Topological Data Analysis (TDA) for Time Series

Nalini Ravishanker∗1 and Renjie Chen†1

1
Department of Statistics, University of Connecticut
arXiv:1909.10604v1 [stat.AP] 23 Sep 2019

Abstract
The study of topology is strictly speaking, a topic in pure mathematics. However in only a
few years, Topological Data Analysis (TDA), which refers to methods of utilizing topological
features in data (such as connected components, tunnels, voids, etc.) has gained considerable
momentum. More recently, TDA is being used to understand time series. This article provides
a review of TDA for time series, with examples using R functions. Features derived from TDA
are useful in classification and clustering of time series and in detecting breaks in patterns.

Keywords Periodicity, Persistence Diagram, Persistence Landscape, Point clouds, Sublevel sets
on functions, Supervised learning, Takens’s embedding, Unsupervised learning.

1 Introduction
Topological Data Analysis (TDA) is now an emerging area for analyzing complex data. TDA refers
to a class of methods that garner information from topological structures in data that belong to a
topological space, i.e., a mathematical space that allows for continuity, connectedness, and conver-
gence (Carlsson, 2009; Edelsbrunner and Harer, 2010). Output from TDA may then be used for
effective statistical learning about the data. TDA combines algebraic topology and other tools from
pure mathematics to allow a useful study of shape of the data. The most widely discussed topologies
of data include connected components, tunnels, voids, etc., of a topological space. Computational
(or algorithmic) topology, is an overlap between the mathematical underpinnings of topology with
computer science, and consists of two parts, i.e., measuring the topology of a space and persis-
tent homology (Chazal and Michel, 2017). Using computational topology, TDA aims at analyzing
topological features of data and representing these features using low dimensional representations
(Carlsson, 2009). In particular, the space must first be represented as simplicial complexes, the
Vietoris-Rips complex and the Čech complex being the most common pathways to obtaining out-
put to characterize the topology.
Persistent homology refers to a class of methods for measuring topological features of shapes
and functions. It converts the data into simplicial complexes and describes the topological structure
of a space at different spatial resolutions. Topologies that are more persistent are detected over a
wide range of spatial scales and are deemed more likely to represent true features of the underlying
space rather than sampling variations, noise, etc. Persistent homology therefore elicits persistence of
∗
[email protected]
†
[email protected]

1
essential topologies in the data and outputs the birth and death of such topologies via a persistence
diagram, which is a popular summary statistic in TDA. Data inputs for persistent homology are
usually represented as point clouds or as functions, while the outputs depend on the nature of
the analysis and commonly consist of either a persistence diagram, or a persistence landscape. A
point cloud of data represents a sample of points from an underlying manifold and its persistent
homology approximates the topological information of the manifold. If data is represented as
a Morse function (i.e., a smooth function on the manifold such that all critical points are non-
degenerate), the persistent homology of the function is mathematically equivalent to analyzing
the topological information of the manifold. For rigorous expositions on algebraic topology and
computational homology, see Munkres (1993) and Edelsbrunner and Harer (2010).
TDA has been used in cosmic web (Van de Weygaert et al., 2011), shape analysis (Carlsson
et al., 2004; Chazal et al., 2009; Di Fabio and Landi, 2011, 2012; Chazal et al., 2014; Li et al., 2014;
Carrière et al., 2015; Bonis et al., 2016), biological data analysis (DeWoskin et al., 2010; Nicolau
et al., 2011; Heo et al., 2012; Kovacev-Nikolic et al., 2016; Bendich et al., 2016; Wang et al., 2018),
sensor networks (Silva and Ghrist, 2007; De Silva and Ghrist, 2007; Adams and Carlsson, 2015), as
well as other fields.
Development of TDA for time series is a relatively new and fast growing area, with many in-
teresting applications in several different domains. Berwald et al. (2013) discussed the use of TDA
in climate analysis. Khasawneh and Munch (2016) used notions of persistence of 1-th homology
groups of point clouds (obtained via Takens’s embedding) within multiple windows of time series
to track the stability of dynamical systems, while Seversky et al. (2016) explored stability of var-
ious single-source and multi-source signals. Perea and Harer (2015) used the notion of maximum
persistence of homology groups to quantify periodicity of time series. Pereira and de Mello (2015)
used features derived from persistent homology to cluster populations of Tribolium flour beetles.
Umeda (2017) used topological features of one and two dimensional homology groups as inputs into
convolutional neural networks for classification of time series in three different domains, showing
that their approach outperformed the baseline algorithm in each case. One illustration consisted
of motion sensor data of daily and sports activities, an area also investigated using TDA by Stolz
et al. (2017). Truong (2017) as well as Gidea (2017); Gidea and Katz (2018) and Gidea et al. (2018)
explored the use of TDA on financial time series. We discuss some of these applications in detail
later in this paper.
It is well known that time series do not naturally have point cloud representations. Transforma-
tion from a time series to a point cloud is implemented through Takens’s embedding (Takens et al.,
1981), guaranteeing the preservation of topological properties of the time series. The approach
consists of transforming a time series {xt , t = 1, 2, . . . , T }, into its phase space, i.e., a point cloud
or a set of points vi = {xi , xi+τ , . . . , xi+dτ }, i = 1, 2, . . . , T − dτ , where τ is a delay parameter and
d specifies the dimension of the point cloud. We discuss Taken’s embedding in Section 2.3 and
the selection of d and τ in Section 2.3.1. TDA of time series through suitable functions is much
less explored. Wang et al. (2018) proposed TDA on weighted Fourier series representations (Morse
functions) of electroencephalogram (EEG) data. They used a randomness test approach to examine
properties of the proposed method and show its robustness to different transformations of the data.
TDA of time series through sublevel set filtration of functions is discussed in Section 3.
The format of this paper follows. Section 2 provides a review of TDA from point clouds and
then describes TDA for time series via the Takens’s embedding method. Section 3 provides a
review of persistent homology on functions and then describes TDA for time series analysis starting
from second-order spectra or Walsh Fourier transforms. Section 4 discusses constructing TDA
based features which are then used in learning about time series, with applications on classification,

2
clustering and detecting changes in patterns. Section 5 gives a discussion and summary.

2 Persistent Homology Based on Point Clouds

In the section, we first describe persistent homology of a manifold starting from point cloud data,
followed by its construction and use in time series analysis using Takens’s embedding.

2.1 Point Clouds to Persistence Diagrams - A Basic Review

Starting from a point cloud, we show the procedure to elicit topological features of data. Denote
the point cloud as P = {vi : i = 1, 2, . . . , N }, where vi ∈ Rd . When d = 2, the points lie on the
plane. Let DE = {DE (vi , vj )} be the N × N matrix of Euclidean distances, for i, j = 1, . . . , N .
For each vi ∈ P, let Bλ (vi ) = {x : DE (x, vi ) ≤ λ/2, x ∈ Rd } denote a closed ball with radius
λ/2; here, 0 ≤ λ ≤ U , where the upper-bound U is usually pre-determined as the maximum of
the distances in DE . A Vietoris-Rips simplex (Edelsbrunner and Harer, 2010) corresponding to a
given λ is defined as the set of points PV (λ) ⊂ P such that any points vi1 , vi2 in PV (λ) satisfy
DE (vi1 , vi2 ) ≤ λ, 1 ≤ i1 , i2 ≤ N . For a given λ value, a simplicial complex κ̃(λ) denotes the set of
(1) (2)
Vietoris-Rips simplexes such that for any two Vietoris-Rips simplexes PV (λ), PV (λ) ∈ κ̃(λ), we
(1) (2) (1)
have (i) PV (λ) ∩ PV (λ) ∈ κ̃(λ) and (ii) if P 0 ⊂ PV , then P 0 ∈ κ̃(λ).
A simplicial complex consisting of (p̃ + 1) points (from different Vietoris-Rips simplexes) is a p̃-
dimensional simplicial complex. In algebraic topology, p̃ is at most N −1 when the point cloud had N
points. The topology of the point cloud is studied through the topology of the simplicial complexes,
denoted by {α̃p̃,k : k = 1, 2, . . . , kp̃ }, and α̃p̃,k is a homology group, consisting of a set of p̃-dimensional
simplicial complexes which are homomorphic. For the theory and computation of homomorphisms,
refer to Munkres (1993); Carlsson (2014) and Edelsbrunner and Harer (2010). As the parameter λ
gradually increases, the birth and death of homology groups {τ̃p̃,k = (λp̃,k,1 , λp̃,k,2 ) : k = 1, 2, . . . , kp̃ }
are recorded in the persistence diagram. A p̃-th Betti number of λ is the number of p̃-th homology
(λ)
groups at λ, denoted as kp̃ .
Computation of the topological features are summarized in the following steps.
Step 1. Compute the Euclidean distance matrix DE = {DE (vi1 , vi2 )} for i1 , i2 ∈ {1, . . . , N }; this
is the default distance for a point cloud in R-TDA.
Step 2. Construct birth and death of homology groups for increasing values of λ. For each λ,
compute α̃p̃,k from κ̃(λ) using closed balls Bλ (vi ) of vi with radius λ/2. If an elder topology
α̃p̃,k1 and a younger one α̃p̃,k2 merge into a single α̃p̃,k at some λ, α̃p̃,k1 would become α̃p̃,k and
α̃p̃,k2 would die.
Step 3. The persistence diagram is an output of the set of points representing birth-death of
homology groups from the point cloud and is denoted as Ω̃ = {τ̃p̃,k = (λp̃,k,1 , λp̃,k,2 ) : p̃ =
0, 1, . . . ; k = 1, 2, . . . , kp̃ }. We plot λp̃,k,1 on the x-axis and λp̃,k,2 on the y-axis (Edelsbrunner
and Harer, 2010).

