Interval modules, presentations and rank-exact lifting
The persistence-module API supports constructions on an arbitrary preorder, including coordinatewise ordered real vector spaces. The results below provide interval geometry, Hall neighbor bounds, two-sided matchings and distance bounds for finite rectangle sums, together with free presentations and lifting lemmas. Rank-exact resolutions are not covered here; minimal finite rank-projective resolutions are constructed in finite rank-projective resolutions.
Supports and rectangle transitions
Fix a ring R and an order-convex subset S of a preorder P. The module
SupportInterval.module S hS has fiber R on S and zero elsewhere, even when S
is empty. Its structure map from a to b preserves coefficients when both points
lie in S and is zero if either point lies outside S. For a nontrivial ring, it
is zero exactly in that case. Over a field, naturality also supplies coefficient
and support-boundary restrictions on morphisms between these modules.
For the rectangle results below, take coefficients in a field k.
A BoundedInterval records two real endpoints and whether each is included.
Nonempty intervals include closed singletons. A Rectangle n is a family of n
such intervals, with product support. For δ ≥ 0,
Rectangle.diagonal_maps_eq_zero_iff proves that every structure map from p to
p + δ·1 is zero exactly when some coordinate interval has length less than δ,
or has length δ and at least one endpoint excluded. At equality, a closed
interval retains its endpoint-to-endpoint map. In particular, a closed singleton
has a nonzero identity map at δ = 0. Rectangle.hom_ext shows
that a morphism is determined by its scalar at one point of a nonempty overlap.
Consumers use open scoped TDA.Persistence to select the coordinatewise preorder category on
Fin n → ℝ, rather than mathlib's category of families of category objects.
These are local geometry and morphism results used in Bjerkevik's rectangle stability proof. They do not extract a matching from an interleaving.
Hall bounds for finite rectangle sums
Let k be a field, n ≥ 1 and ε ≥ 0. Take two finite families Rᵢ and Sⱼ of
nonempty bounded rectangles in ℝⁿ, retaining the inclusion choice at every
endpoint. RectangleFamily.module is their finite biproduct of interval
modules. Indices label occurrences, so repeated rectangles remain distinct.
A rectangle is δ-significant if some p and p + δ·1 both belong to its support.
RectangleHall.significant_iff_transition_ne_zero identifies this condition
with nonvanishing of its actual δ transition. Put ρ = (2n−1)ε. A neighbor of
Rᵢ is an occurrence Sⱼ whose interval module is ρ-interleaved with that of Rᵢ.
For a selected set A of source indices, let μ(A) consist of all their neighbor
indices. Given an actual ε-interleaving of the two family modules,
RectangleHall.card_le_neighbors proves
|A| ≤ |μ(A)|
whenever every Rᵢ selected by A is 2ρ-significant. The inequality counts occurrences in both families. It includes empty A and ε = 0; at the latter parameter closed singletons are significant. The radius is ε in one coordinate and 3ε in two coordinates.
The proof orders the selected occurrences by the sum of their lower and upper endpoints, breaking equal sums by the number of plus decorations. A lower endpoint is plus-decorated when excluded, and an upper endpoint when included. A nonzero shifted rectangle map has one scalar on its overlap. For ordered significant endpoints of a two-step path, the geometric bounds give a common source, middle and target grade; this proves that composition multiplies their scalars, including closed-endpoint equality cases.
Expanding the interleaving equations through the finite biproduct gives sums of these component paths. In the ordered entries, every intermediate occurrence outside μ(A) contributes zero. The restricted scalar matrices therefore multiply to a triangular matrix with diagonal one. Its full rank is at most the number of intermediate occurrences, proving the inequality. This implements the finite bounded part of Bjerkevik's Lemmas 4.7–4.9.
Two-sided occurrence matchings
A partial occurrence matching is an injection from selected source indices to
target indices; its unselected source and target indices are unmatched. At
radius ρ it is admissible when paired rectangle modules are actually
ρ-interleaved and every unmatched module on either side is actually
ρ-interleaved with zero. The existing OccurrenceMatching type records the
partial injection, while RectangleMatching.IsMatching states these module
conditions. Repeated rectangles retain their distinct occurrence indices.
A rectangle module is ρ-interleaved with zero exactly when its transition by 2ρ vanishes, equivalently when it is not 2ρ-significant. The zero natural maps give the forward construction; the composite equation forces the reverse implication. This equivalence retains closed-endpoint equality and ε = 0.
For finite bounded rectangle families over any field, n ≥ 1 and an actual
ε-interleaving, set ρ = (2n−1)ε. The Hall inequality above and Mathlib's Hall
theorem give an injection from the 2ρ-significant source occurrences to
admissible target partners. Reversing the interleaving supplies the corresponding
injection for significant target occurrences. These injections need not agree.
Relation-preserving Schröder–Bernstein combines them into one partial matching
covering both significant sets. Its auxiliary dummy labels allow deletion only
of non-significant occurrences and are discarded from the resulting matching.
RectangleMatching.exists_matching exports the actual pair and deletion
interleavings, with no Hall or matching premise.
This is the finite bounded matching-existence conclusion of Bjerkevik's Theorem 4.3, using Theorem 4.10 and the following Cantor–Bernstein argument.
Assembly and distance bounds
Let M and N be the finite biproduct families associated with R and S.
An admissible radius-ε matching supplies actual interleavings of the paired
rectangle modules and actual interleavings with zero for all unmatched modules
on both sides. IsMatching.interleaving assembles these into an actual
ε-interleaving of M and N, in every coordinate dimension.
The assembled maps have a finite occurrence matrix. A matched edge contributes
the forward and reverse maps of the same chosen pair interleaving; every other
entry is zero. The partial injection gives at most one path through each row
and column of a composite. For a matched diagonal entry, the pair's composite
equation gives the summand transition by 2ε. An unmatched diagonal entry is zero,
as required by its interleaving with zero; off-diagonal entries are zero.
The public RectangleFamily.component_ext and reconstruction lemmas lift these
component equations to the two full family equations. The underlying realization
remains the existing finite biproduct.
