Topology and Data: Bulletin of The American Mathematical Society April 2009

See discussions, stats, and author profiles for this publication at: https://www.researchgate.
net/publication/243073634
Topology and Data
Article in Bulletin of the American Mathematical Society · April 2009

DOI: 10.1090/S0273-0979-09-01249-X
CITATIONS READS
556 2,156
1 author:
Gunnar Carlsson
Stanford University
157 PUBLICATIONS 6,648 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Ayasdi View project
Tropical Coordinates on the Spaces of Barcodes View project
All content following this page was uploaded by Gunnar Carlsson on 28 March 2015.
The user has requested enhancement of the downloaded file.

BULLETIN (New Series) OF THE
AMERICAN MATHEMATICAL SOCIETY
Volume 46, Number 2, April 2009, Pages 255–308
S 0273-0979(09)01249-X
Article electronically published on January 29, 2009
TOPOLOGY AND DATA
GUNNAR CARLSSON
1. Introduction
An important feature of modern science and engineering is that data of various
kinds is being produced at an unprecedented rate. This is so in part because of
new experimental methods, and in part because of the increase in the availability
of high powered computing technology. It is also clear that the nature of the data
we are obtaining is significantly different. For example, it is now often the case
that we are given data in the form of very long vectors, where all but a few of the
coordinates turn out to be irrelevant to the questions of interest, and further that
we don’t necessarily know which coordinates are the interesting ones. A related
fact is that the data is often very high-dimensional, which severely restricts our
ability to visualize it. The data obtained is also often much noisier than in the
past and has more missing information (missing data). This is particularly so in
the case of biological data, particularly high throughput data from microarray or
other sources. Our ability to analyze this data, both in terms of quantity and the
nature of the data, is clearly not keeping pace with the data being produced. In this
paper, we will discuss how geometry and topology can be applied to make useful
contributions to the analysis of various kinds of data. Geometry and topology are
very natural tools to apply in this direction, since geometry can be regarded as the
study of distance functions, and what one often works with are distance functions
on large finite sets of data. The mathematical formalism which has been developed
for incorporating geometric and topological techniques deals with point clouds, i.e.
finite sets of points equipped with a distance function. It then adapts tools from
the various branches of geometry to the study of point clouds. The point clouds are
intended to be thought of as finite samples taken from a geometric object, perhaps
with noise. Here are some of the key points which come up when applying these
geometric methods to data analysis.
• Qualitative information is needed: One important goal of data analysis
is to allow the user to obtain knowledge about the data, i.e. to understand
how it is organized on a large scale. For example, if we imagine that
we are looking at a data set constructed somehow from diabetes patients,
it would be important to develop the understanding that there are two
types of the disease, namely the juvenile and adult onset forms. Once
that is established, one of course wants to develop quantitative methods for
distinguishing them, but the first insight about the distinct forms of the
disease is key.
Received by the editors August 1, 2008.

Research supported in part by DARPA HR 0011-05-1-0007 and NSF DMS 0354543.
2009
c American Mathematical Society
255
256 GUNNAR CARLSSON
• Metrics are not theoretically justified: In physics, the phenomena

studied often support clean explanatory theories which tell one exactly
what metric to use. In biological problems, on the other hand, this is much
less clear. In the biological context, notions of distance are constructed
using some intuitively attractive measures of similarity (such as BLAST
scores or their relatives), but it is far from clear how much significance to
attach to the actual distances, particularly at large scales.
• Coordinates are not natural: Although we often receive data in the form
of vectors of real numbers, it is frequently the case that the coordinates,
like the metrics mentioned above, are not natural in any sense, and that
therefore we should not restrict ourselves to studying properties of the
data which depend on any particular choice of coordinates. Note that the
variation of choices of coordinates does not require that the coordinate
changes be rigid motions of Euclidean space. It is often a tacit assumption
in the study of data that the coordinates carry intrinsic meaning, but this
assumption is often unjustified.
• Summaries are more valuable than individual parameter choices:
One method of clustering a point cloud is the so-called single linkage clus-
tering, in which a graph is constructed whose vertex set is the set of points
in the cloud, and where two such points are connected by an edge if their
distance is ≤ , where is a parameter. Some work in clustering theory has
been done in trying to determine the optimal choice of , but it is now well
understood that it is much more informative to maintain the entire den-
drogram of the set, which provides a summary of the behavior of clustering
under all possible values of the parameter at once. It is therefore pro-
ductive to develop other mechanisms in which the behavior of invariants or
construction under a change of parameters can be effectively summarized.
In this paper, we will discuss methods for dealing with the properties and prob-
lems mentioned above. The underlying idea is that methods inspired by topology
should address them. For each of the points above, we describe why topological
methods are appropriate for dealing with them.
• Topology is exactly that branch of mathematics which deals with qualita-

tive geometric information. This includes the study of what the connected
components of a space are, but more generally it is the study of connec-
tivity information, which includes the classification of loops and higher
dimensional surfaces within the space. This suggests that extensions of
topological methodologies, such as homology, to point clouds should be
helpful in studying them qualitatively.
• Topology studies geometric properties in a way which is much less sensitive
to the actual choice of metrics than straightforward geometric methods,
which involve sensitive geometric properties such as curvature. In fact,
topology ignores the quantitative values of the distance functions and re-
places them with the notion of infinite nearness of a point to a subset in
the underlying space. This insensitivity to the metric is useful in studying
situations where one only believes one understands the metric in a coarse
way.
TOPOLOGY AND DATA 257
• Topology studies only properties of geometric objects which do not depend

on the chosen coordinates, but rather on intrinsic geometric properties of
the objects. As such, it is coordinate-free.
• The idea of constructing summaries over whole domains of parameter val-
ues involves understanding the relationship between geometric objects con-
structed from data using various parameter values. The relationships which
are useful involve continuous maps between the different geometric objects,
and therefore become a manifestation of the notion of functoriality, i.e, the
notion that invariants should be related not just to objects being studied,
but also to the maps between these objects. Functoriality is central in al-
gebraic topology in that the functoriality of homological invariants is what
permits one to compute them from local information, and that functoriality
is at the heart of most of the interesting applications within mathematics.
Moreover, it is understood that most of the information about topological
spaces can be obtained through diagrams of discrete sets, via a process of
simplicial approximation.
The last point above, concerning functoriality, is critical. In developing methods
to address the first two points, we find that we are forced to make functorial geo-
metric constructions and analyze their behavior on maps even to obtain information
about single point clouds. Functoriality has proven itself to be a powerful tool in
the development of various parts of mathematics, such as Galois theory within al-
gebra, the theory of Fourier series within harmonic analysis, and the applicaton of
algebraic topology to fixed point questions in topology. We argue that, as suggested
in [46], it has a role to play in the study of point cloud data as well, and we give
two illustrations of how this could happen, within the context of clustering.
Informally, clustering refers to the process of partitioning a set of data into
a number of parts or clusters, which are recognizably distinguishable from each
other. In the context of finite metric spaces, this means roughly that points within
the clusters are nearer to each other than they are to points in different clusters.
Clustering should be thought of as the statistical counterpart to the geometric con-
struction of the path-connected components of a space, which is the fundamental
building block upon which algebraic topology is based. There are many schemes
which construct clusterings based on metric information, such as single, average,
and complete linkage clustering, k-means clustering, spectral clustering, etc. (see
[31]). Although clustering is clearly a very important part of data analysis, the
ways in which it is formulated and implemented are fraught with ambiguities. In
particular, the arbitrariness of various threshhold choices and lack of robustness
are difficulties one confronts. Much of current research efforts are focused in this
direction (see, e.g., [43] and [39]), and functoriality provides the right general math-
ematical framework for addressing them. For example, one can construct data sets
which have been threshholded at two different values, and the behavior of clusters
under the inclusion of the set with tighter threshhold into the one with the looser
threshhold is informative about what is happening in the data set. We present two
additional examples of how functoriality could be used in analyzing some questions
related to clustering.
Example. In the case of very large X, it may often be difficult to apply the
clustering algorithm to a full data set, and one may instead find it desirable to
cluster subsamples from X. One is then confronted with the task of attempting to
258 GUNNAR CARLSSON
verify that the clustering of the subsample is actually representative of a clustering

of the full data set X. One way of proceeding is to construct two samples from X,
hoping that they are consistent in an appropriate sense. One version of this idea
would be to consider the subsamples X1 and X2 , together with their union X1 ∪X2 .
One could apply the clustering scheme to each of these sets individually: suppose
we denote the set of clusters for the three sets X1 , X2 , and X1 ∪ X2 by C(X1 ),
C(X2 ), and C(X1 ∪ X2 ), respectively. If the clustering scheme were functorial, i.e.
if inclusions of data sets induced maps of the collections of clusters, then one would
have a diagram of sets
C(X1 ∪ X2 )
q8 fMMM
qqqq
q MMM
MMM
qq M
qq
C(X1 ) C(X2 ).
If the clusterings are consistent, i.e. if the clusters in C(X1 ) and C(X2 ) in
C(X1 ∪ X2 ) correspond well under these maps, then one can regard that as ev-
idence that the subsample clusterings actually correspond to clusterings on the full
data set X. Of course, what the phrase “correspond well” means is not well de-
fined here. Later in the paper, we will discuss a way to attach more quantitative
information to questions of this type.
Example. Suppose that we have data X which varies with time. One could then
ask for information concerning the behavior of clusterings produced by clusterings
over time. Clusters can appear, vanish, merge, or split into separate clusters. The
analysis of this behavior can be studied using functoriality. For t0 < t1 , we let
X[t0 , t1 ] denote the set of points in the data set occurring between times t0 and t1 .
If we have t0 < t1 < t2 < t3 , then we have a diagram of point cloud data sets
X[t1 , t2 ]
ss LLL
ss LLL
sss LLL
ysss L%
X[t0 , t2 ] X[t1 , t3 ].
If the clustering scheme is functorial in the sense of the preceding example, we
obtain the corresponding diagram of sets
C(X[t1 , t2 ])
oo OOO
ooo OOO
oo OOO
oo OOO
woo '
C(X[t0 , t2 ]) C(X[t1 , t3 ]).
This set contains information about the behavior over time of the clusters. For
example, the diagram
would correspond to a single cluster at time t0 , which breaks into two clusters in
the interval [t1 , t2 ], which in turn merge back again in the interval [t2 , t3 ].
This paper will deal with a number of methods for thinking about data using
topologically inspired methods. We begin with a discussion of persistent homology,
which is a mathematical formalism which permits us to infer topological information
from a sample of a geometric object, and we show how it can be applied to particular
data sets arising from natural image statistics and neuroscience. Next, we show that
topological methods can produce a kind of imaging of data sets, not by embedding in
Euclidean space but rather by producing a simplicial complex associated to certain
initial information about the data set. We then demonstrate that persistence can be
generalized in several different directions, providing more structure and information
about the data sets in question. We then show that the philosophy of functoriality
can be used to reason about the nature of clustering methods, and we conclude
by speculating about theorems one might hope to prove and discussing how the
subject might develop more generally.
The author is extremely grateful to his collaborators A. Blumberg, A. Collins,
V. de Silva, L. Guibas, T. Ishkanov, F. Memoli, M. Nicolau, D. Ringach, G. Sapiro,
H. Sexton, G. Singh, and A. Zomorodian. In addition, he is grateful to a number
of people for very useful conversations on this subject. They include T. Beke, P.
Diaconis, R. Ghrist, S. Holmes, K. Mischaikow, P. Niyogi, S. Oudot, S. Smale, and
S. Weinberger.
2. Persistence and homology

2.1. Introduction. In thinking about qualitative properties of spaces X, an obvi-
ous one which comes to mind is its decomposition into path-connected components.
It is a partition of X, and the cardinality of the collection of blocks in this partition
is an invariant of X and surely deserves to be called a qualitative invariant of X.
There are other qualitative properties to be considered.
Example. Consider the two spaces in the image below.
We note that both spaces are path-connected, but we can see that they are quali-
tatively distinct in that the letter “B” on the right has two different loops and the
“A” on the left has only one. This property is preserved under continuous defor-
mations, and so if one can formalize it into a precise mathematical statement, one
can then rigorously distinguish between the two spaces. We refer to information
about loops and higher dimensional analogues in a space as “connectivity informa-
tion”. The decomposition into path connected components would be regarded as
zeroth-level connectivity information, loops as level-one connectivity information,
260 GUNNAR CARLSSON
and so forth. The mathematical formalism which makes these notions precise is al-
gebraic topology. To give an idea of what it achieves, we first consider the following
situation.
The space in question is a rectangle with a circular disc removed. There are three
loops in the picture. The two left-hand loops are regarded as being essentially
the same because it is possible to deform or drag the outer loop onto the inner
loop without the outer loop leaving the space by crossing the disc. On the other
hand, the outer left-hand loop and the right-hand loop are considered essentially
different because in order to drag the left-hand loop onto the right-hand loop,
one must either cross the deleted disc or tear the loop. This notion of essential
sameness and essential difference can be encoded into an equivalence relation, and
algebraic topology succeeds in counting or parametrizing the equivalence classes
of loops under this equivalence relation. It also provides invariants which carry
out the same parametrization in the case of higher dimensional “loops”, by which
we mean maps of higher dimensional spheres into the space in question. The set
of equivalence classes is referred to as a homotopy group, since it turns out that
juxtaposition of loops and higher dimensional versions puts a group structure on the
set of equivalence classes. Unfortunately, these groups are frequently very difficult
to compute, since the computations involve generator and relation computations
in free groups. There are more computable invariants, though, called homology
groups, in which (roughly) the equivalence relation on loops is extended by declaring
that two loops are equivalent if there is a surface with boundary equal to the union
of the two loops. These homology groups then provide an infinite vector of integers,
the zeroth of which counts the number of connected components, and so that the
i-th one counts the number of equivalence classes (under the second equivalence
relation) of i-dimensional surfaces in the space. We describe the formalism in more
standard terms. See [33] for a thorough treatment. We recall that two continuous
maps f, g : X → Y are said to be homotopic if there is a continuous map H :
X × [0, 1] → Y so that H(x, 0) = f (x) and H(x, 1) = g(x). We also say that a map
f : X → Y is a homotopy equivalence if there is a map G : Y → X so that f ◦ g
is homotopic to the identity map on Y and g ◦ f is homotopic to f . Two spaces
X and Y for which there exists a homotopy equivalence f : X → Y are said to be
homotopy equivalent. A space that is homotopy equivalent to the one point space
is said to be contractible.
• Definition: For any topological space X, abelian group A, and integer
k ≥ 0, there is assigned a group Hk (X, A).
• Functoriality: For any A and k as above, and any continuous map f :

X → Y , there is an induced homomorphism Hk (f, A) : Hk (X, A) →
Hk (Y, A). One has Hk (f ◦ g, A) = Hk (f, A) ◦ Hk (g, A) and Hk (IdX ; A) =
IdHk (X,A) . These conditions are called collectively functoriality. We refer
the reader to [44] for a treatment of categories and functors.
• Homotopy invariance: If f and g are homotopic, then Hk (f, A) =
Hk (g, A). It follows that if X and Y are homotopy equivalent, then
Hk (X, A) is isomorphic to Hk (Y, A).
• Normalization: H0 (∗, A) ∼ = A, where ∗ denotes the one point space.
• Betti numbers: For any field F , Hk (X, F ) will be a vector space over
F . Its dimension, if it is finite dimensional, will be written as βk (X, F ),
and will be referred to as the k-th Betti number with coefficients in F .
The k-th Betti number corresponds to an informal notion of the number of
independent k-dimensional surfaces. If two spaces are homotopy equivalent,
then all their Betti numbers are equal.
Example. For any topological space X with a finite number of path components,
β0 (X) is the number of path components.
Example. The first Betti number β1 of the letter “B” above is two, and for the
letter “A” it is one, for any field F . In this case, it provides a formalization of the
count of the number of loops present in the space.
Remark. The Betti numbers can vary with the choice of the coefficient field F .
Most calculations in this paper will be with coefficients in the two element field F2 .
The actual definition of homology that applies to all topological spaces (singular
homology) was introduced in [27]. It relies on the linear algebra of infinitely gen-
erated modules over the ring Z in defining homology groups, and for this reason
it is not useful from a computational point of view. Computations can be carried
out by hand using a variety of techniques (long exact sequence of a pair, long exact
Mayer-Vietoris sequence, excision theorem, spectral sequences), but direct compu-
tation from the definition is not feasible for general spaces. However, when one
is given a space equipped with particular structures, there are often finite linear
algebra problems which produce correct answers, i.e. answers which agree with the
singular technique. A particularly nice example of this applies when the space in
question is described as a simplicial complex.
Definition 2.1. An abstract simplicial complex is a pair (V, Σ), where V is a finite
set, and Σ is a family of non-empty subsets of V such that σ ∈ Σ and τ ⊆ σ implies
that τ ∈ Σ. Associated to a simplicial complex is a topological space |(V, Σ)|, which
using a bijection φ : V → {1, 2, . . . , N } as the subspace of R given
N
may be defined
by the union σ∈Σ c(σ), where c(σ) is the convex hull of the set {eφ(s) }s∈σ , where
ei denotes the ith standard basis vector.
Intuitively, a simplicial complex structure on a space is an expression of the
space as a union of points, intervals, triangles, and higher dimensional analogues.
Simplicial complexes provide a particularly simple combinatorial way to describe
certain topological spaces. For this reason, one often attempts to approximate (in
various senses) topological spaces by simplicial complexes. The key point for this
section, though, is that for simplicial complexes, the homology can be computed
using only the linear algebra of finitely generated Z-modules. We describe this in
262 GUNNAR CARLSSON
detail. Given any simplicial complex X = (V, Σ), we write Σk for the subset of Σ
consisting of all σ ∈ Σ for which #(σ) = k + 1. Elements of Σk are referred to
as k-simplices. We define the group of k-chains in X as the group of formal linear
combinations of elements in Σk , or equivalently the free abelian group on the set
Σk , and denote it by Ck (X). If we impose a total order on the vertex set V , we
define set operators di : Σk → Σk−1 , for 0 ≤ i ≤ k, by letting di (σ) = σ − {si },
where si denotes the i-th element in σ, under the given total ordering. We now
define linear operators ∂k : Ck (X) → Ck−1 (X) by

