Log-scale diagrams in degrees n ≥ 1
This note gives the log-scale modules of real filters on a finite complex a persistence diagram in every degree n ≥ 1. The module is shown to be finite-constructible on the logarithms of the positive critical values, so the finite-constructible isometry applies to it, as it does to the radius-level sublevel modules. The log-scale Čech–Rips interleaving then becomes a bottleneck bound log 2 between log-scale diagrams.
Conventions
The log-scale module of a real filter f at t is the homology Hₙ(f_{exp t}; R) of the sublevel complex at exp t, with homology inclusions as structure maps. Vertex values are not assumed to be 0. The filter f is positive off vertices when every face with at least two vertices has positive value; by monotonicity this is the same as positivity on edges. Positivity passes from g to every filter f with g ≤ f.
The positive-log grid of f is the increasing enumeration of the finite set
{log c | c a critical value of f with c > 0}. Nonpositive critical values, in particular the
vertex value 0 of Rips and Čech filters, are excluded, since their logarithm is undefined.
The log-scale diagram logDgm k f n is the module diagram of the log-scale module on this
grid, over a field k. It is defined for every filter and degree. It represents the module
only when the module is constructible on the grid, which is proved below for n ≥ 1 under
positivity off vertices.
Vanishing of higher homology
The n-chains of a complex are finitely supported functions on its faces with n + 1 vertices. If there is no such face, the chain module is zero, and so is the homology in degree n, over every ring. A face with n + 1 vertices has at least two vertices when n ≥ 1, so a complex whose faces have at most one vertex has zero homology in every degree n ≥ 1.
Constructibility on the positive-log grid
Let s ≤ t lie in the same cell of the positive-log grid. Then f_{exp s} ⊆ f_{exp t}. Conversely, let σ be a face of f_{exp t} with value v. If v ≤ 0, then v < exp s. If v > 0, then log v is a grid value at most t, so it is counted in the cell of t, which equals the cell of s; hence log v ≤ s and v ≤ exp s. The two sublevels are equal. This step needs no positivity.
Let t lie below every grid value and assume positivity off vertices. A face of f_{exp t} with at least two vertices has positive value v, so log v is a grid value and exp t < v, which is impossible. The sublevel at exp t therefore contains only vertices, and its homology vanishes in every degree n ≥ 1. This is the zero left tail. Together with the equality of sublevels within a cell, the log-scale module is finite-constructible on the positive-log grid in every degree n ≥ 1, over every ring.
Over a field, every fiber is finite-dimensional, since the sublevel has finitely many faces.
The log identity in degrees n ≥ 1
For a real filter f, positiveGrid f is the exponential image of logGrid f. Its range is
exactly the positive critical values of f, and it refines the full critical-value grid. Under
1 ≤ n and PositiveOnEdges f, sublevelModule_isConstructible_positiveGrid proves that the
radius-level sublevel module is constructible on this positive grid.
The order embedding used to compare the two diagrams is
expEmb : ℝ ↪o ℝ. The reindexing identity for module diagrams,
moduleDiagram_reindex_map, and refinement to the full critical-value grid give
(logDgm k f n).map expEmb = dgm k f n.
This is logDgm_map_exp; it holds for a field k, 1 ≤ n, and
PositiveOnEdges f. Thus the radius-level births and finite deaths are obtained by
exponentiating the log-scale coordinates. Equivalently, logDgm_points_eq_map_log applies
Real.log to both coordinates of every radius-level point and leaves an essential death at
⊤. Under the same hypotheses, dgm_birth_pos and dgm_death_pos show that every birth and
death is positive, with ⊤ included among the positive deaths.
These comparison, constructibility, and positivity results are library extensions; no source theorem is cited for them.
Bottleneck bounds
For two filters f and g on the same finite complex, positive off vertices, the isometry gives, in every degree n ≥ 1,
d_B(logDgm k f n, logDgm k g n) = d_I(log-scale module of f, log-scale module of g).
The two positive-log grids are unrelated. If g ≤ f ≤ c · g with 1 ≤ c and g is positive off vertices, then f is positive off vertices too, and the log c-interleaving of the log-scale modules bounds the bottleneck distance by log c. For labelled points p in a pseudometric space and a label metric d with d(v, w) = dist(p v, p w), the Rips filter of d is positive off vertices, because distinct labels have positive distance and the diameter dominates every pairwise distance. The Čech filter is positive off vertices because rips ≤ 2 · čech. With čech ≤ rips ≤ 2 · čech, the log-scale Rips and Čech diagrams are within log 2 in every degree n ≥ 1, without properness of the ambient space. The resulting embedding, Gaussian-kernel, and persistence-image consumers are summarized in Log-scale Rips versus Čech.
The degree-0 obstruction
In degree 0 the argument fails at the zero left tail. Let some vertex have value at most 0. For any finite grid, choose t below every grid value. The vertex lies in f_{exp t} because exp t > 0, and it gives a nonzero class in H₀(f_{exp t}; k). Hence the fiber at t is not zero, and the log-scale module in degree 0 is not constructible on any finite grid. For Rips and Čech filters every vertex has value 0, so every vertex class is born at −∞ on the log scale.