Write d_I(M,N) for the extended infimum of finite parameters admitting an actual
interleaving. Define d_B(R,S) as the extended infimum of finite parameters
admitting an actual occurrence matching, including its deletion witnesses.
RectangleMatching.bottleneckDistance implements this second value. Both take
values in the extended nonnegative reals, with empty infima equal to infinity.
Assembly and the matching-existence theorem give, for n ≥ 1,
d_I(M,N) ≤ d_B(R,S) ≤ (2n−1) d_I(M,N).
The first inequality follows by taking the infimum of all assembled witnesses.
For the second, each actual ε-interleaving yields a matching at (2n−1)ε.
Multiplication by this finite positive coefficient commutes with both infima,
even when the witness set is empty. Thus no optimal radius or attainment is
assumed. This is the finite bounded distance conclusion of
Bjerkevik's Theorem 4.3, using
Definition 2.9 and the following matching-to-interleaving observation.
Endpoint decorations can prevent attainment. Over ℚ, the singleton-supported interval module and the zero module have both distances zero: they are interleaved, and admit a deletion matching, at every positive radius. They have neither a zero-radius interleaving nor a zero-radius matching. Zero distance is therefore not identified with a parameter-zero witness.
For n = 1 the coefficient is one, giving equality for this decorated occurrence family distance. The common half-open subclass connects to the existing one-dimensional endpoint-diagram API below. In dimension two, the following example attains the factor three.
Sharpness in two coordinates
Fix any field k. RectangleSharpExample.source and .target are the two
Fin 3-indexed families of open rectangles in
Bjerkevik's Example 5.2, pp. 117–118:
| Occurrence | Source Rᵢ | Target Sᵢ |
|---|---|---|
| 0 | (0,10) × (1,11) | (1,11) × (0,10) |
| 1 | (0,12) × (-1,11) | (1,9) × (0,12) |
| 2 | (2,10) × (1,9) | (-1,11) × (2,10) |
Let M and N be their existing finite biproduct modules. The forward matrix F has rows indexed by R and columns indexed by S; the reverse matrix G has rows indexed by S and columns indexed by R:
F = [1 1 1] G = [-1 1 1]
[1 1 0] [ 1 0 -1]
[1 0 1] [ 1 -1 0].
Every nonzero entry is an admissible shifted rectangle map. All endpoint
sums are 22, and every rectangle is 6-significant, so the proved scalar
composition law applies to each two-step path. The identities FG = GF = I
give both family interleaving equations by component extensionality.
RectangleSharpExample.interleaving is this actual 1-interleaving. Its public
component laws recover F and G. The integer identities hold in every field,
including characteristic two.
For ρ < 3, the occurrences R₁ and R₂ are 2ρ-significant, hence neither can be
unmatched. Nonzero forward and reverse maps impose endpoint inequalities.
Pairing either occurrence with S₁ or S₂ would force ρ ≥ 3. Both would
therefore need S₀, contrary to the matching's injection. Thus every actual
matching has radius at least 3. Conversely all endpoint differences are at
most 3, and matching equal indices gives an actual 3-matching with no
deletions. bottleneckDistance_eq_three proves d_B(R,S) = 3.
The proved factor-three matching theorem sends every actual ε-interleaving
of M and N to a 3ε-matching. The obstruction then gives ε ≥ 1. Together
with the explicit witness this proves interleavingDistance_eq_one, namely
d_I(M,N) = 1. Both lower bounds test all witnesses before taking infima.
The two distances are attained in this example. fails_smaller_constant
shows that every coefficient below three fails on this same pair, and
interleavingDistance_lt_bottleneckDistance gives a finite rectangle example
where the two distances differ. These results concern dimension two; they
make no optimality assertion for dimensions three and above.
One-dimensional barcode compatibility
The common objects are finite occurrence-indexed families of intervals [b,d)
with real endpoints b < d. Their one-coordinate rectangles have included lower
endpoints and excluded upper endpoints. The existing RealBarcode.intervalModule
uses exactly this support; its endpoint-pair constructor also permits empty
supports and essential right tails. The bridge restricts to positive length
and finite death, as required for these nonempty bounded rectangles.
RectangleBarcodeBridge.halfOpenRectangle constructs each rectangle.
familyIso identifies its finite categorical biproduct with the reindexing of
the existing real barcode coordinate product along evaluation at the unique
coordinate. diagram places each endpoint pair in the existing diagram
multiset, with no essential occurrences. realizationIso extends the comparison
to the existing diagram realization by a value-preserving occurrence bijection.
Equal endpoint values remain separate coordinates; an empty occurrence family
is allowed. These are natural isomorphisms of modules, including all structure
maps and inactive fibers.
For endpoint families x and y, let D_x and D_y be these diagrams. The original diagram bottleneck takes the infimum of maximum endpoint-pair costs and diagonal deletion costs, the latter equal to half the interval length. The rectangle definition instead asks for actual pair and zero interleavings. The bridge proves
d_B(rectangles x, rectangles y) = bottleneck(D_x,D_y)
= realInterleavingDistance(realize D_x, realize D_y).
Passing from endpoint costs to actual witnesses uses the half-open endpoint
comparisons and the closed deletion threshold d−b ≤ 2ε. The reverse passage
requires care: two short intervals can be ε-interleaved even when their endpoint
pair cost exceeds ε. For example, [0,1) and [100,101) have such a radius-one
interleaving with zero maps, whereas pairing their endpoints costs 100. If either
interval has a nonzero 2ε transition, the two actual composite laws force both
pair maps nonzero, and the proved Hom endpoint bounds force pair cost at most ε.
Consequently every excessive-cost pair has both ends deletable. Removing precisely
those pairs gives a numerical matching at the same radius. Taking infima proves
the equality without selecting an optimal radius. The actual witness conversions
also yield interleaving_iff_cost_matching at every ε ≥ 0.
This is the finite bounded half-open common subclass of the interval and matching
conventions in Bauer–Lesnick, §§2.1–2.3 and 3.1–3.3.