k
∂k = (−1)i di .
i=0
Since the groups Ck (X) are equipped with the bases Σk , these operators can be
expressed as matrices D(k) whose columns are parametrized by Σk , whose rows
are parametrized by Σk−1 , and where for σ ∈ Σk and τ ∈ Σk−1 , the entry D(k)τ σ
is = 0 if τ ⊂ σ, and = (−1)i if τ ⊆ σ and if τ is obtained by removing the i-th
member of the subset σ. The key observation is now that ∂k ◦ ∂k+1 ≡ 0. It follows
that Image(∂k+1 ) ⊆ Kernel(∂k ), and that one can therefore define Hksimp (X, Z)
by
Hksimp (X, Z) ∼
= Kernel(∂k )/Image(∂k+1 ).
This basis-independent version of the definition can be replaced by the result of
matrix manipulations on the collection of matrices {D(k)}k≥0 , as in [22]. The end
results of these calculations are always the Smith normal form of various matrices
constructed out of the D(k)’s. It turns out that Hksimp (X, Z) is always canonically
isomorphic to the singular homology of the space |X|. The conclusion is that for
simplicial complexes, homology is algorithmically computable, although standard
methods for computing Smith normal form can often be painfully slow.
2.2. Building coverings and complexes. Since the homology of simplicial com-
plexes is algorithmically computable, it is frequently desirable to construct simpli-
cial complexes which compute the homology of an underlying space X, or at least
have a strong relationship with it. One way to guarantee that the simplicial com-
plex computes the homology of X is to produce a homotopy equivalence from X to
the simplicial complex, or a homotopy equivalence from a space homotopy equiv-
alent to X to the simplicial complex. There are a number of simplicial complexes
which can be constructed from X together with additional data attached to X. We
begin with the nerve of a covering of a topological space. Let X be a topological
space, and let U = {Uα }α∈A be any covering of X.
Definition 2.2. The nerve of U, denoted by N U, will be the abstract simplicial
complex with vertex set A, and where a family {α0 , . . . , αk } spans a k-simplex if
and only if Uα0 ∩ . . . ∩ Uαk = ∅.
This is an extremely useful construction in homotopy theory. One reason is
that one has the following “nerve theorem” (see [5]), which provides criteria which
guarantee that N (U) is homotopy equivalent to the underlying space X.
Theorem 2.3. Suppose that X and U are as above, and suppose that the covering
consists of open sets and is numerable
(see [51] for a definition). Suppose further
that for all ∅ = S ⊆ A, we have that s∈S Us is either contractible or empty. Then
N (U) is homotopy equivalent to X.
One now needs methods for generating coverings. When the space in question is
a metric space, one covering is given by the family B (X) = {B (x)}x∈X , for some
> 0. More generally, for any subset V ⊆ X for which X = v∈V B (v), one can
construct the nerve of the covering {B (v)}v∈V . We will denote this construction
by Č(V, ), and refer to it as the Čech complex attached to V and . This is a useful
theoretical construction, in view of the following theorem.
Theorem 2.4. Let M be a compact Riemannian manifold. Then there is a positive
number e so that Č(M, ) is homotopy equivalent to M whenever ≤ e. Moreover,
for every ≤ e, there is a finite subset V ⊆ M so that the subcomplex of Č(V, ) ⊆
Č(M, ) is also homotopy equivalent to M .
One problem with this construction is that it is computationally expensive, in
that it requires the storage of simplices of various dimensions. An idea for dealing
with that problem is to construct a simplicial complex which can be recovered
solely from the edge information. This suggests the following variant of the Čech
construction, referred to as the Vietoris-Rips complex.
Definition 2.5. Let X denote a metric space, with metric d. Then the Vietoris-
Rips complex for X, attached to the parameter , denoted by V R(X, ), will be
the simplicial complex whose vertex set is X, and where {x0 , x1 , . . . , xk } spans a
k-simplex if and only if d(xi , xj ) ≤ for all 0 ≤ i, j ≤ k.
We note that the vertex sets of the two constructions are identical, so they
can both be viewed as subcomplexes of the complete simplex on the set X. The
following diagram indicates the differences between the complexes.
The leftmost figure shows the covering, the middle the Čech complex, and the
rightmost the Vietoris-Rips. Note that the parameter for computing the Rips
complex corresponds to the radii of the balls in question, while the parameter for
the Vietoris-Rips complex is the distance between centers of the balls. This leads
to the following comparison between the two complexes. We will see how to make
use of it in the next section.
Proposition 2.6. We have the inclusions
Č(X, ) ⊆ V R(X, 2) ⊆ Č(X, 2).
Even the Vietoris-Rips complex is computationally expensive, though, due to the
fact that its vertex set consists of the entire metric space in question. A solution
to this problem which has been used to study subspaces of Euclidean space is the
Voronoi decomposition. Let X be any metric space, and let L ⊆ X be a subset,
called the set of landmark points. Given λ ∈ L, we define the Voronoi cell associated
to λ, Vλ , by
Vλ = {x ∈ X | d(x, λ) ≤ d(x, λ )}
264 GUNNAR CARLSSON
for all λ ∈ L. It is immediate that the Voronoi cells form a covering of X, and
we define the Delaunay complex attached to L to be the nerve of this covering.
We remark that the generality of this definition is somewhat non-standard. In
particular, what we call Voronoi “cells” are not cells, but are just regions in a
general metric space. When the underlying space is Euclidean space, though, the
Voronoi decomposition gives rise to an extremely useful decomposition of the space,
and in favorable cases the Delaunay complex gives a triangulation of the convex hull
of L, referred to as the Delaunay triangulation [21]. For submanifolds of Euclidean
space, one may construct the restricted Delaunay triangulation as in [25]. In our
level of generality, one does not obtain any actual geometric control on the metric
space, just a simplicial complex which is often is well related to the homotopy type
of the underlying metric space. The value of this construction is that it produces
very small simplicial complexes, whose dimension is often equal to the dimension of
the manifold under consideration. Both the the Čech and Vietoris-Rips complexes
typically produce simplices in dimensions much higher than the dimension of the
space. The definition of the Delaunay complex makes sense for any metric space, in
particular for finite metric spaces. However, for finite metric spaces, it generically
produces degenerate (i.e. discrete) complexes, with no one dimensional simplices.
This is due to the fact that for finite metric spaces, it is generically the case that
each value of the distance is taken only for one pair of points, so one does not
have any points which are equidistant between a pair of landmarks. In order to
make the method useful for finite metric spaces, we therefore modify the definition
of the Delaunay complex to accommodate pairs of points which are “almost” (as
permitted by the introduction of a parameter ) equidistant from a pair of landmark
points. Precisely, we have the following definition from [8].
Definition 2.7. Let X be any metric space, and suppose we are given a finite
set L of points in X, called the landmark set, and a parameter > 0. For every
point x ∈ X, we let mx denote the distance from this point to the set L, i.e. the
minimum distance from x to any point in the landmark set. Then we define the
strong witness complex attached to this data to be the complex W s (X, L, ) whose
vertex set is L, and where a collection {l0 , . . . , lk } spans a k-simplex if and only if
there is a point x ∈ X (the witness) so that d(x, li ) ≤ mx + for all i. We can also
consider the version of this complex in which the 1-simplices are identical to those
of W (X, L, ), but where the family {l0 , . . . , lk } spans a k-simplex if and only if all
the pairs (li , lj ) are 1-simplices. We will denote this by WVs R .
There is a modified version of this construction, which is quite useful, called the
weak witness construction. Suppose we are given a metric space X, and a set of
points L ⊆ X. Let Λ = {l0 , . . . , lk } denote a finite subset of a metric space L. We
say a point x ∈ X is a weak witness for Λ if d(x, l) ≥ d(x, li ) for all i and all l ∈
/ Λ.
Given ≥ 0, we will also say that x is an -weak witness for Λ if d(x, l) + ≥ d(x, li )
for all i and all l ∈
/ Λ.
Definition 2.8. Let X, L, and be as above. We construct the weak witness

complex W w (X, L, ) for the given data by declaring that a family Λ = {l0 , . . . , lk }
spans a k-simplex if and only if Λ and all its faces admit weak witnesses. This
complex also clearly has a version in which a k-simplex is included as a simplex if
and only if all its 1-faces are, and we will denote this version by WVwR .
It is clearly the case that if we have 0 ≤ ≤ , then we have an inclusion
W s (X, L, ) → W s (X, L, )
and similarly for WVs R , W w , and WVwR .
2.3. Persistent homology. Let X be a subspace of Rn . Suppose further that we

have a method of sampling points from X, perhaps with noise. By sampling with
noise, we will mean that we are sampling points from a probability distribution
concentrated near X. Let X be one sample as described above. An interesting
question is to what extent it is possible to infer the Betti numbers of X from X.
In general, of course, the answer is no. For example, it will clearly be necessary
to assume something about the density of the sampling. Niyogi, Smale, and Wein-
berger in [52] provide sufficient hypotheses which guarantee that this is possible for
Riemannian manifolds. Their method is to prove that the Čech complex associated
to a covering by balls of a fixed radius is homotopy equivalent to the underlying
manifold. If one is interested in studying data from experiments, though, one typ-
ically cannot assume that the data lies along a submanifold. Further, even if one
could assume that the data lies along a manifold, one is usually not in a position to
verify that the stringent hypotheses of [52] are satisfied. The key to obtaining the
desired homological information is to avoid selecting a fixed value of the threshhold
, and instead to obtain a useful summary of the homological information for all
the different values of at once. This philosophy is referred to as persistence, and
was first introduced in [24].
We begin with the set X. It is of course a finite metric space, and we may
consider the Čech complexes Č(X, ), attached to the collection of balls of radius
with centers at the points of X. Note that if the centers actually lie on a submanifold
M ⊆ Rn , and the set X is sufficiently dense in M , then this complex is the Čech
complex attached to a covering of M . If further is sufficiently small, then all
the balls will be geodesically convex, and the complex will compute the homology
of M correctly. This connection provides heuristic justification for the use of this
Čech complex as a method for approximating the homology of M . Now recall that
whenever we have ≤ , we have an inclusion of complexes Č(X, ) ⊆ Č(X, ).
Consider the picture below, which is that of a Čech complex constructed on a finite
collection of points in the Euclidean plane.
1=3
266 GUNNAR CARLSSON
We note that the main shape of the set is concentrated around a circle. However,
if we compute the homology of this complex, it will yield a first Betti number of
three, namely including the large main loop and secondly the two smaller loops
corresponding to the two smaller holes in the complex. Intuitively, we regard these
two holes as coming from faulty sampling or other errors in the recovery of the data.
One could then argue that this comes from an incorrect choice of the parameter
, and that one should simply increase the parameter value to obtain a complex
with the correct higher connectivity structure. This would give rise to the following
picture.
1 =2
Note that while the two smaller holes have now been “filled in”, a new hole has been
introduced in the lower right-hand portion of the figure. Consequently, if we were
to compute the homology of this complex, we would obtain a first Betti number of
two. The result is incorrect for either of the parameter values. We now observe,
though, that there is an inclusion of the upper complex into the lower complex,
since the upper one corresponds to a smaller parameter value than the lower one.
We can therefore ask about the image of the homology of the upper complex in the
homology of the lower complex. The two small cycles in the upper complex vanish
in the lower complex, since they are filled in. On the other hand, the small cycle in
the lower complex is not in the image of the homology of the upper complex, since
the required edge is not filled in in the upper complex. We see therefore that the
image consists exactly of the larger cycle, which is what we regard as the “correct”
answer in this case. The goal of this section is to make this observation into a
systematic computational scheme, which will provide the desired summary of the
behavior of homology under all choices of values for the scale parameter .
We begin with the following definition. Again, refer to [44] for material on
categories, functors, and natural transformations.
Definition 2.9. Let C be any category, and P a partially ordered set. We regard P
as a category P in the usual way, i.e. with object set P, and with a unique morphism
from x to y whenever x ≤ y. Then by a P-persistence object in C we mean a functor
Φ : P → C More concretely, it means a family {cx }x∈P of objects of C together
with morphisms φxy : cx → cy whenever x ≤ y, such that φyz ◦ φxy = φxz whenever
x ≤ y ≤ z. Note that the P-persistence objects in C form a category in their own
right, where a morphism F from Φ to Ψ is a natural transformation. Again, in
more concrete terms, a morphism from a family {cx , φxy } to a family {dx , ψxy } is
a family of morphisms {fx }, with fx : cx → dx , and where the diagrams

φxy
cx −−−−→ cy
⏐ ⏐
⏐
fn
⏐f
y
ψxy
dx −−−−→ dy
all commute. We will denote the category of P-persistence objects in C by Ppers (C).
We note finally that if f : P → Q is a partial order preserving map, then we obtain
an evident functor f ∗ : Qpers (C) → Ppers (C) defined by f ∗ (Ψ) = Ψ ◦ f , where f is
f regarded as a functor P → Q.
We let R and N denote the partially ordered sets of real numbers and non-
negative integers, respectively. We now observe that all of the constructions of the
previous subsection (Čech, Vietoris-Rips, witness) yield an R-persistence simplicial
complex attached to X. We can now construct the associated chain complexes
and homology groups and obtain R-persistence chain complexes and R-persistence
groups. What makes homology useful as a discriminator between topological spaces
is the fact that there is a classification theorem for finitely generated abelian groups.
If one had a classification theorem for R-persistence abelian groups, then it could
act as a summary of the behavior of the homology of all the complexes Č(X, ).
However, we do not have such a theorem. However, it turns out that there is a
classification theorem (see [64]) for a subcategory of the category of N-persistence
F -vector spaces, where F is a field.
To understand this classification, we observe that N-persistence abelian groups
can be identified with a more familiar notion, namely that of a graded module over
a graded ring. Let {An } be any N-persistence abelian group. We will define an
associated graded module θ({An }) over the graded polynomial ring Z[t], where t is
assigned degree 1, as follows. We set

θ({An }) = As ,
s≥0
where the n-th graded part is the group An . The action of the polynomial generator
t is given by
t · {αn } = {βn }, where βn = ψn−1,n (αn−1 ).
It is readily checked that θ is a functor from Npers (Ab) to the category of graded
Z[t]-modules, and is in fact an equivalence of categories, since an inverse functor can
be given by M∗ → {Mn }, where the morphisms ψmn are given by multiplication by
tn−m . The conclusion is that the category Npers (Ab) is equivalent to the category
of non-negatively graded modules over Z[t]. Now, there is still no classification
theorem for graded Z[t]-modules. However, if we let F denote any field, then there
is a classification theorem for finitely generated graded F [t]-modules. See [23] for
the non-graded case. The graded case is proved in an identical fashion.
Theorem 2.10. Let M∗ denote any finitely generated non-negatively graded F [t]-
module. Then there are integers {i1 , . . . , im }, {j1 , . . . , jn }, {l1 , . . . , ln }, and an
isomorphism
m n
M∗ ∼
= F [t](is ) ⊕ (F [t]/(tlt ))(jt ),
s=1 t=1
268 GUNNAR CARLSSON
where for any graded F [t]-module N∗ , the notation N∗ (s) denotes N∗ with an upward
dimension shift of s. So, N∗ (s)l = Nl−s . The decomposition is unique up to
permutation of factors.
It is therefore a useful question to ask which N-persistence F -vector spaces cor-
respond under θ to finitely generated non-negatively generated F [t]-modules. We
have the following.
Proposition 2.11. We say an N-persistence F -vector space {V }n is tame if every
vector space Vn is finite dimensional and if ψn,n+1 : Vn → Vn+1 is an isomorphism
for sufficiently large n. Then we have that θ({Vn }n ) is a finitely generated F [t]-
module if and only if {Vn }n is tame.
We now have an easy translation of the classification result in Theorem 2.10.
For any 0 ≤ m ≤ n, we define an N-persistence F -vector space U (m, n) by setting
U (m, n)t = 0 for t < m and t > n, U (m, n) = F for m ≤ t ≤ n, and ψs,t = IdF for
m ≤ s ≤ t ≤ n. We extend this definition to the value n = +∞ in the evident way.
Proposition 2.12. Any tame N-persistence F -vector space {Vn }n can be decom-
posed as
N
{Vn }n ∼
= U (mi , ni ),
i=0
where each mi is a non-negative integer, and ni is a non-negative integer or +∞.
The decomposition is unique in the sense that the collection of pairs {(mi , ni )}i is
unique up to an ordering of factors.
We can reformulate this result as follows. By a bar code, we mean a finite set
of pairs (m, n), where m is a non-negative integer, and n is a non-negative integer
or +∞. We can now restate Proposition 2.11 as the assertion that just as finite
dimensional vector spaces are classified up to isomorphism by their dimension, so
tame N-persistence vector spaces are classified by their associated bar codes.
Returning to the R-persistence simplicial complexes we construct, we may use
any partial order preserving map N → R to obtain an N-persistence simplicial
complex. There are at least two useful ways to construct such maps. The first
would be to choose a small number ε and define a map fε : N → R by fε (n) = nε.
A second method would be as follows. Given a finite point cloud as above, it is
clear that there are only finitely many real values at which there are transitions
in the complex. This follows from the nature of the conditions together with the
fact that the distance function takes only finitely many values on X. Letting these
transition values be enumerated in increasing order as {t0 , t1 , . . . , tN }, we define
an order preserving map g : N → R by g(n) = tn for n ≤ N , and g(n) = tN
for n ≥ N . The first construction can be interpreted as sampling values of the
persistence parameter from a uniform lattice. Of course, the sampling is finer as ε
decreases. The second method is more efficient, since it is precisely adapted to the
complex at hand. Furthermore, it contains complete information about the original
R-vector space.
The methodology we use to study the homology of the complexes constructed
above is as follows.
• Construct the R-persistence simplicial complex {C } using Čech, Vietoris-
Rips, or witness methods. We will denote it by Φ.
• Select a partial order preserving map f : N → R.

