Same-label displacement stability of Čech values and diagrams
Fix a pseudometric space X and two labelled families p, q : V → X on the same label type. Faces are finite label sets, as for the ambient Čech complex. Repeated labels and non-injective families are allowed. Points are compared label by label; the witness space X stays fixed.
Pointwise and face estimates
Let s be a finite label set, x ∈ X and ε ≥ 0, and suppose dist (p v) (q v) ≤ ε for every v ∈ s. For v ∈ s the triangle inequality gives
dist x (p v) ≤ dist x (q v) + dist (q v) (p v) ≤ radius_q(x, s) + ε.
The zero seed of the radius needs radius_q(x, s) + ε ≥ 0, which is where ε ≥ 0 enters. Hence radius_p(x, s) ≤ radius_q(x, s) + ε, and by symmetry the radii differ by at most ε. The sign condition matters only for s = ∅: there both radii are zero, the displacement hypothesis holds vacuously for ε = −1, and the inequality 0 ≤ 0 − 1 fails.
For a face s, nonempty by definition, the hypothesis at any v₀ ∈ s forces ε ≥ 0, and p v₀ shows that X is nonempty. For every x the Čech value of s under p is at most radius_p(x, s) ≤ radius_q(x, s) + ε. Taking the infimum over x gives čech_p(s) ≤ čech_q(s) + ε, and by symmetry |čech_p(s) − čech_q(s)| ≤ ε. The hypothesis concerns only the labels of s.
Displacement
For a finite V the displacement δ(p, q) is the largest of the distances dist (p v) (q v), computed as a maximum with zero seed. It is zero on an empty carrier and uses no supremum. It is nonnegative and symmetric, vanishes for p = q, bounds every label-wise distance, and satisfies δ(p, q) ≤ ε exactly when ε ≥ 0 and every label moves by at most ε. Hence every Čech value moves by at most δ(p, q).
Diagrams
For a finite linearly ordered V, the largest face difference of the two Čech filters is at most δ(p, q). The existing finite-face stability theorem then bounds the bottleneck distance between the Čech diagrams by δ(p, q) in every homological degree. The diagrams keep their occurrence multiplicities and essential bars at infinity. No injectivity, properness or cardinality hypothesis is added. An explicit bound ε ≥ 0 on every label-wise distance gives the same diagram bound with ε.
The finite diagram bound composes with the weighted Gaussian kernel embedding and with finite-part persistence images, without a label metric; see arctan persistence images and Čech consumers.
Comparison with Rips
Let d and e be label metrics agreeing with the ambient distances of p and q. For labels v, w the reverse triangle inequality for distances gives
|dist (p v) (p w) − dist (q v) (q w)| ≤ dist (p v) (q v) + dist (p w) (q w) ≤ 2δ(p, q).
So the Rips pairwise discrepancy is at most 2δ(p, q), every Rips diameter moves by at most 2δ(p, q), and the existing Rips stability theorem bounds the Rips diagrams by 2δ(p, q). Čech values are 1-Lipschitz and Rips values 2-Lipschitz in the displacement.
Fixtures
- Three points. On the line take p = (0, 1, 2) and q = (−1, 1, 3), so δ = 1. The triangle has Čech values 1 and 2, centred at 1, and Rips diameters 2 and 4. The Čech change equals δ and the Rips change and the pairwise discrepancy equal 2δ, so both constants are attained. The diagram bounds are 1 for Čech and 2 for Rips.
- One moved point. For p = (0, 2) and q = (0, 3), δ = 1. The edge has Čech values 1 and 3/2, so its change 1/2 is strictly below δ. The diagrams are within 1.
- Zero displacement. δ(p, p) = 0, so the diagram bound is zero. On the empty carrier δ = 0 and the diagrams coincide in every degree.
- Same image. For p = (0, 2) and q = (2, 0) the images coincide but δ = 2. The estimate depends on labels and is not a bound in terms of the images.
- Empty set. For ε = −1 on s = ∅ the radius estimate fails, so its sign condition cannot be dropped.
Interfaces
TDA.Filtration.CechDisplacement contains the radius and value estimates,
displacement with its laws, the displacement-form face estimate and the Rips comparison
pairwiseDiscrepancy_le_two_mul_displacement. TDA.Persistence.CechStability contains
faceSupDifference_filter_le, bottleneck_dgm_le_displacement,
bottleneck_dgm_le_of_dist_le and bottleneck_rips_dgm_le_two_mul_displacement. The
fixtures are in TDA.Filtration.CechDisplacementExamples and
TDA.Persistence.CechStabilityExamples.
Source correspondence and scope
Chazal–de Silva–Oudot prove in Corollary 4.10 that ambient Čech persistence modules of landmark sets L, L′ with common witnesses are ε-interleaved for ε > d_H(L, L′), and deduce in Theorem 5.6 the bound d_H(L, L′) on the bottleneck distance of their diagrams under total boundedness. The results here are a library extension consistent with Theorem 5.6: for a nonempty finite label type the Hausdorff distance between the images of p and q is at most δ(p, q). The library objects differ, since faces are label sets and labels may repeat, and the bound is proved through the finite-face stability theorem rather than exported from Corollary 4.10 or Theorem 5.6. The Rips factor 2 matches the factor of Theorem 5.2 informally. No Hausdorff or Gromov–Hausdorff statement, interleaving, change of witness space or relabelling is formalized by the same-label modules. The separate ambient relation and Hausdorff results compare different label types in a fixed ambient space and retain repeated coordinates.