It connects Bjerkevik's n=1 conclusion to the library's original endpoint diagram
and real barcode distance APIs. It does not identify all decorated interval modules
with half-open modules. In the raw half-open constructor, b ≥ d gives the zero
module; a nonempty half-open rectangle therefore requires b < d. In contrast,
the decorated rectangle [b,b] is a nonzero singleton module with zero distance
from zero and no parameter-zero interleaving. On the common positive half-open
subclass, real_distance_zero_iff_diagram_eq and real_distance_zero_iff_iso
derive separation from the existing endpoint diagram result. That conclusion
does not apply to the larger decorated domain.
Graded free modules and finite presentations
For a ring R and a graded generator family W, GradedFree.module W at a consists
of finite linear combinations of generators whose grades are at most a.
Structure maps retain the coefficients. Choosing an image for each generator
at its birth grade gives a unique natural transformation from this free module.
Consequently a map from a free module lifts through a componentwise surjective
natural transformation; two lifts differ by a kernel element on each generator
at its birth grade. A structure map is an isomorphism if exactly the same generators are
available at its source and target.
A presentation is the cokernel of a homogeneous map from a free relation module
to a free generator module. GradedFree.FinitelyPresented asserts the existence
of such a presentation with finite generator and relation sets and a natural
isomorphism to the given module. The quotient universal property, invariance
under isomorphism and redundant zero relations are supported. These constructions
implement the free-cover ingredients of
Lesnick, §4.1. The finite-presentation
characterization in Theorem 4.4 relates interleavings to simultaneous
presentations with shared shifted generator and relation families.
Over a field, a finite generator family gives finite-dimensional values of its
graded-free module at every grade of any preorder. The canonical cover of a
presentation is surjective at each grade, so finite generators imply
finite-dimensional fibers of the presented quotient; the relation family is
not needed for this consequence. A natural isomorphism transfers the result
to any module finitely presented by that quotient. Thus
GradedFree.finitelyPresented_finiteDimensional proves pointwise finite
dimensionality without requiring a finite indexing preorder.
The free module on a disjoint union of graded families has the corresponding
sum universal property. A presentation can adjoin an arbitrary graded family
of redundant generators by adding the relations new_v − chosenLift_v.
Presentation.adjoinRedundantIso proves that the enlarged quotient is naturally
isomorphic to the original one; finite families remain finite.
For any morphism between two presented modules,
Presentation.exists_chainMap_lift supplies maps between their generator and
relation free modules making both presentation squares commute. The proof
derives the fiberwise kernel of the cover as the image of the relation map
from its cokernel, then chooses lifts of relation generators. These
statements work over any ring and preorder without finiteness assumptions.
For real coordinate grades, W(σ) has generator grades gr(w) − σ, so
W(−ε) delays every birth by ε. GradedFree.shiftIso identifies
F(W(ε)) with F(W)[ε], and delayedInclusion embeds F(W(−ε)) in F(W).
To give cross-presentations at ε ≥ 0, choose graded families W₁, W₂, Y₁, Y₂
and homogeneous maps from F(Y₁) to F(W₁ ⊔ W₂(−ε)) and from F(Y₂) to
F(W₁(−ε) ⊔ W₂). These are the fields of GradedFree.CrossData. Each
primitive relation also gives a relation on the opposite side born ε later:
the free cross map retains its coefficients and advances its availability
grade. The resulting quotients are
M = ⟨W₁, W₂(−ε) | Y₁, Y₂(−ε)⟩,
N = ⟨W₁(−ε), W₂ | Y₁(−ε), Y₂⟩.
CrossData.quotientInterleaving constructs an actual ε-interleaving of M and N.
Both free cross maps kill both relation families in the opposite quotient,
so the cokernel universal property supplies the quotient maps. Their two
composites are the 2ε transitions: this follows from the free cross-map
identities by cancellation of the surjective free covers.
CrossData.interleavingOfIso transports the result to modules with these
presentations. Finite primitive families give finite presentations, and at
ε = 0 the construction gives a natural isomorphism. The proof works over any
ring, for arbitrary families and dimension zero as well.
For finitely presented M and N, the existence of an ε-interleaving is
equivalent to the existence of finite W₁, W₂, Y₁, Y₂ with the two displayed
presentations. finitelyPresented_interleaving_iff_crossPresentation proves
this finite-presentation version of
Lesnick, Theorem 4.4, over any ring
and including dimension zero. The construction accepts arbitrary specified
presentations; finiteness of the original families ensures finiteness of all
four shared families.
To construct them from an interleaving, take W₁ and W₂ to be the original
generator families and write Z₁ and Z₂ for the original relation families.
The interleaving maps give values for W₂(−ε) in M and W₁(−ε) in N at their
delayed births. Lift these values through the free covers and add the
homogeneous graph relations new generator − chosen free lift. Thus
Y₁ = Z₁ ⊔ W₂(−ε) and Y₂ = Z₂ ⊔ W₁(−ε). Presentation.crossData builds the
primitive maps with precisely these relations.
The two collapse maps send the old generators through their free covers and
the added delayed generators to their prescribed interleaving images.
Both interleaving equations show that crossing and then using the shifted
opposite collapse agrees with collapse followed by the interleaving map.
Consequently the opposite delayed relations also vanish. The quotient maps
then descend by the cokernel universal property. Redundant-generator
elimination and the quotient universal property in both directions prove
Presentation.crossDataLeftIso and crossDataRightIso, the natural
isomorphisms to M and N. No relation-vanishing or quotient-isomorphism
condition is included in the input data.
Presentation.shift shifts both original indexing families without changing
their cardinalities. Its relation map is conjugated through the free-shift
isomorphisms. Presentation.presentedShiftIso identifies the resulting
cokernel with the shift of the original quotient, using Mathlib's preservation
of cokernels under reindexing. Its free-cover compatibility is
Presentation.cover_presentedShiftIso; finitelyPresented_shift gives the
corresponding closure of finite presentation.
A map from a presented module lifts through a natural transformation ρ when ρ is surjective at every generator grade and injective at every relation grade. The proof chooses preimages of the generators, then uses injectivity to show that the relations vanish. If ρ is also injective at generator grades, the lift is unique. Thus postcomposition with ρ gives a bijection on these morphisms. These assertions hold over any ring and preorder; the presentation need not be finite. They test only its declared grades.