Example 2.1. Point Cloud to Persistence Diagram. We illustrate construction of the persis-
tence diagram for a point cloud with N = 60 points sampled from the unit circle x21 + x22 = 1:
set.seed(1); PC <- circleUnif(n = 60, r = 1)
plot(PC, main = "(a)")

3
The point cloud is shown in Figure 1(a). We expect to see a total of k0 = 60 values of α̃0,k and k1 = 1
value of α̃1,k . We use the function ripsDiag from the R-TDA package for constructing the persistence
diagram (Fasy et al., 2014a). In the R code chunk shown below, PC denotes the input point cloud,
maxdimension is the maximum dimension p̃ of points τ̃p̃,k to be calculated, and maxscale is the
maximum value that the filtration parameter λ can assume. We set maxdimension to be 1. The
default dist is the Euclidean distance. The output pers.diag.1 returns the persistence diagram,
as a matrix with three columns which summarize topological features of the point cloud.

(pers.diag.1 <- ripsDiag(X=PC, maxdimension = 1, maxscale = max(dist(PC))) )

$diagram
dimension Birth Death
[1,] 0 0.0000000 1.999999902
[2,] 0 0.0000000 0.306978455
[3,] 0 0.0000000 0.245260715
.....
[60,] 0 0.0000000 0.027923092
[61,] 1 0.3190835 1.737696840

The first row with τ̃0,1 = (0, 2.00) in the output records that there is a 0-th homology group
(connected component) whose birth happens at λ = 0 and whose death happens at about λ = 2.00.
The second row with τ̃0,1 = (0, 0.31) records that the second connected component is born at λ = 0
and is dead at λ = 0.31, etc. We see that all 0-th homology groups have birth time 0, and all N = 60
points start as connected components. These 60 connected components are shown in decreasing
order of persistence (slower death). Row 61 with τ̃1,1 = (0.32, 1.74) describes the birth and death
of a 1-th homology group (tunnel) at λ = 0.32 and λ = 1.74 respectively. Figure 1(b) corresponds
to the filtration parameter λ = 0 and is obtained using this code:

plot(x=1,y=1,type="n",ylim=c(0,2),xlim=c(0,2),ylab="death",xlab="birth",main="(b)")
abline(v = 0, lty = 2)

The dashed vertical line indicates the birth time of connected components; the plot has no points
because none of the connected components has died. Figure 1(c) corresponds to the point cloud
when λ = 0.1:

plot(PC, pch = 16, cex = 5, col = "blue", main = "(c)")

The blue balls Bλ (vi ) around each point enlarge and connect with others, resulting in fewer con-
nected components. The black dots in Figure 1(d) denote the birth-death times of the merged
connected components which have died before λ = 0.1:

death.time = sort(pers.diag.1$diagram[pers.diag.1$diagram[, 1]==0, 3])

plot(x = rep(0, sum(death.time<=0.1)), y = death.time[which(death.time<=0.1)],
ylim = c(0, 2), xlim = c(0, 2), ylab = "death", xlab = "birth", main = "(d)")
abline(v = 0, lty = 2); abline(h = 0.1, lty = 2)

When λ = 0.32 in Figure 1(e), all points connect together and a tunnel emerges, which is the white
area surrounded by the blue circle:

plot(PC, pch = 16, cex = 12, col = "blue", main = "(e)")

4
 1.0
(a) (b) (c) (d)

1.0
2.0

2.0
1.5

1.5
0.5

0.5
death

death
1.0

1.0
0.0

0.0
x2

x2
-0.5

-0.5
0.5

0.5
-1.0

-1.0
0.0

0.0
-1.0 -0.5 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 -1.0 -0.5 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0

x1 birth x1 birth

(e) (f) (g) (h)

1.0

1.0
2.0

2.0
1.5

1.5
0.5

0.5

Death
death

1.0

1.0
0.0

0.0
x2

x2
-0.5

-0.5
0.5

0.5
-1.0

-1.0
0.0

0.0
-1.0 -0.5 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0 -1.0 -0.5 0.0 0.5 1.0 0.0 0.5 1.0 1.5 2.0
Birth
x1 birth x1

Figure 1: Persistence diagram corresponding to a point cloud. (a) shows the raw point cloud and
(h) shows the persistence diagram. (c), (e) and (g) are intermediate steps for the filtration by
varying λ, while (b), (d) and (f) are intermediate steps for constructing the persistence diagram.

The birth time of this tunnel is recorded as λ = 0.32, which is shown as the red dashed line
in Figure 1(f). Further, there are more black dots in this figure since there are more connected
components that have died before λ = 0.32:
plot(x = rep(0, sum(death.time<=0.32)), y = death.time[which(death.time<=0.32)],
ylim = c(0, 2), xlim = c(0, 2), ylab = "death", xlab = "birth", main = "(f)")
abline(h = pers.diag.1$diagram[pers.diag.1$diagram[, 1]==1, 2], lty = 2)
abline(v = 0, lty = 2); abline(v = 0.32, col = "red", lty = 1)
When λ reaches its maximum value of 2 (which is the largest value in DE (vi1 , vi2 )), the algorithm
stops and outputs the persistence diagram (see Figure 1(h)) which finally shows the birth-death
times for all connected components (the dots) and the tunnel (the red triangle):
plot(PC, pch = 16, cex = 40, col = "blue", main = "(g)")
plot(pers.diag.1$diagram, main = "(h)")

R-TDA also supports construction of a persistence diagram given an arbitrary distance matrix as
input, as shown in the example below.

Example 2.2. Distance Matrix to Persistence Diagram. The input to ripsDiag can be a
distance matrix computed from the point cloud generated in Example 2.1. Here, we use the default
dist="euclidean". Other options are "manhattan", "maximum", etc.

dist.PC <- dist(PC)

(pers.diag.2=ripsDiag(X=dist.PC,dist="arbitrary",maxdimension=1,maxscale=max(dist.PC)))
$diagram

5
 dimension Birth Death
[1,] 0 0.0000000 1.999999902
[2,] 0 0.0000000 0.306978455
[3,] 0 0.0000000 0.245260715
.....
[60,] 0 0.0000000 0.027923092
[61,] 1 0.3190835 1.737696840

2.2 Distances Between Persistence Diagrams

Two distance metrics are commonly used to quantify the dissimilarity between two persistence
diagrams Ω̃1 and Ω̃2 , the Wasserstein distance and the bottleneck distance (Mileyko et al., 2011).
We define these distances and describe their computation using the R-TDA package.
The q-Wasserstein distance between two persistence diagrams is defined by
X 1/q
|τ̃p̃,k − η(τ̃p̃,k )|q∞

Wq,p̃ (Ω̃1 , Ω̃2 ) = inf , q = 1, 2, . . . , (1)
η:Ω̃1 →Ω̃2
τ̃p̃,k ∈Ω̃1

where p̃ is referred to as its dimension and q is its degree. When q = ∞, (1) is the bottleneck
distance of dimension p̃ defined by

W∞,p̃ (Ω̃1 , Ω̃2 ) = inf sup |τ̃p̃,k − η(τ̃p̃,k )|∞ . (2)

η:Ω̃1 →Ω̃2 τ̃p̃,k ∈Ω̃1

The bottleneck distance is obtained by minimizing the largest distance of any two corresponding
points of diagrams, over all bijections between Ω̃1 and Ω̃2 and is less sensitive to details in the
diagrams.

Example 2.3. Wasserstein and Bottleneck Distances. Let Ω̃1 be pers.diag.2, the per-
sistence diagram obtained in Example 2.2. Let Ω̃2 be pers.diag.3, the persistent diagram we
construct from a different point cloud from the same unit circle. We use R-TDA to compute the
Wasserstein distance with q = 1 (denoted by the argument p=1 below):

set.seed(2); PC2 <- circleUnif(n = 60, r = 1)

pers.diag.3 <- ripsDiag(X=PC2, maxdimension = 1,maxscale = max(dist(X))$diagram
wasserstein(pers.diag.2, pers.diag.3, p=1, dimension = 0)
1.034579

The function bottleneck enables us to compute the bottleneck distance between the two per-
sistence diagrams. The Wasserstein distance is larger than the bottleneck distance since the former
measures more detailed difference between the diagrams.

bottleneck(pers.diag.2, pers.diag.3, dimension = 0)

0.06954618

It is important to construct persistence diagrams using the same distance functions (Chazal
et al., 2018), as we show below. For instance, we can construct a persistence diagram pers.diag.4
for the point cloud in Example 2.1 using the Manhattan distance (DM (vi , vj ) = d`=1 |vi,` −vj,` |, 1 ≤
P
i, j ≤ N ) instead of the Euclidean distance.

6
dist.PC.man <- dist(PC, method = "manhattan"); max.dist=max(dist.PC.man)
pers.diag.4=ripsDiag(X=dist.PC.man,dist="arbitrary",maxdimension=1,
maxscale=max.dist)$diagram
wasserstein(pers.diag.2, pers.diag.4, p=1, dimension = 0)
2.145083
bottleneck(pers.diag.2, pers.diag.4, dimension = 0)
0.8279462

2.3 TDA of Time Series via Point Clouds

Time series do not naturally have point cloud representations, and are transformed to point clouds
using Takens’s Embedding Theorem (Takens et al., 1981) before we can do TDA as discussed in
Section 2.1. This approach has been used in the literature mostly for quantifying periodicity in
time series (Perea and Harer, 2015), clustering time series (Seversky et al., 2016), classifying time
series (Umeda, 2017), or finding early signals for critical transitions (Gidea, 2017; Gidea and Katz,
2018). Takens’s embedding guarantees the preservation of topological properties of a time series
but not its geometrical properties.

2.3.1 Takens’s Delay Embedding for Time Series

Let {xt , t = 1, 2, . . . , T } denote an observed time series. We use Takens’s embedding to convert
the time series into a point cloud with points vi = (xi , xi+τ , . . . , xi+(d−1)τ )0 , where d specifies the
dimension of the points and τ denotes a delay parameter. For example, if d = 2 and τ = 1, then,
vi = (xi , xi+1 )0 , whereas if d = 15 and τ = 2, vi = (xi , xi+2 , . . . , xi+28 )0 . Both d and τ are unknown
and must be determined in practice.

