Arctan persistence images and Čech kernel and image consumers
This note evaluates the persistence-image stability constant at the arctan weight and composes the same-label Čech displacement bound with the weighted Gaussian kernel embedding and finite-part persistence images, in the same way the Rips consumers compose the Rips bottleneck bound.
The image constant
The persistence-image stability constant is 2A + 3B·√(2/π)/σ, where A is the birth–persistence Lipschitz constant of the weight, B its sup bound and √(2/π)/σ the translation constant of the 2D Gaussian density. The arctan weight has A = C and B = π/2. Since (π/2)·√(2/π) = √(π/2), the constant is
2C + 3·√(π/2)/σ.
Both forms agree for every σ, including σ = 0, where Lean's division by zero makes both terms zero. The arctan image bounds against the finite first-order matching cost and, for positive diagrams with finite bottleneck, against the occurrence count times the bottleneck distance, are instances of the existing image theorems with this constant. The constant is not sharp. At C = 0 the weight is zero, so every image and every embedding vanishes, while the image constant 3·√(π/2) and the kernel constant √2·π/2 at σ = 1 are positive.
Bandwidth hypotheses
The image theorems keep 0 < σ. It is used where the Gaussian density integrates to one, which bounds each pixel vector by one, and where the translation constant is nonnegative. The kernel route has no bandwidth sign hypothesis, since its kernel-function distance bound holds for every σ. The image items below therefore carry 0 < σ, and the kernel items do not.
Čech consumers
Fix labelled configurations p, q : V → X in one pseudometric space on the same finite ordered label type, a coefficient field and a degree. Let D_p, D_q be the Čech diagrams, N_p, N_q their numbers of finite occurrences, and δ the same-label displacement, the largest distance dist(p v, q v). The Čech bound gives full bottleneck at most δ, so the bottleneck is finite and its real value is at most δ. With the count-times-bottleneck theorems this gives
‖Φ(D_p) − Φ(D_q)‖ ≤ (2A + √2·B/|σ|)(N_p + N_q) δ
for the weighted-kernel embeddings Φ and every weight, the same bound divided by |τ| for the second-level Gaussian features and for |K(D, D_p) − K(D, D_q)| with an arbitrary fixed diagram D, and
‖I(D_p) − I(D_q)‖ ≤ (2A + 3B·√(2/π)/σ)(N_p + N_q) δ
for the finite-part persistence images I with 0 < σ and measurable pairwise disjoint pixels. The arctan instances replace the constants by 2C + √2·π/(2|σ|) and 2C + 3·√(π/2)/σ. There is no arctan instance for the second-level Gaussian features, as on the Rips side.
Essential bars stay in the bottleneck comparison but are omitted from the embedding and the image. Occurrence multiplicities are retained throughout. The count factor N_p + N_q is the one of the Rips consumers. Zero displacement, either from a configuration against itself or from the empty label type, makes all three bounds zero.
No distance-compatible label metric is assumed. The Rips route needs metrics d, e on the labels with d(v, w) = dist(p v, p w), which fail to exist when two labels sit at the same point. For the configuration with both labels at 0 there is no such metric, so the Rips consumers do not apply; against the points 0, 2, with displacement 2, the Čech consumers do.
The size-uniform Čech weighted-kernel and persistence-image corollaries are summarized in Essential bars on the full complex.
Sup-norm consumers
In the sup-metric product ι → ℝ with a possibly empty finite ι, let d and e be distance-compatible metrics for p, q : V → ι → ℝ, so d.dist v w = dist (p v) (p w) and likewise for e and q. With a finite ordered label type V, a field k and degree n, write N_p and N_q for the finite-occurrence counts of the two Čech diagrams and put
Δ = Rips.pairwiseDiscrepancy d e / 2.
