Stability of Interval Decomposable Persistence Modules: Håvard Bakke Bjerkevik January 21, 2020

Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

Stability of interval decomposable persistence modules

Håvard Bakke Bjerkevik


January 21, 2020
arXiv:1609.02086v3 [math.AT] 21 Jan 2020

Abstract
The algebraic stability theorem for R-persistence modules is a fundamental result in topological
data analysis. We present a stability theorem for n-dimensional rectangle decomposable persistence
modules up to a constant (2n − 1) that is a generalization of the algebraic stability theorem, and also
has connections to the complexity of calculating the interleaving distance. The proof given reduces to a
new proof of the algebraic stability theorem with n = 1. We give an example to show that the bound
cannot be improved for n = 2. We apply the same technique to prove stability results for zigzag modules
and Reeb graphs, reducing the previously known bounds to a constant that cannot be improved, settling
these questions.

1 Introduction
Persistent homology is a tool in topological data analysis used to determine the structure or shape of data
sets. For example, given a point cloud X ⊂ Rn sampled from a subspace S of Rn , we want to guess at
the homology of S, which tells us something about how many “holes” S has in various dimensions. We can
do this by defining B() to be the union of the (open or closed) balls of radius  centered at each point
in X. Calculating homology, we get a group or vector space Hn (B()) for each  ≥ 0, and the inclusions
B() ,→ B(0 ) induce morphisms Hn (B()) → Hn (B(0 )) for  ≤ 0 . Such a collection of vector spaces
and morphisms is called a persistence module, or R-module, as the vector spaces are parametrized over R.
Under certain assumptions, we can decompose an R-module into interval modules [15], which gives us a set
of intervals uniquely determining the persistence module up to isomorphism. This set of intervals is the
barcode of the persistence module. The intervals in the barcode are interpreted as corresponding to possible
features of the space S, where one might interpret long intervals as more likely to describe actual features
of S and short intervals as more likely to be the result of noise in the input data. In other words, we have
an algorithm with a data set as input and a barcode as output. As data sets always carry a certain amount
of noise, we would like this algorithm to be stable in the sense that a little change in the input data, or in
the persistence modules, should not result in a big change in the barcode.
We measure the difference between persistence modules with the interleaving distance dI , and the differ-
ence between barcodes with the bottleneck distance dB . Proving stability then becomes a question of proving
that the bottleneck distance is bounded by the interleaving distance, i.e. dB ≤ CdI for some constant C.
Stability has been proved for persistence modules over R in [14, 12, 13, 5] in what is called the algebraic
stability theorem, which implies the isometry theorem dI = dB .
Persistence modules can also be parametrized over other posets. A pair of filtrations f, g : S → R of a
topological space S gives rise to an R2 -module which has a vector space Vp for each point in p ∈ R2 and
linear maps V(a,b) → V(c,d) whenever a ≤ c and b ≤ d, for instance by letting V(a,b) be Hn (f −1 (−∞, a) ∩
g −1 (−∞, b)). Again, inclusions induce the linear maps on homology. With n filtrations instead of 2, we
get an Rn -module. Another example is zigzag modules, which are popular objects of study in topological
data analysis [10, 20, 19]. These can arise from a sequence of subspaces Si ⊂ S, where we also consider the
intersections Si ∩ Si+1 (or unions). In this case, we have
· · · ⊆ Si−1 ⊇ Si−1 ∩ Si ⊆ Si ⊇ Si ∩ Si+1 ⊆ . . . ,
which again gives rise to linear maps on homology. Defining interleavings and thus the interleaving distance is
trickier than for R-modules, but in fact one can do this by associating R2 -modules called block decomposable

1
modules to the zigzag modules. One can also associate block decomposable modules to Reeb graphs, which
are of interest because of their ability to present geometrical information despite being relatively simple
objects. See Section 3.
All these examples serve as motivation for why one would like to talk about stability for multi-parameter
modules (that is, persistence modules parametrized over Rn for n ≥ 2). Unfortunately, no isometry theorem
is possible even for general R2 -modules, because there is no nice decomposition theorem like in the one-
parameter cases, meaning that dB is not defined. The block decomposable modules, however, decompose
nicely, and dB ≤ 52 dI has been shown for these [3]. This carries over to stability results for zigzag modules
and Reeb graphs.
Our main contribution is a new method of proving stability for interval decomposable modules. We
demonstrate several applications of this method. The first is Theorem 4.2:
Theorem. Let M = I∈B(M ) II and N = J∈B(N ) IJ be rectangle decomposable Rn -modules. If M and
L L

N are δ-interleaved, there exists a (2n − 1)δ-matching between B(M ) and B(N ).
This is a generalization of the algebraic stability theorem for R-modules, which is the case n = 1.
For n ≥ 2, the result is new. There already exist several proofs of the algebraic stability theorem, but
our approach is different from the ones taken before, which allows this more general theorem, as well
as the results below. Our method is combinatorial, which in our opinion reflects the true nature of the
problem once some of the algebraic technicalities are stripped away. Also, our proof is fairly short in the
case n = 1 compared to earlier proofs of the algebraic stability theorem. In Example 5.2, we construct
rectangle decomposable modules M and N over R2 for which dI (M, N ) = 1 and dB (M, N ) = 3, disproving
a conjecture made in an earlier version of [3] claiming that dB (M, N ) = dI (M, N ) holds for all n-dimensional
interval decomposable modules M and N whose barcodes only contain convex intervals. The example also
shows that the bound in the theorem cannot be improved for n = 2. It is an open question if the bound
can be improved for n ≥ 3.
We do not know of any examples of rectangle decomposable modules arising naturally from real-world
data sets. But as we discuss in Section 6, there is a strong link between the stability of these modules and the
recent proof that calculating the interleaving distance between multi-parameter modules is NP-hard [9]. In
particular, our way of viewing interleavings as pairs of matrices and our observation in Example 5.2 that the
interleaving and bottleneck distances differ for rectangle decomposable modules served as inspiration for the
approach used in [9]. The question of whether the hardness results can be strengthened is closely related to
the question of whether Theorem 4.2 can be improved. Thus, even if rectangle decomposable modules never
arise directly from data sets, the type of questions we consider can have an impact on practical applications.
Another reason why we give the proof in detail for rectangle decomposable modules instead of, say, block
decomposable modules, is that this case demonstrates very well exactly when our method works and when
it fails. The lesson to take home is that the method gives a bound dB ≤ cdI with a c that increases with the
freedom we have in defining the intervals we consider. You need 2n coordinates to define an n-dimensional
(hyper)rectangle, which gives a constant 2n − 1 in the theorem.
Another application of our proof method gives Theorem 4.11:

Theorem. Let M and N be δ-interleaved triangle decomposable modules. Then there is a δ-matching
between B(M ) and B(N ).
This is more immediately connected to practical applications. Theorem 4.11 implies dB ≤ dI for block
decomposable modules, which is an improvement on the previous best known bound, dB ≤ 25 dI . Since the
opposite inequality dI ≤ dB holds trivially, our bound is the best possible. We discuss how stability results
for zigzag modules and Reeb graphs follow in Section 3. The fact that our bound is optimal means that
these stability problems are now settled.
We finish off Section 4 by showing stability for free modules.
We assume that all modules are pointwise finite dimensional (p.f.d.). In a previous version of this paper
[8], we strengthened the theorems by removing this assumption.

2
2 Persistence modules, interleavings, and matchings
In this section we introduce some basic notation and definitions that we will use throughout the paper. Let
k be a field that stays fixed throughout the text, and let vec be the category of finite dimensional vector
spaces over k. We identify a poset with its poset category, which has the elements of the poset as objects,
a single morphism p → q if p ≤ q and no morphism if p  q.

Definition 2.1. Let P be a poset category. A P -persistence module is a functor P → vec.