Choice of τ . Researchers have used different approaches for choosing the delay parameter τ . It
may be selected as the smallest time lag h where the sample autocorrelation function (ACF) ρ̂h
becomes insignificant, i.e., smaller in absolute value than the critical bound √2T (Khasawneh and
Munch, 2016). Truong (2017) also used the ACF, but in a slightly different way. He chose τ as the
smallest lag for which (ρ̂τ − ρ̂τ −1 )/ρ̂τ > 1/e and ρ̂τ < √2T . Pereira and de Mello (2015) determined
τ using the first minimum of the auto mutual information (the mutual information between the
signal and its time delayed version).

Choice of d. Truong (2017) and Khasawneh and Munch (2016) used the false nearest neighbor
method (Kennel et al., 1992) to determine the embedding dimension d as the integer such that the
nearest neighbors of each point in dimension d remain nearest neighbors in dimension d + 1, and
the distances between them also remain about the same. Alternately, an R function false.nearest
in the package tseriesChaos which implements an approach due to Hegger et al. (1999) may be
used. Some authors (Pereira and de Mello, 2015; Seversky et al., 2016), simply assume d to be 2 or
3, while Perea has suggested the use of d = 15 on time series after a cubic spline interpolation (see
Section 2.3.2).

The choice of d and τ then determine the number of points N in the point cloud. In Example
2.4, we illustrate one approach for constructing a point cloud from pure periodic signals with no
noise and then obtaining a persistence diagram. In Example 2.5, we discuss another approach
described in Perea et al. (2015) for noisy time series, when the focus is on finding series with the
same periodicity.

7
Example 2.4. Pure Signals to Persistence Diagrams. We generate point clouds from three
periodic cosine signals of length T = 480 with periods 12, 48, and 96 respectively, and then construct
their persistent diagrams. We set d = 2, and use the ACF method discussed above to choose τ . We
show R code for the time series ts1:
per1=12;ts1 = cos(1:T*2*pi/per1);d=2;
tau <- which(abs(acf(ts.ex, plot = F)$acf) < 2/sqrt(T))[1]-1
PC=t(purrr::map_dfc(1:(T-(d-1)*tau+1),~ts.ex[seq(from=.x, by=tau, length.out=d)]))
diag=ripsDiag(PC, maxdimension=1, maxscale=max(dist(PC)))
ts.plot(ts.ex);plot(PC,xlab ="x1",ylab="x2",main="PC");plot(diag$diagram)
In Figure 2, the top row shows the signals, the middle row shows the point clouds and the bottom
row shows the persistence diagrams. The black dots represent the birth-death of 0-th homology
groups and their persistence shows the dispersion of the points in the point cloud. When there are
more black dots close to the diagonal, the point cloud is more dispersed. Particularly, the point
cloud PC.3 from the time series with period 96 have points close to each other compared with PC.1,
so that it has more black dots in the persistence diagram closer to the diagonal.
The red triangles represent the birth-death of 1-th homology groups, indicating circles in the
point cloud. The red triangle from the time series with period 96 is further away from the diagonal
compared to the series with period 12, and thus has longer persistence in the 1-th homology group.
Seeing a circle indicates that the time series is periodic. This is in contrast to the persistence
diagram for the same time series based on sublevel set flitration of a function, as discussed in
Example 3.2.
1.0

1.0

1.0
0.0

0.0

0.0
ts1

ts2

ts3
-1.0

-1.0

-1.0

0 100 200 300 400 0 100 200 300 400 0 100 200 300 400

Time Time Time

PC.1 PC.2 PC.3

1.0

1.0

1.0
0.0

0.0

0.0
x2

x2

x2
-1.0

-1.0

-1.0

-1.0 -0.5 0.0 0.5 1.0 -1.0 -0.5 0.0 0.5 1.0 -1.0 -0.5 0.0 0.5 1.0

x1 x1 x1
2.0

2.0
2.0
Death

Death

Death
1.0

1.0
1.0 0.0

0.0

0.0

0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0 0.0 0.5 1.0 1.5 2.0
Birth Birth Birth

Figure 2: Pure Periodic Signals to Persistence Diagrams.

Pairwise bottleneck distances between the three persistence diagrams are shown in Table 1,
computed using code as shown below:
round(bottleneck(diag1$diagram, diag2$diagram, dimension = 0), digits = 2)
To study the effect of d, we repeat the computations for d = 3 and d = 15 and also show all pairwise
bottleneck distances in the table. While the values of the distances between the diagrams change as

8
d changes, the relative behavior is preserved, independent of d. Specifically, the bottleneck distances
between Ω̃1 and Ω̃2 and between Ω̃1 and Ω̃3 are larger than the distance between Ω̃2 and Ω̃3 in the
0-th and 1-th homology groups.

Table 1: Pairwise Bottleneck Distances Between Persistence Diagrams for Different d.

d=2 d=3 d = 15
(1,2) (1,3) (2,3) (1,2) (1,3) (2,3) (1,2) (1,3) (2,3)
0th 0.26 0.26 0.07 0.36 0.36 0.09 0.73 0.73 0.18
1th 0.39 0.45 0.06 0.53 0.63 0.09 1.09 1.28 0.19

2.3.2 Point Cloud Construction using SW1PerS

The SW1PerS (Sliding Windows and 1-Persistence Scoring) method is an alternate, more compre-
hensive approach proposed by Perea et al. (2015) to detect periodicity from noisy time courses
whose underlying signals may have different shapes. The approach addresses the following items.

Denoising. The approach considers two types of denoising that are left as options to the user.
The first type smooths the raw time series by a moving average in order to make it easier to detect
the signal. The second type is a moving average on the point cloud. As an alternative to moving
averaging, Pereira and de Mello (2015) used the Empirical Mode Decomposition (EMD) (Huang
et al., 1998) on the raw time series.

Spline Interpolation. The spline interpolation allows handling unevenly spaced time series, or
time series with low temporal resolution.

Point Cloud Standardization. Standardization helps with signal dampening and to make the
procedure amplitude blind.

The pipeline for this approach is described in the following steps.

Step 0. Optionally (Perea and Harer, 2015), denoise the observed time series {xt , t = 1, 2, . . . , T }
using a simple moving average whose window size is no higher than one third of the selected
dimension d. They recommended an embedding dimension of d = 15 and N = 201 as the size
of the point cloud; then T1 = N + d = 216.
Step 1. For selected values of d and τ (see below), create a point cloud from the (possibly denoised)
time series using Steps 1.1 and 1.2.
Step 1.1. Recover a continuous function g : [0, 2π] → R by fitting a cubic spline to the denoised
time series {xt , t = 1, 2, . . . , T }.
Step 1.2. Using values g(t1 ), g(t2 ), . . . , g(tT1 ) from the continuous spline fit g(.) at evenly spaced
time points 0 = t1 ≤ t2 ≤ . . . ≤ tT1 = (T1 − 1)τ = 2π, construct a point cloud with
(0)
N = T1 −d points vt = (g(t), g(t+τ ), . . . , g(t+(d−1)τ ))0 ∈ Rd , t = 0, τ, . . . , 2π−(d−1)τ
2π
and so τ = N +d−1 .
Step 2. Pointwise Point cloud standardization:
v
(0) (0) d u d
vt − v̄t 1 (0)
X (0) (0) (0)
uX (0) (0)
vt = (0) (0)
; v̄t = vt,i /d, ||vt − v̄t 1|| = t (vt,i − v̄t )2 , (3)
||vt − v̄t 1|| i=1 i=1

9
 (0) (0) (0) (0)
where vt = (vt,1 , vt,2 , . . . , vt,d )0 and 1 is the d-dimensional vector of 1’s.

Step 3. Construct the persistence diagram from the point cloud as shown in Section 2.1.

This method is powerful for detecting periodicity in time series. To develop a score for quanti-
fying the periodicity, Perea et al. (2015) first found the longest persistence of the birth-death of the
1-th homology groups (λ1,kM ,1 , λ1,kM ,2 ), where kM = arg maxk (λ1,k,2 − λ1,k,1 ) is chosen to indicate
maximum persistence, and used it to compute

λ21,kM ,2 − λ21,kM ,1
S =1− .
3
√
Since 0 ≤ λ1,kM ,1 ≤ λ1,kM ,2 ≤ 3, for periodic (nonperiodic) time series, the score is close to zero
(one). The R code for implementing Step 1-Step 3 for Case 1 is shown below (the code for other
cases is similar):

x.ts = ts1; d=15; N=201; T1 = 216;

x.ts <- pracma::movavg(x.ts, 5, type = "s") #step 0
sp.ts <- stats::spline(1:T*2*pi/T, x.ts, n=T1)$y #step 1.1
PC <- plyr::ldply(map(1:N, ~sp.ts[.x:(.x+d-1)]))#step 1.2
X.PC=t(apply(PC,1,FUN=function(x){(x-mean(x))/sqrt(sum((x-mean(x))^2))})) #step2
diag <- ripsDiag(X=x.PC, maxdimension = 1, maxscale = sqrt(3))#step 3

The main contributions of Perea et al. (2015) are the use of extensive simulation studies to show
that topological features of time series are largely the same under various non-sinusoidal shapes as
well as under differences in amplitude, phase, mean, frequency, or trend. The results may be affected
by a differences in noise variances, as well as by the shapes of the noise and signal distributions.

Example 2.5. Using SW1PerS for Periodicity Quantification. We generate white noise t
with variance σ2 = 0.64, and then generate periodic time series signals each of length T (= 480),
denoted by ts1, ts2, ts3, ts4, and shown in the top row of Figure 3.

• Case 1: xt = cos(2πt/12) + t ,

• Case 2: xt = 0.05t + 10(cos(2π(t − ϕ)/48) + t ),

• Case 3: xt = 10(cos(2πt/12) + t ) exp (−0.01t),

• Case 4: xt = t

The series ts2 differs from ts1 in frequency, phase, and linear trend, whereas ts3 only differs
from ts1 in shape and amplitude; ts4 is the white noise series. Figure 3 shows that the method in
Perea et al. (2015) is insensitive to different periodicities. The top row shows the simulated time
series under the four cases. The middle row presents a two dimensional view of the point clouds
constructed with d = 15 and τ = 1, via the first two principal components. The bottom row shows
the persistence diagrams, along with the periodicity score S. For periodic (nonperiodic) time series,
the score is close to zero (one), so that the scores for Case 1 to Case 3 are close to 0, while Case 4
has a score close to 1.