Cech.bottleneck_dgm_le_half_pairwiseDiscrepancy gives the full bottleneck bound, and
Cech.bottleneck_dgm_lt_top_supNorm and
Cech.toReal_bottleneck_dgm_le_half_pairwiseDiscrepancy give its finiteness and real
conversion. The embedding, second-level Gaussian feature and fixed-diagram kernel-value
consumers are Cech.norm_diagramEmbedding_sub_le_half_pairwiseDiscrepancy,
Cech.norm_diagramGaussianFeature_sub_le_half_pairwiseDiscrepancy and
Cech.abs_gaussianDiagramKernel_sub_le_half_pairwiseDiscrepancy. Each has the matching
Rips constant times (N_p + N_q) times Δ; the feature and kernel-value constants also carry
the factor 1/|τ|. All consumer items assume hA : 0 ≤ A, hB : 0 ≤ B and a weight
PersistenceImage.Weight A B. The image consumer
Cech.norm_finiteDiagramImage_sub_le_half_pairwiseDiscrepancy has the matching image
constant and requires 0 < σ, measurable pixels and pairwise disjoint pixels. The kernel,
embedding and feature items require only nonnegative weight bounds, with no sign hypothesis
on σ or τ. The arctan instances
Cech.norm_diagramEmbedding_sub_le_half_pairwiseDiscrepancy_arctan,
Cech.abs_gaussianDiagramKernel_sub_le_half_pairwiseDiscrepancy_arctan and
Cech.norm_finiteDiagramImage_sub_le_half_pairwiseDiscrepancy_arctan have constants
2C + √2·π/(2|σ|), (2C + √2·π/(2|σ|))/|τ| and
2C + 3·√(π/2)/σ; the image instance also keeps 0 < σ. There is no arctan feature instance.
The displacement route above needs no label metric. This sup-norm route needs d and e, but
Cech.pairwiseDiscrepancy_le_two_mul_displacement gives Δ ≤ δ, where δ is the same-label
displacement. The moved diagonal pair has Δ = 1/2 and δ = 1, so the halved constants are
strictly smaller there. On the line, only the bottleneck statement
Cech.bottleneck_dgm_le_half_pairwiseDiscrepancy_real is provided; no line consumer
theorems are stated.
Rips versus Čech of one configuration
For a real diagram D and c > 0, bottleneck_self_scale_le bounds the bottleneck distance
from D to D scaled by c by ENNReal.ofReal (|1 − c| · M), when M bounds the absolute values
of both coordinates of every finite point and of every essential birth coordinate. The
proof matches each occurrence to its own scaled image. The bound is sharp for the single
finite bar (0, 2) at c = 1/2: the scaled point is (0, 1), and the bottleneck distance is
exactly 1.
The coordinates of diagrams are critical values by
mem_criticalValues_of_mem_finitePart_dgm and
mem_criticalValues_of_mem_essentialPart_dgm. For a Rips filter, every such value lies in
[0, diameter of the whole carrier] by Rips.filter_value_le_diameter_univ. Combining this
with the sup-norm scaling identity gives
Cech.bottleneck_rips_cech_dgm_le_half_diameter in ι → ℝ and
Cech.bottleneck_rips_cech_dgm_le_half_diameter_real on the line: the Rips and Čech
diagrams of one configuration are at bottleneck distance at most half the carrier
diameter.
In the sup-norm product, let N_d and N_p count finite occurrences of the Rips and Čech
diagrams. The consequences
Cech.norm_diagramEmbedding_rips_sub_cech_le,
Cech.abs_gaussianDiagramKernel_rips_sub_cech_le and
Cech.norm_finiteDiagramImage_rips_sub_cech_le bound the embedding, fixed-diagram
kernel-value and finite-part image differences by the corresponding constant times
(N_d + N_p) times Rips.diameter d Finset.univ / 2. The embedding and kernel-value items
require nonnegative weight bounds; the image item also requires 0 < σ, measurable pixels
and pairwise disjoint pixels. The Euclidean triangle fixture gives Čech_{6/5} ≠ Rips_{12/5},
so the halving identity and these additive bounds are not stated for EuclideanSpace or a
general ambient.
Boundaries
The diagram-level image bound needs finite bottleneck. A diagram with the finite point (0, 2) and one essential bar, compared with the empty diagram, has bottleneck ⊤. With one pixel covering the plane, σ = 1 and C = 1, its arctan image difference has norm arctan 2, but the real value of ⊤ in Lean is 0, so the count-times-bottleneck bound would read 0. The Čech and Rips consumers discharge the hypothesis by their bottleneck bounds.
The additive Rips-versus-Čech kernel and image bounds are given in Rips versus Čech of one configuration, but only in the sup-norm product and on the line. The log-scale consumers are described in Log-scale Rips versus Čech.
Log-scale Rips versus Čech
TDA.Persistence.CechRipsLogScaleConsumers has a generic layer for two filters on one finite
complex. If g ≤ f ≤ c · g, 1 ≤ c, and g is PositiveOnEdges, then in every degree
n ≥ 1 the log-scale bottleneck is at most log c. The layer gives the corresponding
count-times-log c bounds for the weighted Gaussian embedding, its second-level Gaussian
feature, a fixed-diagram Gaussian kernel value, and the finite-part persistence image.