For coordinatewise ordered Fin n → ℝ, each finite presentation determines a
finite set of critical values in every coordinate: the generator and relation
grade coordinates. If no critical value lies in any interval (aᵢ, bᵢ], the
structure map from a to b is an isomorphism. Evaluation preserves the
presentation cokernel, so the generator and relation isomorphisms induce the
isomorphism of presented fibers. This property transfers through the natural
isomorphism in FinitelyPresented. Taking less than the least positive gap
from a fixed a to the finitely many grade coordinates gives a positive
right-diagonal slack valid for every smaller nonnegative shift, including
the empty-grade and zero-coordinate cases. These are the local results in
Lesnick, §6, Lemmas 6.4 and 6.6.
Finite-grid Kan reconstruction
Let R be a ring and let M be a finitely presented module on coordinatewise ordered ℝⁿ. Choose a homogeneous presentation with finite generator family W and finite relation family Y. For each coordinate, collect the values of all generator and relation grades and add zero. Their product is a finite grid G, and coordinate inclusion gives an order embedding j:G→ℝⁿ. Adding zero handles empty presentations uniformly; no minimal-grid claim is needed. Every occurrence in W and Y is retained, even when grades agree.
The reconstructed module is the actual Mathlib left Kan extension of the
presentation's restriction to G. Presentation.gridKanIso proves its
natural isomorphism with the original presented quotient.
finitelyPresented_exists_finiteGridKan transports this comparison along
the presentation isomorphism to M, so grid constructibility is a conclusion.
The real-indexed module may have unbounded support although G is finite.
This is the finite-grid reconstruction step in
Botnan–Oppermann–Oudot–Scoccola, Appendix A's proof of Theorem 2.7.
The proof uses a more general presentation argument. For an order embedding
j:Q→P containing every generator and relation birth, form the chosen Kan
extension of the restricted quotient and its canonical comparison ρ to
the quotient. Full faithfulness makes the Kan unit invertible on Q;
the comparison is therefore invertible at every embedded grade. In
particular it is surjective at generator births and injective at relation
births. The existing presentation lifting theorem constructs a section
of ρ from precisely these conditions. The Kan universal property and
invertibility on Q prove the other inverse law. Presentation.kanIso
is the resulting comparison, and isLeftKanExtension establishes the
standard Mathlib universal property with identity on the restriction.
Consequently restriction gives a bijection between maps from the quotient
and arbitrary natural maps from its restriction. restrictionHomEquiv
and the selected-grid specialization gridHomEquiv expose this bijection,
with forward restriction, extension/restriction and uniqueness laws.
The inverse construction uses the existing graded-free generator lift
and presentation cokernel descent through the presentation lifting theorem;
an extension or an inverse comparison is not an input.
Over a field, the finite free cover is finite dimensional at each grade,
and its surjectivity proves that the presented quotient is pfd. Only
finiteness of W is needed for this argument. Transporting along the
presentation isomorphism proves finitelyPresented_finiteDimensional;
the field version of the Kan theorem supplies pfd fibers on G.
Consumers use two generators born at the same grade, with a genuine nonzero relation killing the first. The second survives by mapping to a principal upset. Distinct generator vectors remain distinct despite equal grid birth points. Extending the restricted doubling map recovers the nonzero doubling map at every real grade over ℚ; in characteristic two it recovers zero while the relation and quotient remain nonzero. Integer coefficients, the zero ring and empty families are included. In dimension zero, G has one point and the same quotient is nonzero.
Finite-grid reconstruction identifies the presentation with an actual Kan extension. Reconstruction alone gives no hook-projectivity or minimal-resolution claim. The later finite-grid cover result combines this reconstruction with rank exactness and hook lifting; preservation of minimal resolutions remains outside its scope.
Kan fibers on finite product grids
Let G be a finite product of finite coordinate sets, j:G→ℝⁿ its coordinate inclusion, R a ring and A a module on G. At a real grade a, consider the grid points p with j(p)≤a. When this set is nonempty, each coordinate has a greatest selected value below aᵢ. Their tuple is the greatest grid point floor_G(a). It is terminal in the indexing category for the pointwise Kan colimit, so the actual Mathlib left Kan extension has a canonical fiber isomorphism to A(floor_G(a)). The comparison is induced by the Kan unit at the floor followed by the real structure map. If no grid point lies below a, the indexing category is empty and the actual fiber is zero.
These comparisons control maps as well as fibers. A natural transformation between grid modules induces the same map, under the fiber isomorphisms, as its component at that grade's greatest lower grid point. A real structure map from a to b becomes the grid structure map from floor_G(a) to floor_G(b). Thus vectors, maps and death grades can be followed through the actual Kan extension. These claims use the finite product grid, whose coordinatewise floors are again grid points; arbitrary order embeddings need not have this pointwise exactness.
Rank exactness and hook lifting
Consider a pointwise short exact sequence L → M → N. At a ≤ b, assume only that
M(a) is finite dimensional. IsPointwiseShortExact.rank_add_iff_lift_kernel
proves that rank additivity at this pair is equivalent to the following lifting
property: every vector of N(a) killed at b has a preimage in M(a) also killed at b.
IsRankExact requires pointwise finite-dimensional middle fibers. Naturally split
pointwise short exact sequences in this domain are rank exact, and reindexing
preserves rank exactness.
For a finite product grid, the canonical fiber and map comparisons transfer
these statements to the actual left Kan extension. Over a field, lan_finiteDimensional
proves pfd for every grid module with pfd fibers. For every original
IsRankExact sequence, lan_isPointwiseShortExact derives ordinary
pointwise exactness of the actual Lan maps, while lan_isRankExact proves
pfd middle fibers and rank additivity. Its lan_lift_kernel consequence
transports a quotient vector killed by a real transition to a middle vector
with the same image that is killed by that same transition. This preservation
theorem has no finite-presentation or bounded-support hypothesis.
This is the finite-product-grid case of the pointwise Kan preservation result
in Botnan–Oppermann–Oudot, §4.3, Lemma 4.15.