• Construct the associated N-persistence simplicial complex.
• Construct the associated N-persistence chain complex {C∗ (n)}n with co-
efficients in F . (It is evident from the finiteness hypotheses on X and
the nature of the constructions that the associated N-persistence F -vector
spaces are tame.)
• Compute the barcodes associated to the N-persistence F -vector spaces
{Hi (C∗ (n), F )}n .
The last step turns out be tractable due to the translation into commutative
algebraic terms above. We recall (see [22]) that the homology computation can be
performed by putting a matrix in Smith normal form. The algorithms for Smith
normal form are typically applied for matrices over Z, but they are applicable in
the context of any principal ideal domain, of which F [t] is one. The fact that the
ring is graded and the boundary matrices are homogeneous makes the application
simpler. This is the observation made in [64], where the algorithm that we use for
computing persistent homology is described in detail.
In interpreting the output, one now finds that long intervals in the output bar-
code indicate the presence of a homology class which “persists” over a long range of
parameter values, while short intervals indicate cycles which are “born” at a given
parameter value and then “die” at a nearby parameter value. The pictures and
discussion above allow us to formulate the intuition that long intervals correspond
to large scale geometric features in the space and that short intervals correspond
to noise or inadequate sampling. Of course, what is short and what is long is very
problem dependent. Also, in some cases, it may be a false dichotomy, and the more
useful point of view is that the barcodes represent the space at various scales, and
the whole multiscale version of the space may actually be of interest.
2.4. Example: Natural image statistics. Images taken with a digital camera
can be viewed as vectors in a very high dimensional vector space. Each image
consists of a number (the gray scale value, which actually takes 255 possible values,
but can usefully be thought of as a continuous variable) attached to each of a large
number of pixels, and therefore we may think of the image as lying in RP , where P
is the number of pixels used by the camera. From this point of view, one can ask
questions about the nature of the collection of all possible images lying within RP .
For example, can it be modeled as a submanifold or subspace of RP ? If it were,
one could conclude on the one hand that it is very high dimensional, since images
are capable of expressing such a wide variety of scenes, and on the other hand that
it would be a manifold of very high codimension, since random pixel arrays will
very rarely approximate an image. David Mumford gave a great deal of thought to
questions such as this one concerning natural image statistics, and he came to the
conclusion that although the above argument indicates that the whole manifold of
images is not accessible in a useful way, a space of small image patches might in fact
contain quite useful information. In [41], A. Lee, D. Mumford, and K. Pedersen
performed an analysis constructed in this way, and we will summarize the results
of that paper. They began with a database of black and white images taken by J.
van Hateren and A. van der Schaaf in [34]. The database consisted of images taken
around Groningen, Holland, in town and in the surrounding countryside. Within
270 GUNNAR CARLSSON
such an image, one can consider 3 × 3 patches, i.e. square arrays of 9 pixels.
Each such patch can now be regarded as a 9-tuple of real numbers (the gray scale
values again), i.e. a vector in R9 . A preliminary observation is that patches which
are constant, or rather nearly constant, will predominate among these patches. The
reason is that most images have large solid regions, where the gray scale intensity
does not change significantly, and these regions will contribute more to the collec-
tion of patches than the patches in which some transitions are occurring. These
nearly constant patches will be referred to as low contrast. Of course, the low con-
trast patches do not carry interesting structure, so Lee, Mumford, and Pedersen
proceeded as follows. They first define the D-norm of a 3 × 3 image patch as a cer-
tain quadratic function of the logs of the gray scale values. This is a way of defining
the contrast of an image patch. Then they select 5,000 patches at random from each
of the images from [34] and select the top 20% as evaluated by the D-norm. This
will now constitute a database of high contrast patches from the patches occurring
in the image database from [34]. They then perform two transformations on the
data, as follows.
(1) Mean center the data. The mean intensity value over all nine pixels is
subtracted from each pixel value to obtain a patch with mean zero. This
means that if a patch is obtained from another patch by adding a constant
value, i.e. “turning up the brightness knob”, then the two patches will be
regarded as the same. Note that this transformation puts all points in an
8-dimensional subspace within R9 .
(2) Normalize the D-norm. Since all the patches chosen will have D-norm
bounded away from zero, one can divide by it to obtain a patch with D-norm
= 1. This means that if one patch is obtained from another by “turning
the contrast knob”, then the two patches will be regarded as identical.
The result of this construction is a database M of ca. 4.5 × 106 points on a
7-dimensional ellipsoid in R8 . The goal of the paper [41] is now to obtain some
understanding of how this set sits within S 7 , and what can be said about the patches
which do appear. A first observation is that the points are scattered throughout
the 7-sphere, in the sense that no point on S 7 is very far from the set, but that the
density appears to vary a great deal. In particular, in [41], indications were found
that the data was concentrated around an annulus. In [11], a systematic study of
the topology of the high density portion was carried out, and this work is what we
will describe in the remainder of this section.
The first issue to be addressed is what is meant by the “high density” portion.
Density estimation is a highly developed area within statistics (see, for example,
[58]). We selected a very crude proxy for density, in the interest of minimizing the
computational burden. It is defined as follows. Fix a positive integer k, and define

the k-codensity function δk of x ∈ X, where X is a set of point cloud data, by
δk (x) = d(x, νk (x)),
where d denotes the distance function in X, and where νk (x) denotes the k-th
nearest neighbor of x in X. Note that δk (−) is inversely related with density, since
a concentrated region will have smaller distances to the k-th nearest neighbor, so
we will be studying subcollections of points for which δk (−) is bounded from above
by a threshhold. Secondly, we also note that each δk yields a different density
estimator. In rough terms, δk for large values of k computes density using points
in large neighborhoods of x, and for small values uses small neighborhoods. So,
δk for large k corresponds to a smoothed out notion of density, and for small k
corresponds to a version which carries more of the detailed structure of the data
set.
For any subset M0 ⊆ M, we now define subsets M0 [k, T ] ⊆ M0 , where k is a
positive integer, and T is a percentage value, by
M0 [k, T ] = {x ∈ M0 | δk (x) lies among the T % lowest values of δk in M0 }.
The goal of the paper [11] is to infer the topology of a putative space underlying
the various sets of points M[k, T ]. In some cases we will approximate this via a
subset M0 [l, T ]. It is fairly direct to see that the variable k scales with the size of
the set, so that if ρ = #(M)/#(M0 ), then M0 [k, T ] is comparable with M[ρk, T ].
We do this by obtaining the data sets M0 [k, T ], then selecting a set of landmark
points, and finding various barcodes attached to witness complexes associated to
the space. Below is the barcode for H1 (W M0 [300, 30])), where M0 is a sample
of 5 × 104 points from M, and W denotes a witness complex constructed with 50
landmark points.
k = 300, T = 30%
We note that there are a number of short lines, and one long one. According to the
philosophy mentioned above, this suggests that the first Betti number should be
estimated to be one. The barcode is stable, in the sense that it appears repeatedly
when the set of landmark points is varied, and when the sample from the full data
set is varied. Therefore, the simplest possible explanation for this barcode is that
the underlying space should be a circle. One can then ask if there is a simple
explanation involving the data which would yield a circle as the underlying space.
272 GUNNAR CARLSSON
The picture below gives such an explanation.
Primary circle
More formally, the picture suggests the explanation that the most frequently oc-
curring patches are those approximating two variable functions which depend only
on a linear projection from the two variable space, and so that the function is in-
creasing in that linear projection. This explanation is consistent with an annulus
conjectured to represent the densest patches in [41].
The next picture is a barcode for H1 (W M0 [15, 30])), where M0 is as above,
again with 50 landmark points chosen.
k = 15, T = 30%
We note that in this case, there are many short segments and 5 longer segments.
It is again the case that this barcode is stable over the choices of landmark points
and samples from M. This suggests that the first Betti number of the putative
underlying space should be five. It now becomes more difficult to identify the
simplest explanation for this result, and a number of such models are possible. The
one which ultimately turns out to fit the data is pictured as follows.
Three circle model
This picture is composed of a primary circle, which surrounds two secondary

circles. Although it is perhaps not clear from the image, the intent is that the
secondary circles each intersect the primary circle in two points, and do not intersect
each other. A space constructed this way can readily be seen to have a first Betti
number of five. An explanation for this geometric object in terms of image patches
is now the following.
PRIMARY
SECONDARY SECONDARY
Three circle model in the data
The secondary circles interpolate between functions which are an increasing func-
tion of a linear projection to functions which are “bump functions” with an internal
local maximum evaluated on the same linear projection. Note that the two sec-
ondary circles each intersect the primary circle in two points, and that they do not
intersect each other. We have informally confirmed that the indicated patches are
the ones which occur in the high density portions of the data set.
Remark. We note that there is a preference within the data for patches which are
aligned in vertical and horizontal directions, since one can construct versions of the
secondary circles which are not aligned in the vertical direction, and they do not
274 GUNNAR CARLSSON
appear. One explanation for this is that patches in natural images are biased in
favor of the horizontal and vertical directions because nature has this bias, since for
example objects aligned in a vertical direction are more stable than those aligned
at a 45 degree angle. Another explanation is that this phenomenon is related to the
technology of the camera, since the rectangular pixel arrays in the camera have the
potential to bias the patches in favor of the vertical and horizontal directions. We
believe that both factors are involved. In [13], we have studied 5 × 5 patches, and
we found the three circles model appearing there as well. In that case, one would
expect to see less bias in favor of the vertical and horizontal directions, since the
pixels give a finer sampling of the image.
One could now ask if there is a larger 2-dimensional space containing the three
circle model, occurring with substantial density. We will first ask to find a natural
embedding of the theoretical three circle model in a 2-manifold. It turns out that
the model embeds naturally in a Klein bottle. (Image courtesy of Tom Banchoff.)
To see this, we first recall that the Klein bottle can be described as an identification
space as pictured below.
R S
P Q
Q P
R S
The colored arrows indicate points being identified using the quotient topology
construction (see [33]), which informally means that the top horizontal edge is
identified with the lower horizontal edge in a way which preserves the x-coordinate,
and the right-hand vertical edge is identified with the left-hand vertical edge after a
twist which changes the orientation of the edge. To understand this a little better,
the identifications encoded by the vertical arrows alone would create the side of a
cylinder, while the identifications encoded by the diagonal arrows create a Möbius
band.
It is convenient to represent the Klein bottle in this way, since it does not embed in
Euclidean 3-space and therefore cannot be precisely visualized, although a useful
visual representation including self-intersections was shown above.
We are interested in finding a sensible embedding of the three circle model in
the Klein bottle. Some experimentation with the three circle model results in the
following picture
in which the horizontal segments form the primary circle and the vertical segments
form the secondary circles, as in the diagram. The horizontal segments form a
single circle since the intersection point of the lower horizontal segment with the
right (respectively left) vertical edge is identified with the intersection point of the
upper horizontal segment with the left (respectively right) vertical edge, and the
vertical segments form circles since the intersection points with the upper horizontal
edge are identified with the corresponding points on the lower horizontal edge. We
note that the vertical circles intersect the horizontal primary circle in two points,
and do not intersect each other, so the picture in question produces a natural
embedding of the three circle model in the Klein bottle. The existence of this
straightforward embedding of the three circle model in the Klein bottle suggests
the possibility that a more relaxed density threshholding would yield a space of
high contrast, high density patches which is homeomorphic to a Klein bottle.
In [11], it was demonstrated that the Klein bottle effectively models a space of
high contrast patches of high density. To understand the results of that paper, it
is necessary to discuss another theoretical version of the Klein bottle. We will be
regarding the 3 × 3 patches as obtained by sampling a smooth real-valued function
on the unit disc at nine grid points, and we study subspaces of the space of all such
functions which have a rough correspondence with the subspaces of the space of
patches we study. We will consider the space Q of all two variable polynomials of
degree 2, i.e. functions
f (x, y) = A + Bx + Cy + Dx2 + Exy + F y 2 .
276 GUNNAR CARLSSON
The set Q is a 6-dimensional real vector space. We now consider the subspace
P ⊆ Q consisting of functions f so that

(2–1) f =0 and f 2 = 1.
D D
The first condition is the analogue of the mean centering condition imposed on the
patches, and the second is the analogue of the normalization condition for the con-
trast. Imposing only the first condition, which is linear, produces a 5-dimensional
vector subspace, and the second, which is quadratic in character, produces a 4-
dimensional ellipsoid within this vector space. We now consider the subspace
P0 ⊆ P, consisting of all functions within P which are of the form
f (x, y) = q(λx + µy),
where q is a single variable quadratic function, and where λ2 + µ2 = 1. The space of
all functions of this form within Q is 4-dimensional (three variables parametrize q,
and (λ, µ) lies on the (one-dimensional) unit circle). The two additional constraints
in (2–1) above imposed on it now yield the 2-dimensional complex P0 . We claim
that P0 is homeomorphic to the Klein bottle. To see this, we let A denote the space
of single variable polynomials q(t) = c0 + c1 t + c2 t2 satisfying the two conditions
1 1
q(t) = 0 and q(t)2 = 1.
−1 −1
It is easy to check that, regarded as a subspace of R3 , this subspace is an ellipse

and therefore is homeomorphic to a circle. For any unit vector v in R2 and any
q ∈ A, we let q v : R2 → R be defined by q v (w)
= q(v · w).
It is easy to check that
for q ∈ A and v a unit vector, we have that

q v = 0 and q v2 = 0
D D
and that therefore the formula
q v
(q, v ) →
||q v ||2
defines a continuous map θ from A × S 1 to P0 . The map θ is however not a
homeomorphism, which one can check as follows. Let ρ : A → A be the involution
defined by ρ(c0 + c1 t + c2 t2 ) = c0 − c1 t + c2 t2 . Then we have the relation
θ(q, v ) = θ(ρ(q), −v)
so that the map θ factors through the space of orbits under the involution. It is
easy to check (a) that the factorization is a homeomorphism and (b) that the orbit
space is homeomorphic to a Klein bottle.
We now ask to what extent we can “see” a Klein bottle in the data. A naive
approach to this question would be to simply perform the experiments we did
above for k = 15 but with a less stringent density threshhold, for example 40 or
50. Performing these experiments does not produce a non-trivial β2 . One might
suppose, though, that the set M0 is not large enough to provide sufficient resolution
in the density estimation, and that if one uses the full set M that one might obtain
different answers. With this larger set, one might also vary the estimator parameter
k to obtain a finer estimation of the density. After some experimentation, one can
construct a data set S by sampling 104 points from the data set M[100, 10]. The
set S exhibits the three circle model clearly, but enlarging it still does not exhibit
the non-trivial β2 which would be characteristic of the Klein bottle. In order to