10
 Case 1 Case 2 Case 3 Case 4

40

15
2

0.5
10
0

-1.5
-2

-20

-15
0 300 0 300 0 300 0 300

Time Time Time Time

1.0

1.0

1.0

1.0
PC2

PC2

PC2

PC2
-1.0 0.0

-1.0 0.0

-1.0 0.0

-1.0 0.0
-1.0 0.5 -1.0 0.5 -1.0 0.5 -1.0 0.5

PC1 PC1 PC1 PC1

S=0.219 S=0.157 S=0.191 S=0.85

Death

Death

Death

Death
1.0

1.0

1.0

1.0
0.0

0.0

0.0

0.0

0.0 1.0 0.0 1.0 0.0 1.0 0.0 1.0

Birth Birth Birth Birth

Figure 3: Persistence Diagrams of Periodic Time Series with Different Shapes

3 Persistent Homology Based on Functions

We first give a basic review of TDA based on functions, followed by the use of frequency domain
representations of time series as starting points for TDA.

3.1 Function to Persistence Diagram - A Basic Review

When data is in the form of a continuous function f : Rd → R, or can be converted to such a
function, TDA using sublevel set filtration is carried out by discretizing the function into grids and
then implementing computational homology on the discretized function. Suppose the components

11
of the function are z = (`1 δ, `2 δ, . . . , `d δ), for `1 , `2 , . . . , `d = 0, ±1, ±2, . . ., where d > 0 and δ > 0.
The sublevel set of the function is defined as

Lλ (f ) = {z : f (z) ≤ λ, z ∈ Rd }, (4)

where 0 ≤ λ ≤ maxz f (z). Define a simplex as a set of components in Lλ (f ) which are “neighbors”,
i.e., z1 , z2 ∈ Lλ (f ), and |z1,j − z2,j | ≤ δ, j = 1, 2, . . . , d.
Recall from Section 2.1 that a simplex with (p̃ + 1) components is called a p̃-simplex. Since only
adjacent points on the grid can be neighbors, the function f (z) can admit at most a d-simplex, so
that p̃ ≤ d − 1 (Edelsbrunner and Harer, 2010). When d = 1, there are only connected components,
whose births and deaths are given by τ̃0,k = (λ0,k,1 , λ0,k,2 ), k = 1, 2, . . . , k0 . The computations in the
steps that are summarized below are done using the R function gridDiag (Fasy et al., 2014a).

Step 1. Assume a filtration parameter starting at λ = minz f (z) and let Lλ (f ) = {z : f (z) =
minz f (z)}.

Step 2. Construct topological features for increasing values of λ. For each λ, simplicial complexes
can be constructed from the sublevel set Lλ (f ), and α̃p̃,k can be computed using computation
homology, where 0 ≤ p̃ ≤ d − 1. If an elder topology α̃p̃,k1 and a younger topology α̃p̃,k2 merge
into a single α̃p̃,k at some λ value, then α̃p̃,k1 becomes α̃p̃,k and α̃p̃,k2 dies at λ, using the Elder
Rule (Edelsbrunner and Harer, 2010).

Step 3. The persistence diagram is the output of the set of points representing birth-death of
homology groups {τ̃p̃,k = (λp̃,k,1 , λp̃,k,2 ) : p̃ = 0, 1, . . . ; k = 1, 2, . . . , kp̃ }.

Example 3.1. Discretized Function to Persistence Diagram. We present an example of

using TDA on a one-dimensional discretized real function generated using the code chunk below,
where funval contains values of the function taken over a grid:

funval = c(1, 0.5, 1, 1.5, 0.5, 0, 1, 1, 0.5, 1)

(pers.diag.4 <- gridDiag(FUNvalues = funval, sublevel = TRUE) )
$diagram
dimension Birth Death
[1,] 0 0.0 1.5
[2,] 0 0.5 1.5
[3,] 0 0.5 1.0

The persistence diagram contains three connected components born at λ = 0 and λ = 0.5, corre-
sponding to the function having three local minima at these two distinct λ values. In Figure 4(a),
a connected component emerges when λ = 0 and is marked as a blue dot (it is the earliest/oldest
connected component).

plot(funval, x=1:10, type = "l",yaxt=’n’,ylim = c(0,2), cex.axis=1.4, xlab = "z",

ylab= "y", cex.lab= 1.3, cex=1.2, lwd=1.5, lty=1, pch=1, bty=’n’, main="(a)")
points(x=6, y=0, pch=16, col="blue", type = "p")
ticks<-c(0, 0.5, 1, 1.5, 2); axis(2,at=ticks,labels=ticks)
abline(a=0, b=0, lty=2, lwd=1,pch=1)

12
 The vertical dashed line in Figure 4(b) corresponds to the birth time λ = 0, while the horizontal
dashed line tracks the current filtration parameter λ. There is no point on the birth/death plot yet,
as no connected components have died at λ = 0.

plot(rep(0,11),x=0:10/5,type = "l",cex.axis=1.4,xaxt=’n’,xlab= "birth",yaxt=’n’,

ylim=c(0,2),ylab="death",cex.lab=1.3,cex=1.2,lwd=1.5,lty=2,pch=1,bty=’n’,main="(b)")
axis(1,at=ticks,labels=ticks); axis(2,at=ticks,labels=ticks); abline(v=0, lty=2)

Figure 4(c) corresponds to λ = 0.5. There are two more connected components indicated by the
blue dots at (2, 0.5) and (9, 0.5). The blue dot (6, 0) in the middle with a blue line connecting it to
the dot (5, 0.5) indicates that the oldest connected component enlarges.

plot(funval, x=1:10, type = "l",yaxt=’n’,ylim = c(0,2),cex.axis=1.4,xlab = "z",

ylab = "y",cex.lab= 1.3, cex=1.2,lwd=1.5, lty=1, pch=1, bty=’n’, main = "(c)")
axis(2,at=ticks,labels=ticks); abline(a=0.5, b=0, lty=2, lwd=1,pch=1)
points(x=6, y=0, pch=16, col="green4", type = "p",xlab = "z", ylab= "y")
points(x=c(2,5,9), y=rep(0.5,3), pch=16, col="blue", type = "p",xlab="z",ylab="y")
segments(x0=6, y0=0, x1=5, y1=0.5, lty = 1, pch=1, lwd=2.5, col = "blue")

Figure 4(d) has one more vertical dashed line, which gives the birth time for the other two new
connected components. There is no connected component dead yet, and so no points are shown on
the second plot either. The code chunks for plotting these are similar and are not shown due to
space limitations. When λ = 1, in Figure 4(e), all components enlarge and one newer component
is killed by the elder one because they are merged. There is a black dot at (0.5, 1) in Figure 4(f),
which indicates the newer connected component that is born at λ = 0.5 and is dead at λ = 1. When
λ = 1.5 reaching the maximum of the function in Figure 4(g), the last component is killed. The
black dot at the location (0, 1.5) in Figure 4(h) is for the last component. The other black dots
corresponding to (0.5, 1.5) and (0.5, 1) show the birth and death of other connected components.

Example 3.2. Morse Function to Persistence Diagram. An alternate technique for persis-
tent homology which is robust to noisy point cloud data uses the R function gridDiag (Chazal
et al., 2018). As mentioned in the introduction, a point cloud is assumed as a sample from an
underlying manifold. To learn the topology of the manifold, gridDiag enables us to construct a
Morse function such as the distance-to-measure (DTM) function from the point cloud using the
sublevel set filtration. Suppose the point cloud is {xi ∈ Rd , i = 1, 2, . . . , N }. Represent the DTM
function from Rd → R as s X
(DT M )
f (x) = ||xi − x||2 /k,
xi ∈Nk (x)

where k = [mN ] (m is the parameter m0) and Nk (x) is the set containing the k nearest neighbors
of x in the point cloud.
A higher value of f (DT M ) (x) means that x is further away from most of the points. DTM is
also robust to outliers (Chazal et al., 2018). We illustrate using the same point cloud data from
Example 2.1:

m0=0.05; by <- 0.065; Xlim <- range(PC[,1]); Ylim <- range(PC[,2])

(pers.diag.5 = gridDiag(X=PC, FUN=dtm, lim=cbind(Xlim, Ylim), by=by, m0=m0) )
$diagram
dimension Birth Death

13
 2
(a) (b) (c) (d)

2
1.5

1.5

1.5

1.5
death

death
y

y
1

1
0.5

0.5

0.5

0.5
0

0
2 4 6 8 10 0 0.5 1 1.5 2 2 4 6 8 10 0 0.5 1 1.5 2

z birth z birth

(e) (f) (g) (h)

2

2
1.5

1.5

1.5

1.5
death

death
y

y
1

1
0.5

0.5

0.5

0.5
0

0
2 4 6 8 10 0 0.5 1 1.5 2 2 4 6 8 10 0 0.5 1 1.5 2

z birth z birth

Figure 4: Construction of a persistence diagram corresponding to a one-dimensional continuous

real function. (a) is the function and (h) is the persistence diagram. (c), (e) and (g) show the
sublevel set filtration procedure, while (b), (d) and (f) are the intermediate steps for constructing
the persistence diagram.