For example, on the grid with coordinate values 0 and 2, a finite-death hook
from (0,0) to (2,2) supplies a nonzero vector at real grade (1/2,1/2). The
actual transition to (1,1) preserves it, and the transition to (5/2,5/2)
kills it. Applying the finite-hook rank-exact deflation on the grid and
lan_lift_kernel produces a nonzero lift in the actual real-grade Lan fiber
that dies at the same later grade. The fixture also checks the induced
quotient map through the canonical fiber comparison.
For i < j in a partial order, the hook has support {a | i ≤ a and j ≰ a}.
Hook.homEquivKernel identifies maps from this hook to N with vectors of N(i)
killed at j. Hook.exists_lift therefore lifts an entire hook morphism through
a rank-exact quotient with pointwise finite-dimensional middle fibers. The support may
be unbounded. This proves the finite-death hook lifting ingredient of
Botnan–Oppermann–Oudot–Scoccola, §2.14 and Appendix A.
The principal-upset module, representing a hook with infinite death, is the
graded free module on one generator. HookInfinity.homEquivFiber identifies
its morphisms with the target's birth fiber. An epimorphism of persistence
modules is epi at each grade, and module-category epimorphisms are surjective,
so the birth value lifts and the lifted map factors the original morphism.
Thus, for R : Type u and P : Type v with [Ring R] and [Preorder P],
the principal-upset module is an ordinary projective object of
PModule.{u,v,u} R P, with no finite-dimensionality assumptions. The
canonical instance is available from
TDA.Persistence.HookInfinityProjective. In particular, rank-exact quotient
maps over a field are among the categorical epimorphisms covered here.
Individual hook endomorphisms and indecomposability
Fix a field k and a partially ordered grade set. A finite-death hook with
birth i < j has a unit vector at its birth, and every supported vector is a
scalar multiple of its transport from i. Naturality therefore forces an
endomorphism to act by the same scalar throughout the support. The same
argument applies to a principal upset through its single free generator.
Hook.end_eq_smul_id and HookInfinity.end_eq_smul_id express these actual
natural transformations, and endRingEquiv in each namespace identifies
the endomorphism ring with k. The comparison reads the coefficient at birth;
its inverse is scalar multiplication of the identity. The multiplication
law uses the actual categorical composition, whose order is reversed in
Mathlib's endomorphism ring.
Since the field has only zero and one as idempotents, these endomorphism
rings have only the zero map and identity as idempotents. A decomposition
into two nonzero summands would give a further idempotent by projecting
onto either summand. Thus Hook.indecomposable and
HookInfinity.indecomposable prove indecomposability in Mathlib's existing
categorical sense. Neither a finite grade set nor bounded support is needed.
Consumers include the two-coordinate hook with birth (0,0) and death (1,1),
whose support contains the whole ray {(t,0) | t ≥ 0}, and its genuinely
decomposable sum with a second identical occurrence.
An isomorphism of interval modules identifies their supports, including the
empty support: a nonzero supported unit cannot map into an inactive fiber.
SupportInterval.support_eq_of_iso proves this without assuming connectedness.
For individual nonzero hooks the least support point determines birth, and
the excluded principal upset determines finite death. Consequently
Hook.nonempty_iso_iff recovers both endpoints, while
HookInfinity.nonempty_iso_iff recovers birth. Finite and infinite death are
never isomorphic. In contrast, death at or before birth has empty support
and is a zero object, so the nonzero hook assertions require i < j.
These results supply the individual-hook ingredients for Botnan–Oppermann–Oudot–Scoccola, §2.6 and Theorem 2.7. Order convexity alone does not assert indecomposability for an arbitrary support. The finite-poset classification and finite-sum multiplicity uniqueness that use these individual-hook results are established in the next section; minimal rank-projective resolutions remain separate.
Enough rank projectives on finite posets
For a field k : Type u and partial order P : Type in universe zero, the
field-side modules have type PModule.{u,0,u} k P. IsRankProjective Q means
that every map from Q lifts along each quotient in an actual rank-exact
sequence. It requires neither pfd fibers of Q nor [Fintype P]; the
universe-zero restriction on P does not assert that P is finite. The hook
lifting results above give this property for finite-death hooks and principal
upsets. It is preserved by finite categorical biproducts, natural isomorphisms
and retracts. This relative property is distinct from ordinary projectivity
against all epimorphisms.
Now assume [Fintype P] and let M be a pointwise finite-dimensional
P-module. For each grade a, choose a basis of M(a); for every a < b,
choose a basis of the kernel of M(a→b).
FiniteHookDeflation.Index M labels each basis vector separately. A fiber
basis vector contributes a principal upset born at a; a kernel basis vector
contributes the hook with birth a and death b. The number of occurrences is
sum over a<b of dim ker M(a→b) + sum over a of dim M(a).
Their existing categorical biproduct is FiniteHookDeflation.module M.
The corresponding generator maps assemble to the natural transformation
FiniteHookDeflation.π M. Fiber bases make it surjective at every grade.
For any vector killed by M(a→b), its expansion in the kernel basis gives
a lift in the hook sum also killed at b. At a = b the only such vector is
zero. Thus lift_kernel proves the dying-element criterion at every pair,
and FiniteHookDeflation.rankExact derives rank additivity for the actual
categorical kernel sequence. Both the hook sum and its kernel have
finite-dimensional fibers. No rank equality or already existing deflation is
an input.
The hook sum is rank projective. isRankProjective_iff_retract characterizes
each pfd module M on this finite poset: it is rank projective exactly when
its constructed deflation admits a natural section. Thus every such
rank-projective M is a retract of its finite hook sum. This supplies enough
relative projectives in the finite-poset part of
Botnan–Oppermann–Oudot–Scoccola, Theorem 2.7(1).
The construction retains repeated occurrences and includes zero modules and
the empty poset. Consumers also give a branching-poset pfd module whose
rank-exact deflation does not split, and check the construction in
characteristic two.