If the choice of poset is obvious from the context, we usually write ‘persistence module’ or just ‘module’
instead of ‘P -persistence module’.
For a persistence module M and p ≤ q ∈ P , M (p) is denoted by Mp and M (p → q) by φM (p, q).
We refer to the morphisms φM (p, q) as the internal morphisms of M . M being a functor implies that
φM (p, p) = idMp , and that φM (q, r) ◦ φM (p, q) = φM (p, r). Because the persistence modules are defined as
functors, they automatically assemble into a category where the morphisms are natural transformations.
This category is denoted by P -mod. Let f : M → N be a morphism between persistence modules. Such
an f consists of a morphism associated to each p ∈ P , and these morphisms are denoted by fp . Because f
is a natural transformation, we have φN (p, q) ◦ fp = fq ◦ φM (p, q) for all p ≤ q.

Definition 2.2. An interval is a subset ∅ 6= I ⊆ P that satisfies the following:


• If p, q ∈ I and p ≤ r ≤ q, then r ∈ I.
• If p, q ∈ I, then there exist p1 , p2 , . . . , p2m ∈ I for some m ∈ N such that p ≤ p1 ≥ p2 ≤ · · · ≥ p2m ≤ q.

We refer to the last point as the connectivity axiom for intervals.


Definition 2.3. An interval persistence module or interval module is a persistence module M that
satisfies the following: for some interval I, Mp = k for p ∈ I and Mp = 0 otherwise, and φM (p, q) = Idk
for points p ≤ q in I. We use the notation IJ for the interval module with J as its underlying interval.
The definitions up to this point have been valid for all posets P , but we need some additional structure
on P to get a notion of distance between persistence modules, which is essential to prove stability results.
Since we will mostly be working with Rn -persistence modules, we restrict ourselves to this case from now
on. We define the poset structure on Rn by letting (a1 , a2 , . . . , an ) ≤ (b1 , b2 , . . . , bn ) if and only if ai ≤ bi
for 1 ≤ i ≤ n. For  ∈ R, we often abuse notation and write  when we mean (, , . . . , ) ∈ Rn . We call an
interval I ⊂ Rn bounded if it is bounded as a subset of Rn in the usual sense. That is, it is contained in a
ball with finite radius.
Definition 2.4. For  ∈ [0, ∞), we define the shift functor (·)() : Rn -mod → Rn -mod by letting M () be
the persistence module with M ()p = Mp+ and φM () (p, q) = φM (p + , q + ). For morphisms f : M → N ,
we define f () : M () → N () by f ()p = fp+ .
We also define shift on intervals I by letting I() be the interval for which II() = II ().
Define the morphism φM, : M → M () by (φM, )p = φM (p, p + ).
Definition 2.5. An -interleaving between Rn -modules M and N is a pair of morphisms f : M →
N (), g : N → M () such that g() ◦ f = φM,2 and f () ◦ g = φN,2 .
If there exists an -interleaving between M and N , then M and N are said to be -interleaved. An
interleaving can be viewed as an ‘approximate isomorphism’, and a 0-interleaving is in fact an isomorphism.
We call a module M -significant if φM (p, p + ) 6= 0 for some p, and -trivial otherwise. M is 2-trivial if
and only if it is -interleaved with the zero module. We call an interval I -significant if II is -significant,
and -trivial otherwise.
Definition 2.6. We define the interleaving distance dI on persistence modules M and N by

dI (M, N ) = inf{ | M and N are -interleaved}. (1)

3
The interleaving distance intuitively measures how close the modules are to being isomorphic. The
interleaving distance between two modules might be infinite, and the interleaving distance between two
different, even non-isomorphic modules, might be zero. Apart from this, dI satisfies the axioms for a metric,
so dI is an extended pseudometric.
Definition 2.7. Suppose M = ∼L I 1
I∈B I for a multiset B of intervals. Then we call B the barcode of M ,
and write B = B(M ). We say that M is interval decomposable.
Since the endomorphism ring of any interval module is isomorphic to k, it follows from Theorem 1 in [2]
that if a persistence module M is interval decomposable, the decomposition is unique up to isomorphism.
Thus the barcode is well-defined, even if we let M be a P -module for an arbitrary poset P . If M is a
p.f.d. R-module, it is interval decomposable [15], but this is not true for R-modules or p.f.d. Rn -modules
in general. [21] gives an example showing the former, and the following is an example of a P -module for a
poset P with four points that is not interval decomposable.
 
 1
1 0 1
k k2 k

 (2)
0 1

A corresponding R2 -module that is not interval decomposable and is at most two-dimensional at each point
can be constructed.
For multisets A, B, we define a partial bijection as a bijection σ : A0 → B 0 for some subsets A0 ⊂ A and
B ⊂ B, and we write σ : A 9 B. We write coim σ = A0 and im σ = B 0 .
0

Definition 2.8. Let A and B be multisets of intervals. An -matching between A and B is a partial
bijection σ : A 9 B such that
• all I ∈ A \ coim σ are 2-trivial
• all I ∈ B \ im σ are 2-trivial
• for all I ∈ coim σ, II and Iσ(I) are -interleaved.
If there is an -matching between B(M ) and B(N ) for persistence modules M and N , we say that M
and N are -matched.
We have adopted this definition of -matching from [3], which differs from e.g. the one in [13], which
allows two intervals I and J to be matched if dI (II , IJ ) ≤  (or rather, this is equivalent to their definition).
Conveniently, with the definition we have chosen, an -interleaving is easily constructed given an -matching.
We feel that this is the more natural definition for this paper, as several of our results are phrased as
statements about matchings and interleavings, and the interleaving distance might not come into the picture
at all. The other definition is perhaps more natural in the context of ‘persistence diagrams’, where intervals
are identified with points in a diagram, and the interleaving distance between the corresponding modules is
simply the distance between the points. This is irrelevant to us, however, as we never consider persistence
diagrams.
We can also define -matchings in the context of graph theory. A matching in a graph is a set of edges
in the graph without common vertices, and a matching is said to cover a set S of vertices if all elements in
S are adjacent to an edge in the matching. Let G be the bipartite graph on A t B with an edge between
I ∈ A and J ∈ B if II and IJ are -interleaved. Then an -matching between A and B is a matching in G
such that the set of 2-significant intervals in A t B is covered.
1 We will not be rigorous in our treatment of multisets. A multiset may contain multiple copies of one element, but we will

assume that we have some way of separating the copies, so that we can treat the multiset as a set. If e.g. I and J are intervals
in a multiset and we say that I 6= J, we mean that they are “different” elements of the multiset, not that they are different
intervals.

4
(b, b) (b, b) (b, b) (b, b)

(a, a) (a, a) (a, a) (a, a)

Figure 1: The intervals [a, b]BL , [a, b)BL , (a, b]BL and (a, b)BL .

Definition 2.9. The bottleneck distance dB is defined by

dB (M, N ) = inf{ | M and N are -matched} (3)

for any interval decomposable M and N .


We might abuse notation and talk about dB (C, D), where C and D are barcodes.

3 Zigzag modules and Reeb graphs


In this section we will give some intuition for how block decomposable modules relate to Reeb graphs and
zigzag modules. We refer to [3] for a more detailed and rigorous treatment.
When explaining the connection to Reeb graphs and zigzag modules, it is more convenient to flip one
of the axes in R2 , so that we work with Rop × R instead. This way, (a, b) ≤ (c, d) iff c ≤ a and b ≤ d, or,
equivalently, if (a, b) ⊆ (c, d) as intervals, assuming a < b. Let U = {(a, b) ∈ Rop × R | a ≤ b}.
Definition 3.1. An interval decomposable Rop × R-module is called block decomposable if its barcode
only contains intervals of the following types:
• [a, b]BL = {(c, d) ∈ U | c ≤ b, d ≥ a}
• [a, b)BL = {(c, d) ∈ U | a ≤ d < b}
• (a, b]BL = {(c, d) ∈ U | a < c ≤ b}
• (a, b)BL = {(c, d) ∈ U | c > a, d < b}
We call these intervals blocks. Each interval intersects the diagonal in an R-interval that is open, closed
or half-open one way or the other depending on the type of the block.