begin to understand the situation, one should then ask what the least frequently
appearing patches on the Klein bottle would be. In thinking about the polynomial
model, one expects that there should be a preference for the linear polynomials,
and the experience with the three circle model suggests that there is an additional
preference for the patches which are lined up with the vertical directions. This
suggests that the least frequently appearing patches
√
would be the pure√quadratics
√ √
composed with the linear functions (x, y) → 22 x + 22 y and (x, y) → 22 x − 22 y.
Slightly more frequently appearing would be pure quadratics composed with any
linear function which is not a projection on the x and y-axes. This set of pure
quadratics forms a pair of open one-dimensional arcs within the Klein bottle. These
arcs are indicated by the central horizontal segments in the image below, where as
before the horizontal arcs form the primary circle, and the vertical arcs form the
secondary circles.
The idea will be that we wish to include some points corresponding to the central
arcs to the data set, but which turn out not to satisfy the density threshhold. We
will find that once these points are added, the set then carries the topology of a
Klein bottle. This would then give a reasonable qualitative description for a portion
of the density distribution of the high contrast 3 × 3 image patches, in that it can
be said to concentrate around a Klein bottle, but with strongly reduced density
around the non-vertical and non-horizontal pure quadratic patches. In order to
find such a map, we identify the pixels in the 3 × 3 array with the points in the set
L = {−1, 0, 1} × {−1, 0, 1} in the Euclidean plane. For any given function f ∈ P0 ,
we define the associated patch by evaluating f at the nine points of L, and then
mean centering and D-norm normalizing. This produces a map h from the Klein
bottle P0 to the normalized patch space. We then obtain additional points to add
to the data set by selecting 30 points at random from the central horizontal arcs in
the Klein bottle, computing h on them, and then for each of the 30 points selecting
the points of M[100, 10] which lie closest to them, and finally adjoining them to
the set S to obtain an enlarged set S . Witness complexes with 50 landmark points
computed for S now display the barcodes which would be associated to a Klein
bottle. Here is a typical picture. We remind the reader that the barcodes are for
mod 2 homology.
We note that the β0 barcode (the upper picture) clearly shows the single component,
the β1 shows two lines from threshhold parameter values .15 to .35, and finally the
β2 barcode shows a single line on roughly the same interval. This gives β0 = 1,
βi = 2, and β2 = 1. These are the mod 2 Betti numbers for the Klein bottle.
278 GUNNAR CARLSSON
Betti0
1
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
20
Betti1
10
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
10
Betti2
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45
Remark. There are actually two two-dimensional manifolds with these mod 2 Betti
numbers: one is the Klein bottle, and the other is a torus. These two are distin-
guished by mod 3 homology, and we have performed the computation to show that
the mod 3 homology is consistent with the Klein bottle and not with the torus.
Remark. One can also ask if the space we are studying is in fact closely related
to the theoretical Klein bottle defined above using quadratic polynomials. That
this is so is strongly suggested in [11] by a comparison with data sets constructed
by adjoining additional points obtained from the whole theoretical Klein bottle to
the set S. The resulting space also shows a strong indication of the same Betti
numbers, indicating that the spaces represent essentially the same phenomenon.
2.5. Example: Electrode array data from primary visual cortex. The goal
of neuroscience is, of course, to obtain as complete as possible an understanding of
how the nervous system operates in performing all its tasks, including vision, motor
control, higher level cognition, olfactory sensation, etc. One aspect of this kind of
understanding is the analysis of the structure and function of individual neurons,
and the creation of an associated taxonomy of individual neurons. Another aspect
is the analysis of how families of neurons cooperate to accomplish various tasks,
which could be referred to as the study of populations of neurons. The second
problem appears to be very amenable to geometric analysis, since it will involve
the activities of several neurons at once. In the paper [59], a first attempt at
topological analysis of data sets constructed out of the simultaneous activity of
several neurons is carried out, with encouraging results, and we will describe the
results of that paper in this section.
The arrays of neurons studied in [59] are from the primary visual cortex or V1
in Macaque monkeys. The primary cortex is a component in the visual pathway,
which begins with the retinal cells in the eye, proceeds through the lateral geniculate
locus, then to the primary visual cortex, and then through a number of higher level
processing units, such as V2, V3, V4, middle temporal area (MT), and others. See
[36] and [63] for useful discussions. It is known that V1 performs low level tasks,
such as edge and line detection, and its output is then processed for higher level and
larger scale properties further along the visual pathway. However, the mechanism
by which it carries out these tasks is not understood. A very interesting series
of experiments was conducted in the papers [62] and [38]. These authors study
the behavior of the V1 in Macaque monkeys by injecting a voltage sensitive dye
in it, and then performing optical imaging of small regions of the cortex. Voltage
changes in this portion of the cortex will then give rise to color differences in the
imaging. Since the voltages change over time, so will the optical images. These
papers study the behavior of the optical images under two separate conditions, one
the evoked state, in which stimuli are being supplied to the eye of the monkey, and
the unevoked or spontaneous state, in which no stimulus is being supplied. It was
observed in [38] that in an informal sense, the images in the different conditions
appeared to be quite similar, and a statistical analysis strongly suggested that the
behavior of V1 in the spontaneous condition was consistent with a behavior which
consisted in moving through a family of evoked images corresponding to responses
to angular boundaries, without any particular order.
Another method for studying the behavior of neurons in V1 and other parts of the
nervous system is the method of embedded electrode arrays. In this case, arrays of
up to ca. 100 regularly spaced electrodes are implanted in the V1 (or whatever other
portion of the nervous system one is studying) . The voltages at the electrodes are
then recorded simultaneously, so one obtains a voltage level at each of the electrodes
at each point in time. Sophisticated signal processing techniques are then used to
obtain an array of N (where N is the number of electrodes) spike trains, i.e. lists
of firing times for N neurons. This experimental setup provides another view into
the behavior of the neurons in V1, and the idea of the paper [59] was to attempt to
replicate the results of [38], which were carried out using the voltage sensitive dye
technology, in the embedded electrode setting and to attempt to refine the results
presented there. We now describe the experiments that were carried out and the
results of our analysis of the data obtained from them.
10 × 10 electrode arrays were used in recording output from the V1 in Macaque
monkeys who view a screen. Two segments of recording, each of roughly 20–30
minutes, were done under two separate conditions. In the first, the eyes of the
animal were occluded, so no stimulus was presented to the visual system. In the
second, a video sequence was obtained by sampling from different movie clips. We
refer to the data obtained in the first setting as spontaneous, and in the second
setting as evoked. A signal processing methodology called spike sorting was then
applied to the data, so that one could identify neurons and firing times for each
neuron. Next, the data was broken up into ten second segments in both cases.
Each such segment was next divided up into 200 50ms bins, and for each neuron
one is able to count the number of firings within each such 50ms bin. The five
neurons with the highest firing rate were selected in each ten second window, and
for each bin one can now obtain the 5-vector of number of firings of each of these
five neurons. By performing this construction over all 200 bins, one obtains a set
of point cloud data consisting of 200 points in R5 . Of course, we have many such
data sets, coming from different choices of ten second segments and from different
choices of “regime” (spontaneous or evoked).
Beginning with these point clouds, witness complexes based on 35 landmark
points were constructed. The landmark sets were constructed by the “maxmin”
280 GUNNAR CARLSSON
procedure, a procedure designed to ensure that the landmark points are well dis-
tributed throughout the point cloud. This procedure begins with a seed point, and
then constructs the rest of the points deterministically from it. For each data set,
we constructed witness complexes from all the possible seed points. In order to
derive topological signatures from each such witness complex, one selects a thresh-
hold as a fraction of the covering radius of the point cloud, and then one determines
the Betti numbers β0 , β1 , and β2 of the witness complex with this given threshhold
value of . Thus, for each witness complex, one can now obtain a vector or signature
of integers (β0 , β1 , β2 ). The observed signatures are listed below, with pictures of
simple models of possible geometries which they represent.
By far the most frequently occurring signatures were (1, 1, 0) and (1, 0, 1), cor-
responding to a circle and a sphere, respectively. The picture below shows the
distribution of occurrences of these two under various choices of the threshholds.
In order to validate the significance of the results, we ran the identical procedures
with data generated at random for firings with a Poisson model. Monte Carlo
simulation shows that the probability of obtaining segments for β1 or β2 longer

than 30% of the diameter of the point cloud is < .005.
The summary of the results of the experiment are that topological methods
clearly distinguish the data in both regimes from a Poisson null hypothesis model.
They also suggest that there is a similarity in the spontaneous and evoked regimes,
since the same topological signatures occur. Further, though, the statistics of the
signatures occurring are also able to distinguish between the two regimes. We do not
yet understand the nature of the topological phenomenon, which is something which
should be addressed by mapping algorithms, perhaps along the lines of the following
section. One aspect we have addressed in this direction, though, is the question of
whether a simple periodic phenomenon associated with the body’s natural rhythms
is responsible for the topology. Such a phenomenon would likely create peaks in
the amplitude spectrum of the segments of the data we study. No statistically
significant peaks of this type were observed.
3. Imaging: mapper
3.1. Visualization. So far we have discussed the attachment of homological sig-
natures to point clouds in an attempt to obtain a geometric understanding of them.
Frequently, though, it is possible to find images of various kinds attached to point
cloud data which allow one to obtain a qualitative understanding of them through
direct visualizaton. One such method is the projection pursuit method (see [37]),
which uses a statistical measure of information contained in a linear projection
to select a particularly good linear projection for data which is embedded in Eu-
clidean space. Another method is multidimensional scaling (see [1]), which begins
from an arbitrary point cloud and attempts to embed it isometrically in Euclidean
spaces of various dimensions with minimum distortion of the metric. Related de-
velopments are the Isomap (see [61]) and locally linear embedding (see [56]). In
all cases, the methodologies result in a point cloud in R2 or R3 , which can then
be visualized by the investigator. There are, however, other possible avenues for
visualization and qualitative representation of geometric objects. One such possi-
bility is representation as a graph or as a higher dimensional simplicial complex.
Such combinatorial representations can lead to useful qualitative understanding in
their own right, but graph visualization software such as Graphviz (available at
http://www.graphviz.org/) can provide useful visualizations. In thinking about
how to develop such a representation, it is useful to keep in mind what character-
istics would be desirable. Here is a list of some such properties.
• Insensitivity to metric: As mentioned in the introduction, metrics used
in analyzing many modern data sets are not derived from a particularly
refined theory, but instead are constructed as a reasonable quantitative
proxy for an intuitive notion of similarity. Therefore, imaging methods
should be relatively insensitive to detailed quantitative changes.
• Understanding sensitivity to parameter changes: Many algorithms
require parameters to be set before an outcome is obtained. Since set-
ting such parameters often involves arbitrary choices, it is desirable to use
methods which provide useful summaries of the behavior under all choices
of parameters if possible.
• Multiscale representations: It is desirable to understand sets of point
clouds at various levels of resolution, and to be able to provide outputs
282 GUNNAR CARLSSON
at different levels for comparison. Features which are seen at multiple

scales will be viewed as more likely to be actual features as opposed to
more transient features which could be viewed as artifacts of the imaging
method.
The rest of this section will be devoted to the description of a method which
addresses each of these points. We have named it Mapper, and it is described in
detail in [60].
3.2. A topological method. We begin with a topological construction based on
a covering of a topological space X.
Definition 3.1. Let X be a topological space, and let U = {Uα }α∈A be a finite cov-
ering of the space X (so the set A is finite). Let ∆[A] denote the standard simplex
with vertex set A, so dim(∆[A]) = #(A) − 1. Further, for any non-empty subset
S ⊆ A, we let ∆[S] ⊆ ∆[A] denote the face spanned by the vertices corresponding
to elements of S, and we let X[S] = s∈S Us ⊆ X. By the Mayer-Vietoris blowup
of X associated to U, denoted by M(X, U), we mean the subspace

∆[S] × X[S] ⊆ ∆[A] × X.
∅=S⊆A
We note that there are natural projection maps f : M(X, U) → X and g :

M(X, U) → ∆[A], which have the following properties.
• The map f is a homotopy equivalence when X has the homotopy type of
a finite complex and the covering consists of open sets. In fact, using a
partition of unity subordinate to the covering U, one can obtain an explicit
homotopy inverse ϕ : X → M(X, U).
• The map g is a homotopy equivalence onto its image (which is the nerve of
the covering, or the Čech complex Č(U)) when all the sets X[S] are either
empty or contractible. This is how the nerve theorem (Theorem 2.3) is
proved.
These two observations demonstrate that one obtains a map from X to Č(U)
for any finite complex X. Such a map can be viewed as a kind of coordinatization
of the space X. Ordinary coordinatizations provide maps to Euclidean spaces
of various dimensions, and they often provide useful insights into the spaces in
question. Simplicial complexes, particularly low dimensional ones, can also often be
readily visualized, and can therefore also be expected to provide useful information
about a space. This is so even if the map is not a homeomorphism, so it does not
provide complete information about a space. We will next observe that there is
a variant of the Č(U) construction which is a somewhat more sensitive target for
this kind of coordinatization map. Let X be any topological space, and let U be
any covering of X. We will now define a simplicial complex Č π0 (U) to be the nerve
of the covering of X by sets which are path connected components of a set of the
form Uα , so the covering is indexed by the set of pairs {(α, ξ)}, where ξ is a path
component of Uα . The set map (α, ξ) → α yields a map of simplicial complexes
Č π0 (U) → Č(U). It is further clear that we have a map M(X, U) → Č π0 (U), which
is induced by the projection Uα → π0 (Uα ) for each α. Finally, the composite
M(X, U) → Č π0 (U) → Č(U)
is the earlier defined map g.
Example. Let X denote the unit circle, and let a covering U of X be given by the
three sets A = {(x, y) | y < 0}, B = {(x, y) | y > 0}, and C = {(x, y) | y = ±1}.
We note that π0 (A) and π0 (B) consist of a single point, and π0 (C) consists of two
points.
The simplicial complexes Č π0 (U) and Č(U) are now given by the following picture.
Note that Č π0 (U) is actually homeomorphic to X, while Č(U) is not. This is an

example of the fact that Č π0 is more sensitive than Č.
In order for this construction to be useful, one must develop methods for con-
structing coverings of topological spaces. Earlier we have looked at constructions
where the sets are open balls in a metric space, and where we have versions of the
Voronoi decomposition adapted to general metric spaces. We now suppose that we
are given a reference map ρ from our space X to a metric space Z, and further that
we are given a finite open or closed covering U of Z. Then we may consider the
covering ρ∗ U given by ρ∗ U = {ρ−1 Uα }α∈A , when U = {Uα }α∈A . This is clearly a
covering of X. Typical examples of useful metric spaces Z include R, Rn , and S 1 .
These spaces admit natural coverings.
Example. For X = R, let R and e be positive real numbers. Then we may
construct the covering U[R, e] which consists of all intervals of the form
[kR − e, (k + 1)R + e]. This is a two parameter family of coverings, and as long
as e < R2 , it has covering dimension 1, in the sense that no non-trivial threefold
overlaps are non-empty. Products of these intervals would give a corresponding
covering of Rn .
Example. Let X = S 1 , N be an integer ≥ 2, and > 0 be a real number. Then
we can form a covering U[N, ] = {Uj }0≤j<N of X by setting
2πj 2πj
Uj = {(cos(x), sin(x)) | x ∈ [ − , + ]}
N N
π
whenever > N .
284 GUNNAR CARLSSON
Remark. When the reference space is R, our construction is closely related to the
Reeb graph of a real-valued function on a manifold (see [55]). The actual Reeb
graph should be viewed as a limiting version of the construction as one studies the
coverings U[R, ] with R and tending to zero.
We must now describe a method for transporting this construction from the
setting of topological spaces to the setting of point clouds. The notion of a covering
makes sense in the point cloud setting, as does the definition of coverings of point
clouds using maps from the point cloud to a reference metric space, by “pulling
back” a predefined covering of the reference space. The notion which does not
make immediate sense is the notion of π0 , i.e. constructing connected components
of a point cloud. The notion of clustering (see [31]) turns out to be the appropriate
analogue. Our main example of such a clustering algorithm will be the so-called
single linkage clustering. It is defined by fixing the value of a parameter , and
defining blocks of a partition of our point cloud as the set of equivalence classes
under the equivalence relation generated by the relation ∼ defined by x ∼ x if
and only if d(x, x ) ≤ . Note that the set of clusters in this setting is precisely π0
applied to the Vietoris-Rips complex V R(X, ), and that each “cluster” corresponds
to the set of vertices in a single connected component. Now, our version of the
construction Č π0 in this context is obtained as follows.
(1) Define a reference map f : X → Z, where X is the given point cloud and
Z is the reference metric space.
(2) Select a covering U of Z. If Z = R, then U can be obtained by selecting R
and e as above, and constructing the covering U[R, e].
(3) If U = {Uα }α∈A , then construct the subsets Xα = f −1 Uα .
(4) Select a value as input to the single linkage clustering algorithm above,
and construct the set of clusters obtained by applying the single linkage
algorithm with parameter value to the sets Xα . At this point, we have
a covering of X parametrized by pairs (α, c), where α ∈ A and c is one of
the clusters of Xα .
(5) Construct the simplicial complex whose vertex set is the set of all possible
such pairs (α, c), and where a family {(α0 , c0 ), (α1 , c1 ), . . . , (αk , ck )} spans a
k-simplex if and only if the corresponding clusters have a point in common.
This construction is a plausible analogue of the continuous construction described
above. We note that it depends on the reference map, a covering of the reference
space, and a value for . We observe that in fact any clustering algorithm could be
used to cluster the sets Xα , and one could still obtain a sensible construction. We
note that if the covering U has covering dimension ≤ d, i.e. if whenever we are given
a family {α0 , α1 , . . . , αt } of distinct elements of α with t > d, then Uα0 ∩ . . . ∩ Uαt =
∅, then the dimension of the simplicial complex we construct will be ≤ d as well.
This follows immediately from the definitions.
We note that this construction readily produces a multiresolution or multiscale
structure which allows one to distinguish actual features from artifacts. To see
this, we begin with the definition of a map of coverings. Let U = {Uα }α∈A and
V = {Vβ }β∈B be coverings of a space Z. By a map of coverings from U to V we
will mean a set map θ : A → B so that for all α ∈ A, we have Uα ⊆ Vθ(α) .
Example. Consider the coverings U[R, e] of R defined above. The indexing set in
this case consists of the integers. It is clear from the definition that the identity
map from Z to itself yields a map of coverings U[R, e] → U[R, e ] whenever e ≤ e .