For this pfd module on the finite poset, the classification theorem
rankProjective_iff_exists_canonical_subfamily strengthens the retract
statement: M is rank projective exactly when it is isomorphic to a finite
selection of the actual occurrences in its canonical hook cover. Thus
repeated hooks are retained as distinct cover indices. The public
RankProjectiveHookTag distinguishes a finite hook by its strict birth and
death pair and a principal upset by its birth. The theorem
rankProjectiveHookSum_multiplicity_unique proves that any isomorphic finite
sums of these nonzero hooks have equal multiplicity for every tag. The
nonzero-hook and zero-module consumers verify the theorem at both boundaries;
the zero module selects the empty occurrence family.
The deflation can have redundant summands; it is not asserted to be a minimal projective cover. An exact-category instance is not claimed; minimal resolutions and signed barcode stability are proved elsewhere (see finite rank-projective resolutions and signed hook bottleneck stability).
Finite-hook covers after grid Kan extension
Let G be a finite coordinate grid with its inclusion into ℝⁿ. For any
grid point pair i,d, CoordinateGrid.hookLanIso identifies the actual
left Kan extension of Hook.module i d with the real hook at the embedded
endpoints. It is valid even when d ≤ i, including coincident endpoints,
where the support and both module fibers vanish; no nonzero-hook premise is
hidden in the comparison. The
ring-level infinityLanIso does the same for the principal upset born at
i, which has unbounded support. Their coefficient laws are expressed using
the canonical grid-floor fiber comparison and the existing coefficient
maps. The finite-biproduct comparison is also ring-level and its projection
and inclusion laws are the actual images of the grid maps.
Over a field, CoordinateGrid.lan_isRankProjective proves preservation of
the relative lifting property. Given a rank-projective grid module and an
actual rank-exact sequence on ℝⁿ, restrict the sequence along the grid
inclusion, use the grid module's own lift, and transfer that lift back through
Mathlib's Lan/restriction adjunction. The result does not assume projectivity
of the target or a lift on ℝⁿ as input, and it is not ordinary categorical
projectivity.
For a pfd grid module A, embeddedCoverFamily A is indexed by exactly the
existing FiniteHookDeflation.Index A: every kernel-basis and fiber-basis
occurrence is retained, even if two endpoints agree. The finite occurrence
sum has its actual finiteHookDeflationLanIso comparison with the biproduct
of the embedded finite-death and principal-upset hooks. Its projection and
inclusion equations identify each summand map with the corresponding Kan
image followed by that summand's actual hook isomorphism.
For any module M satisfying the original
GradedFree.FinitelyPresented M predicate, the finite-presentation grid
theorem supplies G, a pfd grid module A, and an actual isomorphism
e : G.lan.obj A ≅ M. The constructed cover term is
H = ⨁ fun t => G.embeddedCoverFamily A t; its projectivity follows from
the grid deflation's projectivity and preservation under Lan. With
c : G.lan.obj (FiniteHookDeflation.module A) ≅ H, the maps are the actual
conjugated composites
G.lan.map (kernel.ι (FiniteHookDeflation.π A)) ≫ c.hom
c.inv ≫ G.lan.map (FiniteHookDeflation.π A) ≫ e.hom.
finitelyPresented_exists_finiteHookRankExactCover proves these maps are
rank exact and that the displayed kernel term has pfd fibers. The kernel's
finite presentation is not proved here. Neither a full exact structure on
finitely presented modules, classification of all rank-projectives,
minimality, resolution length, decomposition uniqueness nor signed stability
follows from this cover. This finite-coordinate-inclusion case implements
the hook-image and cover ingredients related to
Botnan–Oppermann–Oudot, §4.3, Lemma 4.16 and Proposition 4.17
and Botnan–Oppermann–Oudot–Scoccola, Theorem 2.7 and Appendix A;
it does not claim the full generality of the source's poset-map lemma.
The focused fixtures verify a nonzero planar vector on an unbounded arm and
its death under a genuine real transition, the infinite principal-upset tail,
separate equal-endpoint occurrences and mixed finite/infinite summands, and
the actual identity lift supplied by relative projectivity. They also cover
empty-family comparison, a zero finite hook and nonzero principal upset in
dimension zero, characteristic two, and cover consumers on the original
finite presentations over ℚ and ZMod 2.
Rank-projectives among finitely presented real modules
Fix a field k and n : ℕ (no positivity is needed). For a module M on Fin n → ℝ satisfying
the original GradedFree.FinitelyPresented M predicate,
GradedFree.finitelyPresented_rankProjective_iff_exists_hook_sum proves
IsRankProjective M ↔
∃ m : ℕ, ∃ t : Fin m → RankProjectiveHookTag (Fin n → ℝ),
Nonempty (M ≅ ⨁ fun i => rankProjectiveHookModule (t i)).
A finite tag stores a strict birth/death pair, while an infinite tag stores the
birth of a principal upset. Indexing the sum by Fin m retains every repeated
occurrence; the zero module is represented by the empty sum. Neither direction
assumes a decomposition of M: the reverse direction uses lifting for each
hook, finite-biproduct closure and transport across the given isomorphism.
For the forward direction, finitelyPresented_exists_finiteHookRankExactCover
provides an actual finite grid G, a pfd grid module A, a reconstruction
isomorphism e : G.lan.obj A ≅ M, and the actual rank-exact cover map. Relative
projectivity lifts 𝟙 M through that displayed right map. The resulting
section is transported back through e and the finite-hook biproduct
comparison. It is then descended to a section of
FiniteHookDeflation.π A: the Kan unit is an isomorphism on the full grid
embedding, so a map between these Kan extensions has its grid map recovered
and lifted through that unit. CoordinateGrid.lan_map_injective records the
faithfulness needed to verify the descended section against the actual Kan
image. Thus A is a retract of its actual finite-hook cover, hence rank
projective. The finite-poset
rankProjective_iff_exists_canonical_subfamily theorem selects actual cover
occurrences, which are transported to real hooks by the existing Lan and
biproduct comparisons.
This formalizes the finite-coordinate, finitely-presented projective
classification component related to
Botnan–Oppermann–Oudot–Scoccola, Theorem 2.7(1), v3.
