finiteFirstOrderMatchingCost_le_card_mul_bottleneck. PersistenceImageDiagram supplies finiteDiagramImage, its empty-finite-part formula and count-dependent bottleneck estimate. RipsPersistenceImage supplies norm_finiteDiagramImage_sub_le_pairwiseDiscrepancy, with the same LinearOrder V, Fintype V and k : Type u field assumptions as the existing Rips API for V : Type u. Ordinary TDA.Persistence includes these noncomputable mathematical APIs.

Adams et al., §4 Definitions 1–2 and §5 Theorems 9–10, provide the Gaussian PI construction and first-order stability route. The library's maximum-norm weight condition and nonsharp C are its proved variant. Remark 6's lack of a uniform bottleneck constant is consistent with the count factor here. Divol–Lacombe corrected v4, formulas (1)–(2) and Remark 1.2, distinguish finite matching costs and essential contributions. Chazal–de Silva–Oudot v3, §4.2.1, supplies the nonnegative-scale diameter convention; the library retains its stated different negative-time convention. The finite count comparison and composed Rips-to-PI estimate are library results, not literal theorems attributed to these papers. Registered versions and rights remain those in the source register.

Numerical integration, executable distance-matrix reduction, first-order essential costs, measure-coupling equality, general p, different-point-set/Gromov–Hausdorff stability and Čech comparisons require separate results. The same composition for the RKHS norm of the weighted Gaussian kernel embedding is in the arctan weight and Rips consumers. The arctan image constant, the arctan Rips image instance and the Čech image consumer are in arctan persistence images and Čech consumers.

The size-uniform Rips persistence-image corollary is summarized in Essential bars on the full complex.