For one finite configuration in a general pseudometric ambient, with a compatible label metric,
the exact log-scale Rips-versus-Čech bottleneck gives the same four consumer bounds, and the
log 2 bottleneck bound gives versions stated with the radius-level occurrence counts through
the identity
Fintype.card (logDgm k f n).finiteOccurrences = Fintype.card (dgm k f n).finiteOccurrences
for n ≥ 1 and PositiveOnEdges f. The arctan embedding, fixed-diagram kernel-value, and
image instances are included; there is no arctan feature instance. The restriction n ≥ 1
is required because the degree-zero log-scale module is not constructible on a finite positive
log grid, so no degree-zero log 2 consumer is available.
TDA.Persistence.CechRipsLogScaleConsumersExamples records the two-point, triangle, filled
triangle, zero-weight, degree-zero, and repeated-label fixtures for these statements.
Lean interfaces and sources
TDA.Persistence.ArctanPersistenceImage proves PersistenceImage.stabilityConstant_arctan,
the two arctan image bounds, PersistenceImage.finiteDiagramImage_eq_sum and the arctan
Rips image consumer. TDA.Persistence.CechWeightedKernel proves
Cech.bottleneck_dgm_lt_top, Cech.toReal_bottleneck_dgm_le_displacement, the three kernel
consumers and the arctan embedding and kernel-value instances.
TDA.Persistence.CechPersistenceImage proves the image consumer and its arctan instance.
TDA.Persistence.CechSupNormWeightedKernel proves
Cech.bottleneck_dgm_lt_top_supNorm,
Cech.toReal_bottleneck_dgm_le_half_pairwiseDiscrepancy, the embedding, feature and
kernel-value consumers, their arctan embedding and kernel-value instances, and
Cech.bottleneck_dgm_le_half_pairwiseDiscrepancy_real for compatible label metrics in
ι → ℝ and on the line. TDA.Persistence.CechSupNormPersistenceImage proves
Cech.norm_finiteDiagramImage_sub_le_half_pairwiseDiscrepancy and
Cech.norm_finiteDiagramImage_sub_le_half_pairwiseDiscrepancy_arctan.
TDA.Persistence.CechRipsAdditive proves the additive
scaling bound, the critical-value coordinate lemmas, the carrier-diameter bound for Rips
values, and the half-diameter Rips-versus-Čech bottleneck bounds in the sup-norm product
and on the line, with the consumer consequences in the product. Its consumers use the
corresponding nonnegative weight bounds, and its image consumers additionally require
0 < σ and measurable pairwise disjoint pixels.
TDA.Persistence.CechSupNormConsumersExamples covers the moved diagonal pair, strict
Δ ≤ δ, zero weight, zero discrepancy, the line half-discrepancy bound, the sharp single
bar, two points, the sup-norm triangle and the Euclidean counterexample. The additive
scaling bound, critical-value description and Rips-versus-Čech comparison are elementary
library results and are not attributed to a paper.
TDA.Persistence.CechRipsLogScaleConsumers proves the generic log-scale layer, the exact
log-scale bottleneck consumers, the log 2 Rips-versus-Čech consumers, and their arctan
instances in a general pseudometric ambient. TDA.Persistence.CechRipsLogScaleConsumersExamples
provides the corresponding fixtures.
Kusano–Fukumizu–Hiraoka (CC-BY 4.0; see the source register for attribution and license) is used for statement shapes: §3.2 p. 14 (w_arc), §3.2 p. 16 (assumption (W2), eq. (10), the p = 1 remark B₂ = π/2, L₂ = 2C, and the w_arc line of Corollary 11) and §3.3 eq. (13). Adams et al. (CC-BY 4.0; see the source register for attribution and license), §4 Definitions 1–2 and the §5 stability route, supply the image construction. AE17 uses no arctan weight: the arctan persistence image and its constant are a library composition of KFH18's weight with AE17's construction and the library's nonsharp constant. The ambient Čech complex is that of Chazal–de Silva–Oudot §4.2.3, as in the displacement note. The Čech-to-kernel and Čech-to-image compositions are library results and are not attributed to any paper. The statements adapt both sources: the weight is truncated below the diagonal, C = 0 is allowed, the constant is the library's nonsharp one, and the costs are the library's finite matching cost and count-times-bottleneck rather than a 1-Wasserstein distance. Proofs are independent.
Not covered here: assumption (W1) and total-persistence bottleneck stability (Proposition 8, Theorem 9, Corollary 12), Propositions 6–7, the w_pers and w_one lines of Corollary 11, p ≠ 1, line consumers of the half-discrepancy bound, a per-point sharp additive bound, Rips-versus-Čech bounds outside the sup-norm product and the line, explicit Čech or Rips diagram points, and Hausdorff, Gromov–Hausdorff or different-label stability.