In this case, the map of coverings consists simply of the inclusion of an interval into
an interval with the same center but with larger diameter.
Example. The map of integers k → k2 defines a map of coverings U[R, e] →
U[2R, e], which is two to one in the sense that every interval in U[2R, e] contains
two intervals from U. In order to use these maps to obtain a multiresolution version
of the Mapper construction, we need a definition.
Definition 3.2. A clustering algorithm is said to be functorial if whenever one has
an inclusion X → Y of point clouds, i.e. a set map preserving distances, then the
image of each cluster constructed in X under f is included in one of the clusters
in Y . It follows from the fact that the clustering algorithm produces a partition of
the point cloud in question that each cluster is contained in a unique cluster, and
therefore that we have an induced map of sets from the clusters in X to the clusters
in Y .
Now suppose we are given data for applying Mapper, i.e. a point cloud X together
with a reference map ρ to a metric space Z. Suppose further that we are given
two coverings U = {Uα }α∈A and V = {Vβ }β∈B of Z, and a map of coverings
θ : A → B. Since we have a map of coverings, it is clear that we obtain inclusions
ρ−1 Uα ⊆ ρ−1 Vθ(α) for all α ∈ A. If we apply a functorial clustering scheme to each
of the sets ρ−1 Uα and ρ−1 Vβ , it is clear from the definition that we will obtain
a map from the set of clusters obtained by applying the clustering algorithm to
ρ−1 Uα to the set of clusters obtained by applying it to ρ−1 Vβ , and therefore a
map from the vertex set of M(X, U) to the vertex set of M(X, V). One readily
checks that it is actually a simplicial map, so we obtain an associated simplicial
map Θ : M(X, U) → M(X, V). So now, for example, we will always obtain a
diagram of simplicial complexes
· · · → M(X, U[R/4, e]) → M(X, U[R/2, e]) → M(X, U[R, e]).
As one moves to the left, the coverings of R (and therefore of X) become more
refined, and are presumed to give a picture with finer resolution of the space in
question. Studying the behavior of the features under such maps will allow one
to get a sense of which observed features are real geometric features of the point
cloud, and which are artifacts, since the intuition is that features which appear at
several levels in such a multiresolution diagram would be more intrinsic to the data
set than those which appear at a single level.
3.3. Filters. An important question, of course, is how to generate useful reference
maps ρ. If our reference space Z is actually Rn , then this means simply generating
real valued functions on the point cloud. To emphasize the way in which these
functions are being used, we refer to them as filters. Frequently one has interesting
filters, defined by a user, which one wants to study. However, in other cases one
simply wants to obtain a geometric picture of the point cloud, and it is important
to generate filters directly from the metric which reflect interesting properties of
the point cloud. Here are some important examples.
• Density: Consider any density estimator applied to a point cloud X. It
will produce a non-negative function on X, which reflects useful information
about the data set. Often, it is exactly the nature of this function that is
of interest.
286 GUNNAR CARLSSON
• Data depth: The notion of data depth refers to any attempt to quantify
the notion of nearness to the center of a data set. It does not necessarily
require the existence of an actual center in any particular sense, although a
point which minimizes the quantity in question could perhaps be thought of
as a choice for a center. In our group’s work, we have referred to quantities
of the form
1
ep (x) = d(x, x )p
#(X)
x ∈X
(with an obvious generalization to p = ∞) as eccentricity functions, and we
have used them as filters. Other notions could equally well be used. The
main point is that the Mapper output based on such functions can identify
the qualitative structure of a particular kind. For example, if the space
were as pictured below,
then Mapper would recover the structure of the three flares coming out
from the central point.
• Eigenvectors of graph Laplacians: Graph Laplacians are interesting
linear operators attached to graphs (see [40]). They are the discrete ana-
logues of the Laplace-Beltrami operators that appear in the Hodge theory
of Riemannian manifolds, whose null spaces provide particularly nice repre-
sentative forms for de Rham cohomology. In particular, their eigenfunctions
produce functions on the vertex set of the graph. They can be used, for ex-
ample, to produce cluster decompositions of data sets when the graph is the
1-skeleton of a Vietoris-Rips complex. We find that these eigenfunctions
(again applied to the 1-skeleton of the Vietoris-Rips complex of a point
cloud) also can produce useful filters in the Mapper analysis of data sets.
3.4. Scale space. The construction from the previous subsection depends on cer-
tain inputs, including a parameter . The decision of how to choose this parameter
is in principle a difficult one, for which one has little guidance. Further, it may
often be desirable to broaden the definition of the complex to permit choices of
which vary with α, i.e. over the reference space Z. In this section, we discuss
a systematic way of considering such varying choices of “scale”. We first note
that because the single linkage procedure applied to a point cloud X can be in-
terpreted as computing connected components of V R(X, ), the persistence bar-
code for β0 yields interesting information about the behavior of the components
(or clusters) for all values of . To be explicit about this, we consider the subset
E(X) ⊂ R+ = [0, +∞) consisting of all the endpoints of the intervals occurring
in the barcode. E(X) is a finite set on the non-negative real line, and there is
consequently a total ordering on it induced from the total ordering on R+ , and we
write E(X) = {e1 , . . . , et }, with ei < ej whenever i < j. From the definition, it is
clear that whenever we have ei < η < η < ei+1 , the natural map on H0 induced
by the inclusion V R(X, η) → V R(X, η ) is an isomorphism, and that therefore the
inclusion induces a bijection on connected components. For this reason, we call
each of the intervals (ei , ei+1 ) a stability interval, or an S-interval for X. We now
have the following definition.
Definition 3.3. Given a point cloud X, a reference map ρ : X → Z to a met-

ric space, and a covering U = {Uα }α∈A , we define a simplicial complex SS =
SS(X, ρ, U) as follows. The vertices of SS are pairs (α, I), where α ∈ A, and
where I is a stability interval for the point cloud Xα = ρ−1 (Uα ). A (k + 1)-tuple
{(α0 , I0 ), (α1 , I1 ), . . . , (αk , Ik )} spans a k-simplex in SS if (a) Uα0 ∩ . . . ∩ Uαk = ∅
and (b) I0 ∩ . . . ∩ Ik = ∅. The vertex map (α, I) → α induces a map of simplicial
complexes p : SS → Č(U). By a scale choice for X and U, we will mean a section
of the map p, i.e. a simplicial map s : Č(U) → SS such that p ◦ s = IdČ(U) .
Given any scale choice s for X and U, we set s(α) = (α, Iα ). Now, for any scale
choice s and α ∈ A, we choose α ∈ Iα . This gives a choice of the scale parameter
varying with α, and we can build a new complex whose vertex set consists of pairs
(α, c), where c is a cluster in the single linkage clustering applied to Xα = ρ−1 Uα
with perimeter value α . From the definition of the stability intervals, it is clear
that the complex is independent of the choice of α ∈ Iα .
Remark. The intuition behind the definition of scale choice is the following. We
wish to permit a choice of scale parameter which varies with α. Of course, the
set of all such choices is too large to contemplate using any kind of exhaustive
enumeration of the possible values and will in any case not be useful since we will
not have any criteria to determine which choices are more plausible than others.
The definition given above incorporates two different heuristics which permit us
to restrict the choices of α which we make, as well as to evaluate various choices
relative to each other.
• From the fact that the scale choice s is a simplicial map, it follows that
whenever Uα and Uα have a non-empty intersection, we also have Iα ∩ Iα .
This means that the choices of the parameters α have a certain kind of
continuity in the variable α, which is surely a desirable feature of a varying
choice of scales.
• The fact that the stability intervals have a notion of length allows us to
evaluate the scale choices. The general rule of thumb is that choices of scale
which are stable over a large range of parameter values are to be preferred
over those with stability over a shorter range. This permits various notions
of numerical weights (such as, for example, α l(Iα ), where l(−) denotes
length) which allow one to compare scale choices.
3.5. Examples. We show the outputs from Mapper applied to various data sets.
The first example comes from a six dimensional data set constructed by G.M.
Reaven and R.G. Miller from a diabetes study conducted at Stanford University in
the 1970’s. 145 patients were included. Details of the study and of the construction
of the data set can be found in [48]. Below is the output of Mapper applied to this
data set, with two different levels of resolution.
288 GUNNAR CARLSSON
The filter is in this case a density estimator, and high values occur at the dark
nodes at the top, while low density values occur on the lower flares. At both scales,
there is a central dense core, and two “flares” consisting of points with low density.
The core consists of normal or near-normal patients, and the two flares consist of
patients with the two different forms of diabetes. An imaging coming from the
projection pursuit method is given below.
A second example comes from the paper [61] and consists of scanned images of
handdrawn copies of the digit “two”.
The images are compared using a simple L2 -metric, and Mapper is applied using
a density filter. One can observe that the dominant feature, which is changing as
one moves along the graph, is the increasing presence of a loop in the lower left-
hand corner of the digit. This result is consistent with what was obtained by using
ISOMAP in [61].
23%
98% 100% 100%
100% 99% 100% 98% 44%
3%
G1 G1 G1 G1
C6 A7
G2 G2 G5 A8
G2 G2
G3 G3 G3 G3 C4 G9
C4 C4 C4 C4 G3 0.41 C6 A7 C6 A7 C6
C10 A7
0.51 0.58
G5 A8 G5 0.41
G5 G5 G5 G5 G2 G5 A8
C11 A8
0.96 C4
G1 U12 C4 0.71G9
C6 C6 C6 C6 C4 0.62G9 0.75
0.46 G9
G3
A7 A7 A7 A7 G3 C10 0.80 C10
0.72
C6 A7 G3 C10
0.79 0.51 0.57 G2 0.72
A8 A8 A8 A8 G2 C11 C11
G2 C11
G5 A8
0.50 G1 0.45
0.75
G9 G9 G9 G9 G1 U12 G1 U12 U12
0.70
C4 G9
C10 C10 C10 C10
G3 C10
C11 C11 C11 C11 G2
C11
G1
U12
U12 U12 U12 U12
C6
G5
A7
C4
0.42 A8
0.50
G3
G9
G2 G1 0.50 C10
C11
U12
The picture above is constructed using a data set constructed with the folding home
project, with website (http://folding.stanford.edu/), by simulating the folding of so-
called “RNA hairpins”. These are relatively small so-called motifs occurring within
larger RNA molecules, and are among the most frequently occurring such motifs.
The actual data set is obtained from a probabilistic simulation of the dynamics,
based on the notion of a contact map. The contact map is simply an array of zeroes
and ones, whose slots correspond to the residues along the molecule, and where an
entry is one if the corresponding pair of residues are in contact with one another,
by which we mean that they are within a fixed threshhold of each other. One can
impose a Hamming style metric on the set of these contact maps. At this point,
given a family of such contact maps generated by simulations, we can employ a
filter that is a good proxy for density and apply Mapper. The output from this
application is the colored graph displayed above. One notes that in the middle, one
has some slightly complicated behavior among the nodes, in particular a loop in the
corresponding graph. The contact maps corresponding to members of the clusters
corresponding to these nodes are displayed below the Mapper output. The contact
maps are displayed by inserting edges between residues which are in contact. We
note that given only the Mapper output, one might suspect that the small feature
(the array of orange nodes) could simply be an artifact, but examination of the data
shows that they correspond to essentially distinct contact maps. Note also that the
data are obtained by simply examining the states occurring in the simulation, and
that it does not include any dynamic information which would show how the states
are traversed in the folding process. This example points out an advantage of the
method, in that it is capable of locating small features within a larger data set.
The results described above appear in [6].
290 GUNNAR CARLSSON
4. Generalized forms of persistence

4.1. Multidimensional persistence. We have studied families of spaces para-
metrized by a single parameter as a way of extracting connectivity information
from a point cloud or finite metric space. It turns out that it is often useful to be
able to analyze the behavior of increasing families of spaces parametrized by more
than one variable.
Example. Given a point cloud X, one often attempts to understand the nature
of an underlying probability distribution which may have given rise to it. This
was clearly the case in the example of the image patch data described in Section
2.4 above. One way to do this is to estimate the density function using one of
many possible density estimators (see [58]). Given such an estimator, one can
now construct the family of spaces X[T ], where T is a percentage parameter, and
X[T ] ⊆ X is the subset of points which lie within the T -th percentile of density as
measured by the given density estimator. Clearly, if T ≤ T , we have an inclusion
X[T ] ⊆ X[T ], and one is often interested in understanding the geometric evolution
of this set as T increases. In the image patch example above, we chose a few possible
values of T to locate levels for T in which we saw interesting topological behavior,
but if one were to be able to study all the values of T simultaneously, and obtain a
summary of the behavior across all values of T , then one would not have to search
at random through the different values of T . In this case, we note that when we
apply, for example, the Vietoris-Rips construction to these sets, one obtains a two
parameter family of simplicial complexes {V R(X[T ], )},T . The parameter is
used to introduce geometry into the discrete sets X[T ], and the parameter T is the
function value we wish to track.
Example. Persistent methods can be used to study qualitative properties of shapes
which are not directly topological. For example, if one has a manifold, one can
study the filtration on the manifold by the value of the scalar curvature, and the
evolution of the topology of the sublevel sets of this filtration can reflect interesting
properties of the shape and can provide the basis for methods for discriminating
between shapes. In addition, one can build associated spaces to manifolds or other
complexes by studying various versions of the tangent bundle or the tangent cone
of geometric measure theory, which can also be equipped with interesting filtrations
which provide interesting information about the shape, and which can also be used
in locating features such as singularities in the space. See [10], [17], [28], and [7]
for details of these lines of research. In order to study this kind of persistence
for spaces given as point clouds, it is necessary to find discrete versions of the
geometric quantities (such as curvature) which are relevant, but it is also necessary
to use multiple persistence based on the geometric quantity and the scale parameter
simultaneously. As in the density example above, one needs the parameter in
order to impose some geometry on the discrete point cloud one is given.
Example. Suppose one is interested in studying the qualitative behavior of a real-
valued function f on Rn , in terms of local maxima, minima, saddle points, etc. An
efficient way of doing this is to study the evolution of the topology of the sublevel
sets SR = {x ∈ Rn | f (x) ≤ R}, as is done in Morse theory (see [49]). If one does
not have the explicit form of f , but only the values of f on some grid or other
sample S of points in Rn , one can approximate the topology of the sublevel sets
by the Vietoris-Rips complexes of S ∩ SR and study their evolution as R increases.
Of course, to extract the topology, one also needs to allow the scale parameter
in the Vietoris-Rips complexes to vary, and one obtains a two parameter family of
simplicial complexes {V R(S ∩ SR , )},R .
It is clear from these examples that it is very desirable to obtain useful and
computable summaries of the evolution of topology in situations where there is
more than one persistence parameter. A theory which describes how this can be
done is developed in [12]. We now describe this theory.
We recall from Definition 2.9 the notion of a P-persistence object in a category
C, where P is a partially ordered set, as a functor P → C, where P regards P as
a category in the usual way. The morphisms of P-persistence objects are natural
transformations of functors. Suppose we are given a family of topological spaces
(or simplicial complexes) {Xs,t }, with inclusions Xs,t → Xs ,t whenever s ≤ s and
t ≤ t . By choosing any order preserving map N × N → R × R, we obtain an N × N-
persistence object in the category of topological spaces (simplicial complexes). Of
course, this can be carried out for more than two variables in the obvious way. One
can now apply any of the homology functors Hi (−) with coefficients in a field F to
obtain an N × N-persistence F -vector space. We recall from Section 2.3 that the
category of N-persistence vector spaces over a field F is equivalent to the category of
non-negatively graded F [t]-modules. There is a corresponding statement concerning
Ns -persistence F -vector spaces.
Definition 4.1. By an n-graded ring, we will mean a ring A together with a direct
sum decomposition of abelian groups

A∼= At1 ,t2 ,...,tn ,
t1 ,t2 ,...,tk
where each of the ti ’s varies over N, and where the multiplication in the ring A
satisfies the requirement
As1 ,s2 ,...,sn · At1 ,t2 ,...,tn ⊆ As1 +t1 ,...,sn +tn .
Similarly, an n-graded module over an n-graded ring A is an A-module M equipped
with a direct sum decomposition

M∼ = Mti ,t2 ,...,tk
t1 ,t2 ,...,tk
so that the requirement

As1 ,s2 ,...,sn · Mt1 ,t2 ,...,tn ⊆ Ms1 +t1 ,...,sn +tn
is satisfied. Notions of homomorphism and isomorphism of n-graded rings and
modules are defined in the obvious ways, making the collection of n-graded A-
modules into a category.
The following proposition from [12] is now a straightforward observation.
Proposition 4.2. The category of Nn -persistence vector spaces over a field F
is equivalent to the category of n-multigraded modules over the polynomial ring
A(n) = F [x1 , x2 , . . . , xn ], where the multigrading structure on A(n) is given by
A(n)t1 ,t2 ,...,tn = F · xt11 xt22 · · · · xtnn .
As is well known to algebraists, the classification of finitely generated mod-
ules over multivariable polynomial rings is much more complicated than the cor-
responding result for single variable polynomial rings, and in fact no reasonable
292 GUNNAR CARLSSON
parametrization is known. This situation is also valid in the graded and multi-
graded cases. The classification of finitely generated graded modules in the single
variable case is parametrized by a set which is independent of the field in question,
while examples show that in the case of more than one variable, the classification of
multigraded modules definitely depends on the field in question. In fact, the classi-
fication in the multivariable case is parametrized by points in moduli varieties over
the ground field, and we therefore say that the one variable classification is discrete
while the classification in the multivariable case is continuous. This observation
is initially disappointing, since it suggests that useful classification results are not
likely to be available. However, it turns out that there are useful invariants, even
though they are not complete.
Definition 4.3. Let M be any finitely generated n-graded A(n)-module. Then for
any vector
t = (t1 , t2 , . . . , tn ) ∈ Nn ,
we define d(t) to be the dimension of the vector space M t. Similarly, for any pair
of vectors t, t ∈ Nn , with t ≤ t in the sense that ti ≤ ti for all i, we define r(t, t )
to be the rank of the multiplication map
t −t1 t2 −t2 t −tn

x11 x2 · · · xnn · : M t → M t .
The assignments t → d(t) and (t, t ) → r(t, t ) can be regarded as N-valued
functions on the sets Zn and Pn = {(t, t ) ∈ Zn × Zn |t ≤ t }, respectively. These
functions are clearly invariants of the isomorphism class of the module M , and we
refer to them as the dimension and rank invariants, respectively.
The following result is proved in [12].
Proposition 4.4. In the case n = 1, the rank invariant is a complete invariant of