Fixtures
- Positivity. The two-point and triangle Rips and Čech filters are positive off vertices.
- Comparison. Over ZMod 2, in every degree n ≥ 1, the triangle log-scale Rips and Čech modules are constructible on their positive-log grids, and their diagrams are within log 2; the same bound holds on two points over ZMod 2 and ℚ.
- Triangle at log (6/5). No face with two vertices lies in the triangle Rips sublevel at 6/5, so the degree-1 log-scale Rips module vanishes at log (6/5). At the same time the three Čech edges are present and the Čech 2-face is absent.
- Above the dimension. On two vertices no face has three vertices, so in every degree n ≥ 2 the log-scale modules vanish, both diagrams are empty and their bottleneck distance is 0.
- Dimension filters. The filled and hollow triangles, with value card − 1 cast to ℝ, are positive off vertices and their degree-1 log-scale modules are constructible. Scaling the filled triangle by 2 moves its degree-1 log-scale diagram by at most log 2.
- Degree 0. The two-point Rips and Čech log-scale modules in degree 0 are constructible on no finite grid, in particular not on their positive-log grids.
- Repeated labels. Two distinct labels at one point give a Čech edge of value 0, so the Čech filter is not positive off vertices. No label metric is compatible with this configuration.
Explicit two-point diagrams
The two-point radius-level diagrams over ZMod 2 are
(dgm (ZMod 2) (Rips.filter TwoPoints.metric) 0).points = {(0, ⊤), (0, 2)},
(dgm (ZMod 2) (Cech.filter TwoPoints.pts) 0).points = {(0, ⊤), (0, 1)}.
These are twoPoints_rips_dgm_zero and twoPoints_cech_dgm_zero. The positive-degree
radius-level emptiness is stated by twoPoints_rips_dgm_pos and twoPoints_cech_dgm_pos;
the corresponding ZMod 2 log-scale statements are twoPoints_rips_logDgm_pos and
twoPoints_cech_logDgm_pos.
In every degree n ≥ 1, the two radius-level point multisets are empty over every field. The
log-scale point multisets are also empty in every such degree for the ZMod 2 Rips and Čech
fixtures. For these ZMod 2 diagrams, twoPoints_bottleneck_zero gives the
exact radius-level bottleneck 1 in degree 0, while twoPoints_bottleneck_pos gives 0 in
every degree n ≥ 1. The exact log-scale value is 0 in every degree n ≥ 1, by
twoPoints_logBottleneck_pos. No degree-0 log-scale bottleneck is stated.
The fixtures also record the hypotheses needed by the comparison. On the edge fixture,
constEdge_dgm_zero shows that the monotone law must filter points that become diagonal:
the finite bar (5, 9) maps to (0, 0), leaving {(0, ⊤)}. The translated filled triangle
has shiftedReal_dgm_one equal to {(-1, 0)}, shiftedReal_not_positive, and
shiftedReal_birth_neg; positivity is therefore needed for the positive-coordinate conclusion.
Finally,
twoPoints_rips_dgm_zero_birth exhibits a birth at 0 even though the Rips filter is positive
on edges; the hypothesis 1 ≤ n therefore cannot be dropped.
The two-point diagrams, bottleneck values, and necessity fixtures in this section are library extensions; no source theorem is cited for them.
Explicit triangle diagrams
The acute library triangle Triangle.pts has vertices (0,0), (2,0), (1,2),
circumcentre (1, 3/4), and circumradius 5/4. Its exact face values are
| face | Čech | Rips |
|---|---|---|
| a vertex | 0 |
0 |
{0,1} |
1 |
2 |
{0,2} |
√5/2 |
√5 |
{1,2} |
√5/2 |
√5 |
{0,1,2} |
5/4 |
√5 |
The entries are proved by cech_value_singleton, triangle_cech_value_edge01,
triangle_cech_value_edge02, triangle_cech_value_edge12, triangle_cech_value_tri,
triangle_rips_value_edge01, triangle_rips_value_edge02, triangle_rips_value_edge12,
and triangle_rips_value_tri.
Over ZMod 2, the certified filled-triangle pivot computation gives the radius-scale
diagrams (the cechT and ripsT filters are defined in the examples module):
(dgm (ZMod 2) cechT 0).points = {(0, ⊤), (0, 1), (0, √5/2)},
(dgm (ZMod 2) cechT 1).points = {(√5/2, 5/4)},
(dgm (ZMod 2) ripsT 0).points = {(0, ⊤), (0, 2), (0, √5)},
(dgm (ZMod 2) ripsT 1).points = 0.
These are triangle_cech_dgm_zero, triangle_cech_dgm_one, triangle_rips_dgm_zero,
and triangle_rips_dgm_one. The Rips degree-1 pivot pair is (√5, √5) and is removed
by the diagonal filter, as recorded by triangle_rips_pair_diagonal. At log scale in
degree 1, triangle_cech_logDgm_one gives {(½ log (5/4), log (5/4))}, while
triangle_rips_logDgm_one gives the empty diagram. No degree-0 log-scale statement is made.