[1,] 0 0.02414385 0.96912324

[2,] 0 0.02623549 0.15885911
[3,] 0 0.03489488 0.15662338
...
[20,] 0 0.14527336 0.15092905
[21,] 1 0.20234735 0.96912324

Here, lim specifies the range of the point cloud in different dimensions, by is the step size for
increasing λ, and m0 (which lies in (0, 1) with 0.05 as the default value) is the smoothing parameter
of the DTM method. The code for constructing Figure 5 is shown below:

par(mfrow = c(1,2))
Xseq <- seq(from = Xlim[1], to = Xlim[2], by = by)
Yseq <- seq(from = Ylim[1], to = Ylim[2], by = by)
Grid <- expand.grid(Xseq, Yseq); DTM = dtm(X = PC, Grid = Grid, m0 = m0)
persp(x = Xseq, y = Yseq,z = matrix(DTM, nrow = length(Xseq), ncol = length(Yseq)),
xlab = "", ylab = "", zlab = "", theta = -20, phi = 35, scale = FALSE,
expand = 2, col = "red", border = NA, ltheta = 50, shade = 0.5,main = "(a)")
plot(pers.diag.5[["diagram"]], main = "(b)")

Figure 5(a) shows the DTM function of the point cloud. The function has a peak in the middle
of the plot, surrounded by a rough circle of local minima (since the original point cloud is from the
unit circle and DTM is a smoothed distance function) Figure 5(b) shows the persistence diagram.
The dots in the persistence diagram are the birth-death points of 0-th homology groups and the

14
 (a) (b)

1.0
0.8
0.6
Death
0.4
0.2
0.0
0.0 0.2 0.4 0.6 0.8 1.0
Birth

Figure 5: Distance-to-Measure (DTM) function and the persistence diagram.

triangle denotes the birth-death point of 1-th homology group. There are only 20 birth-death points
of 0-th homology groups instead of 60 as we saw in Example 2.1, because the DTM smooths the
distance function, as mentioned earlier, resulting in a different output from those in Examples 2.1
and 2.2. The point at (0.2, 0.97) indicates a big circle in the data (representing the topology of the
underlying unit circle).

3.2 TDA of Time Series via Frequency Domain Functions

Section 3.2.1 discusses TDA starting from variations of the Fourier transform for continuous-valued
time series, while Section 3.2.2 shows how to build persistence diagrams based on Walsh Fourier
transforms for categorical time series.

3.2.1 Discrete Fourier Transforms to Persistence Diagrams

In this section, we look at topological properties of time series through their frequency domain
representations such as second-order spectra. We construct a persistence diagram using sublevel
set filtration on the smoothed tapered estimate of the second-order spectrum of the time series
{xt , t = 1, . . . , T }. The modified DFT (Stoffer, 1991) and corresponding periodogram with tapering
are defined as
T
X
−1/2
dh (ωj ) = T ht xt e−2πiωj t and
t=1
2
Ih (ωj ) = |dh (ωj )| ,

for t = 1, 2, . . . , T , where ht is a taper function. In Example 3.3, we show the use of the R function
gridDiag to construct the persistence diagram starting from a smoothed version of Ih (ωj ), using
the Daniel window for smoothing.

15
Example 3.3. Smoothed Tapered Second-order Spectrum to Persistence Diagram. We
use the R function gridDiag to construct the persistence diagrams for the same three periodic time
series signals shown in Example 2.4. For Case 1, spc.t1 denotes the smoothed tapered periodogram
(taper= 0.1 as default in the R function spec.pgram, and smoothing via the modified Daniel window
(0.25, 0.5, 0.25)) of the time series ts1.
x.1 = 1:length(ts1); ts1 = lm(ts1~x.1)$residuals; ts1 = ts1/sd(ts1) #Step 1
spc.t1=spec.pgram(ts1, kernel=kernel("modified.daniell", c(1)), plot = F)#Step 2
#Step 3
PD.t1=gridDiag(FUNvalues=spc.t1$spec,location = FALSE, sublevel=TRUE)$diagram

Case 1 Case 2 Case 3

0.0 1.0

0.0 1.0

0.0 1.0
ts1

ts2

ts3
-1.5

-1.5

-1.5
0 100 200 300 400 0 100 200 300 400 0 100 200 300 400

Time Time Time

80

80

80
spec

spec

spec
40

40

40
0

0
0.0 0.1 0.2 0.3 0.4 0.5 0.0 0.1 0.2 0.3 0.4 0.5 0.0 0.1 0.2 0.3 0.4 0.5

freq freq freq

40 80

40 80

40 80
Death

Death

Death
0

0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100

Birth Birth Birth

Figure 6: Persistence diagrams using second-order spectrum.

In Figure 6, the top row shows the time series signals, the middle row shows the second-order
spectra and the bottom row shows the persistence diagrams. The peaks in the spectra occur at
different frequencies and correspond to the contributions at these frequencies to the total variance
of xt . The three persistence diagrams in the bottom row are similar to each other, since this method
is insensitive to differences in the periodicity of the time series.
In contrast with the persistence diagrams in Figure 3 for the same signals, we no longer see
red triangles for 1-th homology groups, indicating that there is a difference between constructing
persistence diagrams from point clouds versus the spectrum.
We compute the bottleneck distances between the three persistence diagrams. The distance
between Case 1 and Case 2 is 0.01, which is much smaller than the distance between Case 1 and
Case 3 which is 2.77, or the one between Case 2 and Case 3 which is 2.76. The code for computing
the distances is similar to the one shown in Example 2.4.

3.2.2 Walsh-Fourier Transforms to Persistence Diagrams

Stoffer (1991) suggested that Walsh spectral analysis is suited to the analysis of discrete-valued
and categorical-valued time series, and of time series that contain sharp discontinuities. The fast

16
Walsh-Fourier Transform construction uses the method of Shanks (1969) to decompose a time series
{xt , t = 1, . . . , T } into a sequence of Walsh functions, each representing a distinctive binary sequency
pattern. If the time series length T is not a power of 2, let T2 denote the next power of 2. For
example, if T = 1440, then T2 = 211 = 2048. We use zero-padding to obtain a time series of length
T2 by setting set xT +1 , xT +2 , . . . , xT2 = 0.
For j = 0, . . . , T2 − 1, let λj = j/T2 denote the jth sequency. Let W (t, j) denote the t-th Walsh
function value in sequency λj . Walsh functions are iteratively generated as follows (Shanks, 1969):

W (0, j) = 1, j = 0, 1, . . . , T2 − 1,
(
1 j = 0, 1, . . . , (T2 )/2 − 1
W (1, j) =
−1 j = (T2 )/2, (T2 )/2 + 1, . . . , T2 − 1
W (t, j) = W ([t/2], 2j) × W (t − 2[t/2], j), (5)
t = 2, . . . , T2 − 1, j = 0, 1, . . . , T2 − 1,

where [a] denotes the integer part of a. For more details on Walsh functions, refer to Stoffer (1991).
The Walsh-Fourier Transform (WFT) of the time series is computed as
T2
1 X
dT (λj ) = √ xt W (t, j)), 0 ≤ j ≤ T2 − 1. (6)
T2 t=1

The computational complexity is O(T log(T )) (Shanks, 1969). In Example 3.4, we illustrate the
construction of a persistence diagram for a categorical time series with two levels.

Time series WFT PD

0.8
0.6
1

0.6
0.4

0.4
0.2

death

0.2
0.0

0.0
-0.2

-0.4
-0.4
0

0 20 40 60 80 100 120 0 20 40 60 80 100 -0.4 0.0 0.2 0.4 0.6 0.8

time birth

Figure 7: Persistence diagrams using Walsh-Fourier Transforms.

Example 3.4. Walsh-Fourier Transform to Persistence Diagram. In Figure 7, a simulated

categorical time series of length T = 120 with two levels, 0 or 1, is shown in the first column. Level
1 only occurs in the period between t = 21 and t = 100. The middle column shows the WFT of

17
the time series, while the third column shows its persistence diagram. There is one point in the
diagram away from the diagonal line, which is a significant birth-death point of the 0-th homology
groups.
The R code for simulating the time series and converting the WFT into a persistence diagram
is shown below.
x.ts = c(rep(0, 20), rep(1, 80), rep(0, 20))
# create WFT using C++ code
x.diag=gridDiag(FUNvalues = x.WFTs, location = FALSE, sublevel = TRUE)$diagram

4 Feature Construction Using TDA

Unlike a vector space, the space of persistence diagrams is not easy to work with. For instance,
a set of persistence diagrams may not have a unique mean (Mileyko et al., 2011). The bottleneck
or Wasserstein distances are also more complicated than the Euclidean distance in practice. This
section discusses an alternative.

4.1 Persistence Landscapes - A Basic Review

Bubenik (2015) introduced persistence landscapes as useful statistical summaries which build topo-
logical features and are easy to combine with tools from statistics and machine learning. This
section reviews persistence landscapes while Section 4.2 describes their construction for time series,
with examples.
The νth order persistence landscape of p̃-th homology groups is defined as

PLp̃,ν (`) = {min(` − λp̃,k,1 , λp̃,k,2 − `)+ : k = 1, 2, . . .}(ν) (7)

where λp̃,k,1 and λp̃,k,2 were introduced under Step 3 of Section 2.1, ` ∈ R, min(a, b)+ denotes
the smaller value if both a and b are positive, or zero if neither value is positive, and {A}(ν) is
the ν-th order statistic of the set A. Bubenik (2015) proved that a set of persistence landscapes
admits a unique mean and preserves statistical stability of the data distribution. If we assume that
the observed data X is a random draw from an underlying space X equipped with a probability
measure, and assume multiple copies X1 , X2 , . . . , XN , then the mean of their persistence landscapes
(N ) (i)
PLp̃,ν (`) = N
P
i=1 PLp̃,ν (`)/N would converge almost surely (as N → ∞) to the expectation of the
persistence landscapes E(PLX p̃,ν (`)) if and only if
Z
E||PLp̃,ν (`)|| = |PLX
X
p̃,ν (`)|d` < ∞,
`

i.e., ` |PLX
R
p̃,1 (`)|d` < ∞ for all p̃. This means that the set of persistence landscapes from the
observed data is a good representative for the underlying distribution of true persistence landscapes
and it is possible to do statistical inference for these using the sample persistence landscapes.
Another important statistical property is the stability in terms of using persistence landscapes versus
using persistence diagrams. Suppose there are two persistence diagrams Ω̃1 and Ω̃2 together with
(1) (2) (1) (2)
the corresponding persistence landscapes PLp̃,ν (`) and PLp̃,ν (`). Then, ||PLp̃,ν (`) − PLp̃,ν (`)||q =
R (1) (2) 1/q
ν
|PLp̃ − PLp̃ |q is no larger than a function of the q-Wasserstein distance Wq,p̃ (Ω̃1 , Ω̃2 ).
Intuitively, it means that using persistence landscapes could preserve differences in the persistence
diagrams.