the isomorphism class of a finitely generated graded F [t]-module for F a field.
This suggests that the computation of the rank invariant could be regarded as a
suitable generalization of single variable persistence, even though it clearly ignores
potentially interesting information. One question one could then ask is if there is
an interesting generalization of barcodes, which could be used to understand the
rank invariant in the same way as barcodes do for the single variable case. This
can be done as follows.
Suppose we are given an n-graded A(n)-module M , and we have computed the
dimension and rank invariants dM and rM . We say two elements s and t of Pn are
ρ-related, and write s ∼ρ t, if (a) dM (s) = dM (t) and (b) rM (s, t) = dM (s) = dM (t).
We then let ρ denote the equivalence relation generated by ∼ρ . This equivalence
now gives a partition of the set Nn . When we imagine the module as arising from
a multidimensional persistence vector space, we can imagine the various nodes in
the persistence diagram (i.e. the elements of Nn ) as embedded in Rn , and we can
assume that they are embedded quite densely, so that adjacent points are very close
in the actual Euclidean space. One can now color code the regions in the partition
associated to ρ , and if the dimension is ≤ 3, obtain a kind of image describing
the regions of the partition. A typical output might look like the image below.
0.8932 23
0.8937 22
0.9019 21
0.9043 20
1.1840 19
1.1888
18
1.2062
17
1.2076
1.2093 16
1.2335 15
1.2796 14
eccentricity
1.2806 13
1.3106 12
1.3142 11
1.3303 10
1.3309 9
1.3466 8
1.3467
7
1.3503
6
1.3814
1.3997 5
1.4050 4
1.4355 3
1.4809 2
1.4951 1
0.0000
1.7455
1.8429
1.8505
1.8519
1.8607
1.8843
1.8923
1.9041
1.9346
1.9663
2.0098
2.0224
2.0300
2.0389
2.0505
2.0636
2.1438
2.1572
2.1933
2.3211
2.3839
2.3894
2.3985
2.5019
2.5271
2.6651
2.6986
2.7430
2.8070
3.0776
3.1957
5.7746
9.3399
9.8997
distance
Regions of constant coloring then correspond to regions in which the vector space is
constant, i.e. all the members of a single color region have the same dimension, and
further can be connected by a sequence of isomorphisms associated to comparable
members of Nn . These regions correspond to intervals of constancy in the barcode,
i.e. intervals which contain no endpoints of any intervals occurring in the barcode.
A difficulty which must be addressed in working with multidimensional persis-
tence is the computational efficiency, and indeed setting up a viable computational
framework. In the case of single variable persistence, we use algorithms developed
for computing the Smith normal form of a matrix over a principal ideal domain.
This machinery is not available in the multivariable case, but it turns out that it
can be replaced by the Gröbner basis methodology (see [18] or [47]). The part
of that methodology which is relevant is the multigraded version of the notion of
a Gröbner basis for a submodule of a free finitely generated module, the Buch-
berger algorithm for constructing such a basis, and the algorithm for constructing
syzygies attached to homomorphisms of free multigraded modules (Schreyer’s al-
gorithm). These results are developed in [14]. The Gröbner basis provides a very
compact description which contains all the information about the multidimensional
persistence problem. In particular, it permits the reconstruction of the rank invari-
ant, which naively would have to be stored as sets of values for very large sets of
inputs.
4.2. Quivers and zigzags. In Section 2.3, we developed the notion of a P-persis-
tence object in a category C, where P is a partially ordered set. We then developed
the theory of such persistence objects in the case P = N to define and analyze
persistent homology. In the previous section, we extended this development to
294 GUNNAR CARLSSON
the case P = Nn , and we saw that it permitted the study of additional kinds of
problems not addressed in the case n = 1. In this section, we wish to develop the
notion of P-persistence for another class of partially ordered sets and show that
it allows us to address some interesting classes of problems. The results of this
section will appear in joint work with V. de Silva. We begin by listing the types of
problems we wish to study.
Example. Suppose that we are given a large data set X, and we wish to study
its homological invariants by studying the corresponding invariants of subsamples
from X. So, for example, if one wanted to estimate the first Betti number of a
putative space X underlying X, one might build a Vietoris-Rips complex with a
fixed for a collection of many samples, and if sufficiently many of them compute
the first Betti number to be n, then one might guess that the first Betti number
for X is n. However, the picture below suggests what might go wrong with such an
approach. It is a schematic picture of two different data sets.
Note that in the leftmost data set, the dominant qualitative picture is that of a sin-
gle loop, and one can expect that with reasonable frequency, samples may produce
point clouds which capture the circular structure through a barcode computation,
in the way illustrated by the two similar loops. In the rightmost data set, though,
one sees many different smaller circles, and one can imagine that each of the dif-
ferent samples might compute a first Betti number of one, but where each one
corresponds to a different loop, as is again indicated in the right-hand picture. One
can attempt to distinguish these by insisting that there be some measurable notion
of compatibility between the computations. Here is one such notion. Suppose we
have a family of samples Si ⊆ X, for i = 0, 1, . . . , N . For each i, with 0 ≤ i ≤ N −1,
we consider also the sample Ti = Si ∪Si+1 and note that we have inclusions Si → Ti
and Si+1 → Ti . This means that we actually have a diagram of samples from X:
T0 T1 · · · bE TN −1
~~> `@@@ ~~> ÀAA EE
EEjN −2 vv; cFF
FFjN −1
io ~
~ @@j0 i1 ~
~ AAj1 EE
iN −1
vv FF
~~ @@
@ ~~ AA EE vvv FF
~~ ~~ A vv
S0 S1 ··· SN −1 SN .
The value of a diagram of this form is that if we are given elements x0 ∈ Hk (V R(S0 ))
and x1 ∈ Hk (V R(S1 )), we can obtain information relating the two classes by con-
sidering their images i0 (x0 ) and j0 (x1 ) in the group Hk (V R(T0 )). If we find that
i0 (x0 ) = j0 (x1 ), then this acts as confirmation that the two elements correspond
to an element arising from the full data set X. Of course, there is no certainty
arising from this, but it suggests the likelihood that this occurs. Note that in the
example above, we will find this kind of compatibility in the left-hand set, and not
in the right. We would like to develop a systematic methodology which assesses

the frequency of these kinds of compatibilities.
Remark. The idea of recovering information about a large data set by studying
the behavior of various statistics on subsamples is an example of the bootstrap
method due to B. Efron [26]. In that context, one studies means, variances, and
other quantities evaluated on samples, and one then assesses how these statistics
vary over the samples. This is regarded as more informative than simply evaluating
those statistics directly on the full data set. We can view the kind of analysis we are
proposing above as a version of the bootstrap idea which is adapted to structural
information, such as presence of loops or cluster decompositions, rather than to
numerical values.
Example. Suppose that we are given a data set X equipped with a map τ to the
real line R. For example, if we have data containing a time component, then a
choice of τ would be the time coordinate. For any t0 ≤ t1 , we let X[t0 , t1 ] denote
the set {x ∈ X | t0 ≤ ρ(x) ≤ t1 }, and let Z(i) = X[i, i + 1] ∪ X[i + 1, i + 2]. Then
we have a diagram of point cloud data sets.
dII : Z(i) eLLL fMMM 9 ···

qq8
···
ss
Z(i + 1)
II
II uuu LLL qq MMM ss
II uu LLL qqq MMM ss
s
uu q
I uu L qqq MM sss
X[i, i + 1] X[i + 1, i + 2] X[i + 2, i + 3]
We note that the “shape” of this diagram is identical to the shape of the diagram in
the previous example. The description of the behavior of the diagram of F -vector
spaces obtained by applying homology with coefficients in F to the individual terms
should be a useful way of tracking how the data behaves dynamically, assuming one
can find a useful summary of the nature of the diagram of vector spaces.
Example. Density estimation is an important subject in statistics, and much of
what one wants to do in analyzing a data set is to describe in a useful way the
behavior of functions which estimate density in some way (see [58]). One way to
do this is via kernel density estimators. Suppose that we are given a data set X
embedded in Euclidean space Rn . For any point v ∈ Rn , and positive number ,
we let γv, denote a spherically symmetric Gaussian distribution with center at v
and with variance . Then one can construct the function
1
δ = γx,
#(X)
x∈X
as an estimate of the density of the distribution from which X arises by sampling.

The resulting function depends on the parameter . For large values of , one is
estimating density in a way which assigns significant weight to points which are far
from the given point x, and for smaller values of , one is estimating density where
one weights much more heavily the points in a smaller neighborhood of x. If < ,
then δ is a smoothed out version of δ . For a given X, it is an interesting question
to determine which choice of is “correct”, and statisticians have developed useful
heuristics along these lines. Another approach, though, is to attempt to provide
a summary of the behavior of invariants over the full range of values of at once.
Fixing a percentage threshhold T , we define the set X[T, ] to be the set of points
lying in the T % densest points as measured by the estimator δ . If we now fix a
296 GUNNAR CARLSSON
sequence of values 0 < 1 < · · · < k , and set Zi = X[T, i ] ∪ X[T, i+1 ], we can
now construct the following diagram of data sets.
· · · cG Z ···
u:
Zi dI
GG
GG xx; II
I tt9 i+1 eKK
K uu
x II t KK u
GG xx II tt KK uu
GG x xx II t tt KK uuu
x t u
X[T, i ] X[T, i+1 ] X[T, i+2 ]
If we apply the Vietoris-Rips construction with a fixed parameter value to the
diagram, and then apply Hj for some j, we will obtain another diagram of F -
vector spaces of the same shape as what we have been looking at in the earlier
examples. It should then be helpful in tracking the qualitative behavior of the
density estimator with varying .
Example. We recall the witness complex construction introduced in Section 2.2.
Recall that this was a complex W w (X, L, ) constructed on the point cloud X,
using a finite “landmark set” L and a positive parameter . It would clearly be
informative to understand the extent to which homology calculations or clustering
depends on the choice of the landmark set L, but there is no apparent relationship
between the complexes W w (X, L, ), even if the one landmark set is contained in
the other. One can, however, proceed as follows. Given a point cloud X and
two landmark sets L and L , we construct a two-variable version of the witness
complexes, denoted W w (X, L, L , ). Its vertex set is L × L , and we declare that
a collection {(l0 , l0 ), . . . , (lk , lk )} spans a k-simplex if and only if there exist -weak
witnesses x and x in X for the collections {l0 , . . . , lk } and {l0 , . . . , lk }, respectively.
It is roughly analogous to the Čech complex of the covering by intersections of
Voronoi cells in two different Voronoi decompositions of a metric space. We note
that we have projections
W w (X, L, L , ) → W w (X, L, ) and W w (X, L, L , ) → W w (X, L , )
induced by the vertex maps L × L → L and L × L → L . Suppose now that
we have a family of landmark sets Li for X. We let Wi = W w (X, L, ) and Ui =
W w (X, Li , Li+1 ). Then we have the following diagram.
· · ·aD Wi+1 ···

z=
Wi À
DD
DD zz< AA zz< cGG
G zz
DD zz AA z GG zz
DD zz AA zz GG zz
zz A zzz G zz
Ui−1 Ui Ui+1
Once again, one can apply homology to the diagram, and the compatibility of classes
under the maps in this diagram should be a useful indication that the classes are
intrinsic to X and do not depend on the choices of landmark points. This kind of
compatibility information would, for example, be extremely useful in interpreting
the neuroscience results described in Section 2.5.
The question we now face is how to formalize the notion of compatibility in these
diagrams. Here is how one can proceed. We first define a partial ordering on the
set of integers Z. We will declare that for every integer k, we have 2k + 1 > 2k
and 2k + 1 > 2k + 2. Thus, every odd number is greater than its two adjacent even
numbers, and except for identities, there are no other comparabilities.
··· 2k − 1 2k + 1 2k + 3 ···
Q > > Q > > Q > > Q > >
Q
s
Q +
Q
s
Q
+
QQ
s +
QQ
s +

2k − 2 2k 2k + 2 2k + 4
We will denote this partially ordered set by Z, and we will denote by Z[m, n] the
subset of all integers i with m ≤ i ≤ n, where m and n are integers with m ≤ n. We
note that all the examples above are either Z-persistence sets or Z[m, n]-persistence
sets, and when one applies a Vietoris-Rips construction one obtains corresponding
Z or M(m, n)-persistence simplicial complexes. Applying Hi with coefficients in
a field F for some i to each of these diagrams then gives Z or Z[m, n]-persistence
F -vector spaces. The key ingredient in ordinary persistence is the observation
that there is a classification of N-persistence vector spaces, and it turns out that
there is a classification of the isomorphism classes of Z[m, n]-persistence F -vector
spaces, which is proved and discussed in [29]. We will describe the structure of this
classification.
Fix m ≤ n. For any m0 ≤ n0 , with m ≤ m0 ≤ n0 ≤ n, we will define an
elementary object E(m0 , n0 ) as follows. In order to specify a Z[m, n]-persistence
vector space, it is sufficient to specify (a) a vector space Vi for all m ≤ i ≤ n
and (b) linear transformations V2k+1 → V2k and V2k+1 → V2k+2 whenever the
two vector spaces Vi involved in the linear transformation are defined. To define
E = E(m0 , n0 ), we set Ei = F for m0 ≤ i ≤ n0 , and Ei = {0} otherwise, and we
declare that all transformations are the identity if they can be. If they cannot, then
they involve a vector space which contains only the zero vector, and are therefore
of necessity equal to zero. See below for a picture of E(1, 4).
{0} {0} F F {0} {0}

@ @ @id id @ id
@
@R
@ R
@ @R
@ @R
@ @R
@
··· {0} F F ···
The upper row consists of the vector spaces for the odd integers, and the lower
row consists of the ones for the even integers. The three dots indicate an array
of zero vector spaces. Now, as in Section 2.3, we have a notion of morphism of
Z[m, n]-persistent vector spaces, and we can therefore ask for the classification up
to isomorphism of them. Here is the theorem, which can be extracted from [29].
Theorem 4.5. Let M denote any Z[m, n]-persistent F -vector space, so that every
vector space Vi is finite dimensional. Then there is an isomorphism

t
M∼
= E(mj , nj )
j=1
for some t and a family of pairs of integers (mj , nj ), such that m ≤ mj ≤ nj ≤ n

for all j. Moreover, this decomposition is essentially unique, in the sense that t is
the same for all such decompositions, and further that the pairs (mj , nj ) are also
unique up to a reordering of elements.
One can use this result to obtain the following straightforward consequence.
Corollary 4.6. We say that a Z-persistence F -vector space M is finite if each Mi
is finite dimensional and there exists an integer N such that Mi = {0} if |i| ≥ N .
298 GUNNAR CARLSSON
Any finite Z-persistence F -vector space M can be decomposed as

t
M∼
= E(mj , nj )
j=1
for some t and a family of pairs of integers (mj , nj ), such that mj ≤ nj . The
decomposition is again essentially unique.
Remark. Z[m, n]-persistence vector spaces are examples of quiver representations,
a highly developed area of algebra. See [29] for a complete description.
The families of intervals in each of these decompositions can be interpreted as
barcodes with integer valued endpoints. We argue that the analysis of these dia-
grams should be useful in the analysis of the kinds of problems we have discussed
above. We give intuitive illustrations of how this might work. We suspect that
there are theorems which could be proved in this direction, and we hope that that
will be the subject of future work.
Example. In the context of clustering samples as above, suppose that one obtains
a β0 -barcode with two long lines and a family of shorter lines. This outcome would
suggest the possibility that one is really looking at two essential clusters in the full
data set, with others arising out of the sampling. We view this idea as a potential
contributor to the study of consistency or stability of clustering ([4], [43]).
Example. In the context of variable landmark sets as above, suppose one obtains
a β1 -barcode with one long line. This would be confirming evidence toward the
hypothesis that the original data set has a β1 of one, and that what one is seeing
in each of the witness complexes is a reflection of that qualitative feature.
Example. In considering dynamic data, the barcodes obtained should be useful in
understanding the topological transitions occurring in a changing data set. They
should, for example, give a guide to the behavior of clusters over time.
Remark. It is interesting to note that the problems mentioned here are interesting
and difficult even in the analysis of β0 , so that the methods we propose should give
interesting new information about the behavior of clustering.
4.3. Tree based persistence. As we have seen above, algebraic topology is ca-
pable of producing signatures which indicate the presence of topological features
within a space. As it stands, however, it is not capable of describing the source of
the feature, i.e. where in the space the hole or other feature is located. By using
persistence, one is able to develop a systematic way of addressing such questions.
That methodology has been described in [65]. In this section, we describe what was
done there and also suggest another persistence framework into which it might fit.
We suppose that we are given a simplicial complex Σ and a covering S =
{Σα }α∈A of Σ by subcomplexes. For example, if we suppose that we are given
a set of point cloud data X and a covering U = {Uα }α∈A of X, one could obtain a
covering of V R(X, ) by the full subcomplexes of V R(X, ) on the vertex sets Uα .
The use of filters as described in Section 3 above can provide coverings of X, or
one could cover using balls with centers distributed through the space. One might
also cover X by versions of Voronoi cells, as in the discussion of witness complexes
in Section 2.2.
Definition 4.7. Let x ∈ Hi (Σ). We say x is S-small if x ∈ im(iα : Hi (Σα ) →
Hi (Σ)) for some α.
Once one determines that a class is S-small, and for which α the class x is in
the image of iα , one has in effect shown that the feature at least can be represented
within the given subset of Σ, and so has information about the source of the class.
We recall the Mayer-Vietoris blow-up construction M(|Σ|, S) from Section 3.2.
This is a space equipped with projection maps π∆ : M(|Σ|, S) → ∆[A], where A
is the indexing set for the covering S, and pΣ : M(|Σ|, S) → |Σ|. It can be shown
that pΣ is a homotopy equivalence (see [65]). One can consider the skeletal filtration
{∆[A](k) } on ∆[A], where ∆[A](k) is the subspace consisting of the union of all faces
−1
of dimension ≤ k, and the corresponding filtration π∆ (∆[A](k) ) on M(|Σ|, S). The
following propositions are now easy to verify.
Proposition 4.8. A homology class x ∈ Hi (|Σ|) is S-small if and only if x is in
−1
the image of the homomorphism Hi (π∆ (∆[A](0) )) → Hi (M(|Σ|, S)).
Proposition 4.9. If a homology class x ∈ Hi (|Σ|) is in the image of
Hi (|Σα0 | ∪ |Σα1 | ∪ · · · ∪ |Σαk |) → Hi (|Σ|)
for some (k + 1)-fold collection of elements of S, then x is in the image of the
homomorphism
−1
Hi (π∆ (∆[A](k) )) → Hi (M(|Σ|, S)).
Note that this means that we have a {0, 1, . . . , #(A) − 1}-persistence vector
−1
space {Hi ((π∆ (∆[A](j) ))}j , which contains information about where the homology
classes in Σ arise. If we are asking whether the class arises from an individual set in
S, then the persistence vector space gives us a complete answer to this question. If
the question is instead whether or not an element arises from a union of k elements
of S, then we do not obtain complete information this way, but we do obtain partial
information in that we can preclude the possibility that a class arises from such a
union. Examples of the application of this approach are given in [65].
We will also suggest the possibility of another approach to the question of de-
termining the origin of homology classes. Consider first any rooted tree (T, v0 ),
where v0 is the root. The vertex set of T can now be given a partial ordering
≤T defined by v1 ≤T v2 if and only if the shortest path from v1 to v0 contains
v2 . The properties of trees guarantee that ≤T is a partial ordering. Next, suppose
that we have a simplicial complex Σ with a family of coverings Si = {Σα }α∈Ai by
subcomplexes, equipped with functions θi : Ai → Ai+1 such that for any simplex
σ ∈ Σ and α ∈ Ai , we have that σ ∈ Σα implies that σ ∈ Σθi (α) . We suppose that
there is an integer N so that AN consists of a single element, and that therefore
the covering SN consists only of Σ itself. This covering data gives rise to a rooted