With Rips first and Čech second, the exact bottleneck values are √5/2 in degree 0 at
radius scale (triangle_bottleneck_zero), (5/4 − √5/2)/2 in degree 1 at radius scale
(triangle_bottleneck_one), and ¼ log (5/4) in degree 1 at log scale
(triangle_logBottleneck_one). The latter is also the log-scale interleaving distance by
triangle_logInterleaving_one, and is strictly below log 2 by
triangle_logBottleneck_one_lt_log_two.
The right-triangle control RightTriangle.pts has vertices (0,0), (2,0), (2,2);
its Čech 2-face enters with the hypotenuse (rightTriangle_cech_value_tri_eq_edge02),
both degree-1 diagrams are empty (rightTriangle_cech_dgm_one and
rightTriangle_rips_dgm_one), and its degree-1 log-scale bottleneck is 0
(rightTriangle_logBottleneck_one). The value ¼ log (4/3) belongs to the equilateral
triangle, which is not a fixture. These face values, diagrams, and distances are library
extensions with no cited source theorem.
Scaling and translation
For 0 < c, 1 ≤ n and PositiveOnEdges f,
logDgm_scale gives
logDgm k (f.scale c hc) n = translateDiagram (logDgm k f n) (Real.log c).
The proof proceeds through the pointwise logarithm identity and dgm_scale; it has no
1 ≤ c restriction. Positivity is scale invariant by positiveOnEdges_scale_iff, while
logCriticalValues_scale and logGrid_scale_apply say that the positive-log grid translates
by log c.
The bottleneck is jointly invariant under translateDiagram by bottleneck_translateDiagram,
including the value ⊤. The module statement has two cases: for 1 ≤ c, the merged
logModule_scale_shiftIso applies; for c ≤ 1,
logModule_scale_shiftIso_of_le_one supplies the inverse-factor shift. The restriction in the
first statement is an artefact of the ℝ≥0-valued shift parameter: logParam c = log c only
when 1 ≤ c.
Consequently, bottleneck_logDgm_scale and, over a field,
realInterleavingDistance_logModule_scale give scale invariance. For a scaled configuration,
Rips.filter_eq_scale_of_dist handles the label-metric relation d' = c · d,
Cech.filter_smul handles a real normed ambient space, and
Cech.bottleneck_logDgm_rips_cech_scale gives the corresponding Rips--Čech invariance.
The fixtures in LogScaleScalingExamples record these cases:
- F1:
filledReal_scale_two_logDgm_onegives a nonvacuous translation bylog 2. - F2:
logDgm_scale_one,filledReal_logDgm_scale_oneandhollowReal_logDgm_scale_onegive the identity control. - F3:
filledReal_scale_half_logDgm_one,hollowReal_scale_half_logDgm_oneandfilledReal_scale_half_birth_negexhibit a negative log shift. - F4:
scaledPts_eq,scaled_rips_eq,scaled_cech_eq,scaled_rips_logDgm_pos,scaled_cech_logDgm_pos,scaled_bottleneck_eqandscaled_bottleneck_le_log_twocover the scaled two-point configuration. - F5:
filled_hollow_scale_threeis a cross-complex bottleneck invariance instance. - F6:
repeated_scale_not_positiveis the negative control for positivity. - F7:
bottleneck_essentialOne_zeroandbottleneck_translate_essentialOneshow that a common translation preserves an infinite bottleneck.
These are library extensions; translation invariance of the bottleneck cost follows directly from the library's common-translation matching construction.
Scope
Degree 0 at log scale is not treated beyond the obstruction; it needs extended-real indexing
or a lower truncation. No degree-0 log-scale identity, diagram or bottleneck value is stated.
For 1 ≤ n and PositiveOnEdges f, the identity
(logDgm k f n).map expEmb = dgm k f n gives the radius-level diagram by applying the
order embedding exp to the log-scale diagram. The embedding is used in this direction
because Real.exp is strictly increasing on all of ℝ; Real.log is not an order embedding
of ℝ, so the reverse comparison is expressed by the pointwise log form rather than by a
diagram map along log. Explicit diagrams are stated only for the two-point fixtures, as in
their section, for the scaled filled/hollow and scaled two-point fixtures in §Scaling and
translation, and for the triangle fixtures: degrees 0 and 1 at radius scale and degree 1 at
log scale over ZMod 2. For the triangle, other fields and degrees n ≥ 2 are not treated;
other configurations are outside this note. Jung's constant, Hausdorff and
Gromov–Hausdorff statements and general positive scaling of real modules remain
outside this note; positive filter scaling is treated in the section above. The
log-scale module, constructibility, the module diagram and the isometry are used
unchanged.
Sources
The isometry used is the merged finite-constructible form of the theorem of Bauer–Lesnick. The Rips and ambient Čech definitions match Chazal–de Silva–Oudot §§4.2.1, 4.2.3, which states no Rips/Čech inclusion. The log-scale comparison of Čech and Rips diagrams, positivity of these filters off vertices, the vanishing of higher homology of a complex of vertices, the exp-map identity and the two-point fixtures are library extensions proved from merged definitions; no source theorem is cited for these results and no theorem of either paper is exported.