3.1 Reeb graphs


There have been proposed several distances on Reeb graphs; see [4] for a summary, as well as references to
various applications. The interleaving distance we consider was introduced in [16].
A Reeb graph is a topological graph G together with a continuous function γ : G → R such that
the level sets of γ are discrete. Let S(γ) = Rop × R → Set be the functor sending (a, b) to the set of
connected components of γ −1 (a, b) and S((a, b) → (c, d)) be induced by the inclusion γ −1 (a, b) ⊆ γ −1 (c, d)
for c ≤ a ≤ b ≤ d. In Figure 2, γ is projection to a horizontal axis. Above the graph, the functor S(γ) is
shown, the shade of grey at (a, b) determined by the size of S(γ)(a,b) .
Given two Reeb graphs (G1 , γ1 ) and (G2 , γ2 ), we get two functors S(γ1 ) and S(γ2 ), and we can talk about
interleavings and interleaving distance by adjusting the definitions in the previous section. It turns out that
this interleaving distance is at least as big as the one we get by replacing S(γ1 ) and S(γ2 ) by corresponding
block decomposable modules Mγ1 and Mγ2 . In Figure 2, the blocks comprising this block decomposable
module are exactly what you would guess by looking at the figure. In [3], dB (M, N ) ≤ 25 dI (M, N ) is proved
for such modules; with Theorem 4.14, we have dB (M, N ) = dI (M, N ).
There is also a barcode L0 (γ) of R-intervals (the level set persistence diagram [11]) associated to a Reeb
graph (G, γ), which we can think of as arising from the intersection of S(γ) with the diagonal x = y. This

5
x=y

Figure 2: A Reeb graph (G, γ) with S(γ) above. Evaulating S(γ) at the point shown, we get the intersection
of G with the red strip, which has two connected components.

barcode is the same as B(Mγ ), except that (a, b]BL is replaced by (a, b], and so on. It is not too hard to see
that dB (L0 (γ1 ), L0 (γ2 )) ≤ 2dB (Mγ1 , Mγ2 ).2 Altogether, this gives

dB (L0 (γ1 ), L0 (γ2 )) ≤ 2dB (Mγ1 , Mγ2 )


= 2dI (Mγ1 , Mγ2 )
≤ 2dI (S(γ1 ), S(γ2 )).

In other words:

Theorem 3.2. For Reeb graphs γ1 , γ2 , the inequality dB (L0 (γ1 ), L0 (γ2 )) ≤ 2dI (S(γ1 ), S(γ2 )) holds.
Thus an easily computed bottleneck distance gives a lower bound for the interleaving distance between
Reeb graphs. This improves the result in [3], which was again an improvement on [6], by lowering the
constant in the inequality from 5 to 2, and this cannot be improved.

3.2 Zigzag modules


A zigzag module is a module over ZZ = {(a, b) ∈ Z2 | a = b ∨ a = b + 1} taken as a sub-poset of Rop × R.
Let ZZ|(a,b) be the sub-poset of ZZ containing the elements {(c, d) ∈ ZZ | a ≤ c, d ≤ b}. A zigzag module
M gives rise to a block decomposable module MBL defined by letting MBL (a, b) be the colimit of the
restriction of M to ZZ|(a,b) . MBL ((a, b) → (c, d)) is defined to be the induced morphism we get by the
universal property of colimits for (a, b) ≤ (c, d). (This definition is given in [3], but something very similar is
described in the discussions of pyramids in [11] and [7].) This way, we can define interleaving and bottleneck
distance between zigzag modules by letting dI (M, N ) = dI (MBL , NBL ) and dB (M, N ) = dB (MBL , NBL ).
Thus Theorem 4.14 holds if we replace ‘block decomposable modules’ by ‘zigzag modules’:
2 The reason for the constant 2 is that (a, b)BL is (b − a)/4-trivial, while (a, b) is not -trivial for  < (b − a)/2.

6
Figure 3: Three rectangles, where the left and middle rectangles are of the same type (unbounded down-
wards), while the last is of a different type (unbounded upwards and to the right). Assuming that it contains
its boundary, the rightmost rectangle is also an example of a free interval, which we will define in a later
section.

Theorem 3.3. Let M and N be zigzag modules. If M and N are δ-interleaved, there exists a δ-matching
between B(M ) and B(N ).
This implies an isometry theorem for zigzag modules: dI (M, N ) = dB (M, N ).

4 Higher-dimensional stability
The algebraic stability theorem for R-modules states that an -interleaving between R-modules M and N
induces an -matching between B(M ) and B(N ), implying dI (M, N ) = dB (M, N ), the isometry theorem.
The main purpose of this paper is to find out when similar results for Rn -modules hold. Our first result,
Theorem 4.2, is a generalization of the algebraic stability theorem for R-modules. Variations of the algebraic
stability theorem have been proved several times already [14, 12, 13, 5], but this is a new proof with ideas
that are applicable to more than just R-modules.

4.1 Rectangle decomposable modules


For any interval I ⊂ Rn , we let its projection on the i’th coordinate be denoted by Ii .
Definition 4.1. A rectangle is an interval of the form R = R1 × R2 × · · · × Rn .
Two rectangles R and S are of the same type if Ri \ Si and Si \ Ri are bounded for every i. For n = 1,
we have four types of rectangles:
• intervals of finite length
• intervals of the form (a, ∞) or [a, ∞)
• intervals of the form (−∞, a) or (−∞, a]
• (−∞, ∞),
for some a ∈ R. We see that for n ≥ 1, rectangles R and S are of the same type if Ri and Si are of the
same type for all 1 ≤ i ≤ n. Examples of 2-dimensional rectangles are given in Figure 3.
In [13], decorated numbers were introduced. These are endpoints of intervals ‘decorated’ with a plus or
minus sign depending on whether the endpoints are included in the interval or not. Let R = R ∪ {−∞, ∞}.
A decorated number is of the form a+ or a− , where a ∈ R.3 The notation is as follows for a, b ∈ R:
• I = (a+ , b+ ) if I = (a, b]
• I = (a+ , b− ) if I = (a, b)
3 The decorated numbers −∞− and ∞+ are never used, as no interval contains points at infinity, but it does not matter

whether we include these two points in the definition.

7
• I = (a− , b+ ) if I = [a, b]
• I = (a− , b− ) if I = [a, b).
We define decorated points in n dimensions for n ≥ 1 as tuples a = (a1 , a2 , . . . , an ), where all the ai ’s are dec-
orated numbers. For an n-dimensional rectangle R and decorated points (a1 , a2 , . . . , an ) and (b1 , b2 , . . . , bn ),
we write R = ((a1 , a2 , . . . , an ), (b1 , b2 , . . . , bn )) if Ri = (ai , bi ) for all i. We define minR and maxR as
the decorated points for which R = (minR , maxR ). We write a∗ for decorated numbers with unknown
‘decoration’, so a∗ is either a+ or a− .
There is a total order on the decorated numbers given by a∗ < b∗ for a < b, and a− < a+ for all a, b ∈ R.
This induces a poset structure on decorated n-dimensional points given by (a1 , a2 , . . . , an ) ≤ (b1 , b2 , . . . , bn )
if ai ≤ bi for all i. We can also add decorated numbers and real numbers by letting a+ + x = (a + x)+ and
a− + x = (a + x)− for a ∈ R, x ∈ R. We add n-dimensional decorated points and n-tuples of real numbers
coordinatewise.
If M is an interval decomposable Rn -module and all I ∈ B(M ) are rectangles, M is rectangle decompos-
able.
Our goal is to prove the following theorem:
Theorem 4.2. Let M = I∈B(M ) II and N = J∈B(N ) IJ be rectangle decomposable Rn -modules. If M
L L

and N are δ-interleaved, there exists a (2n − 1)δ-matching between B(M ) and B(N ).
The inequality dB (M, N ) ≤ (2n−1)dI (M, N ) for rectangle decomposable modules M and N immediately
follows.
Fix 0 ≤ δ ∈ R. Assume that M and N are δ-interleaved, with interleaving morphisms f : M → N (δ)
and g : N → M (δ). Recall that this means that g(δ) ◦ f = φM,2δ and f (δ) ◦ g = φN,2δ . For any I ∈ B(M ),
ιI πI
we have a canonical injection II −→ M and projection M −→ II , and likewise, we have canonical morphisms
J ιJ π J J
I −→ N and N −−→ I for J ∈ B(N ). We define

fI,J = πJ (δ) ◦ f ◦ ιI : II → IJ (δ)


(4)
gJ,I = πI (δ) ◦ g ◦ ιJ : IJ → II (δ).