tree whose vertex set is N i=1 Ai , and where the edges are all of the form (α, θi α),
for some α ∈ Ai . The single element in AN is a root for the tree. This kind of
family of coverings can arise in natural ways from certain coverings of complexes
or spaces.
Example. Consider the unit interval I = [0, 1] and fix N . Let Si denote the
N −i
N −i ], where 0 ≤ k ≤ 2
covering of I given by the family of intervals [ 2Nk−i , 2k+1 −1
N −i
is an integer. Let Ai denote {k ∈ Z | 0 ≤ k ≤ 2 − 1}, and define θi : Ai → Ai+1
to be the function k → k2 . We now have a family of coverings, which become
increasingly “coarse” as i increases. The associated tree is a binary tree with 2N −1
leaves.
300 GUNNAR CARLSSON
Example. If our space is I n instead, we may cover by products of the sets in the
coverings Si .
Let Si = {Uα }α∈Ai be a family of coverings as above, and θi : Ai → Ai+1 be
maps of coverings as above. Let (T, v0 ) denote the associated rooted tree, and VT
its vertex set equipped with the partial ordering ≤T . Then we define an associated
(VT , ≤T )-persistence vector space {Wt }t∈VT as follows. Given α ∈ Ai ⊆ VT , we set
Wα = Hi (Uα ), and whenever α ≤T α , then the associated linear transformation
from Hi (Uα ) to Hi (Uα ) is the map induced from the inclusion Uα → Uα = Uθi (α) .
The idea of studying the source of homology classes should now be rephrased in
terms of invariants of (VT , v0 )-persistent vector spaces. This situation is a bit
like the situation encountered in Section 4.1 in that the classification will involve
points on positive dimensional varieties over the ground field. Nevertheless, it
appears plausible that one can construct useful invariants, such as the rank invariant
discussed there.
5. Reasoning about clustering

As we have noted above, clustering algorithms are methods which take as input
a finite metric space (X, d) and produce as output a partition Π(X, d) of the un-
derlying set X. In [39], J. Kleinberg proves a non-existence theorem for clustering
algorithms satisfying certain properties, in a spirit similar to that of the Arrow
Impossibility Theorem. We will enumerate three properties which may be satisfied
by clustering algorithms.
(1) Scale invariance: Given a finite metric space (X, d), the partitions of X
associated to (X, d) and (X, rd), where r > 0, are identical.
(2) Richness: Any partition of a finite set X can be realized as Π(X, d) for
some metric d on X.
(3) Consistency: Suppose that d and d are two metrics on a finite set X,
and suppose that (a) for any x, x contained in one of the clusters attached
to d (i.e. blocks in the partition Π(X, d)) we have d (x, x ) ≤ d(x, x ) and
(b) for any x, x which belong to distinct clusters attached to d, we have
d (x, x ) ≥ d(x, x ). Then Π(X, d) = Π(X, d ).
Kleinberg’s theorem is now
Theorem 5.1 (Kleinberg, [39]). There are no clustering algorithms that satisfy
scale invariance, richness, and consistency.
This interesting result is disappointing in that it does not give guidance con-
cerning the choices of clustering algorithms, but rather points out deficiencies from
which all clustering algorithms must suffer. It is therefore interesting to identify
situations in which one can prove existence, and perhaps existence and unique-
ness given certain properties. As Kleinberg points out, it is possible to do so in
a number of different ways by specifying cost functions on particular clusterings,
and prove a uniqueness result for optimal choices of clusterings with respect to
this cost function, but that this is perhaps less interesting in that cost functions
can often be defined which will isolate a particular algorithm. In [9], such a con-
text is developed which is not dependent on the choice of a cost function but is
rather dependent on “structural” criteria which have a great deal in common with
Kleinberg’s requirements. We now describe the main result of [9].
We begin with the informal observation that clustering for finite metric spaces
can be thought of as the statistical version of the geometric notion of forming the
set π0 (X) of connected components of a topological space X. We note that the
correspondence X → π0 (X) can actually be viewed as a functor (see [44]) from the
category of topological spaces to the category of sets, in the sense that a continuous
map f : X → Y induces a map of sets π0 (f ) : π0 (X) → π0 (Y ), satisfying certain
obvious conditions on composite maps and identity maps. This observation is much
more than a curiosity. It is the basis for many comparison theorems in topology,
and in fact underlies the combinatorization of topology obtained via simplicial sets
([45], [19]). It is also what underlies the theoretical constructions underlying the
Mapper algorithm discussed in Section 3. Finally, it is also the basis for étale
homotopy theory, which adapts topological methods to the study of number theo-
retic problems. The naturality of this condition as well as its utility in many other
mathematical contexts suggests that it is very natural to formulate such a condition
for clustering algorithms as well. We will therefore attempt to describe clustering
algorithms as functors between two categories, where the domain category has as
its objects the collection of finite metric spaces. One must therefore first define a
notion of morphism of finite metric spaces. We define several such notions.
(1) Isometry: An isometry from a finite metric space (X, dX ) to another finite
metric space (Y, dY ) is a bijective map of sets f : X → Y so that
dY (f (x), f (x )) = dX (x, x ) for all x, x ∈ X.
(2) Embeddings: An embedding from a finite metric space (X, dX ) to another
finite metric space (Y, dY ) is a map of sets f : X → Y so that
dY (f (x), f (x )) = dX (x, x ) for all x, x ∈ X.
(3) Monomorphisms: A monomorphism from a metric space X to another
metric space Y is a monic set map f : X → Y so that for all x, x ∈ X, we
have dY (f (x), f (x )) ≤ d(x, x ).
(4) Distance non-increasing maps: A distance non-increasing (dni) map
from a finite metric space (X, dX ) to another finite metric space (Y, dY )
is a map of sets f : X → Y so that dY (f (x), f (x )) ≤ dX (x, x ) for all
x, x ∈ X.
Each of these notions creates the structure of a category whose objects are the
finite metric spaces, since one can readily observe that each of the classes of mor-
phisms is closed under composition and contains the identity. We denote each of
these categories by Miso , Memb , Mmon , and Mgen , respectively. One could ini-
tially hope to study clustering algorithms as functors from each of these categories
to sets. Thinking in these terms, though, it is first clear that not every functor
on one of these categories deserves the name “clustering functor”. To see this, we
return to the geometric notion of connected components. Not only is the correspon-
dence π0 (−) a functor from spaces to sets, it is equipped with a natural surjective
map of sets ηX : X → π0 (X). Further, these surjective maps have the property
that for every continuous map f : X → Y , the diagram of sets (below) commutes.
f
X −−−−→ Y
⏐ ⏐
⏐
ηX
⏐ηY

π0 (f )
π0 (X) −−−−→ π0 (Y )
302 GUNNAR CARLSSON
Remark. In formal terms, the maps ηX , as X runs over all topological spaces, form
a natural transformation from S to π0 (X), where S denotes the “underlying set”
functor from the category of topological spaces to the category of sets.
By arguing with the analogy with the connected component construction, we will
begin with a provisional definition of a C-functorial clustering algorithm, where C is
one of the above mentioned categories. It will be a functor χ from C to the category
of sets together with a family of surjective maps of sets η(X,d) : X → χ(X, d) so
that the diagrams
f
X −−−−→ Y
⏐ ⏐
⏐
ηX
⏐
ηY
χ(f )
χ(X, dX ) −−−−→ χ(Y, dY )
commute for every morphism f : (X, dX ) → (Y, dY ) in C.
It is clear that a C-functorial clustering algorithm is a clustering algorithm in
the sense of Kleinberg, since the surjective map of sets from X to χ(X, dX ) yields a
−1
partition of X, namely the partition whose blocks are the sets ηX (z), as z ranges
over all elements of χ(X, dX ). This means that Kleinberg’s conditions also make
sense in this context. We examine the possible clustering functors in two of these
cases.
Example. Miso -functorial clustering algorithms are very simple to describe. Let
I denote the collection of isometry classes of finite metric spaces, and for each
ι let (Xι , dι ) denote an element of the isomorphism class ι. Let Gι denote the
automorphism group of (Xι , dι ), and let Pι denote the set of all possible partitions
on Xi . Clearly the group Gι acts on the set Pι , and we let PιGι denote the fixed point
set of that action. A Miso -functorial clustering algorithm determines an choice
pι ∈ PιGι for every ι, and conversely an arbitrary choice of such pι ’s determines an
Miso -functorial clustering algorithm. If we impose Kleinberg’s scaling condition, we
must instead decompose the set of all finite metric spaces into equivalence classes,
where two metric spaces are in the same equivalence class if and only if one is
isometric to a rescaling of the other. The classification is now determined exactly as
above, except that one needs only to make a choice for each of these new equivalence
classes.
Example. The study of general Mgen -functorial and Mmon clustering algorithms
is more subtle, and we do not yet have a general classification. We will instead give
some examples.
• Fix a threshhold parameter and, for a finite metric space (X, dX ), let ∼
denote the equivalence relation generated by the relation R on X defined
by xR x if and only if d(x, x ) ≤ . Then the clustering algorithm which
assigns to each (X, dX ) the partition associated to ∼ is clearly Mgen -
functorial. This example corresponds to single linkage clustering with a
fixed parameter value .
• Consider the finite metric space ∆[n] whose elements are {1, 2, . . . , n}, and
where d(i, j) = for all pairs 1 ≤ i < j ≤ n. For any finite metric space
(X, dX ) we define a new relation R on X by the requirement that xR x if
and only if there is a distance non-increasing inclusion j : ∆[n] → (X, dX )
such that j(1) = x and j(2) = x . We then let E denote the equivalence
relation generated by R . The clustering algorithm which assigns to each

(X, dX ) the partition of X into the blocks of the equivalence relation E is
now clearly Mmon -functorial. This notion of clustering is closely related to
clique clustering algorithms in network clustering [53].
• More generally, for any family Φ of finite metric spaces, one can define
clustering algorithms attached to Φ by analogy with the previous example,
where the relation involves injective maps from elements of the family or
possible arbitrary maps of finite metric spaces from the family.
• For any finite metric space (X, dX ), define µ(X) to be the minimal non-
zero distance between distinct points of X. The clustering algorithm which
assigns to each metric space (X, dX ) the clustering associated to the equiva-
lence relation generated by R µ(X)
1 is readily checked to be Mmon -functorial.
It is now easy to show that if one imposes the scale invariance condition of
Kleinberg, one finds that there are no non-trivial Mgen -functorial claustering algo-
rithms, where the trivial ones are understood to mean the discrete one (with one
element clusters) and the indiscrete one (in which X forms the single cluster for
every (X, dX )).
Non-existence results are of course interesting, but more useful from the point
of view of applying and using clustering algorithms are situations where existence
and uniqueness can be proved. The non-existence result mentioned in the previous
example suggests that one should look for a more relaxed framework. In order to do
this, we will change the target category for clustering algorithms from the category
of sets to the category of R+ -persistent sets, where R+ denotes the non-negative
real numbers. This is not an unreasonable thing to do in view of the fact that
hierarchical clustering algorithms do not report single partitions but rather den-
drograms, which are roughly speaking R+ -persistent sets. We have already defined
the notion of morphisms of R+ -persistent objects in any category. A morphism of
persistent sets is said to be surjective if each of the individual morphisms which
make it up is surjective. For any finite metric space (X, dX ), we associate to it the
R+ -persistent set α(X) = {α(X)r }r for which α(X)r = X for all r ∈ R+ , and such
that all the morphisms α(X)r → α(X)r , for r ≤ r , are the identity morphisms on
X. Then by a persistent clustering algorithm we will mean an assignment to every
finite metric space (X, dX ) of a persistent set ξ(X, dX ) and a surjective morphism
α(X) → ξ(X, dX ) of persistent sets for every finite metric space (X, dX ). Letting
C be any of the category structures on the collection of finite metric spaces given
above, one can now define a corresponding notion of C-functorial persistent clus-
tering algorithms as follows. Such a clustering algorithm is a functor ξ from C to
the category of persistent sets, equipped with a surjective morphism of persistent
sets ηX : α(X) → ξ(X, dX ) for every (X, dX ), so that the diagrams
α(f )
α(X) −−−−→ α(Y )
⏐ ⏐
⏐
ηX
⏐ηY

ξ(f )
ξ(X, dX ) −−−−→ ξ(Y, dY )
commute for every morphism f in C. In this context, Kleinberg’s scale invariance

condition takes a different form. For any R+ -persistent set U = {Ur } and any
304 GUNNAR CARLSSON
positive real number t, we define the persistent set tU by tUr = U rt , and by re-
quiring that for any r ≤ r , the homomorphism tUr → tUr be identified with the
corresponding homomorphism U rt → U r . The correspondence U → tU is clearly
t
a functor from the category of persistent sets to itself, which we will write as δr .
We also have the rescaling functor ρr from the category C to itself, which simply
multiplies all distances by r. We now say that a persistent clustering algorithm ξ is
scale invariant if (a) χ(ρr (X, dX )) = δr (χ(X, dX )) and (b) ηρr (X,dX ) = δr (η(X,dX ) ).
In [9], the following result is proved.
Theorem 5.2. Let E denote a metric space with two points, and such that the
distance between those two points is = 1. Let P denote the persistent set for which
Pr consists of two points for r < 1, of one point for r ≥ 1, and such that all the
maps Pr → Pr are surjective for all r ≤ r . There is a unique Mgen -functorial
persistent clustering algorithm Ξ which is scale invariant and which satisfies the
requirement that Ξ(E) = P . The algorithm Ξ is the algorithm which associates to
a finite metric space (X, dX ) the persistent set {π0 (V R(X, ))}≥0 .
The proof is not difficult. This result is in the spirit of Kleinberg’s result in
that the requirements which define it are structural rather than requiring the opti-
mization of some cost function, but yields an existence and uniqueness result. This
theorem therefore precludes a number of interesting algorithms such as average
linkage and complete linkage clustering from being functorial. Users of clustering
algorithms often assert that average linkage and complete linkage clustering are to
be favored over single linkage clustering because the clusters tend to be more “com-
pact” in an intuitive sense. We believe that their observation should be interpreted
as saying that in clustering one needs to take into account more than just the metric
as a geometric object, but in addition some notion of density. This suggests the
possibility of including density into notions of functoriality, which we will explore
in a future paper.
6. What should the theorems be?