The source theorem also asserts exact-structure axioms and finite minimal
rank-projective resolutions; this result proves neither. It does not cover
arbitrary pfd real-indexed modules, minimal covers or resolutions, or
choice-independent decomposition. It also does not by itself establish the 2n−2 degree bound
on the nonzero terms of minimal resolutions or signed-bottleneck stability, which are proved
separately (see the API guide's real-coordinate degree bound and
signed hook bottleneck stability). The focused
FinitePresentationRankProjectiveClassificationExamples module exercises a
nonzero finite-death hook, a principal upset with nonzero fibers above every
grade, repeated hook occurrences and the zero/empty-sum boundary.
Pullbacks and composites of rank-exact deflations
For a field k : Type u and partial order P : Type in universe zero, the
field-side deflation APIs use modules PModule.{u,0,u} k P. A rank-exact
deflation is a natural map π:M→N that is the quotient of an actual rank-exact
short exact sequence; IsRankDeflation expresses this condition. Its
IsRankDeflation.rankExact_kernel theorem identifies the actual categorical
kernel sequence as rank exact. Ordinary pointwise surjectivity alone is insufficient. For
example, the principal upset born at zero on the two-grade chain surjects
onto the hook dying at grade one. The target's unit dies, but every
transition of the principal upset preserves its coefficient, so that unit
has no dying lift. epi_not_rankDeflation verifies this obstruction.
Given a rank-exact deflation π:M→N and any f:T→N with T pointwise finite
dimensional, form the existing categorical pullback X. At grade p, a pair
(m,t) is compatible when πₚ(m)=fₚ(t). Maps from the principal upset born at
p represent m and t. Their compatibility gives a map into X by the
pullback universal property, and its value at birth constructs the pair
in X(p). PullbackFiber.mk and its projection laws express this
construction. Conversely, the two projections distinguish every vector
in X(p), by the same generator property and categorical extensionality.
PullbackFiber.mk, its projection laws and PullbackFiber.element_ext are
defined for a ring R : Type u and a preorder P : Type v, with modules
PModule.{u,v,u} R P. The finite-dimensionality conclusion here is the
field result and requires pfd fibers of both M and T. Thus X(p) embeds
linearly into M(p)×T(p), proving it finite dimensional.
The projection X→T is surjective: lift fₚ(t) through πₚ and pair that lift
with t. If t is killed between a≤b, naturality shows that fₐ(t) is killed
as well. Rank exactness of π gives a lift m killed at b. The compatible
pair (m,t) is killed in X because both projected images vanish. The kernel
criterion therefore proves rank additivity for the actual kernel of X→T.
IsRankDeflation.pullback_snd establishes this closure without assuming an
exact-category instance or the desired lifting property.
IsRankDeflation.id makes identity on any pfd module a rank-exact deflation.
For two composable deflations, lift each dying vector through the second and
then through the first, retaining the same death at both steps. Surjectivity
also composes, so the kernel criterion proves IsRankDeflation.comp for the
actual composite kernel sequence. It takes the two deflation certificates
directly and has no additional external pfd premise. Consumers apply these
laws to successive finite hook deflations and to a nonidentity pullback, with
genuinely nonzero dying images, zero base change and characteristic two.
These prove the deflation-side closure laws in Botnan–Oppermann–Oudot–Scoccola, §2.12, including the pullback step used in Appendix A's Lemma A.1. They do not establish the full exact structure; pushout closure for rank-exact sequences is proved separately below.
Pushouts of rank-exact sequences
For a field k : Type u and partial order P : Type in universe zero, take
the modules in PModule.{u,0,u} k P. Suppose
0 → L --ι→ M --π→ N → 0 is an actual rank-exact row, and let
f : L → L' with L' pointwise finite dimensional. This result has no
[Fintype P] premise; P need not be finite. The actual pushout of ι along
f has a canonical row
The last map is induced by π on M and zero on L'. At each grade, the
actual ModuleCat pushout is the quotient of M(p) ⊕ L'(p) by the relation
coming from l ↦ (ι_p(l), -f_p(l)). The original row is short exact, so this
quotient map has kernel equal to the image of L'(p), and its map from L'(p)
is injective. The quotient description also proves that each pushout fiber is
finite dimensional.
For a ≤ b, take n ∈ N(a) killed by the structure map to N(b). The
original rank-exact sequence lifts n to m ∈ M(a) killed at b. Its image
in the pushout still dies at b by naturality and maps to n under the new
quotient. Thus the new row has the same-death lifting property. Since its
middle fibers are finite dimensional, the equivalence between this lifting
property and rank additivity proves the new row rank exact. This is
IsRankExact.pushout; it does not assume exactness or rank additivity of the
output row.
This argument is fiberwise and uses no finiteness assumption on P. It proves
the pushout-closure law alone. The exact-structure theorem in
Botnan–Oppermann–Oudot–Scoccola, §2.14, Theorem 2.7
also includes other exact-category axioms and has its stated finite-poset or
finitely presented real-coordinate scope; those conclusions do not follow from
this local closure theorem.
Finite-presentation closure
A finite presentation consists of finite graded-free generator and relation
families. For a ring R and preorder P, with compatible universes
PModule.{u,v,max u w} and FinitelyPresented.{u,v,w}, these modules are
closed under actual categorical cokernels and binary biproducts. For a map
f : M → N, lift the source generators through the target's free cover and
add the lifted generators to the target relations, retaining the target
generators. A biproduct uses disjoint unions of both occurrence families. The
comparison isomorphisms identify these presentations with the actual
categorical objects, and repeated grades remain distinct occurrences. These
constructions are finitelyPresented_cokernel
and finitelyPresented_biprod.
The actual pushout of f : X → Y and g : X → Z is the cokernel of
(f, -g) : X → Y ⊞ Z; hence
finitelyPresented_pushout
requires finite presentations of exactly X, Y and Z, without monicity
or rank hypotheses.