We prove the theorem by a mix of combinatorial and geometric arguments. First we show that it is
enough to prove the theorem under the assumption that all the rectangles in B(M ) and B(N ) are of the
same type. Then we define a real-valued function α on the set of rectangles which in a sense measures, in
the case n = 2, how far ‘up and to the right’ a rectangle is. There is a preorder ≤α associated to α. The
idea behind ≤α is that if there is a nonzero morphism χ : II → IJ () and I ≤α J, then I and J have to be
close to each other. Finding pairs of intervals in B(M ) and B(N ) that are close is exactly what we need to
construct a (2n − 1)δ-matching. Lemmas 4.6 and 4.7 say that such morphisms behave nicely in a precise
sense that we will exploit when we prove Lemma 4.8. If we remove the conditions mentioning ≤α , Lemmas
4.6 and 4.7 are not even close to being true, so one of the main points in the proof of Lemma 4.8 is that we
must exclude the cases that are not covered by Lemmas 4.6 and 4.7. We do this by proving that a certain
matrix is upper triangular, where the ‘bad cases’ correspond to the elements above the diagonal and the
‘good cases’ correspond to elements on and below the diagonal.
Lemma 4.8 is what ties together the geometric and combinatorial parts of the proof of Theorem 4.2.
While we prove Lemma 4.8 by geometric arguments, by Hall’s marriage theorem the lemma is equivalent to
a statement about matchings between B(M ) and B(N ). We have to do some combinatorics to get exactly
the statement we need, namely that there is a (2n − 1)δ-matching between B(M ) and B(N ), and we do
this after stating Lemma 4.8.
We begin by describing morphisms between rectangle modules.
Lemma 4.3. Let χ : II → IJ be a morphism between interval modules. Suppose A = I ∩ J is an interval.
Then, for all a, b ∈ A, χa = χb as k-endomorphisms.
Proof. Suppose a ≤ b and a, b ∈ A. Then χb ◦φII (a, b) = φIJ (a, b)◦χa . Since the φ-morphisms are identities,
we get χa = χb as k-endomorphisms. By the connectivity axiom for intervals, the equality extends to all
elements in A.

8
Since the intersection of two rectangles is either empty or a rectangle, we can describe a morphism
between two rectangle modules uniquely as a k-endomorphism if their underlying rectangles intersect. A
k-endomorphism, in turn, is simply multiplication by a constant. Note that we could have relaxed the
assumptions in the proof above and assumed that a is in I instead of in A, and still have gotten χa = χb .
In particular, this means that if 0 6= χ : II → IJ , and I and J are rectangles, then minJi ≤ minIi for all i,
which gives minJ ≤ minI . Similarly, maxJ ≤ maxI , and one can also see that minI < maxJ must hold, or
else I ∩ J = ∅. We summarize these observations as a corollary of Lemma 4.3:
Corollary 4.4. Let R and S be rectangles, and let χ : IR → IS be a nonzero morphism. Then minS ≤ minR
and maxS ≤ maxR .
This will come in handy when we prove Lemmas 4.5, 4.6, and 4.7.
We define a function w : (B(M ) × B(N )) t (B(N ) × B(M )) → k by letting w(I, J) = x if fI,J is given
by multiplication by x, and w(I, J) = 0 if fI,J is the zero morphism. w(J, I) is given by gJ,I in the same
way.
With the definition of w, it is starting to become clear how combinatorics comes into the picture. We
can now construct a bipartite weighted directed graph on B(M ) t B(N ) by letting w(I, J) be the weight of
the edge from I to J. The reader is invited to keep this picture in mind, as a lot of what we do in the rest
of the proof can be interpreted as statements about the structure of this graph.
The following lemma allows us to break up the problem and focus on the components of M and N with
the same types separately.
Lemma 4.5. Let R and T be rectangles of the same type, and S be a rectangle of a different type. Then
ψχ = 0 for any pair χ : IR → IS , ψ : IS → IT of morphisms.
Proof. Suppose ψ, χ 6= 0. By Corollary 4.4, minR ≥ minS ≥ minT and maxR ≥ maxS ≥ maxT . We get
minRi ≥ minSi ≥ minTi and maxRi ≥ maxSi ≥ maxTi for all i, and it follows that if R and T are of the
same type, then S is of the same type as R and T .
Let f 0 : M → N (δ) be defined by fI,J
0
= fI,J for I ∈ B(M ) and J ∈ B(N ) if I and J are of the same
type, and fI,J = 0 if they are not, and let g 0 : N → M (δ) be defined analogously. Here f 0 and g 0 are
0
0 0
assembled from fI,J and gJ,I the same way f and g are from fI,J and gJ,I . Suppose I, I 0 ∈ B(M ). Then
we have
X X
0 0
gJ,I 0 (δ)fI,J = gJ,I 0 (δ)fI,J .
(5)
J∈B(N ) J∈B(N )

When I and I 0 are of different types, the left side is zero because f and g are δ-interleaving morphisms, and
all the summands on the right side are zero by definition of f 0 and g 0 . When I and I 0 are of the same type,
the equality follows from Lemma 4.5. This means that g 0 (δ)f 0 = g(δ)f . We also have f 0 (δ)g 0 = f (δ)g, so f 0
and g 0 are δ-interleaving morphisms. In particular, f 0 and g 0 are δ-interleaving morphisms when restricted
to the components of M and N of a fixed type. If we can show that f 0 and g 0 induce a (2n − 1)δ-matching
on each of the mentioned components, we will have proved Theorem 4.2. In other words, we have reduced
the problem to the case where all the intervals in B(M ) and B(N ) are of the same type.
For a decorated number a∗ , let u(a∗ ) = a if a 6= ±∞ and u(a∗ ) = 0 otherwise. Let a = (a1 , a2 , . . . , an )
be a decorated point. We P define P (a) to be the number of the decorated numbers ai decorated with +,
and we also define α(a) = 1≤i≤n u(ai ). What we really want to look at are rectangles and not decorated
points by themselves, so we define P (R) = P (minR ) + P (maxR ) and α(R) = α(minR ) + α(maxR ) for any
rectangle R. Define an order ≤α on decorated points given by a ≤α b if either
• α(a) < α(b), or
• α(a) = α(b) and P (a) ≤ P (b)
This defines a preorder. In other words, it is transitive (R ≤α S ≤α T implies R ≤α T ) and reflexive
(R ≤α R for all R). We write R <α S if R ≤α S and not R ≥α S.
The order ≤α is one of the most important ingredients in the proof. The point is that if there is a
nonzero morphism from IR to IS () and R ≤α S, then R and S have to be close to each other. If  = 0, R

9
Figure 4: Rectangles R = (0, 4) × (0, 4) (purple), S = (2, 5) × (2, 5) (pink), S(1) = (1, 4) × (1, 4) (dotted
border), and S(2) = (0, 3) × (0, 3) (dotted border).