We have presented some examples of how topological methods can be applied
to study data sets. The methods provide signatures which yield information about
the shape of the data sets being studied, and they can also provide useful methods
for visualizing data sets. However, there are many outstanding questions of a
statistical nature. One situation which illustrates some of these questions was that
which occurred in the discussion of the neuroscience data in Section 2.5 above. In
that case, one found that the data exhibited barcodes which appeared to indicate
the presence of non-zero β1 and β2 in the data sets. While the observation is of
course suggestive, to prove that it indicates structure in the data set one must show
that segments of the indicated length cannot occur in a random model associated to
a null hypothesis. In the paper [59], this was shown by simulation, which is of course
always an option. However, it is clear that if it were the case that such results could
actually be proved, one would avoid the time and effort spent in simulation, and it
would also provide the basis for thinking more systematically about the significance
of questions for the barcodes arising out of persistent homology. We outline how
one might begin formulating such results.
Suppose we are given a metric space X and a probability measure on X. Then
we can consider an experiment consisting of selecting n points {x1 , x2 , . . . , xn }
i.i.d. from X. For each time this experiment is carried out, we can then form
the R+ -persistent simplicial complex {V R({x1 , . . . , xn }, )}≥0 , the corresponding
homology vector spaces {Hi (V R({x1 , . . . , xn }, ))}≥0 , and finally therefore a bar-
code. We note that the barcode is simply a family of intervals with endpoints in R+ .
Each such interval can be considered as a point in the set D = {(x, y) ∈ R2 |x ≤ y},
and the output is therefore a finite subset of the set D. As such, we now have
a finite spatial point process (see [20], [3]). Roughly, a finite spatial point process
with values in a locally compact second countable Hausdorff space S is a probability
measure on a σ-algebra constructed on the collection of finite counting measures on
S. The σ-algebra is the minimal one which makes all the counting maps ΦB mea-
surable, where B is in the Borel σ-algebra associated to S, and ΦB evaluates the
counting measure on B. Point processes are a heavily studied area within statistics.
We believe that theorems which describe these point processes will be very useful
in applying algebraic topology to the qualitative analysis in many areas of science.
Here are some suggestions which would be interesting.
• Determine the distributions of various statistics, such as maximum and
average distance to the diagonal, on the point processes attached to the
Vietoris-Rips complexes obtained by sampling sets of points from various
probability distributions on Euclidean space. Gaussian distributions are a
good place to start.
• Similarly, determine distributions of various statistics on point processes
obtained by selecting sets of landmark points using various strategies from
probability distributions on Euclidean space and computing the persistent
witness complexes.
• Study consistency of clustering problems by studying the distributions of
various statistics on the zigzag barcodes obtained by sampling from a larger
data set.
• Develop a method for studying the output of multidimensional persistent
homology probabilistically.
We point out that in the context of ordinary homology (i.e. not persistent homol-
ogy), M. Penrose and others have developed a theory of “geometric random graphs”
and have proved various results concerning the Betti numbers of complexes attached
to such graphs (see [54]). Also, results about barcodes under the general heading
have now begun to appear (see [15] and [16]). These results should be an excellent
starting point for the development of theorems of the type mentioned above.
About the author

Gunnar Carlsson is the Swindells Professor of Mathematics at Stanford Univer-
sity. His research is supported by major grants from DARPA and the National
Science Foundation.
References
[1] H. Abdi, Metric multidimensional scaling, in Encyclopedia of Measurement and Sta-
tistics, Sage, Thousand Oaks, CA, (2007), pp. 598-605.
[2] H. Adams and G. Carlsson, On the non-linear statistics of range image patches, preprint,
(2007), available at http://comptop.stanford.edu/preprints/.
[3] A. Baddeley, Spatial point processes and their applications, appears in A. Baddeley, I. Bárány,
R. Schneider, and W. Weil, editors, Stochastic Geometry: Lectures given at the
C.I.M.E. Summer School held in Martina Franca, Italy, September 13-18, 2004,
306 GUNNAR CARLSSON
Lecture Notes in Mathematics 1892, Springer-Verlag, Berlin, 2007. ISBN 3-540-38174-0, pp.
1–75. MR2327290 (2008c:60045)
[4] S. Ben-David, U. von Luxburg, D. Pál, A sober look on clustering stability, in G. Lugosi and
H. Simon, editors, Proceedings of the 19th Annual Conference on Learning Theory (COLT),
pages 5–19, Springer-Verlag, Berlin, 2006. MR2277915 (2007i:68102)
[5] A. Björner, Topological methods, appears in Handbook of Combinatorics, Vols. 1, 2,
1819–1872, Elsevier, Amsterdam, 1995. MR1373690 (96m:52012)
[6] G. R. Bowman, X. Huang, Y. Yao, J. Sun, G. Carlsson, L. J. Guibas, and V. S. Pande,
Structural insight into RNA hairpin folding intermediates, Journal of the American Chemical
Society Communications, July, 2008.
[7] E. Carlsson, G. Carlsson, and V. de Silva, An algebraic topological method for feature identi-
fication, International Journal of Computational Geometry and Applications, 16 (2006), no.
4, pp. 291-314. MR2250511 (2007c:52015)
[8] G. Carlsson and V. de Silva, Topological estimation using witness complexes, Symposium on
Point-Based Graphics, ETH, Zürich, Switzerland, June 2-4, 2004.
[9] G. Carlsson and F. Memoli, Persistent Clustering and a Theorem of J. Kleinberg , Preprint,
March 2008.
[10] G. Carlsson, A. Zomorodian, A. Collins, and L. Guibas, Persistence barcodes for shapes,
International Journal of Shape Modeling, 11 (2005), pp. 149-187.
[11] G. Carlsson, T. Ishkhanov, V. de Silva, and A. Zomorodian, On the local behavior of spaces
of natural images, International Journal of Computer Vision, (76), 1, 2008, pp. 1-12.
[12] G. Carlsson and A. Zomorodian, The theory of multidimensional persistence, 23rd ACM
Symposium on Computational Geometry, Gyeongju, South Korea, June 6-7, 2007.
[13] G. Carlsson and T. Ishkanov, Local structure of spaces of natural images, preprint, (2007),
available at http://comptop.stanford.edu/preprints/
[14] G. Carlsson, G. Singh, and A. Zomorodian, Computing multidimensional persistence, in
preparation.
[15] D. Cohen-Steiner, H. Edelsbrunner and J. Harer, Stability of persistence diagrams, Discrete
Comput. Geom., 37 (2007), 103–120. MR2279866 (2008i:68130)
[16] D. Cohen-Steiner, H. Edelsbrunner, J. Harer and Y. Mileyko, Lipschitz functions have Lp -
stable persistence, Found. Comput. Math., to appear.
[17] A. Collins, A. Zomorodian, G. Carlsson, and L. Guibas, A barcode shape descriptor for curve
point cloud data, Computers and Graphics, Volume 28, 2004, pp. 881–894.
[18] D. Cox, J. Little, and Donal O’Shea, Using Algebraic Geometry, Graduate Texts in
Mathematics, Springer-Verlag, 1998, xii + 499 pages, ISBN 0-387-98492-5. MR1639811
(99h:13033)
[19] E. Curtis, Simplicial homotopy theory, Advances in Math. 6 (1971), 107-209. MR0279808
(43:5529)
[20] D. Daley and D. Vere-Jones, An Introduction to the Theory of Point Processes, two
volumes, Second Edition, Springer-Verlag, 2003, ISBN 0-387-95541-0. MR950166 (90e:60060)
[21] B. Delaunay, Sur la sphere vide, Izvestia Akademii Nauk SSSR, Otdelenie Matematicheskikh
i Estestvennykh Nauk, 7:793-800 1934.
[22] J.G. Dumas, F. Heckenbach, B.D. Saunders, and V. Welker, Computing simplicial homol-
ogy based on efficient Smith normal form algorithms, In Algebra, Geometry, and Software
Systems (2003), 177-207. MR2011758 (2004i:55009)
[23] D. Dummit and R. Foote, Abstract Algebra. Third edition, John Wiley & Sons, Inc., Hoboken,
NJ, 2004. xii+932 pp. ISBN: 0-471-43334-9 00-01 (16-01 20-01). MR2286236 (2007h:00003)
[24] H. Edelsbrunner, D. Letscher, and A. Zomorodian, Topological persistence and simplification,
Discrete and Computational Geometry 28, 2002, 511-533. MR1949898 (2003m:52019)
[25] H. Edelsbrunner and N.R. Shah, Triangulating topological spaces, Tenth Annual ACM Sym-
posium on Computational Geometry (Stony Brook, NY, 1994). Internat. J. Comput. Geom.
Appl. 7 (1997), no. 4, 365–378. MR1460843 (98f:57038)
[26] B. Efron, Bootstrap methods: Another look at the jackknife, Ann. Statist. 7 (1979), no. 1,
pp. 1-26. MR515681 (80b:62021)
[27] S. Eilenberg, Singular homology theory, Ann. of Math. (2) 45 (1944), 407–447. MR0010970
(6:96f)
[28] P. Frosini and C. Landi, Size theory as a topological tool for computer vision, Pattern Recog-
nition and Image Analysis, vol. 9 (4) (1999), pp. 596-603.
[29] P. Gabriel and A. Roiter, Representations of Finite-Dimensional Algebras. Translated

from the Russian. With a chapter by B. Keller. Reprint of the 1992 English translation.
Springer-Verlag, Berlin, 1997. iv+177 pp. ISBN: 3-540-62990-4. MR1475926 (98e:16014)
[30] P.G. Goerss and J.F. Jardine, Simplicial homotopy theory, Progress in Mathemat-
ics, 174. Birkhäuser Verlag, Basel, 1999. xvi+510 pp. ISBN: 3-7643-6064-X. MR1711612
(2001d:55012)
[31] J. Hartigan, Clustering Algorithms, Wiley, New York. MR0405726 (53:9518)
[32] T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning,
Springer, New York, 2001, ISBN: 0-387-95284-5. MR1851606 (2002k:62048)
[33] A. Hatcher, Algebraic Topology, Cambridge University Press, Cambridge, 2002. xii+544
pp. ISBN: 0-521-79160-X; 0-521-79540-0. MR1867354 (2002k:55001)
[34] J. H. van Hateren and A. van der Schaaf, Independent component filters of natural images
compared with simple cells in primary visual cortex, Proc. R. Soc. Lond., vol. B 265 (1998),
359-366.
[35] J. Headd, Y.-H. A. Ban, H. Edelsbrunner, M. Vaidya and J. Rudolph. Protein-protein inter-
faces: Properties, preferences, and projections, Protein Research, to appear, 2007.
[36] D. Hubel, Eye, Brain, and Vision, Scientific American Library, W. H. Freeman, New York,
1995, viii+242pp. ISBN: 0-716-76009-6.
[37] P.J. Huber, Projection pursuit, Ann. Statistics (13), 2, (1985), pp. 435-525, with discussion.
MR790553 (88b:62118)
[38] T. Kenet, D. Bibitchkov, M. Tsodyks, A. Grinvald, and A. Arieli, Spontaneously emerging
cortical representations of visual attributes, Nature 425 (2003), 954-956.
[39] J.M. Kleinberg, An impossibility theorem for clustering, NIPS 2002: 446-453.
[40] S. Lafon and A.B. Lee, Diffusion maps and coarse-graining: A unified framework for dimen-
sionality reduction, graph partitioning, and data set parametrization, IEEE Transactions on
Pattern Analysis and Machine Intelligence 28, 9 (2006), pp. 1393-1403.
[41] A.B. Lee, K.S. Pedersen, and D. Mumford, The nonlinear statistics of high-contrast patches
in natural images, International Journal of Computer Vision (54), No. 1-3, August 2003, pp.
83-103.
[42] R.Y. Liu, J.M. Parelius and K. Singh, Multivariate analysis by data depth: Descriptive
statistics, graphics and inference (with discussion and a rejoinder by Liu and Singh), Ann.
Statist. Volume 27, Number 3 (1999), 783-858. MR1724033 (2001m:62071)
[43] U. von Luxburg, M. Belkin, and O. Bousquet, Consistency of spectral clustering, Annals of
Statistics, 36 (2), 555-586, 2008. MR2396807
[44] S. Mac Lane, Categories for the Working Mathematician, Second edition. Graduate
Texts in Mathematics, 5. Springer-Verlag, New York, 1998. xii+314 pp. ISBN: 0-387-98403-8.
MR1712872 (2001j:18001)
[45] J.P. May, Simplicial objects in algebraic topology, Reprint of the 1967 original. Chicago
Lectures in Mathematics. University of Chicago Press, Chicago, IL, 1992. viii+161 pp. ISBN:
0-226-51181-2. MR1206474 (93m:55025)
[46] P. McCullagh, What is a statistical model? With comments and a rejoinder by the author.
Ann. Statist. 30 (2002), no. 5, 1225–1310. MR1936320 (2003j:62006)
[47] E. Miller, and B. Sturmfels, Combinatorial Commutative Algebra, Graduate Texts
in Mathematics, 227. Springer-Verlag, New York, 2005. xiv+417 pp. ISBN: 0-387-22356-8.
MR2110098 (2006d:13001)
[48] R. Miller, Discussion of Projection Pursuit, Ann. Statistics 13,2, 1985, pp. 510-513. (with
discussion). MR790553 (88b:62118)
[49] J. Milnor, Morse Theory, Based on lecture notes by M. Spivak and R. Wells. Annals of
Mathematics Studies, No. 51, Princeton University Press, Princeton, N.J., 1963, vi+153 pp.
MR0163331 (29:634)
[50] D. Mumford, The dawning of the age of stochasticity, appears in Mathematics: Fron-
tiers and Perspectives, Amer. Math. Soc., Providence, RI, 2000, 197–218. MR1754778
(2001e:01004)
[51] J. Munkres, Topology: a first course, Prentice-Hall, Inc., Englewood Cliffs, N.J., 1975.
xvi+413 pp. MR0464128 (57:4063)
[52] P. Niyogi, S. Smale, and S. Weinberger, Finding the homology of submanifolds with high
confidence from random samples, Discrete and Computational Geometry, vol. 39, nos. 1-3,
(2008). MR2383768
308 GUNNAR CARLSSON
[53] G. Palla, I. Derènyi, I. Farkas, and T. Vicsek, Uncovering the overlapping community struc-
ture of complex networks in nature and society, Nature, Volume 435, 9 June 2005, pp. 814-818.
[54] M. Penrose, Random Geometric Graphs, Oxford Studies in Probability, 5. Oxford Uni-
versity Press, Oxford, 2003. xiv+330 pp. ISBN: 0-19-850626-0. MR1986198 (2005j:60003)
[55] G. Reeb, Sur les points singuliers d’une forme de Pfaff complètement intégrable ou d’une
fonction numérique, C.R. Acad. Sci. Paris 222 (1946), pp. 847-849. MR0015613 (7:446d)
[56] S.T. Roweis and L.K. Saul, Nonlinear dimensionality reduction by locally linear embedding,
Science 290 (2000) (December), pp. 2323-2326.
[57] V. de Silva, R. Ghrist, Coverage in sensor networks via persistent homology, Algebraic and
Geometric Topology, 7, 2007, pp. 339-358. MR2308949 (2008c:55008)
[58] B.W. Silverman, Density Estimation for Statistics and Data Analysis, Monographs
on Statistics and Applied Probability. Chapman & Hall, London, 1986. x+175 pp. ISBN:
0-412-24620-1. MR848134 (87k:62074)
[59] G. Singh, F. Memoli, T. Ishkhanov, G. Carlsson, G. Sapiro and D. Ringach, Topological
Structure of Population Activity in Primary Visual Cortex, Journal of Vision, Volume 8,
Number 8, Article 11, pp. 1-18, 2008.
[60] G. Singh, F. Memoli and G. Carlsson, Topological Methods for the Analysis of High Dimen-
sional Data Sets and 3D Object Recognition, Point Based Graphics 2007, Prague, September
2007.
[61] J.B. Tenenbaum, V. de Silva and J.C. Langford, A global geometric framework for nonlinear
dimensionality reduction, Science 290 (2000) (December), pp. 2319-2323.
[62] M. Tsodyks, T. Kenet, A. Grinvald, and A. Arieli, Linking spontaneous activity of single
cortical neurons and the underlying functional architecture, Science 286, (1999), pp. 1943-
1996.
[63] B. Wandell, Foundations of Vision, Sinauer Associates, Sunderland, Mass., 1995,
xvi+476pp., ISBN:0-878-93853-2.
[64] A. Zomorodian and G. Carlsson, Computing persistent homology, Discrete and Computa-
tional Geometry, 33 (2), 2005, pp. 247-274. MR2121296 (2005j:55004)
[65] A. Zomorodian and G. Carlsson, Localized homology, Computational Geometry: Theory and
Applications, 41,(3), pp. 126-148, 2008. MR2442490
Department of Mathematics, Stanford University, Stanford, California 94305
View publication stats

Topology and Data: Bulletin of The American Mathematical Society April 2009

Uploaded by

Copyright:

Available Formats

Topology and Data: Bulletin of The American Mathematical Society April 2009

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Topology and Data: Bulletin of The American Mathematical Society April 2009

Uploaded by

Copyright:

Available Formats

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

Topology and Data

Article in Bulletin of the American Mathematical Society · April 2009

Ayasdi View project

Tropical Coordinates on the Spaces of Barcodes View project

The user has requested enhancement of the downloaded file.

TOPOLOGY AND DATA

Received by the editors August 1, 2008.

• Metrics are not theoretically justiﬁed: In physics, the phenomena

• Topology is exactly that branch of mathematics which deals with qualita-

• Topology studies only properties of geometric objects which do not depend

verify that the clustering of the subsample is actually representative of a clustering

2. Persistence and homology

Example. Consider the two spaces in the image below.

• Functoriality: For any A and k as above, and any continuous map f :

Deﬁnition 2.8. Let X, L, and be as above. We construct the weak witness

It is clearly the case that if we have 0 ≤ ≤ , then we have an inclusion

2.3. Persistent homology. Let X be a subspace of Rn . Suppose further that we

a family of morphisms {fx }, with fx : cx → dx , and where the diagrams

• Select a partial order preserving map f : N → R.

computational burden. It is deﬁned as follows. Fix a positive integer k, and deﬁne

The picture below gives such an explanation.

one which ultimately turns out to ﬁt the data is pictured as follows.

Three circle model

This picture is composed of a primary circle, which surrounds two secondary

Three circle model in the data

It is easy to check that, regarded as a subspace of R3 , this subspace is an ellipse

the non-trivial β2 which would be characteristic of the Klein bottle. In order to

simulation shows that the probability of obtaining segments for β1 or β2 longer

at diﬀerent levels for comparison. Features which are seen at multiple

We note that there are natural projection maps f : M(X, U) → X and g :

Note that Č π0 (U) is actually homeomorphic to X, while Č(U) is not. This is an

map from Z to itself yields a map of coverings U[R, e] → U[R, e ] whenever e ≤ e .

Deﬁnition 3.3. Given a point cloud X, a reference map ρ : X → Z to a met-

4. Generalized forms of persistence

so that the requirement

t −t1 t2 −t2 t −tn

The following result is proved in [12].

Proposition 4.4. In the case n = 1, the rank invariant is a complete invariant of

in the right. We would like to develop a systematic methodology which assesses

dII : Z(i) eLLL fMMM 9 ···

as an estimate of the density of the distribution from which X arises by sampling.

· · ·aD Wi+1 ···

{0} {0} F F {0} {0}

for some t and a family of pairs of integers (mj , nj ), such that m ≤ mj ≤ nj ≤ n

Any ﬁnite Z-persistence F -vector space M can be decomposed as

5. Reasoning about clustering

relation generated by R . The clustering algorithm which assigns to each

commute for every morphism f in C. In this context, Kleinberg’s scale invariance

6. What should the theorems be?

About the author

[29] P. Gabriel and A. Roiter, Representations of Finite-Dimensional Algebras. Translated

Department of Mathematics, Stanford University, Stanford, California 94305

View publication stats

You might also like

relation generated by R . The clustering algorithm which assigns to each