For any field k and a partially ordered type P : Type, the field-side results
below use PModule.{u,0,u} and the original
GradedFree.FinitelyPresented.{u,0,0} predicate. Given an ordinary pointwise
short exact row 0 → L → M → N → 0 with quotient π : M → N, and with L
and N finitely presented,
IsPointwiseShortExact.finitelyPresented_middle
proves that M is finitely presented. Choose a finite free cover of N and
pull π back along it. The free generators lift through π and give a
section of the pullback projection, so the pullback is isomorphic to L ⊞ F,
with F the finite free cover. This pulled-back row splits; the original
extension need not. Let d : F(A) → F be the relation map in the chosen
presentation of N, and let c : F → N be its free cover. Set
B := Limits.pullback π c and let q : B → M be the first projection. The
relation map r : F(A) → B has pullback components (0,d); on each fiber it
sends a to (0,d(a)). This is the pullback lift of zero to M and d to
F; the section used to split B → F is separate from r. The image of r
is ker q, and q is an epimorphism, so M is identified with the actual
categorical cokernel of r. The relation map need not be injective, and the
original row needs no
pointwise finite-dimensionality or rank-exactness hypothesis.
For a pointwise short exact row with quotient π : M → N and any f : T → N,
the actual pullback Limits.pullback π f fits in an ordinary short exact row
0 → L → Limits.pullback π f → T → 0; thus finite presentation of L and
T gives finite presentation of the pullback by
IsPointwiseShortExact.finitelyPresented_pullback.
In the rank-exact case,
IsRankExact.pullback_finitelyPresented
also proves finite presentation of the actual kernel of the second projection
and rank exactness of its canonical kernel row. The finite presentation of
T supplies its pointwise finite-dimensionality; no presentation of M or
N is required.
For any composable maps π : M → N and σ : N → O,
kernelCompIsoPullback
identifies the actual kernel of π ≫ σ with the actual pullback of π along
the kernel inclusion of σ. This comparison requires no exactness,
epimorphism or finite-presentation assumptions. When the maps belong to
rank-exact rows with finitely presented left terms L and K,
the second row gives a canonical isomorphism from K to the actual
Limits.kernel σ, transferring its finite presentation before the pullback
closure is applied.
IsRankExact.comp_finitelyPresented
uses pullback closure and this comparison to prove finite presentation of the
actual composite kernel and rank exactness of its canonical kernel row. This
result covers kernels of composites of rank-exact deflations, not kernels of
arbitrary maps.
For a rank-exact row 0 → L → M → N → 0 and a map f : L → L',
IsRankExact.pushout_finitelyPresented
proves finite presentation of the actual pushout and rank exactness of its
canonical row, taking finite presentations of L, M and L' as inputs.
The finite presentation of L' supplies its pointwise finite-dimensionality
needed to apply the existing rank-exact pushout law; N need not be finitely
presented.
These are standalone closure results: they do not give finite presentation
for kernels of arbitrary maps or an exact-category structure. Signed stability
obtains its finite resolutions by finite-grid Kan reconstruction and does not
use these closure laws.
One-step persistent Schanuel
Fix a field k and coordinatewise ordered ℝⁿ. Suppose K→P→M and L→Q→N are rank-exact short exact sequences and P,Q are rank projective. The rank-exactness hypothesis includes pointwise finite dimensionality; no finite grid or finite presentation is required here. An actual ε-interleaving i between M and N, for any finite ε≥0, induces an actual ε-interleaving
Here A[ε](a)=A(a+ε). This is the rank-exact specialization of
Botnan–Oppermann–Oudot–Scoccola, Appendix A Lemma A.1.
RankSchanuel.interleavingOfSequences uses the original K and L; it does
not require chosen sections or identifications with categorical kernels.
Write α:P→M and β:Q→N for the quotient maps. Form the crossed pullbacks X=Q[ε]×{N[ε]}P and Y=P[ε]×{M[ε]}Q, using α followed by i's forward map, and β followed by its reverse map. A vector of X(a) is a compatible pair (q,p), with q in Q(a+ε) and p in P(a). The forward map sends it to (P(a→a+2ε)(p),q) in Y(a+ε). Naturality and the interleaving equations prove compatibility. The symmetric map gives the reverse direction; both composites are the corresponding double transitions. The public projection laws expose these coordinates without unfolding the proofs. This crossed-pullback interleaving only requires a coefficient ring.
Over the field, rank-exact pullback closure makes X→P and Y→Q
rank-exact deflations. Relative projectivity lifts identity to derive
their sections. Mathlib's splitting theorem and its pullback-kernel and
shift-kernel comparisons then identify X with P⊕(ker β)[ε] and Y with
(ker α)[ε]⊕Q. IsPointwiseShortExact.kernelIso identifies K with ker α
and L with ker β, preserving the original inclusions. Transporting the
constructed maps proves the stated interleaving with the same ε.
At ε=0 and a common quotient, the zero-shift comparison gives
RankSchanuel.kernelSumIso. Concrete consumers use two-copy and three-copy
covers of an unbounded planar hook. The first kernel inclusion is the
diagonal (1,−1), so the given-sequence theorem genuinely transports a
noncanonical kernel term. Both actual kernels are nonzero, including in
characteristic two. A principal-upset consumer detects a nonzero forward
map at positive width through its public projection law. Zero modules
and n=0 are also allowed; ℝ⁰ has one grade, rather than no grades.
This one-step theorem is an input to persistent Schanuel for resolutions. It does not construct finite minimal resolutions, bound their length or prove signed barcode stability.
Attainment on real finite-constructible modules
Over a field k, for right-continuous real modules on finite critical grids, with zero left tails and finite-dimensional critical fibers, the barcode isometry and an optimal finite matching give an interleaving at every finite distance bound. The admissible parameters form a closed upper ray when the distance is finite, and are empty when it is infinite. Distance zero is equivalent to a natural isomorphism. See finite-constructible closure.
This assertion depends on the stated constructibility and finiteness hypotheses.
For a field k, RealRay.distance_eq_zero and RealRay.not_nonempty_iso show that k_[0,∞) and
k_(0,∞) have zero distance but are not isomorphic: positive shifts interleave them,
while their fibers at zero differ. The zero distance is not attained. For finitely presented modules on ℝⁿ,
the multiparameter closure theorem gives attainment,
closedness and distance-zero separation directly from finite generators and relations.