18
 The following steps describe the construction of persistence landscapes of p̃-th homology groups
in increasing order starting from the first-order landscape.

Step 1. Extract the 1-th homology groups τ̃p̃,k (k = 1, 2, . . . , kp̃ ) from the persistence diagram Ω̃
and create a set of computation grids, ` = M1 , M1 + δ, M1 + 2δ, . . . , M2 , where the lower and
upper bounds M1 and M2 are usually set as mink λp̃,k,1 and maxk λp̃,k,2 respectively. Here,
δ depends on the degree of resolutions. The R function landscape uses 500 grid points by
default so that δ = (M2 − M1 )/500.

Step 2. Use each τ̃p̃,k to compute PLp̃,k (`) = min(` − λp̃,k,1 , λp̃,k,2 − `)+ for all values of `.
(1) (2) (k )
Step 3. Fixing `, sort plp̃,k (`) in decreasing order, calling them plp̃ (`), plp̃ (`), . . . , plp̃ p̃ (`).

Step 4. Output the νth order persistence landscape of the p̃-th homology groups, PLp̃,ν (`) =
(ν) (ν)
(plp̃ (`))+ , where ν = 1, 2, . . . and set plp̃ (`) = 0 when ν > kp̃ .

Step 3 primarily determines the computational cost of constructing persistence landscapes.

Higher order persistence landscapes require more values sorted for each `, which could be costly
when kp̃ (number of p̃-th homology groups) is large.

Example 4.1. Persistence Landscapes from Persistence Diagrams. The R function

landscape can be used to compute persistence landscapes. We illustrate on the the persistence
diagram from Example 2.3.

Diag = pers.diag.3$diagram; Land <- c(); k=1; threshold=1

while(threshold>0){
Land <- cbind(Land, landscape(Diag = Diag, dimension = 0, KK = k,
tseq = seq(min(Diag[,2:3]), max(Diag[,2:3]), length=500)))
threshold = sum(abs(Land[, ncol(Land)]))
}
[,1] [,2] [,3] [,4]
[1,] 0.000000000 0 0 0
[2,] 0.003006012 0 0 0
[3,] 0.006012024 0 0 0
.....
[499,] 0.003006012 0.003006012 0 0
[500,] 0.000000000 0.000000000 0 0

In the code above, the function landscape takes several arguments. When dimension=0, it takes
0-th homology groups of the persistence diagram pers.diag.3 to compute landscape functions. The
KK argument specifies the order of landscapes to be computed. The tseq argument specifies the
range of the landscape functions. It uses a while-loop to compute persistence landscapes of all
orders (here, four) over 500 grids ranging from the minimum of the birth time to the maximum of
the death time.

19
 label
1st PL
2nd PL

0.6 3rd PL
4th PL
Density

0.4

0.2

0.0

0 100 200 300 400 500

Distance

Figure 8: Persistence landscapes of the persistence diagram pers.diag.3.

Persistence landscapes of any order can be used as features. Lower order persistence landscapes
contain information about important topological features than higher order persistence landscapes
which are closer to zero and handle topological noise. Therefore, selecting the order of the persistence
landscapes to serve as features requires a delicate balance between missing important signals and
introducing too much noise.

4.2 TDA of Time Series via Persistence Landscapes

TDA on functions can be used to construct feature representations for time series analysis, and
persistence landscapes are useful as topological representations for similarity/dissimilarity analysis
on time series. In the literature, different representations of time series have been used, such as the
weighted Fourier transform in Wang et al. (2018) or the Walsh-Fourier transform in Chen et al.
(2019). We describe these situations in the following sections.

4.2.1 Persistence Landscapes for Continuous Time Series

Wang et al. (2018) proposed TDA to measure structural changes in electroencephalogram (EEG)
time series. They first constructed Fourier transforms of the time series, then applied an expo-
nential weighting scheme on the Fourier transforms to focus on the more important low frequency
components of EEG. They further smoothed the weighted Fourier transform in order to make it a
Morse function (Palais, 1963).
The smoothed weighted Fourier series of a time series {xt , t = 1, . . . , T } has the form
2 2
X X
µ̂kTu (t) = e−(2jπ/T ) σ aj cos(2jπt/T ) + e−(2jπ/T ) σ bj sin(2jπt/T ),
j∈I1 j∈I2
p
where I1 = {j = 0, 1, 2, . . . , k : |aj | > Tu }, I2 = {j = 1, 2, . . . , k : |bj | > Tu }, Tu = s 2 log(n),
aj = T2 Tt=1 xt cos(2jπt/T ), bj = T2 Tt=1 xt sin(2jπt/T ) and a0 = Tt=1 xt /T . Here, k is the degree
P P P

20
deciding the highest frequency [k/T ] to be included in the representation (for T = 500, they used
k = 99), n is the number of data points in each phase and s is the median of the absolute deviation
(MAD) of the Fourier coefficients:

a(m) = median{|ai |, i = 1, 2, . . . , k}
b(m) = median{|bi |, i = 1, 2, . . . , k}
s = median{|ai − a(m) |, |bj − b(m) |, i, j = 1, 2, . . . , k} (8)

Using this finite sum of weighted sinusoidal functions, they argued that µ̂kTu (t) becomes a Morse
function. They used this Morse function representation to construct persistence landscapes of all
orders as features to detect possible structural changes. A main contribution of this paper is to
show via simulation studies that the proposed TDA framework is robust to topology-preserving
transformations such as translation and amplitude and frequency scaling, while being sensitive to
topology-destroying transformations. They argued the topological change happens only if there is
a structural change in the time series.

4.2.2 Persistence Landscapes for Categorical Time Series

Chen et al. (2019) described TDA of categorical time series via their Walsh-Fourier transforms
(which are not Morse functions). They constructed first order persistence landscapes based on
Walsh-Fourier transforms of categorical time series, which they then used as features for clustering.
They applied this analysis to a large travel-activity data set, carrying out computations in paral-
lel. They showed that construction of the first order persistence landscape only involves a linear
transformation of the Walsh Fourier transform.
Given a sequence of WFT dT (n, λj ), j = 0, 1, . . . , T2 − 1 of the time series xn,t , n = 1, 2, . . . , N ,
denote the minimum and maximum of the WFT values of the time series xn,t by

dn,min = min dT (n, λj ) and dn,max = max dT (n, λj ).

j j

Let
Dmin = min dn,min and Dmax = max dn,max
n n

denote the minimum and maximum values of the WFTs across all N time series. The first-order
persistence landscape of xn,t is obtained for ` = 1, 2, . . . , L as

PL(n, `) = min(V1 (n, `), V2 (n, `))+ (9)

where
(` − 1)(Dmax − Dmin )
V1 (n, `) = Dmin + − dn,min ,
L−1
(` − 1)(Dmax − Dmin )
V2 (n, `) = dn,max − Dmin − ,
L−1
and (a)+ denotes the positive part of a real number a. For ` = 1, 2, . . . , L and n = 1, . . . , N , the
PL(n, `) are piecewise linear functions that constitute features constructed for each of the N time
series and useful for clustering. Our C++ code is available here: https://github.com/bluemarlon/
TDA-of-K-means-on-1st-PL.

21
 series 1 WFT1 1st PL of 1
2

0.6

0.4
0.4

0.3
1

0.2

0.2
0.0

0.1
−0.2

0.0
0

0 200 400 600 800 1200 0 500 1000 1500 2000 0 20 40 60 80 100

series
time 2 WFT2
Time 1st Time
PL of 2
2

0.15

0.4
0.10

0.3
0.05
0.00
1

0.2
−0.05
−0.10

0.1
−0.15

0.0
0

0 200 400 600 800 1200 0 500 1000 1500 2000 0 20 40 60 80 100

time Time Time

Figure 9: Persistence diagrams using Walsh-Fourier Transforms.

In Figure 9, the first column shows categorical time series on activity-travel behavior of two
randomly chosen adults from the National Household Travel Survey (Chen et al., 2019). The length
of each time series is T = 1440, corresponding to the number of minutes in a day. The response
has three levels for each adult: 0 for staying at home, 1 for travel and 2 for being out of the home.
The middle column shows the WFT of the time series, each of length T2 = 2048. The last column
shows first order persistence landscapes which are quite distinct for the two series.
After the first order persistence landscapes are constructed, Chen et al. (2019) used a divide and
combine K-means approach for clustering a large number of subjects and identified three distinct
temporal patterns among them. The main contribution of this paper was to implement clustering
of a large set of activity-travel time series through TDA of non-Morse functions.

4.3 Other TDA Based Approaches

Numerical summaries other than persistence landscapes have been used for clustering (Berwald
et al., 2013; Pereira and de Mello, 2015; Seversky et al., 2016), classification (Umeda, 2017), and
break detection (Gidea, 2017; Gidea and Katz, 2018) of time series. We give a brief review in the
following sections.

4.3.1 Clustering
TDA based feature construction may be used with classical clustering methods, like K-means clus-
tering for time series. For instance, Pereira and de Mello (2015) considered the model analyzed
by Costantino et al. (1995) for evaluating two regimes of adult Tribolium flour population growth

22
(numerical measures) under stable equilibrium and aperiodic oscillations. They simulated a total
of 400 time series from the model for a period of 240 weeks (2 weeks per unit), consisting of 200
time series for the stable equilibrium regime and the other 200 for the aperiodic oscillations. A brief
description of their approach follows.