and S actually have to be equal. This ‘closeness property’ is expressed in Lemma 4.6, and is also exploited
in Lemma 4.7. Finally, in the proof of Lemma 4.8, we make sure that we only have to deal with morphisms
gJ,I 0 (δ) ◦ fI,J for I ≤α I 0 and not I >α I 0 , so that our lemmas can be applied.
In Figure 4 we see two rectangles R = (0, 4) × (0, 4) and S = (2, 5) × (2, 5). There is no nonzero
morphism from IR to IS or IS(1) , because minR < minS() for all  < 2. This is connected to the fact that
α(R) = 8 < 14 = α(S), which can be interpreted to mean that R is ‘further down and to the left’ than S.
The point of including P (α) in the definition of α is that e.g. (a, b] is a tiny bit ‘further to the right’ than
[a, b), and this is a subtlety that P recognizes, and that matters in the proofs of Lemmas 4.6 and 4.7.
Lemma 4.6. Let R, S, and T be rectangles of the same type with R ≤α T . Suppose there are nonzero
morphisms χ : IR → IS () and ψ : IS → IT (). Then IS is (2n − 1)-interleaved with either IR or IT .
Proof. Since χ 6= 0, we have
• minS ≤ minR + 
• maxS ≤ maxR + .
This follows from Corollary 4.4.
Suppose IR and IS are not (2n − 1)-interleaved. Then either minS + (2n − 1)  minR or maxS +
(2n − 1)  maxR ; let us assume the latter. (The former is similar.) In this case, there is an m such that
maxSm < maxRm − (2n − 1). For i 6= m, we have maxSi ≤ maxRi +  by the second bullet point. We get
 
X X
u(maxSi ) ≤  u(maxRi ) − (2n − 1) + (n − 1)
1≤i≤n 1≤i≤n
  (6)
X
= u(maxRi ) − n.
1≤i≤n

The first bullet point gives us


 
X X
u(minSi ) ≤  u(minRi ) + n, (7)
1≤i≤n 1≤i≤n

10
so we get α(S) ≤ α(R). If the inequality is strict, we have S <α R. If not, we have
• u(minSi ) = u(minRi ) +  for all i
• u(maxSi ) = u(maxRi ) +  for i 6= m
• u(maxSm ) = u(maxRm ) − (2n − 1).
Because of the inequalities minS ≤ minR +  and maxS ≤ maxR +  (recall that these are inequalities of
decorated points with the poset structure we defined earlier), we have P (minSi ) ≤ P (minRi ) for all i and
P (maxSi ) ≤ P (maxRi ) for i 6= m. But since maxSm < maxRm −(2n−1), we have P (maxSm ) < P (maxRm ),
so S <α R. Similarly, we can prove T <α S if IS and IT are not (2n − 1)-interleaved, so we have T <α R,
which is a contradiction.
Lemma 4.7. Let R, S, and T be rectangles of the same type with R and T (4n − 2)-significant and
α(R) ≤ α(T ). Suppose there are nonzero morphisms χ : IR → IS () and ψ : IS → IT (). Then ψ() ◦ χ =
6 0.
The constant (4n − 2) can be improved on for n > 1, but since the constant (2n − 1) in Lemma 4.6 is
optimal, strengthening Lemma 4.7 will not help us get a better constant in Theorem 4.2.
Proof. Suppose that χ and ψ are nonzero, but ψ() ◦ χ = 0. We have

• minR + 2 ≥ minT
• minRm + 2 ≥ maxTm for some m
• maxR + 2 ≥ maxT
• maxRm ≥ maxTm + (4n − 4).

The first and third statements hold because χ, ψ 6= 0. (See Corollary 4.4.) The second is equivalent to
minR ≮ maxT (2) . If this did not hold, R and T (2) would intersect, and ψ() ◦ χ would be nonzero in this
intersection, which is a contradiction. The fourth statement follows from the second and the fact that R is
(4n − 2)-significant.
Since T is (4n − 2)-significant, minT + (4n − 2) < maxT . Thus the second bullet point implies that
minRm + 2 > minTm + (4n − 2). The first point gives minRi ≥ minTi − 2 for i 6= m. In a similar fashion,
we get from the last two points that maxRm ≥ maxT + (4n − 4) and maxRi ≥ maxTi − 2 for i 6= m. From
all this, we get
X
α(R) = u(minRi ) + u(maxRi )
1≤i≤n
X
≥ u(minTm ) + u(maxTm ) + 2(4n − 4) + (u(minTi ) + u(maxTi ) − 4)
(8)
i6=m

= α(T ) + (4n − 4)


≥ α(T ).

Equality only holds if u(minTm )+(4n−2) = u(maxTm ), u(minRm )+(4n−2) = u(maxRm ), and n = 1. This
means that R = R1 = T = T1 = [u(minR ), u(minR ) + 2]. As we see, R ∩ T (2) = [u(minR ), u(minR )] 6= ∅,
so ψ() ◦ χ 6= 0.
We define a function µ by

µ(I) = {J ∈ B(N ) | I and J are (2n − 1)δ-interleaved} (9)

for I in B(M ). In other words, µ(I) contains all the intervals that can be matched with I in a (2n − 1)δ-
matching. Let I ∈ B(M ) be (4n − 2)δ-significant, and pick p ∈ Rn such that p, p + (4n − 2)δ ∈ I.
Then, p + (2n − 1)δ ∈ J for everySJ ∈ µ(I). Since M and N are p.f.d., this means that µ(I) is a finite set.
For A ⊂ B(M ), we write µ(A) = I∈A µ(I).

11
Lemma 4.8. Let A be a finite subset of B(M ) containing no (4n − 2)δ-trivial elements. Then |A| ≤ |µ(A)|.

Before we prove Lemma 4.8, we show that it implies that there is a (2n − 1)δ-matching between B(M )
and B(N ) and thus completes the proof of Theorem 4.2.
Let Gµ be the undirected bipartite graph on B(M ) t B(N ) with an edge between I and J if J ∈ µ(I).
Observe that Gµ is the same as the graph G(2n−1)δ we defined when we gave the graph theoretical definition
of an -matching (in this case, (2n − 1)δ-matching) in section 2. Following that definition, a (2n − 1)δ-
matching is a matching in Gµ that covers the set of all (4n − 2)δ-significant elements in B(M ) and B(N ).
For a subset S of a graph G, let AG (S) be the neighbourhood of S in G, that is, the set of vertices in
G that are adjacent to at least one vertex in S. We now apply Hall’s marriage theorem [18] to bridge the
gap between Lemma 4.8 and the statement we want to prove about matchings.
Theorem 4.9 (Hall’s theorem). Let G be a bipartite graph on bipartite sets X and Y such that AG ({x})
is finite for all x ∈ X. Then the following are equivalent:
• for all X 0 ⊂ X, |X 0 | ≤ |AG (X 0 )|
• there exists a matching in G covering X.
One of the two implications is easy, since if |X 0 | > |AG (X 0 )| for some X 0 ⊂ X, then there is no matching
in G covering X 0 . It is the other implication we will use, namely that the first statement is sufficient for a
matching in G covering X to exist.
Letting X be the set of (4n − 2)δ-significant intervals in B(M ) and Y be B(N ), Hall’s theorem and
Lemma 4.8 give us a matching σ in the graph Gµ covering all the (4n − 2)δ-significant elements in B(M ).4
By symmetry, we also have a matching τ in Gµ covering all the (4n − 2)δ-significant elements in B(N ).
Neither of these is necessarily a (2n − 1)δ-matching, however, as each of them only guarantees that all
the (4n − 2)-significant intervals in one of the barcodes are matched. We will use σ and τ to construct a
(2n − 1)δ-matching. This construction is similar to one used to prove the Cantor-Bernstein theorem [1, pp.
110-111].
Let H be the undirected bipartite graph on B(M ) t B(N ) for which the set of edges is the union of the
edges in the matchings σ and τ . Let C be a connected component of H. Suppose the submatching of σ in C
does not cover all the (4n − 2)δ-significant elements of C. Then there is a (4n − 2)δ-significant J ∈ C ∩ B(N )
that is not matched by σ. If we view σ and τ as partial bijections σ : B(M ) 9 B(N ) and τ : B(N ) 9 B(M ),
we can write the connected component of J, which is C, as {J, τ (J), σ(τ (J)), τ (σ(τ (J))), . . . }. Either this
sequence is infinite, or it is finite, in which case the last element is (4n − 2)δ-trivial. In either case, we get
that the submatching of τ in C covers all (4n − 2)δ-significant elements in C.
By this argument, there is a (2n − 1)δ-matching in each connected component of H. We can piece these
together to get a (2n − 1)δ-matching in B(M ) t B(N ), so Lemma 4.8 completes the proof of Theorem 4.2.

