Total persistence and finite Rips–Čech bounds

For a real persistence diagram D, the degree-p total persistence is

Pers_p(D) = Σ |d − b|^p,

where the sum ranges over the finite bars (b,d) in D. Essential bars are excluded, as in the finite-part kernel embeddings. The exponent p is a natural number, and the absolute value makes the definition total on arbitrary diagrams. On a positive diagram, each summand is (d − b)^p. At p = 0, total persistence is the number of finite bars, including multiplicity.

Diagram laws

Total persistence is nonnegative and is zero when the finite part is empty. For c > 0, positive scaling of a diagram gives

Pers_p(cD) = c^p Pers_p(D),

and translating both endpoints by the same real number leaves Pers_p unchanged for every p. If every finite bar length is at most M, with 0 ≤ M, then

Pers_p(D) ≤ |D_fin| M^p,

where |D_fin| is the finite-bar count. The same estimate follows when both finite coordinates lie in [0,M]; positive bars then have length at most M.

Coordinates and face counts

For a finite sublevel diagram, every finite coordinate is a value of the filter on a face. Thus, if all face values lie in [0,M], finite birth and death coordinates lie in that interval, and essential birth coordinates lie in [0,M] as well.

For a Rips filter on a finite labelled metric space, M can be the diameter of the full carrier. For a Čech filter, M can be any enclosing radius around an ambient point. If the label metric is realized by the ambient distances, the label diameter is another valid choice. The enclosing radius formulation does not require a metric on the labels.

Fix a compatible face order. Each face supplies at most one endpoint occurrence across all degrees. A finite death in degree n can occur only at a face with n + 2 vertices, while a birth in degree n can occur only at a face with n + 1 vertices. Consequently,

# finite bars in degree n ≤ # Face K (n + 1),

and

# essential bars in degree n ≤ # Face K n.

For the full complex on N vertices these face counts are

# Face (full complex) (n + 1) = C(N, n + 2)

and

# Face (full complex) n = C(N, n + 1).

Combining the finite-bar count with the coordinate estimate gives the size-dependent bound

Pers_p(dgm k f n) ≤ C(N, n + 2) M^p.

For Rips diagrams, M is the carrier diameter. For Čech diagrams, it may be an enclosing radius or, under distance compatibility, the label diameter.

Kernel consumer

The Rips weighted-kernel estimate replaces the two diagram-dependent finite-bar counts by 2 C(N, n + 2). Thus, for a weight with stability constant C_σ,A,B, the embedding difference is bounded by

C_σ,A,B · (2 C(N, n + 2)) · pairwiseDiscrepancy(d,e).

This is a size-uniform statement for configurations on a fixed carrier size N; its constant still depends on N.

These bounds supply the card-times-maximum total-persistence factors in KFH18 eqs. (8)–(9). In the p = 1 use of eq. (9), the factor Pers_0 is exactly the finite-bar count. The source's card-independent total-persistence estimate in Lemma 4 is not obtained here. In particular, this note does not formalize assumption (W1), Proposition 8, Theorem 9, or Corollary 12. It also does not cover real exponents. The vanishing of essential bars in degree n ≥ 1 for Rips and Čech diagrams is proved in Essential bars on the full complex.

Essential bars on the full complex

For every filter f whose face values are bounded by a threshold t, card_essentialPart_dgm_eq_finrank proves the public identity

# essential bars in dgm k f n = finrank k H_n(K).

Finite complexes always supply such a threshold. On the full complex, every filter has no essential bars in degree n ≥ 1, by essentialPart_dgm_fullComplex_eq_zero; this applies to Rips and Čech filters, including Čech filters in any pseudometric ambient and with no injectivity or label-metric hypothesis. For a nonempty vertex type there is exactly one essential bar in degree zero, by card_essentialPart_dgm_fullComplex_zero. Thus the general essential face-count bound # Face K n sharpens on the full complex to 0 in positive degree and 1 in degree zero. In positive degree, dgm_points_fullComplex_eq_finite identifies the whole diagram with the finite-part embedding.

At large scale, a Rips sublevel with r ≥ Rips.diameter d Finset.univ is the full complex, and a Čech sublevel with r ≥ Cech.radiusAt p x Finset.univ is the full complex. Their positive-degree homology is therefore zero, by Rips.isZero_sublevelHomology_of_diameter_le and Cech.isZero_sublevelHomology_of_radiusAt_le.

Total persistence and the kernel/image embeddings already exclude essential bars, so their merged bounds do not change value. For Rips and Čech in degree n ≥ 1, the new vanishing theorem proves that this exclusion loses nothing. The count card_finiteOccurrences_dgm_fullComplex_le yields the size-uniform Čech weighted-kernel and persistence-image bounds and the Rips persistence-image bound, each with the factor 2 C(N, n + 2); the consumer declarations are Cech.norm_diagramEmbedding_sub_le_choose_mul_displacement, Cech.norm_finiteDiagramImage_sub_le_choose_mul_displacement, and Rips.norm_finiteDiagramImage_sub_le_choose_mul_pairwiseDiscrepancy. No KFH18 statement beyond the existing consumed bounds is claimed.

TDA.Persistence.EssentialVanishingExamples checks the triangle, the two-point degree-zero control with a nonempty essential part, repeated labels, and the hollow-triangle negative control: it has an essential degree-one bar and is not an apex cone at any vertex. The filled triangle recovers its degree-one rank zero by the essential-bar route and by transport to the full complex. The final fixtures instantiate the size-uniform Čech weighted-kernel and persistence-image consumers.

Fixtures

The checked examples use coefficients in ZMod 2.

  • The two-point Rips degree-zero diagram has one finite bar of length 2; its total persistence is 2^p, and the face-count bound is attained. The two-point Čech diagram has one finite bar of length 1.
  • On the library triangle, the degree-one finite-bar count for both Čech and Rips is at most C(3,3) = 1. The Čech degree-one first-order total persistence is at most 3. In degree two, both finite parts have count zero by C(3,4) = 0.
  • Repeated Čech labels have zero spread, so their finite part and total persistence vanish in every degree and exponent. This uses only the Čech filter and no label metric.
  • The filled real triangle has degree-one bar (1,2). Scaling its filter by 2 changes its first-order total persistence from 1 to 2 and its second-order total persistence to 4.
  • The two-point Rips degree-zero diagram has one essential bar, satisfying the general essential count bound, and its exponent-zero total persistence is the finite-bar count 1.

Lean interfaces

TDA.Persistence.TotalPersistence defines total persistence and its diagram-level laws. TDA.Persistence.TotalPersistenceRipsCech proves coordinate, face-count, Rips, Čech and size-uniform weighted-kernel bounds. TDA.Persistence.TotalPersistenceRipsCechExamples contains the fixtures above. TDA.Persistence.EssentialVanishing proves the essential-count identity, full-complex vanishing and size-uniform consumers, with fixtures in TDA.Persistence.EssentialVanishingExamples. The endpoint-degree facts used by the face-count argument are recorded in TDA.Persistence.FaceEndpointUniqueness.