Step 1. They pre-processed the data using the Empirical Mode Decomposition (EMD) (Huang
et al., 1998) to denoise the data.
Step 2. They constructed point clouds using Takens’ embedding with d = 2 and τ = 3, taking
the time series {xt , t = 1, 2, . . . , T } to {vi = (xi , xi+3 )|i = 1, 2, . . . , T − 3}. The point cloud,
therefore, contains a total of N = T − 3 number of points.
Step 3. They used the Witness complex (De Silva and Carlsson, 2004) for doing persistent homol-
ogy. Witness complex is essentially a different simplicial complex, which is most useful when
dealing with large data sets.
Step 4. After acquiring the persistence diagram Ω̃ = {τ̃p̃,k |k = 1, 2, . . . , kp̃ , p̃ = 0, 1} (τ̃p̃,k =
(λp̃,k,1 , λp̃,k,2 )), they constructed the following set of features:

(i) The number of points on each dimension p̃, i.e., (k0 , k1 , k2 , . . . , kp̃ );
(ii) The maximum lifetime of each p̃, i.e, maxk (λp̃,k,2 − λp̃,k,1 );
(iii) The number of relevant points for each p̃, namely #{τ̃p̃,k |λp̃,k,2 − λp̃,k,1 ≥ 0.5 maxk (λp̃,k,2 −
λp̃,k,1 )}, where # denotes the cardinality of the set;
Pkp̃
(iv) The average lifetime of all homology groups on each p̃, k=1 (λp̃,k,2 − λp̃,k,1 )/kp̃ ;
Pkp̃
(v) The sum of the lifetimes of all homology groups on each p̃, k=1 (λp̃,k,2 − λp̃,k,1 ).

Step 5. Apply all features from Step 4 to do K-means clustering.

The implementation was done in Java. They compared results with true labels using different
metrics from the confusion matrix, and showed that their method achieved a high F1-score of 94%.
In sum, they used EMD to filter out data noise, constructed point cloud using Takens’ embedding
method and used persistent homology to construct features for unsupervised learning. This logic is
applicable to all other research papers.

4.3.2 Classification
Features constructed using TDA can be applied for classification of time series as well. Umeda
(2017) described an example with motion sensor data of daily and sports activities used in Altun
and Barshan (2010); Altun et al. (2010), including both chaotic and non-chaotic time series data.
Each of 19 activities was performed by 8 subjects and 60 signals were obtained for each activity
and each subject, yielding 9120 time series. The sensor frequency was 512Hz and each signal was
collected for 5 minutes. The pipeline of their method which was implemented in MATLAB is shown
below.

Step 1. Construct a point cloud via Takens’ embedding, {xt , t = 1, 2, . . . , T } → {vi } ⊂ R3 , for
i = 1, 2, . . . , T − 2 and vi = (xi , xi+1 , xi+2 ).
Step 2. Convert the point cloud to the persistence diagram Ω̃ = {τ̃p̃,k |k = 1, 2, . . . , p̃, p̃ = 0, 1, 2}
using Rips complex and persistent homology.

23
Step 3. Compute features using Betti sequences of p̃-th homology groups ∆ ˜ p̃ (λ) = #{λp̃,k,1 ≤ λ ≤
λp̃,k,2 |k = 1, 2, . . . , kp̃ } from the persistence diagram, discretized into 300 points for each p̃,
and connected into one feature vector of length 900.

Step 4. Apply a one-dimensional convolutional neural network (CNN) (Krizhevsky et al., 2012)
on the feature vector from Step 3 to do the classification.

They concluded that their approach performed better than an approach using support vector
machine (SVM) (Hastie et al., 2001).

4.3.3 Structural Break Detection

TDA based features for structural break detection has been discussed in bioinformatics (Wang
et al., 2018), financial data analysis (Gidea, 2017; Truong, 2017; Gidea and Katz, 2018; Gidea
et al., 2018), etc. Here we describe the approach in Gidea et al. (2018), who used four major
daily log-price cryptocurrencies (Bitcoin, Ethereum, Litecoin, and Ripple) between 2016-01-01 to
2018-02-28 independently for break detection of critical transitions:

Step 1. They constructed point clouds each with 50 points from each log-price time series xt , t =
1, 2, . . . , T . Each of the T − 52 = 448 point clouds had a total of 50 points, the first point
cloud being (v1 , v2 , . . . , v5 0) and the last point cloud being (v4 48, v4 49, . . . , v4 97), with vi =
(xi , xi+1 , xi+2 , xi+3 ) ∈ R4 }.

Step 2. On each windowing point cloud P (w) , they constructed a persistence diagram Ω̃(w) using
(w)
Rips complex. Since they only focused on the 1-th homology groups, Ω̃(w) = {τ̃1,k , k =
1, 2, . . . , k1 }.
(w)
Step 3. They constructed all persistence landscapes PL1,ν (`) from each Ω̃(w) , ν = 1, 2, . . ., and then
(w) (w)
converted the persistence landscapes PL1,ν (`) to its L1 norm, which is defined as ||PL1 ||1 =
P∞ (w) (w) R (w)
ν=1 ||PL1,ν ||1 , where ||PL1,ν ||1 = R PL1,ν (`)d`.

Step 4. On each window, they combined the log-price time series xw , first difference of the log-price
(w) (w)
values xw+1 − xw and the L1 norm ||PL1 ||1 as the feature vector (xw , xw+1 − xw , ||PL1 ||1 )
and used it to do K-means clustering with number of cluster K = 18.

They applied the method independently to each daily log-price cryptocurrency time series, and
then used the clusters to identify topologically distinct regimes before the crash of each asset. They
argued that this method has the potential to automatically recognize approaching critical transitions
in the cryptocurrency markets, even when the relevant time series exhibit a highly non-stationary,
erratic behavior.

5 Discussion and Summary

This paper gives a comprehensive overview of TDA which consists of a set of powerful tools for mea-
suring topological features of time series and using it for pattern detection, clustering, classification,
and structural break detection. Research extensions in several directions are possible. First, TDA
for multivariate time series analysis has been studied very recently (Gidea and Katz, 2018; Gidea,
2017; Stolz et al., 2017). A second extension consists of using summary statistics and dissimilarity

24
measures for TDA. A review of summary statistics and dissimilarity measures refer to Nanopoulos
et al. (2001); Fulcher (2017); Aghabozorgi et al. (2015). Third, research into improved computation
tools in computational topology and further exploration of statistical properties while using TDA
is a rich research area. Ongoing research can be separated into these different scenarios: computa-
tional homology (Phillips et al., 2015; De Silva and Carlsson, 2004; Liu et al., 2012; Carlsson and
Zomorodian, 2009; Corbet et al., 2019; Ren et al., 2018), study of topological summaries (Bendich
et al., 2016; Biscio and MÞller, 2019; Reininghaus et al., 2015; Kusano et al., 2016; Carrière et al.,
2017), and statistical inference (Fasy et al., 2014b; Phillips et al., 2015; Chazal et al., 2018; Alaa
and Mohamed, 2017; Robinson and Turner, 2017; Brécheteau, 2019).

References
Adams, H. and Carlsson, G. (2015). “Evasion paths in mobile sensor networks.” The International
Journal of Robotics Research, 34, 1, 90–104.

Aghabozorgi, S., Seyed Shirkhorshidi, A., and Ying Wah, T. (2015). “Time-series Clustering - A
Decade Review.” Information Systems, 53, C, 16–38.

Alaa, H. and Mohamed, S. (2017). “On the Topological Data Analysis extensions and comparisons.”
Journal of the Egyptian Mathematical Society, 25, 4, 406 – 413.

Altun, K. and Barshan, B. (2010). “Human activity recognition using inertial/magnetic sensor
units.” International workshop on human behavior understanding, 38–51.

Altun, K., Barshan, B., and Tunçel, O. (2010). “Comparative study on classifying human activities
with miniature inertial and magnetic sensors.” Pattern Recognition, 43, 10, 3605–3620.

Bendich, P., Marron, J. S., Miller, E., Pieloch, A., and Skwerer, S. (2016). “Persistent homology
analysis of brain artery trees.” The annals of applied statistics, 10, 1, 198.

Berwald, J., Gidea, M., and Vejdemo-Johansson, M. (2013). “Automatic recognition and tagging
of topologically different regimes in dynamical systems.” CoRR, abs/1312.2482.

Biscio, C. A. N. and MÞller, J. (2019). “The Accumulated Persistence Function, a New Useful
Functional Summary Statistic for Topological Data Analysis, With a View to Brain Artery Trees
and Spatial Point Process Applications.” Journal of Computational and Graphical Statistics, 0,
0, 1–21.

Bonis, T., Ovsjanikov, M., Oudot, S., and Chazal, F. (2016). “Persistence-Based Pooling for Shape
Pose Recognition.” Proceedings of the 6th International Workshop on Computational Topology
in Image Context - Volume 9667 , 19–29.

Brécheteau, C. (2019). “A statistical test of isomorphism between metric-measure spaces using the
distance-to-a-measure signature.” Electron. J. Statist., 13, 1, 795–849.

Bubenik, P. (2015). “Statistical topological data analysis using persistence landscapes.” The Journal
of Machine Learning Research, 16, 1, 77–102.

Carlsson, G. (2009). “Topology and data.” Bulletin of the American Mathematical Society, 46, 2,
255–308.

25
— (2014). “Topological pattern recognition for point cloud data.” Acta Numerica, 23, 289–368.

Carlsson, G. and Zomorodian, A. (2009). “The theory of multidimensional persistence.” Discrete

& Computational Geometry, 42, 1, 71–93.

Carlsson, G., Zomorodian, A., Collins, A., and Guibas, L. (2004). “Persistence Barcodes for
Shapes.” Proceedings of the 2004 Eurographics/ACM SIGGRAPH Symposium on Geometry Pro-
cessing, 124–135.

Carrière, M., Cuturi, M., and Oudot, S. (2017). “Sliced Wasserstein Kernel for Persistence Dia-
grams.” Proceedings of the 34th International Conference on Machine Learning, 664–673.

Carrière, M., Oudot, S. Y., and Ovsjanikov, M. (2015). “Stable Topological Signatures for Points
on 3D Shapes.” Proceedings of the Eurographics Symposium on Geometry Processing, 1–12.