Proof of Lemma 4.8. Because ≤α is a preorder, we can order A = {I1 , I2 , . . . , Ir } so that Ii ≤α Ii0 for all
i ≤ i0 . Write µ(A) = {J1 , J2 , . . . , Js }. For I ∈ B(M ), we have

φII ,2δ = πI (2δ)g(δ)f |I


 
X
= πI (2δ)  g|J πJ  (δ)f |I
J∈B(N )
X (10)
= πI (2δ)g|J (δ)πJ (δ)f |I
J∈B(N )
X
= gJ,I (δ)fI,J .
J∈B(N )

4 Strictly speaking, Lemma 4.8 says nothing about infinite A, but the case with A countably infinite follows from the finite

cases. Each interval in A contains a rational point, so since M is p.f.d., the cardinality of A is at most finite times countably
infinite, which is countable. Thus we have covered all the possible cases.

12
Also, J∈B(N ) gJ,I 0 (δ)fI,J = 0 for I 6= I 0 ∈ B(M ), since φM,2δ is zero between different components of M .
P

Lemma 4.6 says that if gJ,I 0 (δ)fI,J 6= 0 and I ≤α I 0 , then J is (2n − 1)δ-interleaved with either I or J 0 .
This means that if i < i0 , then
X
0= gJ,Ii0 (δ)fIi ,J
J∈B(N )
X (11)
= gJ,Ii0 (δ)fIi ,J ,
J∈µ(A)

as gJ,Ii0 (δ)fIi ,J = 0 for all J that are not (2n − 1)δ-interleaved with either Ii or Ii0 . Similarly,
X
φIIi ,2δ = gJ,Ii (δ)fIi ,J
J∈B(N )
X (12)
= gJ,Ii (δ)fIi ,J .
J∈µ(A)

Writing this in matrix form, we get


 φM ? ... ? 
 " II1 ,2δ
gJ1 ,I1 (δ) ... gJs ,I1 (δ) fI1 ,J1 ... fIr ,J1 #
0 φM I ... ?
I 2 ,2δ
.. .. .. ..
 
.. .. = .
. . .. .. .. ..
 
. . . .
.
 
gJ1 ,Ir (δ) ... gJs ,Ir (δ) fI1 ,Js ... fIr ,Js . . .
0 0 ... φM I
I r ,2δ

That is, on the right-hand side we have the internal morphisms of the Ii on the diagonal, and 0 below the
diagonal.
Recall that a morphism between rectangle modules can be identified with a k-endomorphism, and that in
our notation, fI,J and gJ,I are given by multiplication by w(I, J) and w(J, I), respectively. For an arbitrary
morphism ψ between rectangle modules, we introduce the notation w(ψ) = c if ψ is given by multiplica-
tion by c, and 0 otherwise. A consequence of Lemma 4.7 is that w(gJ,Ii0 (δ)fIi ,J ) = w(gJ,Ii0 )w(fIi ,J ) =
w(J, Ii )w(Ii0 , J) whenever Ii ≤α Ii0 , in particular if i ≤ i0 . We get

1 = w φII ,2δ
 
X
= w gJ,Ii (δ)fIi ,J 
J∈µ(A)
X (13)
= w(gJ,Ii (δ)fIi ,J )
J∈µ(A)
X
= w(J, Ii )w(Ii , J),
J∈µ(A)

w(J, Ii0 )w(Ii , J) for i ≤ i0 . Again we can interpret this as a matrix equation:
P
and similarly 0 = J∈µ(A)
 
   1 ? ... ?
w(J1 , I1 ) . . . w(Js , I1 ) w(I1 , J1 ) . . . w(Ir , J1 ) 0 1
.. .. .. .. ... ?
 ..  ..  =  .. ..
 
..  .

 . . .  . . . ..
. . . .
w(J1 , Ir ) . . . w(Js , Ir ) w(I1 , Js ) . . . w(Ir , Js )
0 0 ... 1

That is, the right-hand side is an r × r upper triangular matrix with 1’s on the diagonal. The right-hand
side has rank |A| and the left-hand side has rank at most |µ(A)|, so the lemma follows immediately from
this equation.

13
Figure 5: A bounded triangle T .

4.2 Block decomposable modules


Next, we prove stability for block decomposable modules, which, as explained in Section 3, implies stability
for zigzag modules and Reeb graphs. Let R2+ = {(x, y) ∈ R2 | x + y ≥ 0}.
Definition 4.10. A triangle is a nonempty set T of the form {(x, y) ∈ R2 | x < a, y < b} ∩ R2+ for some
(a, b) ∈ (R ∪ {∞})2 with a + b > 0.
It follows that triangles are intervals. For a triangle T = {(x, y) ∈ R2 | x < a, y < b} ∩ R2+ , we write
maxT = (a, b) ∈ (R∪{∞})2 . If T is bounded, maxT is the maximal element in the closure of T , as illustrated
in Figure 5. A triangle decomposable module is an interval decomposable R2 -module whose barcode only
contains triangles. Observe that triangles correspond exactly to blocks of the form (a, b)BL under the poset
isomorphism between Rop × R and R2 flipping the x-axis.
Theorem 4.11. Let M and N be δ-interleaved triangle decomposable modules. Then there is a δ-matching
between B(M ) and B(N ).
To prove this, we can split the triangles into sets of different ‘types’, as we did with the rectangles. We
get four different types of triangles T , depending on whether maxT is of the form (a, b), (∞, b), (a, ∞),
or (∞, ∞) for a, b ∈ R. Now a result analogous to Lemma 4.5 holds, implying that it is enough to show
Theorem 4.11 under the assumption that the barcodes only contain intervals of a single type. The case in
which the triangles are bounded is the hardest one, and the only one we will prove. So from now on, we
assume all triangles to be bounded.
Again, we reuse parts of the proof of Theorem 4.2. For I ∈ B(M ), we define ν(I) = {J ∈ B(N ) |
I and J are δ-interleaved}. The discussion about Hall’s theorem is still valid, so we only need to prove the
analogue of Lemma 4.8 for ν. Define α(T ) = m1 + m2 , where maxT = (m1 , m2 ). The only things we need
to complete the proof of the analogue of Lemma 4.8 for triangle decomposable modules are the following
analogues of Lemmas 4.6 and 4.7:
Lemma 4.12. Let R, S, and T be triangles with α(R) ≤ α(T ). Suppose there are morphisms f : IR → IS ()
and g : IS → IT () such that g() ◦ f 6= 0. Then IS is -interleaved with either IR or IT .
Lemma 4.13. Let R, S, and T be triangles with T 2-significant and α(R) ≤ α(T ). Suppose there are
nonzero morphisms f : IR → IS () and g : IS → IT (). Then g() ◦ f 6= 0.

14
Proof of Lemma 4.12. Suppose IR and IS are not -interleaved. Then maxS  maxR − . But at the same
time, maxR ≥ maxS − , which gives α(R) > α(S). Assuming that IS and IT are not -interleaved, either,
we also get α(S) > α(T ). Thus α(R) > α(T ), a contradiction.
Proof of Lemma 4.13. For all triangles I, we treat minI and maxI as undecorated points. We have maxT −
 ≤ maxS and maxS −  ≤ maxR , so maxT − 2 ≤ maxR . Because T is 2-significant, maxT − 2 − 0 ∈ R2+
for some 0 > 0. Combining these facts, we get maxT − 2 − 0 ∈ R, so (g() ◦ f )maxT −2−0 6= 0.
Theorem 4.11 implies dB (M, N ) = dI (M, N ) for block decomposable M and N such that B(M ) and
B(N ) only have blocks of the form (a, b)BL (so no closed or half-closed blocks). Our proof technique extends
easily to prove the same equality for all block decomposable M and N . In fact, dB (M, N ) ≤ dI (M, N ) in
the case where all the intervals in the barcodes are of the form [a, b]BL follows from Theorem 4.16 below
with n = 2 by the correspondence [a, b]BL ↔ h(−a, b)i, while the two cases with half-open blocks are both
essentially the algebraic stability theorem. In the end we could stitch the cases together by something
similar to Lemma 4.5 and the discussion following it. We omit the details, and anyway the closed and
half-open cases are taken care of in [3]. Thus, either by appealing to previous work for the other cases or
using our own methods, we get
Theorem 4.14. Let M and N be block decomposable modules. If M and N are δ-interleaved, there exists
a δ-matching between B(M ) and B(N ).