Chazal, F., Cohen-Steiner, D., Guibas, L. J., Mémoli, F., and Oudot, S. Y. (2009). “Gromov-
Hausdorff Stable Signatures for Shapes Using Persistence.” Proceedings of the Symposium on
Geometry Processing, 1393–1403.

Chazal, F., de Silva, V., and Oudot, S. (2014). “Persistence stability for geometric complexes.”
CoRR, abs/1207.3885.

Chazal, F., Fasy, B., Lecci, F., Bertr, Michel, Aless, ro Rinaldo, and Wasserman, L. (2018). “Robust
Topological Inference: Distance To a Measure and Kernel Distance.” Journal of Machine Learning
Research, 18, 159, 1–40.

Chazal, F. and Michel, B. (2017). “An introduction to Topological Data Analysis: fundamental
and practical aspects for data scientists.” Journal de la Société Française de Statistique.

Chen, R., Zhang, J., Ravishanker, N., and Konduri, K. (2019). “Clustering Activity-Travel Behavior
Time Series using Topological Data Analysis.” arXiv preprint arXiv:1907.07603 .

Corbet, R., Fugacci, U., Kerber, M., Landi, C., and Wang, B. (2019). “A kernel for multi-parameter
persistent homology.” Computers & Graphics: X , 100005.

Costantino, R. F., Cushing, J. M., Dennis, B., and Desharnais, R. A. (1995). “Experimentally
induced transitions in the dynamic behaviour of insect populations.” Nature, 375, 6528, 227.

De Silva, V. and Carlsson, G. E. (2004). “Topological estimation using witness complexes.” SPBG,
4, 157–166.

De Silva, V. and Ghrist, R. (2007). “Coverage in sensor networks via persistent homology.” Algebraic
& Geometric Topology, 7, 1, 339–358.

DeWoskin, D., Climent, J., Cruz-White, I., Vazquez, M., Park, C., and Arsuaga, J. (2010). “Ap-
plications of computational homology to the analysis of treatment response in breast cancer
patients.” Topology and its Applications, 157, 1, 157–164.

Di Fabio, B. and Landi, C. (2011). “A Mayer–Vietoris Formula for Persistent Homology with an
Application to Shape Recognition in the Presence of Occlusions.” Foundations of Computational
Mathematics, 11, 5, 499.

26
— (2012). “Persistent homology and partial similarity of shapes.” Pattern Recognition Letters, 33,
11, 1445–1450.

Edelsbrunner, H. and Harer, J. (2010). Computational Topology. An Introduction.. American

Mathematical Society.

Fasy, B. T., Kim, J., Lecci, F., and Maria, C. (2014a). “Introduction to the R package TDA.”
CoRR, abs/1411.1830.

Fasy, B. T., Lecci, F., Rinaldo, A., Wasserman, L., Balakrishnan, S., and Singh, A. (2014b).
“Confidence sets for persistence diagrams.” Ann. Statist., 42, 6, 2301–2339.

Fulcher, B. D. (2017). “Feature-based time-series analysis.” Computing Research Repository,

abs/1709.08055.

Gidea, M. (2017). “Topological data analysis of critical transitions in financial networks.” Interna-
tional Conference and School on Network Science, 47–59.

Gidea, M. and Katz, Y. (2018). “Topological data analysis of financial time series: Landscapes of
crashes.” Physica A: Statistical Mechanics and its Applications, 491, 820–834.

Gidea, M., Katz, Y. A., Roldan, P., Goldsmith, D., and Shmalo, Y. (2018). “Topological recognition
of critical transitions in time series of cryptocurrencies.” Social Science Research Network .

Hastie, T., Tibshirani, R., and Friedman, J. (2001). The Elements of Statistical Learning. Springer
Series in Statistics. New York, NY, USA: Springer New York Inc.

Hegger, R., Kantz, H., and Schreiber, T. (1999). “Practical implementation of nonlinear time series
methods: The TISEAN package.” Chaos: An Interdisciplinary Journal of Nonlinear Science, 9,
2, 413–435.

Heo, G., Gamble, J., and Kim, P. T. (2012). “Topological analysis of variance and the maxillary
complex.” Journal of the American Statistical association, 107, 498, 477–492.

Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih, H. H., Zheng, Q., Yen, N.-C., Tung, C. C., and
Liu, H. H. (1998). “The empirical mode decomposition and the Hilbert spectrum for nonlinear
and non-stationary time series analysis.” Proceedings of the Royal Society of London. Series A:
Mathematical, Physical and Engineering Sciences, 454, 1971, 903–995.

Kennel, M. B., Brown, R., and Abarbanel, H. D. (1992). “Determining embedding dimension for
phase-space reconstruction using a geometrical construction.” Physical review A, 45, 6, 3403.

Khasawneh, F. A. and Munch, E. (2016). “Chatter detection in turning using persistent homology.”
Mechanical Systems and Signal Processing, 70, 527–541.

Kovacev-Nikolic, V., Bubenik, P., Nikolić, D., and Heo, G. (2016). “Using persistent homology
and dynamical distances to analyze protein binding.” Statistical applications in genetics and
molecular biology, 15, 1, 19–38.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). “Imagenet classification with deep convo-
lutional neural networks.” Advances in neural information processing systems, 1097–1105.

27
Kusano, G., Hiraoka, Y., and Fukumizu, K. (2016). “Persistence weighted Gaussian kernel for
topological data analysis.” International Conference on Machine Learning, 2004–2013.

Li, C., Ovsjanikov, M., and Chazal, F. (2014). “Persistence-based Structural Recognition.” IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 2014 , 1995–2002.

Liu, X., Xie, Z., Yi, D., et al. (2012). “A fast algorithm for constructing topological structure in
large data.” Homology, Homotopy and Applications, 14, 1, 221–238.

Mileyko, Y., Mukherjee, S., and Harer, J. (2011). “Probability measures on the space of persistence
diagrams.” Inverse Problems, 27, 12, 124007.

Munkres, J. R. (1993). Elements of Algebraic Topology. Addison-Wesley.

Nanopoulos, A., Alcock, R., and Manolopoulos, Y. (2001). “Information Processing and Technol-
ogy.” chap. Feature-based Classification of Time-series Data, 49–61. Commack, NY, USA: Nova
Science Publishers, Inc.

Nicolau, M., Levine, A. J., and Carlsson, G. E. (2011). “Topology based data analysis identifies a
subgroup of breast cancers with a unique mutational profile and excellent survival.” Proceedings
of the National Academy of Sciences of the United States of America, 108 17, 7265–70.

Palais, R. S. (1963). “Morse theory on Hilbert manifolds.” Topology, 2, 4, 299–340.

Perea, J. A., Deckard, A., Haase, S. B., and Harer, J. (2015). “SW1PerS: Sliding windows and
1-persistence scoring; discovering periodicity in gene expression time series data.” BMC bioin-
formatics, 16, 1, 257.

Perea, J. A. and Harer, J. (2015). “Sliding windows and persistence: An application of topological
methods to signal analysis.” Foundations of Computational Mathematics, 15, 3, 799–838.

Pereira, C. M. and de Mello, R. F. (2015). “Persistent Homology for Time Series and Spatial Data
Clustering.” Expert Syst. Appl., 42, 15, 6026–6038.

Phillips, J. M., Wang, B., and Zheng, Y. (2015). “Geometric Inference on Kernel Density Esti-
mates.” 31st International Symposium on Computational Geometry (SoCG 2015), 34, 857–871.

Reininghaus, J., Huber, S. M., Bauer, U., and Kwitt, R. (2015). “A stable multi-scale kernel
for topological machine learning.” 2015 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 4741–4748.

Ren, S., Wu, C., Wu, J., et al. (2018). “Weighted persistent homology.” Rocky Mountain Journal
of Mathematics, 48, 8, 2661–2687.

Robinson, A. and Turner, K. (2017). “Hypothesis testing for topological data analysis.” Journal of
Applied and Computational Topology, 1, 2, 241–261.

Seversky, L. M., Davis, S., and Berger, M. (2016). “On time-series topological data analysis: New
data and opportunities.” 2016 IEEE Conference on Computer Vision and Pattern Recognition
Workshops (CVPRW), 1014–1022.

Shanks, J. L. (1969). “Computation of the Fast Walsh-Fourier Transform.” IEEE Trans. Comput.,
18, 5, 457–459.

28
Silva, V. D. and Ghrist, R. (2007). “Homological sensor networks.” Notices Amer. Math. Soc,
10–17.

Stoffer, D. S. (1991). “Walsh-Fourier Analysis and Its Statistical Applications.” Journal of the
American Statistical Association, 86, 414, 461–479.

Stolz, B. J., Harrington, H. A., and Porter, M. A. (2017). “Persistent homology of time-dependent
functional networks constructed from coupled time series.” Chaos: An Interdisciplinary Journal
of Nonlinear Science, 27, 4, 047410.

Takens, F. et al. (1981). “Detecting strange attractors in turbulence.” Lecture notes in mathematics,
898, 1, 366–381.

Truong, P. (2017). “An exploration of topological properties of high-frequency one-dimensional

financial time series data using TDA.” Ph.D. thesis, KTH Royal Institute of Technology.

Umeda, Y. (2017). “Time Series Classification via Topological Data Analysis.” Transactions of the
Japanese Society for Artificial Intelligence, 32, 3.

Van de Weygaert, R., Vegter, G., Edelsbrunner, H., Jones, B. J. T., Pranav, P., Park, C., Hellwing,
W. A., Eldering, B., Kruithof, N., Bos, E. G. P. P., Hidding, J., Feldbrugge, J., ten Have, E.,
van Engelen, M., Caroli, M., and Teillaud, M. (2011). “Transactions on Computational Science
XIV.” chap. Alpha, Betti and the Megaparsec Universe: On the Topology of the Cosmic Web,
60–101. Berlin, Heidelberg: Springer-Verlag.

Wang, Y., Ombao, H., and Chung, M. K. (2018). “Topological data analysis of single-trial elec-
troencephalographic signals.” The annals of applied statistics, 12, 3, 1506.

29