4.3 Free modules


Definition 4.15. We define a free interval as an interval of the form hpi := {q | q ≥ p} ⊂ Rn .
For a free interval R, we define minR by R = hminR i.5 We define a free Rn -module as an interval
decomposable module whose barcode only contains free intervals. It is easy to see that free intervals are
rectangles, so it follows from Theorem 4.2 that dB (M, N ) ≤ (2n − 1)dI (M, N ) for free modules M , N . But
because of the geometry of free modules, this result can be strengthened.
Theorem 4.16. Let M and N be free δ-interleaved Rn -modules with n ≥ 2. Then there is a (n − 1)δ-
matching between B(M ) and B(N ).
We already did most of the work while proving Theorem 4.2, and there are some obvious simplifications.
Firstly, free intervals are -significant for all  ≥ 0. Secondly, for all nonzero f : IR → IS and g : IS → IT with
R, S, T free, gf is nonzero. For I ∈ B(M ), define ν(I) = {J ∈ B(N ) | I and J are (n − 1)δ-interleaved}.
By the arguments in the proof of Theorem 4.2, we only need to prove Lemma 4.8 with µ replaced by ν.
Lemmas 4.6 and 4.7 still hold for free modules, but we need to sharpen Lemma 4.6.
Lemma 4.17. Let R, S, and T be free intervals with R ≤α T . Suppose there are morphisms 0 6= f : IR →
IS () and 0 6= g : IS → IT (). Then IS is (n − 1)-interleaved with either IR or IT .
Proof. In this proof, we treat minI and maxI as undecorated points for all free intervals I, so that we
can add them. We have minS ≤ minR + . Suppose IR and IS are not (n − 1)-interleaved. Then
minS + (n − 1)  minR , so for some m, we must have minSm < minRm − (n − 1). We get
X
α(S) = minSi
1≤i≤n
X
< minRm − (n − 1) + (minRi + )
i6=m (14)
X
= minRi
1≤i≤n

= α(R).
We can also prove that α(T ) < α(S) if IS and IT are not (n − 1)-interleaved, so we have α(T ) < α(R), a
contradiction.
5 This makes min
R an undecorated point, while we have previously defined min− as decorated points, but this does not
matter, as we will not need decorated points in this subsection.

15
Figure 6: M and N . I1 and I2 are the light purple squares, I3 is deep purple, and J is pink.

5 Counterexamples to a general algebraic stability theorem


Theorem 4.2 gives an upper bound of (2n − 1) on dB /dI for rectangle decomposable modules that increases
with the dimension. An obvious question is whether it is possible to improve this constant, or if for
each C < 2(n − 1) there exist pairs M, N of modules for which dB (M, N ) > CdI (M, N ), in which case
the bound is optimal. We know that dB (M, N ) ≥ dI (M, N ) for any M and N whenever the bottleneck
distance is defined, so for n = 1, the constant is optimal. For n > 1, however, it turns out that the equality
dB (M, N ) = dI (M, N ) does not always hold, and the geometry becomes more confusing when n increases. In
dimension 2, we give an example of rectangle decomposable modules M and N with dB (M, N ) = 3dI (M, N )
in Example 5.2, which means that the bound is optimal for n = 2, as well. This is a counterexample to a
conjecture made in a previous version of [3] which claims that interval decomposable Rn -modules M and
N such that B(M ) and B(N ) only contain convex intervals are -matched if they are -interleaved.
Example 5.1. Let B(M ) = {I1 , I2 , I3 }6 and B(N ) = {J}, where

• I1 = (−3, 1) × (−1, 3)
• I2 = (−1, 3) × (−3, 1)
• I3 = (−1, 1) × (−1, 1)
• J = (−2, 2) × (−2, 2).

See Figure 6. We can define 1-interleaving morphisms f : M → N (1) and g : N → M (1) by letting
w(I1 , J) = w(I2 , J) = w(I3 , J) = w(J, I1 ) = w(J, I2 ) = 1 and w(J, I3 ) = −1, where w is defined as in the
proof of Theorem 4.2. On the other hand, in any matching between B(M ) and B(N ) we have to leave either
I1 or I2 unmatched, and they are -significant for all  < 4. In fact, any possible matching between B(M )
and B(N ) is a 2-matching. Thus dI (M, N ) = 1 and dB (M, N ) = 2.

A crucial point is that even though w(I1 , J), w(J, I2 ), w(I2 , J), and w(J, I1 ) are all nonzero, both
gJ,I2 ◦ fI1 ,J and gJ,I1 ◦ fI2 ,J are zero. To do the same with one-dimensional intervals, we would have to
shrink I1 and I2 so much that they no longer would be 2-significant (see Lemma 4.7), and then they would
not need to be matched in a 1-matching. This shows how the geometry of higher dimensions can allow us
to construct examples that would not work in lower dimensions.
Next, we give an example of rectangle decomposable R2 -modules M and N such that dB (M, N ) =
3dI (M, N ), proving that our upper bound of 2(n − 1) is the best possible for n = 2.
Example 5.2. Let B(M ) = {I1 , I2 , I3 } and B(N ) = {J1 , J2 , J3 }, where
• I1 = (0, 10) × (1, 11)
6 Here we use subscripts to index different intervals, not to indicate projections, as we did earlier.

16
Figure 7: I1 , I2 , and I3 are the filled pink rectangles, and J1 , J2 , and J3 are the black rectangles without
fill.

• I2 = (0, 12) × (−1, 11)


• I3 = (2, 10) × (1, 9)
• J1 = (1, 11) × (0, 10)
• J2 = (1, 9) × (0, 12)
• J3 = (−1, 11) × (2, 10).
The rectangles in B(M ) and B(N ) are shown in Figure 7.
We give an example of 1-interleaving morphisms f and g that we write on matrix form. In the first
matrix, w(Ii , Jj ) is in row i, column j. In the second, w(Jj , Ii ) is in row j, column i.

   
1 1 1 −1 1 1
f : 1 1 0 , g: 1 0 −1 . (15)
1 0 1 1 −1 0

This means that M and N are 1-interleaved, but they are not -interleaved for any  < 1, so dI (M, N ) =
1.
Let  < 3. We see that the difference between maxI2 = (12, 11) and maxJ2 = (9, 12) is 3 in the first
coordinate, so I2 and J2 are not -interleaved, and thus they cannot be matched in an -matching. In fact,
Ii and Jj cannot be matched in an -matching for any i, j ∈ {2, 3} by similar arguments. Since I2 and
I3 cannot both be matched with J1 , one of them has to be left unmatched, but since both I2 and I3 are
6-significant, this means that there is no -matching between B(M ) and B(N ). On the other hand, any
bijection between B(M ) and B(N ) is a 3-matching, so dB (M, N ) = 3.
There is a strong connection between n-dimensional rectangle decomposable modules and 2n-dimensional
free modules. This is related to the fact that we need 2n coordinates to determine an n-dimensional rectangle,
and also 2n coordinates to determine a 2n-dimensional free interval. The following example illustrates this
connection, as we simply rearrange the coordinates of minR , maxR for all rectangles R involved in Example
5.2 to get 4-dimensional free modules with similar properties as in Example 5.2.

17
Example 5.3. Let B(M ) = {I1 , I2 , I3 } and B(N ) = {J1 , J2 , J3 }, where

• I1 = h(0, 1, 10, 11)i


• I2 = h(0, −1, 12, 11)i
• I3 = h(2, 1, 10, 9)i

• J1 = h(1, 0, 11, 10)i


• J2 = h(1, 0, 9, 12)i
• J3 = h(−1, 2, 11, 10)i.
(Compare with the intervals Ii and Jj in Example 5.2.) We have 1-interleaving morphisms defined the same
way as in Example 5.2. Just as in that example, we can deduce that there is nothing better than a 3-matching
between B(M ) and B(N ), so dB (M, N ) = 3 and dI (M, N ) = 1.
As a consequence of this example, we get that our upper bound of dB /dI ≤ n − 1 for free n-dimensional
modules cannot be improved on for n = 4.

6 Relation to the complexity of calculating interleaving distance


The interleaving distance between arbitrary persistence modules is on the surface not easy to find, as naively
trying to construct interleaving morphisms can quickly lead you to consider a complicated set of equations
for which it is not clear that one can decide if there is a solution in polynomial time. For R-modules, however,
the interval decomposition theorem plus the algebraic stability theorem gives us a polynomial time algorithm
to compute dI : decompose the modules into intervals and find the bottleneck distance. Since dI = dB , this
gives us the interleaving distance. When it exists, one can compute the bottleneck distance in polynomial
time also in two dimensions [17], but the approach fails for general Rn -modules already at the first step, as
we do not have a nice decomposition theorem. But in the recent proof that calculating interleaving distance
is NP-hard [9], it is the failure of the second step that is exploited. Specifically, a set of modules that
decompose nicely into interval modules (staircase modules, to be precise) is constructed, but for these, dI
and dB are different. It turns out that calculating dI for these corresponds to deciding whether CI problems
are solvable, which is shown to be NP-hard.
Though rectangle modules are not considered in the NP-hardness proof, they have similar properties
to staircase modules,7 and Example 5.2 is essentially a CI problem with a corresponding pair of modules.
Importantly, it shows that dI = dB does not hold in general for modules corresponding to CI problems.
This crucial observation, which appeared first in a preprint of this paper [8], opened the door to proving
NP-hardness of calculating dI by the approach used in [9].
In [9], it is also shown that c-approximating dI is NP-hard for c < 3, where an algorithm is said to
c-approximate dI if it returns a number in the interval [dI (M, N ), cdI (M, N )] for any input pair M , N of
modules. Whether the approach by CI problems can be used to prove hardness of c-approximation for c ≥ 3
is closely related to Theorem 4.2. It can be shown that if dB (M, N ) ≤ cdI (M, N ) for any pair M , N of
rectangle decomposable modules, the same holds for staircase modules, and therefore there is a polynomial
time algorithm c-approximating dI for these, meaning that the strategy of going through CI problems will
not give a proof that c-approximation of dI is NP-hard. On the other hand, if one can find an example of
rectangle decomposable modules M and N such that dB (M, N ) = cdI (M, N ) for c > 3, one might be able
to use that to increase the constant 3 in the approximation hardness result. Thus there is a strong link
between stability of rectangle decomposable modules and the only successful method so far known to the
author of determining the complexity of computing or approximating multiparameter interleaving distance.
7 The only significant difference in this setting is that in a fixed dimension, rectangle modules are defined by a limited

number of coordinates, or “degrees of freedom”, while there is no such restriction on staircase modules even in dimension 2.

18
7 Acknowledgements
I would like to thank my supervisors Gereon Quick and Nils Baas for invaluable support and help. I would
also like to thank Peter Landweber for detailed comments on several drafts of this text, Steve Oudot for
feedback on the first arXiv version and Magnus Bakke Botnan for interesting discussions.

References
[1] Martin Aigner and Günter M. Ziegler. Proofs from THE BOOK. Springer, Berlin, 4th edition, 2010.
[2] Gorô Azumaya. Corrections and supplementaries to my paper concerning Krull-Remak-Schmidt’s
theorem. Nagoya Mathematical Journal, 1:117–124, 1950.
[3] Magnus Bakke Botnan and Michael Lesnick. Algebraic stability of zigzag persistence modules. Algebraic
& Geometric Topology, 18, 04 2016.
[4] Ulrich Bauer, Xiaoyin Ge, and Yusu Wang. Measuring distance between reeb graphs. In Proceedings
of the thirtieth annual symposium on Computational geometry, page 464. ACM, 2014.

[5] Ulrich Bauer and Michael Lesnick. Induced matchings and the algebraic stability of persistence bar-
codes. Journal of Computational Geometry, 6(2):162–191, 2015.
[6] Ulrich Bauer, Elizabeth Munch, and Yusu Wang. Strong Equivalence of the Interleaving and Functional
Distortion Metrics for Reeb Graphs. In 31st International Symposium on Computational Geometry
(SoCG 2015), volume 34 of Leibniz International Proceedings in Informatics (LIPIcs), pages 461–475.
Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, 2015.
[7] Paul Bendich, Herbert Edelsbrunner, Dmitriy Morozov, Amit Patel, et al. Homology and robustness
of level and interlevel sets. Homology, Homotopy and Applications, 15(1):51–72, 2013.
[8] Håvard Bakke Bjerkevik. Stability of higher-dimensional interval decomposable persistence modules.
arXiv preprint arXiv:1609.02086v2, 2016.
[9] Håvard Bakke Bjerkevik, Magnus Bakke Botnan, and Michael Kerber. Computing the interleaving
distance is np-hard. arXiv preprint arXiv:1811.09165, 2018.
[10] Gunnar Carlsson and Vin De Silva. Zigzag persistence. Foundations of computational mathematics,
10(4):367–405, 2010.
[11] Gunnar Carlsson, Vin De Silva, and Dmitriy Morozov. Zigzag persistent homology and real-valued
functions. In Proceedings of the twenty-fifth annual symposium on Computational geometry, pages
247–256. ACM, 2009.
[12] Frédéric Chazal, David Cohen-Steiner, Marc Glisse, Leonidas J Guibas, and Steve Y Oudot. Proximity
of persistence modules and their diagrams. In Proceedings of the twenty-fifth annual symposium on
Computational geometry, pages 237–246. ACM, 2009.
[13] Frédéric Chazal, Vin De Silva, Marc Glisse, and Steve Oudot. The structure and stability of persistence
modules. Springer, 2016.

[14] David Cohen-Steiner, Herbert Edelsbrunner, and John Harer. Stability of persistence diagrams. Dis-
crete & Computational Geometry, 37(1):103–120, 2007.
[15] William Crawley-Boevey. Decomposition of pointwise finite-dimensional persistence modules. Journal
of Algebra and Its Applications, 14(05):1550066, 2015.
[16] Vin De Silva, Elizabeth Munch, and Amit Patel. Categorified reeb graphs. Discrete & Computational
Geometry, 55(4):854–906, 2016.

19
[17] Tamal K. Dey and Cheng Xin. Computing Bottleneck Distance for 2-D Interval Decomposable Modules.
In 34th International Symposium on Computational Geometry (SoCG 2018), volume 99 of Leibniz In-
ternational Proceedings in Informatics (LIPIcs), pages 32:1–32:15. Schloss Dagstuhl–Leibniz-Zentrum
fuer Informatik, 2018.
[18] Philip Hall. On representatives of subsets. J. London Math. Soc, 10(1):26–30, 1935.

[19] Woojin Kim and Facundo Memoli. Stable signatures for dynamic metric spaces via zigzag persistent
homology. arXiv preprint arXiv:1712.04064, 2017.
[20] Steve Y Oudot and Donald R Sheehy. Zigzag zoology: Rips zigzags for homology inference. Foundations
of Computational Mathematics, 15(5):1151–1186, 2015.
[21] Cary Webb. Decomposition of graded modules. Proceedings of the American Mathematical Society,
94(4):565–571, 1985.

20

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy