Selective review of offline change point detection methods

Charles Truonga,∗, Laurent Oudreb , Nicolas Vayatisa

a CMLA, CNRS, ENS Paris Saclay
b L2TI, University Paris 13

Abstract
This article presents a selective survey of algorithms for the offline detection of multiple change points in
multivariate time series. A general yet structuring methodological strategy is adopted to organize this
vast body of work. More precisely, detection algorithms considered in this review are characterized
by three elements: a cost function, a search method and a constraint on the number of changes.
Each of those elements is described, reviewed and discussed separately. Implementations of the main
algorithms described in this article are provided within a Python package called ruptures.
Keywords: change point detection, segmentation, statistical signal processing

Contents

1 Introduction 2

2 Background 3
2.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Selection criteria for the review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.5 Outline of the article . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Evaluation 6
3.1 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Evaluation metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2.1 AnnotationError . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2.2 Hausdorff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2.3 RandIndex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2.4 F1-score . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4 Models and cost functions 8

4.1 Parametric models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.1.1 Maximum likelihood estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.1.2 Piecewise linear regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
4.1.3 Mahalanobis-type metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 Non-parametric models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2.1 Non-parametric maximum likelihood . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2.2 Rank-based detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2.3 Kernel-based detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.3 Summary table . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

∗ Corresponding author

Accepted to Signal Processing December 8, 2020

5 Search methods 18
5.1 Optimal detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.1.1 Solution to Problem 1 (P1): Opt . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5.1.2 Solution to Problem 2 (P2): Pelt . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Approximate detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2.1 Window sliding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2.2 Binary segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.2.3 Bottom-up segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

6 Estimating the number of changes 25

6.1 Linear penalty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
6.2 Fused lasso . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
6.3 Complex penalties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

7 Summary table 28

8 Presentation of the Python package 31

9 Conclusion 32

1. Introduction

A common task in signal processing is the identification and analysis of complex systems whose
underlying state changes, possibly several times. This setting arises when industrial systems, physical
phenomena or human activity are continuously monitored with sensors. The objective of practitioners
is to extract from the recorded signals a posteriori meaningful information about the different states
and transitions of the monitored object for analysis purposes. This setting encompasses a broad range
of real-world scenarios and a wide variety of signals.
Change point detection is the task of finding changes in the underlying model of a signal or time
series. The first works on change point detection go back to the 50s [1, 2]: the goal was to locate
a shift in the mean of independent and identically distributed (iid) Gaussian variables for industrial
quality control purposes. Since then, this problem has been actively investigated, and is periodically
the subject of in-depth monographs [3–6]. This subject has generated important activity in statistics
and signal processing [7–9] but also in various application settings such as speech processing [10–
13], financial analysis [7, 14, 15], bio-informatics [16–24], climatology [25–27], network traffic data
analysis [28, 29]. Modern applications in bioinformatics, finance, monitoring of complex systems have
also motivated recent developments from the machine learning community [18, 30, 31].
Let us take the example of gait analysis, illustrated on the flowchart displayed on Figure 1. In this
context, a patient’s movements are monitored with accelerometers and gyroscopes while performing
simple activities, for instance walking at preferred speed, running or standing still. The objective is to
objectively quantify gait characteristics [32–36]. The resulting signal is described as a succession of non-
overlapping segments, each one corresponding to an activity and having its own gait characteristics.
Insightful features from homogeneous phases can be extracted if the temporal boundaries of those
segments are identified. This analysis therefore needs a preliminary processing of the signals: change
point detection.
Change point detection methods are divided into two main branches: online methods, that aim to
detect changes as soon as they occur in a real-time setting, and offline methods that retrospectively
detect changes when all samples are received. The former task is often referred to as event or anomaly
detection, while the latter is sometimes called signal segmentation.
In this article, we propose a survey of algorithms for the detection of multiple change points in
multivariate time series. All reviewed methods presented in this paper address the problem of offline
(also referred to as retrospective or a posteriori ) change point detection, in which segmentation is

2
 Subject

Monitoring

Raw signal

Change point detection

Regime 1 Regime 2 Regime 3 Regime 4 Regime 5

Feature extraction on homogeneous regimes

Figure 1: Flowchart of a study scheme, for gait analysis.

performed after the signal has been collected. The objective of this article is to facilitate the search
of a suitable detection method for a given application. In particular, focus is made on practical
considerations such as implementations and procedures to calibrate the algorithms. This review also
presents the mathematical properties of the main approaches, as well as the metrics to evaluate and
compare their results. This article is linked with a Python scientific library called ruptures [37], that
includes a modular and easy-to-use implementation of all the main methods presented in this paper.

2. Background

This section introduces the main concepts for change point detection, as well as the selection criteria
and the outline of this review.

2.1. Notations
In the remainder of this article, we use the following notations. For a given signal y = {yt }Tt=1 , the
(b − a)-sample long sub-signal {yt }bt=a+1 (1 ≤ a < b ≤ T ) is simply denoted ya..b ; the complete signal is
therefore y = y0..T . A set of indexes is denoted by a calligraphic letter: T = {t1 , t2 , . . . } ⊂ {1, . . . , T },
and its cardinal is |T |. For a set of indexes T = {t1 , . . . , tK }, the dummy indexes t0 := 0 and tK+1 := T
are implicitly available.

3
2.2. Problem formulation
Let us consider a multivariate non-stationary random process y = {y1 , . . . , yT } that takes value in
Rd (d ≥ 1) and has T samples. The signal y is assumed to be piecewise stationary, meaning that some
characteristics of the process change abruptly at some unknown instants t∗1 < t∗2 < · · · < t∗K ∗ . Change
point detection consists in estimating the indexes t∗k . Depending on the context, the number K ∗ of
changes may or may not be known, in which case it has to be estimated too.
Formally, change point detection is cast as a model selection problem, which consists in choosing
the best possible segmentation T according to a quantitative criterion V (T , y) that must be minimized.
(The function V (T , y) is simply denoted V (T ) when it is obvious from the context that it refers to
the signal y.) The choice of the criterion function V (·) depends on preliminary knowledge on the task
at hand.
In this work, we make the assumption that the criterion function V (T ) for a particular segmentation
is a sum of costs of all the segments that define the segmentation:
K
X
V (T , y) := c(ytk ..tk+1 ) (1)
k=0

t
where c(·) is a cost function which measures goodness-of-fit of the sub-signal ytk ..tk+1 = {yt }tk+1
k +1
to
a specific model. The “best segmentation” T is the minimizer of the criterion V (T ). In practice,
b
depending on whether the number K ∗ of change points is known beforehand, change point detection
methods fall into two categories.

• Problem 1 : known number of changes K. The change point detection problem with a
fixed number K of change points consists in solving the following discrete optimization problem

min V (T ). (P1)
|T |=K

• Problem 2 : unknown number of changes. The change point detection problem with an
unknown number of change points consists in solving the following discrete optimization problem

min V (T ) + pen(T ) (P2)

T

where pen(T ) is an appropriate measure of the complexity of a segmentation T .

All change point detection methods considered in this work yield an exact or an approximate
solution to either Problem 1 (P1) or Problem 2 (P2), with the function V (T , y) adhering to the
format (1).

2.3. Selection criteria for the review

To better understand the strengths and weaknesses of change point detection methods, we propose to
classify algorithms according to a comprehensive typology. Precisely, detection methods are expressed
as the combination of the following three elements.
• Cost function. The cost function c(·) is a measure of “homogeneity”. Its choice encodes the
type of changes that can be detected. Intuitively, c(ya..b ) is expected to be low if the sub-signal
ya..b is “homogeneous” (meaning that it does not contain any change point), and large if the
sub-signal ya..b is “heterogeneous” (meaning that it contains one or several change points).
• Search method. The search method is the resolution procedure for the discrete optimization
problems associated with Problem 1 (P1) and Problem 2 (P2). The literature contains several
methods to efficiently solve those problems, in an exact fashion or in an approximate fashion.
Each method strikes a balance between computational complexity and accuracy.

4
 Change point detection

Cost function Search method Constraint

Figure 2: Typology of change point detection methods described in this article. Reviewed algorithms are defined by
three elements: a cost function, a search method and a constraint (on the number of change points).

• Constraint (on the number of change points). When the number of changes is unknown (P2),
a constraint is added, in the form of a complexity penalty pen(·) (P2), to balance out the
goodness-of-fit term V (T , y). The choice of the complexity penalty is related to the amplitude
of the changes to detect: with too “small” a penalty (compared to the goodness-of-fit) in (P2),
many change points are detected, even those that are the result of noise. Conversely, too much
penalization only detects the most significant changes, or even none.

This typology of change point detection methods is schematically shown on Figure 2.

2.4. Limitations
The described framework, however general, does not encompass all published change point detection
methods. In particular, Bayesian approaches are not considered in the remainder of this article, even
though they provide state-of-the-art results in several domains, such as speech and sound processing.
The most well-known Bayesian algorithm is the Hidden Markov Model (HMM) [38]. This model was
later extended, for instance with Dirichlet processes [39, 40] or product partition models [41, 42]. The
interested reader can find reviews of Bayesian approaches in [4] and [6].
Also, several literature reviews with different selection criteria can be found. Recent and important
works include [43] which focuses on window-based detection algorithms. In particular, the authors use
the quantity of samples needed to detect a change as a basis for comparison. Maximum likelihood
and Bayes-type detection are reviewed, from a theoretical standpoint, in [8]. Existing asymptotic
distributions for change point estimates are described for several statistical models. In [44], detection
is formulated as a statistical hypothesis testing problem, and emphasis is put on the algorithmic and
theoretical properties of several sequential mean-shift detection procedures.

2.5. Outline of the article

Before starting this review, we propose in Section 3 a detailed overview of the main mathematical
tools that can be used for evaluating and comparing the change point detection methods. The organi-
zation of the remaining of this review article reflects the typology of change point detection methods,
which is schematically shown on Figure 2. Precisely, the three defining elements of a detection algo-
rithm are reviewed separately. In Section 4, cost functions from the literature are presented, along
with the associated signal model and the type of change that can be detected. Whenever possible, the-
oretical results on asymptotic consistency are also given. Section 5 lists search methods that efficiently
solve the discrete optimizations associated with Problem 1 (P1) and Problem 2 (P2). Both exact
and approximate methods are described. Constraints on the number of change points are reviewed in
Section 6. A summary table of the literature review can be found in Section 7. The last section 8 is
dedicated to the presentation of the Python package that goes with this article and propose a modular
implementation of all the main approaches described in this article.

5
3. Evaluation

Change point detection methods can be evaluated either by proving some mathematical properties
of the algorithms (such as consistency) in general case, or empirically by computing several metrics to
assess the performances on a given dataset.

3.1. Consistency
A natural question when designing detection algorithms is the consistency of estimated change
point indexes, as the number of samples T goes to infinity. In the literature, the “asymptotic setting”
is intuitively described as follows: the observed signal y is regarded as a realization of a continuous-
time process on an equispaced grid of size 1/T , and “T goes to infinity” means that the spacing of
the sampling grid converges to 0. Precisely, for all τ ∈ [0, 1], let Y (τ ) denote an Rd -valued random
variable such that
yt = Y (t/T ) ∀t = 1, . . . , T. (2)
The continuous-time process undergoes K ∗ changes in the probability distribution at the time instants
τk∗ ∈ (0, 1). Those τk∗ are related to the change point indexes t∗k through the following relationship:

t∗k = bT τk∗ c. (3)

Generally, for a given change point index tk , the associated quantity τk = tk /T ∈ (0, 1) is referred
to as a change point fraction. In particular, the change point fractions τk∗ (k = 1, . . . , K ∗ ) of the
time-continuous process Y are change point indexes of the discrete-time signal y. Note that in this
asymptotic setting, the lengths of each regime of y increase linearly with T . The notion of asymptotic
consistency of a change point detection method is formally introduced as follows.
Definition 1 (Asymptotic consistency). A change point detection algorithm is said to be asymp-
totically consistent if the estimated segmentation Tb = {t̂1 , t̂2 , . . . } satisfies the following conditions,
when T −→ +∞:
(i) P (|Tb | = K ∗ ) −→ 1,
1 p
(ii) T Tb − T ∗ −→ 0,
∞

where the distance between two change point sets is defined by

Tb − T ∗ := max { max ∗min∗ |t̂ − t∗ |, max

∗ ∗
min |t̂ − t∗ | }. (4)
∞ t̂∈Tb t ∈T t ∈T t̂∈Tb

In Definition 1, the first condition is trivially verified when the number K ∗ of change points is
known beforehand. As for the second condition, it implies that the estimated change point fractions
are consistent, and not the indexes themselves. In general, distances |t̂ − t∗ | between true change point
indexes and their estimated counterparts do not converge to 0, even for simple models [18, 45–47]. As
a result, consistency results in the literature only deal with change point fractions.

3.2. Evaluation metrics

Several metrics from the literature are presented below. Each metric correspond to one of the
previously listed criteria by which segmentation performances are assessed. In the following, the set of
true change points is denoted by T ∗ = {t∗1 , . . . , t∗K ∗ }, and the set of estimated change points is denoted
by Tb = {t̂1 , . . . , t̂Kb }. Note that that the cardinals of each set, K ∗ and K,b are not necessarily equal.

6
 ∆t2 ∆t3
∆t1

Figure 3: Hausdorff. Alternating gray areas mark the segmentation T ∗ ; dashed lines mark the segmentation Tb . Here,
Hausdorff is equal to ∆t1 = max(∆t1 , ∆t2 , ∆t3 ).

3.2.1. AnnotationError
The AnnotationError is simply the difference between the predicted number of change points
|Tb | and the true number of change points |T ? |:

b − K ∗ |.
AnnotationError := |K (5)

This metric can be used to discriminate detection method when the number of changes is unknown.

3.2.2. Hausdorff
The Hausdorff metric measures the robustness of detection methods [47, 48]. Formally, it is
equal to the greatest temporal distance between a change point and its prediction:

Hausdorff(T ∗ , Tb ) := max { max ∗min∗ |t̂ − t∗ |, max

∗ ∗
min |t̂ − t∗ | }.
t̂∈Tb t ∈T t ∈T t̂∈Tb

It is the worst error made by the algorithm that produced Tb and is expressed in number of samples.
If this metric is equal to zero, both breakpoint sets are equal; it is large when a change point from
either T ∗ or Tb is far from every change point of Tb or T ∗ respectively. Over-segmentation as well as
under-segmentation is penalized. An illustrative example is displayed on Figure 3.

3.2.3. RandIndex
Accuracy can be measured by the RandIndex, which is the average similarity between the pre-
dicted breakpoint set Tb and the ground truth T ∗ [30]. Intuitively, it is equal to the number of
agreements between two segmentations. An agreement is a pair of indexes which are either in the
same segment according to both Tb and T ∗ or in different segments according to both Tb and T ∗ .
Formally, for a breakpoint set T , the set of grouped indexes and the set of non-grouped indexes are
respectively gr(T ) and ngr(T ):

gr(T ) := {(s, t), 1 ≤ s < t ≤ T s.t. s and t belong to the same segment according to T },
ngr(T ) := {(s, t), 1 ≤ s < t ≤ T s.t. s and t belong to different segments according to T }.

The RandIndex is then defined as follows:

|gr(Tb ) ∩ gr(T ∗ )| + |ngr(Tb ) ∩ ngr(T ∗ )|)

RandIndex(T ∗ , Tb ) := . (6)
T (T − 1)

It is normalized between 0 (total disagreement) and 1 (total agreement). Originally, RandIndex

has been introduced to evaluate clustering methods [30, 47]. An illustrative example is displayed on
Figure 4.

7
 True partition Computed partition Disagreement

Figure 4: RandIndex. Top: alternating gray areas mark the segmentation T ∗ ; dashed lines mark the segmentation
Tb . Below: representations of associated adjacency matrices and disagreement matrix. The adjacency matrix of a
segmentation is the T × T binary matrix with coefficient (s, t) equal to 1 if s and t belong to the same segment, 0
otherwise. The disagreement matrix is the T × T binary matrix with coefficient (s, t) equal to 1 where the two adjacency
matrices disagree, and 0 otherwise. RandIndex is equal to the white area (where coefficients are 0) of the disagreement
matrix.

3.2.4. F1-score
Another measure of accuracy is the F1-Score. Precision is the proportion of predicted change
points that are true change points. Recall is the proportion of true change points that are well predicted.
A breakpoint is considered detected up to a user-defined margin of error M > 0; true positives Tp are
true change points for which there is an estimated one at less than M samples, i.e.

Tp(T ∗ , Tb ) := {t∗ ∈ T ∗ | ∃ t̂ ∈ Tb s.t. |t̂ − t∗ | < M }. (7)

Precision Prec and recall Rec are then given by

Prec(T ∗ , Tb ) := |Tp(T ∗ , Tb )|/K

b and Rec(T ∗ , Tb ) := |Tp(T ∗ , Tb )|/K ∗ . (8)

Precision and Recall are well-defined (ie. between 0 and 1) if the margin M is smaller than the
minimum spacing between two true change point indexes t∗k and t∗k+1 . Over-segmentation of a signal
causes the precision to be close to zero and the recall close to one. Under-segmentation has the opposite
effect. The F1-Score is the harmonic mean of precision Prec and recall Rec:

Prec(T ∗ , Tb ) × Rec(T ∗ , Tb )
F1-Score(T ∗ , Tb ) := 2 × . (9)
Prec(T ∗ , Tb ) + Rec(T ∗ , Tb )

Its best value is 1 and its worse value is 0. An illustrative example is displayed on Figure 5.

4. Models and cost functions

This section presents the first defining element of change detection methods, namely the cost
function. In most cases, cost functions are derived from a signal model. In the following, models
and their associated cost function are organized in two categories: parametric and non-parametric, as
schematically shown in Figure 6. For each model, the most general formulation is first given, then
special cases, if any, are described. A summary table of all reviewed costs can be found at the end of
this section.

8
 × X X

Figure 5: F1-Score. Alternating gray areas mark the segmentation T ∗ ; dashed lines mark the segmentation Tb ; dashed
areas mark the allowed margin of error around true change points. Here, Prec is 2/3, Rec is 2/2 and F1-Score is 4/5.

4.1. Parametric models

Parametric detection methods focus on changes in a finite-dimensional parameter vector. Histori-
cally, they were the first to be introduced, and remain extensively studied in the literature.

4.1.1. Maximum likelihood estimation

Maximum likelihood procedures are ubiquitous in the change point detection literature. They
generalize a large number of models and cost functions, such as mean-shifts and scale shifts in normally
distributed data [2, 49–51], changes in the rate parameter of Poisson distributed data [39], etc. In
the general setting of maximum likelihood estimation for change detection, the observed signal y =
{y1 , . . . , yT } is composed of independent random variables, such that
∗
K
X
yt ∼ f (·|θk )1(t∗k < t ≤ t∗k+1 ) (M1)
k=0

where the t∗k are change point indexes, the f (·|θ) are probability density functions parametrized by the
vector-valued parameter θ, and the θk are parameter values. In other words, the signal y is modelled
by iid variables with piecewise constant distribution. The parameter θ represents a quantity of interest
whose value changes abruptly at the unknown instants t∗k , which are to be estimated. Under this
setting, change point detection is equivalent to maximum likelihood estimation if the sum of cost
V (T , y) is equal to the negative log-likelihood. The corresponding cost function, denoted ci.i.d. , is
defined as follows.

Cost function 1 (ci.i.d. ). For a given parametric family of distribution densities {f (·|θ)|θ ∈ Θ}
where Θ is a compact subset of Rp (for a certain p), the cost function ci.i.d. is defined by

b
X
ci.i.d. (ya..b ) := − sup log f (yt |θ). (C1)
θ t=a+1

Model M1 and the related cost function ci.i.d. encompasses a large number of change point methods.
Note that, in this context, the family of distributions must be known before performing the detection,
usually thanks to prior knowledge on the data. Historically, the Gaussian distribution was first used,
to model mean-shifts [52–54] and scale shifts [39, 50, 55]. A large part of the literature then evolved
towards other parametric distributions, most notably resorting to distributions from the general expo-
nential family [15, 25, 56].
From a theoretical point of view, asymptotic consistency, as described in Definition 1, has been demon-
strated, in the case of a single change point, first with Gaussian distribution (fixed variance), then for

9
 Change point detection

Cost function Search method Constraint

Parametric Non-parametric

Non-parametric
Maximum
Multiple linear Mahalanobis-type maximum Rank-based Kernel-based
likehood
model metric likehood detection detection
estimation
estimation

ci.i.d. , cL2 , cΣ , clinear , cAR , ckernel , crbf ,

cM cFb crank
cPoisson clinear,L1 cH,M

Figure 6: Typology of the cost functions described in Section 4.

several specific distributions, e.g. Gaussian with mean and scale shifts [3, 6, 51, 57], discrete distribu-
tions [49], etc. The case with multiple change points has been tackled later. For certain distributions
(e.g. Gaussian), the solutions of the change point detection problems (P1) (known number of change
points) and (P2) (unknown number of change points) have been proven to be asymptotically con-
sistent [58]. The general case of multiple change points and a generic distribution family has been
addressed decades after the change detection problem has been introduced: the solution of the change
point detection problem with a known number of changes and a cost function set to ci.i.d. is asymp-
totically consistent [59]. This is true if certain assumptions are satisfied: (i) the signal follows the
model (M1) for a distribution family that verifies some regularity assumptions (which are no different
from the assumptions needed for generic maximum likelihood estimation, without any change point)
and (ii) technical assumptions on the value of the cost function on homogeneous and heterogeneous
sub-signals. As an example, distributions from the exponential family satisfy those assumptions.

Related cost functions.. The general model (M1) has been applied with different families of distribu-
tions. We list below three notable examples and the associated cost functions: change in mean, change
in mean and scale, and change in the rate parameter of count data.

• The mean-shift model is the earliest and one of the most studied model in the change point
detection literature [2, 53, 60–62]. Here, the distribution is Gaussian, with fixed variance. In
other words, the signal y is simply a sequence of independent normal random variables with
piecewise constant mean and same variance. In this context, the cost function ci.i.d. becomes
cL2 , defined below. This cost function is also referred to as the quadratic error loss and has been
applied for instance on DNA array data [15] and geology signals [63].

10
 Cost function 2 (cL2 ). The cost function cL2 is given by
b
2
X
cL2 (ya..b ) := kyt − ȳa..b k2 (C2)
t=a+1

where ȳa..b is the empirical mean of the sub-signal ya..b .

• A natural extension to the mean-shift model consists in letting the variance abruptly change as
well. In this context, the cost function ci.i.d. becomes cΣ , defined below. This cost function
can be used to detect changes in the first two moments of random (not necessarily Gaussian)
variables, even though it is the Gaussian likelihood that is plugged in ci.i.d. [8, 49]. It has been
applied for instance on stock market time series [49], biomedical data [63], and electric power
consumption monitoring [64].

Cost function 3 (cΣ ). The cost function cΣ is given by

b
X
cΣ (ya..b ) := (b − a) log det Σ
b a..b + b −1 (yt − ȳa..b )
(yt − ȳa..b )0 Σ (C3)
a..b
t=a+1

where ȳa..b and Σ

b a..b are respectively the empirical mean and the empirical covariance matrix of
the sub-signal ya..b .

• Change point detection has also be applied on count data modelled by a Poisson distribu-
tion [39, 65]. More precisely, the signal y is a sequence of independent Poisson distributed
random variables with piecewise constant rate parameter. In this context, the cost function
ci.i.d. becomes cPoisson , defined below.

Cost function 4 (cPoisson ). The cost function cPoisson is given by

cPoisson (ya..b ) := −(b − a)ȳa..b log ȳa..b (C4)

where ȳa..b is the empirical mean of the sub-signal ya..b .

Remark 1. A model slightly more general than (M1) can be formulated by letting the signal samples to
be dependant and the distribution function f (·|θ) to change over time. This can in particular model the
presence of unwanted changes in the statistical properties of the signal (for instance in the statistical
structure of the noise [49]). The function f (·|θ) is replaced in (M1) by a sequence of distribution
functions ft (·|θ) which are not assumed to be identical for all indexes t. Changes in the functions
ft are considered nuisance parameters and only the variations of the parameter θ must be detected.
Properties on the asymptotic consistency of change point estimates can be obtained in this context. We
refer the reader to [49, 66] for theoretical results.

4.1.2. Piecewise linear regression

Piecewise linear models are often found, most notably in the econometrics literature, to detect so-
called “structural changes” [67–69]. In this context, a linear relationship between a response variable
and covariates exists, and this relationship changes abruptly at some unknown instants. Formally, the
observed signal y follows a piecewise linear model with change points located at the t∗k :

∀ t, t∗k < t ≤ t∗k+1 , yt = x0t uk + zt0 v + εt (k = 0, . . . , K ∗ ) (M2)

where the uk ∈ Rp and v ∈ Rq are unknown regression parameters and εt is noise. Under this setting,
the observed signal y is regarded as a univariate response variable (ie d = 1) and the signals x = {xt }Tt=1

11
and z = {zt }Tt=1 are observed covariates, respectively Rp -valued and Rq -valued. In this context, change
point detection can be carried out by fitting a linear regression on each segment of the signal. To that
end, the sum of costs is made equal to the sum of squared residuals. The corresponding cost function,
denoted clinear , is defined as follows.
Cost function 5 (clinear ). For a signal y (response variable) and covariates x and z, the cost func-
tion clinear is defined by

b
X
clinear (ya..b ) := min (yt − x0t u − zt0 v)2 . (C5)
u∈Rp ,v∈Rq
t=a+1

In the literature, Model (M2) is also known as a partial structural change model because the linear
relationship between y and x changes abruptly, while the linear relationship between y and z remains
constant. The pure structural change model is obtained by removing the term zt0 v from (M2). This
formulation generalizes several well-known models such as the autoregressive (AR) model [12, 70],
multiple regressions [69, 71], etc. A more general formulation of (M2) that can accommodate a
multivariate response variable y exists [72], but is more involved, from a notational standpoint.
From a theoretical point of view, piecewise linear models are extensively studied in the context of
change point detection by a series of important contributions [14, 67–71, 73–77]. When the number of
changes is known, the most general consistency result can be found in [14]. A multivariate extension
of this result has been demonstrated in [72]. As for the more difficult situation of an unknown number
of changes, statistical tests have been proposed for a single change [78] and multiple changes [74].
All of those results are obtained under various sets of general assumptions on the distributions of the
covariates and the noise. The most general of those sets can be found in [79]. Roughly, in addition
to some technical assumptions, it imposes the processes x and z to be weakly stationary within each
regime, and precludes the noise process to have a unit root.
Related cost functions.. In the rich literature related to piecewise linear models, the cost function
clinear has been applied and extended in several different settings. Two related cost functions are
listed below.
• The first one is clinear,L1 , which was introduced in order to accommodate certain noise distri-
butions with heavy tails [68, 73] and is defined as follows.

Cost function 6 (clinear,L1 ). For a signal y (response variable) and covariates x and z, the
cost function clinear,L1 is defined by

b
X
clinear,L1 (ya..b ) := min |yt − x0t u − zt0 v|. (C6)
u∈Rp ,v∈Rq
t=a+1

The difference between clinear,L1 and clinear lies in the norm used to measure errors: clinear,L1
is based on a least absolute deviations criterion, while clinear is based on a least squares criterion.
As a result, clinear,L1 is often applied on data with noise distributions with heavy tails [8, 25].
In practice, the cost function clinear,L1 is computationally less efficient than the cost function
clinear , because the associated minimization problem (C6) has no analytical solution. Never-
theless, the cost function clinear,L1 is often applied on economic and financial data [67–69]. For
instance, changes in several economic parameters of the G-7 growth have been investigated using
a piecewise linear model and clinear,L1 [80].
• The second cost function related to clinear has been introduced to deal with piecewise autoregres-
sive signals. The autoregressive model is a popular representation of random processes, where
each variable depends linearly on the previous variables. The associated cost function, denoted
cAR , is defined as follows.

12
 Cost function 7 (cAR ). For a signal y and an order p ≥ 1, the cost function cAR is defined by
b
2
X
cAR (ya..b ) := minp kyt − x0t uk (C7)
u∈R
t=a+1

where xt := [yt−1 , yt−2 , . . . , yt−p ] is the vector of lagged samples.

The piecewise autoregressive model is a special case of the generic piecewise linear model, where
the term zt0 v is removed (yielding a pure structural change model) and the covariate signal x is
equal to the signal of lagged samples. The resulting cost function cAR is able to detect shifts in
the autoregressive coefficients of a non-stationary process [21, 70]. This model has been applied
on EEG/ECG time series [72], functional magnetic resonance imaging (fMRI) time series [81]
and speech recognition tasks [12].

4.1.3. Mahalanobis-type metric

The cost function cL2 (C2), adapted for mean-shift detection, can be extended through the use of
Mahalanobis-type pseudo-norm. Formally, for any symmetric positive semi-definite matrix M ∈ Rd×d ,
the associated pseudo-norm k·kM is given by:
2
kyt kM := yt0 M yt (10)

for any sample yt . The resulting cost function cM is defined as follows.

Cost function 8 (cM ). The cost function cM , parametrized by a symmetric positive semi-definite
matrix M ∈ Rd×d , is given by
b
2
X
cM (ya..b ) := kyt − ȳa..b kM (C8)
t=a+1

where ȳa..b is the empirical mean of the sub-signal ya..b .

Intuitively, measuring distances with the pseudo-norm k·kM is equivalent to applying a linear trans-
formation on the data and using the regular (Euclidean) norm k·k. Indeed, decomposing the matrix
M = U 0 U yields:
2 2
kyt − ys kM = kU yt − U ys k . (11)
Originally, the metric matrix M was set equal to the inverse of the covariance matrix, yielding the
Mahalanobis metric [82], ie
M =Σ b −1 (12)
where Σ b is the empirical covariance matrix of the signal y. By using cM , shifts in the mean of the
transformed signal can be detected. In practice, the transformation U (or equivalently, the matrix M )
is chosen to highlight relevant changes. This cost function generalizes all linear transformations of the
data samples. In the context of change point detection, most of the transformations are unsupervised,
for instance principal component analysis or linear discriminant analysis [83]. Supervised strategies
are more rarely found, even though there exist numerous methods to learn a task-specific matrix M
in the context of supervised classification [83–85]. Those strategies fall under the umbrella of metric
learning algorithms. In the change point detection literature, there is only one work that proposes
a supervised procedure to calibrate a metric matrix M [30]. In this contribution, the authors use a
training set of annotated signals (meaning that an expert has provided the change point locations) to
learn M iteratively. Roughly, at each step, a new matrix M is generated in order to improve change
point detection accuracy on the training signals. However, using the cost function cM is not adapted to
certain applications, where a linear treatment of the data is insufficient. In that situation, a well-chosen
non-linear transformation of the data samples must be applied beforehand [30].

13
4.2. Non-parametric models
When the assumptions of parametric models are not adapted to the data at hand, non-parametric
change point detection methods can be more robust. Three major approaches are presented here, each
based on different non-parametric statistics, such as the empirical cumulative distribution function,
rank statistics and kernel estimation.

Signal model.. Assume that the observed signal y = {y1 , . . . , yT } is composed of independent random
variables, such that
K∗
X
yt ∼ Fk 1(t∗k < t ≤ t∗k+1 ) (M3)
k=0

where the t∗k are change point indexes and the Fk are cumulative distribution functions (c.d.f.), not
necessarily parametric as in (M1). Under this setting, the sub-signal yt∗k ..t∗k+1 , bounded by two change
points, is composed of iid variables with c.d.f. Fk . When the Fk belong to a known parametric distri-
bution family, change point detection is performed with the MLE approach described in Section 4.1.1,
which consists in applying the cost function ci.i.d. . However, this approach is not possible when the
distribution family is either non-parametric or not known beforehand.

4.2.1. Non-parametric maximum likelihood

The first non-parametric cost function example, denoted cFb , has been introduced for the single
change point detection problem in [86] and extended for multiple change points in [87]. It relies on the
empirical cumulative distribution function, estimated on sub-signals. Formally, the signal is assumed
to be univariate (ie d = 1) and the empirical cdf on the sub-signal ya..b is given by
b
 X 
1
∀u ∈ R, Fba..b (u) := 1(yt < u) + 0.5 × 1(yt = u) . (13)
b−a t=a+1

In order to derive a log-likelihood function that does not depend on the probability distribution of the
data, ie the f (·|θk ), the authors use the following fact: for a fixed u ∈ R, the empirical cdf Fb of n iid
random variables, distributed from a certain cdf F is such that nFb(u) ∼ Binomial(n, F (u)) [87]. This
observation, combined with careful summation over u, allows a distribution-free maximum likelihood
estimation. The resulting cost function cFb is defined as follows. Interestingly, this strategy was first
introduced to design non-parametric two-sample statistical tests, which were experimentally shown to
be more powerful than classical tests such as Kolmogorov-Smirnov and Cramr-von Mises [86, 88].
Cost function 9 (cFb ). The cost function cFb is given by

T b
X Fa..b (u) log Fba..b (u) + (1 − Fba..b (u)) log(1 − Fba..b (u))
cFb (ya..b ) := −(b − a) (C9)
u=1
(u − 0.5)(T − u + 0.5)

where the empirical cdf Fba..b is defined by (13).

From a theoretical point of view, asymptotic consistency of change point estimates is verified, when
the number of change points is either known or unknown [87]. However, solving either one of the
detection problems can be computationally intensive, because calculating the value of the cost function
cFb on one sub-signal requires to sum T terms, where T is the signal length. As a result, the total
complexity of change point detection is of the order of O(T 3 ) [87]. To cope with this computational
burden, several preliminary steps are proposed. For instance, irrelevant change point indexes can be
removed before performing the detection, thanks to a screening step [87]. Also, the cost function cFb
can be approximated, by summing, in (C9), over a few (carefully chosen) terms, instead of T terms
originally [89]. Thanks to those implementation techniques, the cost function cFb has been applied on
DNA sequences [87] and heart-rate monitoring signals [89].

14
4.2.2. Rank-based detection
In statistical inference, a popular strategy to derive distribution-free statistics is to replaced the
data samples by their ranks within the set of pooled observations [28, 90, 91]. In the context of change
point detection, this strategy has first been applied to detect a single change point [28, 29], and then
has been extended by [92] to find multiple change points. The associated cost function, denoted crank ,
is defined as follows. Formally, it relies on the centered Rd -valued “rank signal” r = {rt }Tt=1 , given by
T
X T +1
rt,j := 1(ys,j ≤ yt,j ) − , ∀1 ≤ t ≤ T, ∀1 ≤ j ≤ d. (14)
s=1
2

In other words, rt,j is the (centered) rank of the j th coordinate of the tth sample, ie yt,j , among the
{y1,j , y2,j , . . . , yT,j }.
Cost function 10 (crank ). The cost function crank is given by

0 b −1
crank (ya..b ) := −(b − a) r̄a..b Σ r r̄a..b (C10)

b r ∈ Rd×d is the following matrix

where the signal r is defined in (14) and Σ
T
b r := 1
X
Σ (rt + 1/2)0 (rt + 1/2). (15)
T t=1

Intuitively, crank measures changes in the joint behaviour of the marginal rank statistics of each
coordinate, which are contained in r. One of the advantages of this cost function is that it is invariant
under any monotonic transformation of the data. Several well-known statistical hypothesis testing
procedures are based on this scheme, for instance the Wilcoxon-Mann-Whitney test [93], the Friedman
test [94], the Kruskal-Wallis test [95], and several others [90, 91]. From a computational point of view,
two steps must be performed before the change point detection: the calculation of the rank statistics,
in O(dT log T ) operations, and the calculation of the matrix Σ b r , in O(d2 T + d3 ) operations. The
resulting algorithm has been applied on DNA sequences [92] and network traffic data [28, 29].

4.2.3. Kernel-based detection

A kernel-based method has been proposed by [96] to perform change point detection in a non-
parametric setting. To that end, the original signal y is mapped onto a reproducing Hilbert space
(rkhs) H associated with a user-defined kernel function k(·, ·) : Rd × Rd → R. The mapping function
φ : Rd → H onto this rkhs is implicitly defined by φ(yt ) = k(yt , ·) ∈ H, resulting in the following
inner-product and norm:
2
hφ(ys )|φ(yt )iH = k(ys , yt ) and kφ(yt )kH = k(yt , yt ) (16)

for any samples ys , yt ∈ Rd . The associated cost function, denoted ckernel , is defined as follows.
This kernel-based mapping is central to many machine learning developments such as support vector
machine or clustering [97, 98].
Cost function 11 (ckernel ). For a given kernel function k(·, ·) : Rd × Rd → R, the cost function
ckernel is given by
b
2
X
ckernel (ya..b ) := kφ(yt ) − µ̄a..b kH (C11)
t=a+1

where µ̄a..b ∈ H is the empirical mean of the embedded signal {φ(yt )}bt=a+1 and k·kH is defined in (16).
Remark 2 (Computing the cost function). Thanks to the well-known “kernel trick”, the explicit
computation of the mapped data samples φ(yt ) is not required to calculate the cost function value [99].

15
Indeed, after simple algebraic manipulations, ckernel (ya..b ) can be rewritten as follows:

b b
X 1 X
ckernel (ya..b ) = k(yt , yt ) − k(ys , yt ). (17)
t=a+1
b − a s,t=a+1

Remark 3 (Intuition behind the cost function). Intuitively, the cost function ckernel is able to
detect mean-shifts in the transformed signal {φ(yt )}t . Its use is motivated in the context of Model M3
by the fact that, under certain conditions on the kernel function, changes in the probability distribution
coincide with mean-shifts in the transformed signal. This connection has been investigated in several
works on kernel methods [97, 98, 100, 101]. Formally, let P denote a probability distribution defined
over Rd . Then there exists a unique element µP ∈ H [100], called the mean embedding (of P), such
that
µP = EX∼P [φ(X)]. (18)
In addition, the mapping P 7→ µP is injective (in which case the kernel is said to be characteristic),
meaning that
µP = µQ ⇐⇒ P = Q, (19)
where Q denotes a probability distribution defined over Rd . In order to determine if a kernel is
characteristic (and therefore, useful for change point detection), several conditions can be found in
the literature [97, 98, 100]. For instance, if a kernel k(·, ·) is translation invariant, meaning that
k(ys , yt ) = ψ(ys − yt ) ∀s, t, where ψ is a bounded continuous positive definite function on Rd , then it
is characteristic [100]. This condition is verified by the commonly used Gaussian kernel. As a conse-
quence, two transformed samples φ(ys ) and φ(yt ) are distributed around the same mean value if they
belong to the same regime, and around different mean-values if they each belong to two consecutive
regimes. To put it another way, a signal that follows (M3) is mapped by φ(·) to a random signal with
piecewise constant mean.

From a theoretical point of view, asymptotic consistency of the change point estimates has been
demonstrated for both a known and unknown number of change points in the recent work of [102].
This result, as well as an important oracle inequality on the sum of cost V (T ) [103], also holds in
a non-asymptotic setting. In addition, kernel change point detection was experimentally shown to
be competitive in many different settings, in an unsupervised manner and with very few parameters
to manually calibrate. For instance, the cost function ckernel was applied on the Brain-Computer
Interface (BCI) data set [96], on a video time series segmentation task [103], DNA sequences [99] and
emotion recognition [104].

Related cost functions.. The cost function ckernel can be combined with any kernel to accommodate
various types of data (not just Rd -valued signals). Notable examples of kernel functions include [101]:
• The linear kernel k(x, y) = hx|yi with x, y ∈ Rd .

• The polynomial kernel k(x, y) = (hx|yi + C)deg with x, y ∈ Rd , and C and deg are parameters.
2
• The Gaussian kernel k(x, y) = exp(−γ kx − yk ) with x, y ∈ Rd and γ > 0 is the so-called
bandwidth parameter.
• The χ2 -kernel k(x, y) = exp(−γ i [(xi − yi )2 /(xi + yi )]) with γ ∈ R a parameter. It is often
P
used for histogram data.
Arguably, the most commonly used kernels for numerical data are the linear kernel and the Gaussian
kernel. When combined with the linear kernel, the cost function ckernel is formally equivalent to cL2 .
As for the Gaussian kernel, the associated cost function, denoted crbf , is defined as follows.

16
Cost function 12 (crbf ). The cost function crbf is given by

b
1 X 2
crbf (ya..b ) := (b − a) − exp(−γ kys − yt k ) (C12)
b − a s,t=a+1

where γ > 0 is the so-called bandwidth parameter.

The parametric cost function cM (based on a Mahalanobis-type norm) can be extended to the non-
parametric setting through the use of a kernel. Formally, the Mahalanobis-type norm k·kH,M in the
feature space H is defined by
2
kφ(ys ) − φ(yt )kH,M = (φ(ys ) − φ(yt ))0 M (φ(ys ) − φ(yt )) (20)

where M is a (possibly infinite dimensional) symmetric positive semi-definite matrix defined on H.

The associated cost function, denoted cH,M , is defined below. Intuitively, using cH,M instead of cM
introduces a non-linear treatment of the data samples.
Cost function 13 (cH,M ). For a given kernel function k(·, ·) : Rd × Rd → R and M a symmetric
positive semi-definite matrix defined on the associated rkhs H, the cost function cH,M is given by
b
2
X
cH,M (ya..b ) := kφ(yt ) − µ̄a..b kH,M (C13)
t=a+1

where µa..b is the empirical mean of the transformed sub-signal {φ(yt )}bt=a+1 and k·kH,M is defined
in (20).

4.3. Summary table

Reviewed cost functions (parametric and non-parametric) are summarized in Table 1. For each
cost, the name, expression and parameters of interest are given.

17
 Name c(ya..b ) Parameters
Pb
ci.i.d. (C1) − supθ t=a+1 log f (yt |θ) θ: changing parameter; density
function: f (·|θ)
Pb
cL2 (C2) t=a+1 kyt − ȳa..b k22 ȳa..b : empirical mean of ya..b

cΣ (C3) b a..b + Pb
(b − a) log det Σ 0 b −1 b a..b : empirical covariance of ya..b
t=a+1 (yt − ȳa..b ) Σa..b (yt − ȳa..b ) Σ

cPoisson (C4) −(b − a)ȳa..b log ȳa..b ȳa..b : empirical mean of ya..b
Pb
clinear (C5) minu∈Rp ,v∈Rq t=a+1 (yt − x0t u − zt0 v)2 xt ∈ Rp , zt ∈ Rq : covariates
Pb
clinear,L (C6) minu∈Rp ,v∈Rq t=a+1 |yt − x0t u − zt0 v| xt ∈ Rp , zt ∈ Rq : covariates
1
Pb
cAR (C7) minu∈Rp t=a+1 (yt − x0t u)2 xt = [yt−1 , yt−2 , . . . , yt−p ]:
lagged samples
Pb
cM (C8) t=a+1 kyt − ȳa..b k2M M ∈ Rd×d : positive semi-definite
matrix

PT b
F b b b
a..b (u) log Fa..b (u)+(1−Fa..b (u)) log(1−Fa..b (u)) b: empirical c.d.f. (13)
cFb (C9) −(b − a) u=1 (u−0.5)(T −u+0.5)
F

crank (C10) 0
−(b − a) r̄a..b b −1 r̄a..b
Σ b r : empirical
r: rank signal (14); Σ
r
covariance of r (15)
Pb Pb
ckernel (C11) t=a+1 k(yt , yt ) − 1
b−a s,t=a+1 k(ys , yt ) k(·, ·) : Rd × Rd 7→ R: kernel
function
Pb
crbf (C12) (b − a) − 1
b−a s,t=a+1 exp(−γ kys − yt k2 ) γ > 0: bandwidth parameter
Pb
cH,M (C13) t=a+1 kyt − ȳa..b k2H,M M : positive semi-definite matrix
(in the feature space H)

Table 1: Summary of cost reviewed functions

5. Search methods

This section presents the second defining element of change detection methods, namely the search
method. Reviewed search methods are organized in two general categories, as shown on Figure 7:
optimal methods, that yield the exact solution to the discrete optimization of (P1) and (P2), and the
approximate methods, that yield an approximate solution. Described algorithms can be combined with
cost functions from Section 4. Note that, depending on the chosen cost function, the computational
complexity of the complete algorithm changes. As a consequence, in the following, complexity analysis
is done with the assumption that applying the cost function on a sub-signal requires O(1) operations.
Also, the practical implementations of the most important algorithms are given in pseudo-code.

5.1. Optimal detection

Optimal detection methods find the exact solutions of Problem 1 (P1) and Problem 2 (P2). A
naive approach consists in enumerating all possible segmentations of a signal, and returning the one
that minimizes the objective function.
However, for (P1), minimization is carried out over the set
T −1
{T s.t. |T | = K} (which contains K−1 elements), and for (P2), over the set {T s.t. 1 ≤ |T | < T }
PT −1 T −1

(which contains K=1 K−1 elements). This makes exhaustive enumeration impractical, in both
situations. We describe in this section two major approaches to efficiently find the exact solutions of
(P1) and (P2).

5.1.1. Solution to Problem 1 (P1): Opt

In (P1), the number of change points to detect is fixed to a certain K ≥ 1. The optimal solution to
this problem can be computed efficiently, thanks to a method based on dynamic programming. The

18
 Change point detection

Cost function Search method Constraint

Optimal Approximate

Opt, Pelt

Binary Bottom-up
Window-sliding
segmentation segmentation

Win BinSeg BotUp

Figure 7: Typology of the search methods described in Section 5.

algorithm, denoted Opt, relies on the additive nature of the objective function V (·) to recursively solve
sub-problems. Precisely, Opt is based on the following observation:
K
X
min V (T , y = y0..T ) = min c(ytk ..tk+1 )
|T |=K 0=t0 <t1 <···<tK <tK+1 =T
k=0
 K−1 
(21)
X
= min c(y0..t ) + min c(ytk ..tk+1 )
t≤T −K t=t0 <t1 <···<tK−1 <tK =T
k=0
 
= min c(y0..t ) + min V (T , yt..T )
t≤T −K |T |=K−1

Intuitively, Equation 21 means that the first change point of the optimal segmentation is easily com-
puted if the optimal partitions with K − 1 elements of all sub-signals yt..T are known. The complete
segmentation is then computed by recursively applying this observation. This strategy, described in
detail in Algorithm 1, has a complexity of the order of O(KT 2 ) [105, 106]. Historically, Opt was intro-
duced for a non-related problem [107] and later applied to change point detection, in many different
contexts, such as EEG recordings [66, 108], DNA sequences [99, 109], tree growth monitoring [20],
financial time-series [49, 76], radar waveforms [110], etc.

19
Algorithm 1 Algorithm Opt
Input: signal {yt }Tt=1 , cost function c(·), number of regimes K ≥ 2.
for all (u, v), 1 ≤ u < v ≤ T do
Initialize C1 (u, v) ← c({yt }vt=u ).
end for
for k = 2, . . . , K − 1 do
for all u, v ∈ {1, . . . , T }, v − u ≥ k do
Ck (u, v) ← min Ck−1 (u, t) + C1 (t + 1, v)
u+k−1≤t<v
end for
end for
Initialize L, a list with K elements.
Initialize the last element: L[K] ← T .
Initialize k ← K.
while k > 1 do
s ← L(k)
t∗ ← argmink−1≤t<s Ck−1 (1, t) + C1 (t + 1, s)
L(k − 1) ← t∗
k ←k−1
end while
Remove T from L
Output: set L of estimated breakpoint indexes.

Related search methods.. Several extensions of Opt have been proposed in the literature. The proposed
methods still find the exact solution to (P1).
- The first extension is the “forward dynamic programming” algorithm [20]. Contrary to Opt,
which returns a single partition, the “forward dynamic programming” algorithm computes the
top L (L ≥ 1) most probable partitions (ie with lowest sum of costs). The resulting computational
complexity is O(LKT 2 ) where L is the number of computed partitions. This method is designed
as a diagnostic tool: change points present in many of the top partitions are considered very likely,
while change points present in only a few of the top partitions might not be as relevant. Thanks
to “forward dynamic programming”, insignificant change points are trimmed and overestimation
of the number of change point is corrected [20], at the expense of a higher computational burden.
It is applied on tree growth monitoring time series [20] that are relatively short with around a
hundred samples.
- The “pruned optimal dynamic programming” procedure [109] is an extension of Opt that relies
on a pruning rule to discard indexes that can never be change points. Thanks to this trick, the
set of potential change point indexes is reduced. All described cost functions can be plugged into
this method. As a result, longer signals can be handled, for instance long array-based DNA copy
number data (up to 106 samples, with the quadratic error cost function) [109]. However, worst
case complexity remains of the order of O(KT 2 ).

5.1.2. Solution to Problem 2 (P2): Pelt

In (P2), the number of changes point is unknown, and the objective function to minimize is the pe-
nalized sum of costs. A naive approach consists in applying Opt for K = 1, . . . , Kmax for a sufficiently
large Kmax , then choosing among the computed segmentations the one that minimizes the penalized
problem. This would prove computational cumbersome because of the quadratic complexity of the
resolution method Opt. Fortunately a faster method exists for a general class of penalty functions,
namely linear penalties. Formally, linear penalties are linear functions of the number of change points,
meaning that
pen(T ) = β|T | (22)

20
where β > 0 is a smoothing parameter. (More details on such penalties can be found in Section 6.1.)
The algorithm Pelt (for “Pruned Exact Linear Time”) [111] was introduced to find the exact solution
of (P2), when the penalty is linear (22). This approach considers each sample sequentially and, thanks
to an explicit pruning rule, may or may not discard it from the set of potential change points. Precisely,
for two indexes t and s (t < s < T ), the pruning rule is given by:
   
if min V (T , y0..t ) + β|T | + c(yt..s ) ≥ min V (T , y0..s ) + β|T | holds,
T T

then t cannot be the last change point prior to T. (23)

This results in a considerable speed-up: under the assumption that regime lengths are randomly drawn
from a uniform distribution, the complexity of Pelt is of the order O(T ). The detailed algorithm can
be found in Algorithm 2. An extension of Pelt is described in [9] to solve the linearly penalized change
point detection for a range of smoothing parameter values [βmin , βmax ]. Pelt has been applied on
DNA sequences [16, 17], physiological signals [89], and oceanographic data [111].

Algorithm 2 Algorithm Pelt

Input: signal {yt }Tt=1 , cost function c(·), penalty value β.
Initialize Z a (T + 1)-long array; Z[0] ← −β.
Initialize L[0] ← ∅.
Initialize χ ← {0}. . Admissible indexes.
for t = 1, . . . , T do 
t̂ ← argmins∈χ Z[s] + c(ys..t ) + β .
 
Z[t] ← Z[t̂] + c(yt̂..t ) + β
L[t] ← L[t̂] ∪ {t̂}.
χ ← {s ∈ χ : Z[s] + c(ys..t ) ≤ Z[t]} ∪ {t}
end for
Output: set L[T ] of estimated breakpoint indexes.

5.2. Approximate detection

When the computational complexity of optimal methods is too great for the application at hand,
one can resort to approximate methods. In this section, we describe three major types of approximate
segmentation algorithms, namely window-based methods, binary segmentation and bottom-up seg-
mentation. All described procedures fall into the category of sequential detection approaches, meaning
that they return a single change point estimate t̂(k) (1 ≤ t̂(k) < T ) at the k-th iteration. (In the
following, the subscript ·(k) refers to the k-th iteration of a sequential algorithm.) Such methods can
be used to solve (approximately) either (P1) or (P2). Indeed, if the number K ∗ of changes is known,
K ∗ iterations of a sequential algorithm are enough to retrieve a segmentation with the correct number
of changes. If K ∗ is unknown, the sequential algorithm is run until an appropriate stopping criterion
is met.

5.2.1. Window sliding

The window-sliding algorithm, denoted Win, is a fast approximate alternative to optimal methods.
It consists in computing the discrepancy between two adjacent windows that slide along the signal y.
For a given cost function c(·), this discrepancy between two sub-signals is given by

d(ya..t , yt..b ) = c(ya..b ) − c(ya..t ) − c(yt..b ) (1 ≤ a < t < b ≤ T ). (24)

When the two windows cover dissimilar segments, the discrepancy reaches large values, resulting in a
peak. In other other words, for each index t, Win measures the discrepancy between the immediate past

21
 Original Signal

Discrepancy Curve

Peak Detection

Figure 8: Schematic view of Win

(“left window”) and the immediate future (“right window”). Once the complete discrepancy curve has
been computed, a peak search procedure is performed to find change point indexes. The complete Win
algorithm is given in Algorithm 3 and a schematic view is displayed on Figure 8. The main benefits
of Win are its low complexity (linear in the number of samples) and ease of implementation.

Algorithm 3 Algorithm Win

Input: signal {yt }Tt=1 , cost function c(·), half-window width w, peak search procedure PKSearch.
Initialize Z ← [0, 0, . . . ] a T -long array filled with 0. . Score list.
for t = w, . . . , T − w do
p ← (t − w)..t.
q ← t..(t + w).
r ← (t − w)..(t + w).
Z[t] ← c(yr ) − [c(yp ) + c(yq )].
end for
L ← PKSearch(Z) . Peak search procedure.
Output: set L of estimated breakpoint indexes.

In the literature, the discrepancy measure d(·, ·) is often derived from a two-sample statistical test
(see Remark 4), and not from a cost function, as in (24). However, the two standpoints are generally
equivalent: for instance, using cL2 , ci.i.d. or ckernel is respectively equivalent to applying a Student
t-test [3], a generalized likelihood ratio (GLR) [112] test and a kernel Maximum Mean Discrepancy
(MMD) test [98]. As a consequence, practitioners can capitalize on the vast body of work in the field
of statistical tests to obtain asymptotic distributions for the discrepancy measure [28, 29, 98, 113],
and sensible calibration strategies for important parameters of Win (such as the window size or the
peak search procedure). Win has been applied in numerous contexts: for instance, on biological
signals [11, 114–117], on network data [28, 29], on speech time series [10, 11, 118] and on financial time
series [3, 119, 120]. It should be noted that certain window-based detection methods in the literature
rely on a discrepancy measure which is not related to a cost function, as in (24) [11, 121–123]. As a
result, those methods, initially introduced in the online detection setting, cannot be extended to work

22
with optimal algorithms (Opt, Pelt).
Remark 4 (Two-sample test). A two-sample test (or homogeneity test) is a statistical hypothesis
testing procedure designed to assess whether two populations of samples are identical in distribution.
Formally, consider two sets of iid Rd -valued random samples {xt }t and {zt }t . Denote by Px the
distribution function of the xt and by Pz , the distribution function of the zt . A two-sample test
procedure compares the two following hypotheses:

H0 : Px = Pz
(25)
H1 : Px 6= Pz .

A general approach is to consider a probability (pseudo)-metric d(·, ·) on the space of probability dis-
tributions on Rd . Well-known examples of such a metric include the Kullback-Leibler divergence, the
Kolmogorov-Smirnov distance, the Maximum Mean Discrepancy (MMD), etc. Observe that, under the
null hypothesis, d(Px , Pz ) = 0. The testing procedure consists in computing the empirical estimates
bx and P
P bz and rejecting H0 for “large” values of the statistics d(P bz ). This general formulation
bx , P
relies on a consistent estimation of arbitrary distributions from a finite number of samples. In the
parametric setting, additional assumptions are made on the distribution functions: for instance, Gaus-
sian assumption [3, 63, 113], exponential family assumption [15, 124], etc. In the non-parametric
setting, the distributions are only assumed to be continuous. They are not directly estimated; instead,
the statistics d(P bz ) are computed [11, 90, 98, 122].
bx , P
In the context of single change point detection, the two-sample test setting is adapted to assess whether
a distribution change has occurred at some instant in the input signal. Practically, for a given index t,
the homogeneity test is performed on the two populations {ys }s≤t and {ys }s>t . The estimated change
point location is given by
t̂ = argmaxt d(P b•>t )
b•≤t , P (26)
where P b•>t are the empirical distributions of respectively {ys }s≤t and {ys }s>t .
b•≤t and P

5.2.2. Binary segmentation

Binary segmentation, denoted BinSeg, is a well-known alternative to optimal methods [53], because
it is conceptually simple and easy to implement [63, 111, 125]. BinSeg is a greedy sequential algorithm,
outlined as follows. The first change point estimate t̂(1) is given by

t̂(1) := argmin1≤t<T −1 c(y0..t ) + c(yt..T ) . (27)

| {z }
V (T ={t})

This operation is “greedy”, in the sense that it searches the change point that lowers the most the
sum of costs. The signal is then split in two at the position of t̂(1) ; the same operation is repeated
on the resulting sub-signals until a stopping criterion is met. A schematic view of the algorithm is
displayed on Figure 9 and an implementation is given in Algorithm 4. The complexity of BinSeg is of
the order of O(T log T ). This low complexity comes at the expense of optimality: in general, BinSeg’s
output is only an approximation of the optimal solution. As argued in [111, 126], the issue is that
the estimated change points t̂(k) are not estimated from homogeneous segments and each estimate
depends on the previous ones. Change points that are close are imprecisely detected especially [8].
Applications of BinSeg range from financial time series [7, 63, 113, 126, 127] to context recognition
for mobile devices [128] and array-based DNA copy number data [19, 125, 129].

Related search methods.. Several extensions of BinSeg have been proposed to improve detection ac-
curacy.
- Circular binary segmentation [125] is a well-known extension of BinSeg. This method is also a
sequential detection algorithm that splits the original at each step. Instead of searching for a

23
 Step 0

Step 1

Step 2

Figure 9: Schematic example of BinSeg

single change point in each sub-signal, circular binary segmentation searches two change points.
Within each treated sub-segment, it assumes a so-called “epidemic change model”: the parameter
of interest shifts from one value to another at the first change point and returns to the original
value at the second change point. The algorithm is dubbed “circular” because, under this model,
the sub-segment has its two ends (figuratively) joining to form a circle. Practically, this method
has been combined with cL2 C2, to detect changes in the mean of array-based DNA copy number
data [125, 130, 131]. A faster version of the original algorithm is described in [132].

- Another extension of BinSeg is the wild binary segmentation algorithm [127]. In a nutshell, a
single point detection is performed on multiple intervals with start and end points that are drawn
uniformly. Small segments are likely to contain at most one change but have lower statistical
power, while the opposite is true for long segments. After a proper weighting of the change score
to account for the differences on sub-signals’ length, the algorithm returns the most “pronounced”
ones, ie those that lower the most the sum of costs. An important parameter of this method is
the number of random sub-segments to draw. Wild binary search is combined with cL2 C2 to
detect mean-shifts of univariate piecewise constant signals (up to 2000 samples) [127].

5.2.3. Bottom-up segmentation

Bottom-up segmentation, denoted BotUp, is the natural counterpart of BinSeg. Contrary to
BinSeg, BotUp starts by splitting the original signal in many small sub-signals and sequentially merges
them until there remain only K change points. At every step, all potential change points (indexes
separating adjacent sub-segments) are ranked by the discrepancy measure d(·, ·), defined in 24, between
the segments they separate. Change points with the lowest discrepancy are then deleted, meaning that
the segments they separate are merged. BotUp is often dubbed a “generous” method, by opposition to
BinSeg, which is “greedy” [133]. A schematic view of the algorithm is displayed on Figure 10 and an
implementation is provided in Algorithm 5. Its benefits are its linear computational complexity and
conceptual simplicity. However, if a true change point does not belong to the original set of indexes,
BotUp never considers it. Moreover, in the first iterations, the merging procedure can be unstable
because it is performed on small segments, for which statistical significance is smaller. In the litera-
ture, BotUp is somewhat less studied than its counterpart, BinSeg: no theoretical convergence study

24
Algorithm 4 Algorithm BinSeg
Input: signal {yt }Tt=1 , cost function c(·), stopping criterion.
Initialize L ← { }. . Estimated breakpoints.
repeat
k ← |L|. . Number of breakpoints
t0 ← 0 and tk+1 ← T . Dummy variables.
if k > 0 then
Denote by ti (i = 1, . . . , k) the elements (in ascending order) of L, ie L = {t1 , . . . , tk }.
end if
Initialize G a (k + 1)-long array. . list of gains
for i = 0, . . . , k do
G[i] ← c(yti ..ti+1 ) − min [c(yti ..t ) + c(yt..ti+1 )] .
ti <t<ti+1
end for
bi ← argmaxi G[i]
t̂ ← argmintbi <t<tbi+1 [c(ytbi ..t ) + c(yt..tbi+1 )].
L ← L ∪ {t̂}
until stopping criterion is met.
Output: set L of estimated breakpoint indexes.

is available. It has been applied on speech time series to detect mean and scale shifts [119]. Besides,
the authors of [133] have found that BotUp outperforms BinSeg on ten different data sets such as
physiological signals (ECG), financial time-series (exchange rate), industrial monitoring (water levels),
etc.

6. Estimating the number of changes

This section presents the third defining element of change detection methods, namely the constraint
on the number of change points. Here, the number of change points is assumed to be unknown (P2).
Existing procedures are organized by the penalty function that they are based on. Common heuristics
are also described. The organization of this section is schematically shown in Figure 11.

6.1. Linear penalty

Arguably the most popular choice of penalty [111], the linear penalty (also known as l0 penalty)
generalizes several well-known criteria from the literature such as the Bayesian Information Criterion
(BIC) and the Akaike Information Criterion (AIC) [134, 135]. The linear penalty, denoted penl0 , is
formally defined as follows.
Penalty 1 (penl0 ). The penalty function penl0 is given by

penl0 (T ) := β|T | (28)

where β > 0 is the smoothing parameter.

Intuitively, the smoothing parameter controls the trade-off between complexity and goodness-of-fit
(measured by the sum of costs): low values of β favour segmentations with many regimes and high
values of β discard most change points.

Calibration.. From a practical standpoint, once the cost function has been chosen, the only parameter
to calibrate is the smoothing parameter. Several approaches, based on model selection, can be found
in the literature: they assume a model on the data, for instance (M1), (M2), (M3), and choose a value
of β that optimizes a certain statistical criterion. The best-known example of such an approach is

25
 Step 0

Steps 1, 2, . . .

step 8 step 1 step 4 step 2 step 3 step 5 step 7

step 6

Result

Figure 10: Schematic view of BotUp

BIC, which aims at maximizing the constrained log-likelihood of the model. The exact formulas of
several linear penalties, derived from model selection procedures, are given the following paragraph.
Conversely, when no model is assumed, different heuristics are applied to tune the smoothing parame-
ter. For instance, one can use a procedure based on cross-validation [136] or the slope heuristics [137].
In [138, 139], supervised algorithms are proposed: the chosen β is the one that minimizes an approxi-
mation of the segmentation error on an annotated set of signals.

Related penalties.. A number of model selection criteria are special cases of the linear penalty penl0 .
For instance, under Model (M1) (iid with piecewise constant distribution), the constrained likelihood
that is derived from the BIC and the penalized sum of costs are formally equivalent, upon setting
c = ci.i.d. and pen = penBIC , where penBIC is defined as follows.
Penalty 2 (penBIC ). The penalty function penBIC is given by
p
penBIC (T ) := log T |T | (29)
2
where p ≥ 1 is the dimension of the parameter space in (M1).
In the extensively studied model of an univariate Gaussian signal, with fixed variance σ 2 and piecewise
constant mean, the penalty penBIC becomes penL2 , defined below. Historically, it was one of the first
penalties introduced for change point detection [134, 140].

Penalty 3 (penBIC,L2 ). The penalty function penBIC,L2 is given by

penBIC,L2 (T ) := σ 2 log T |T |. (30)

where σ is the standard deviation and T is the number of samples.

In the same setting, AIC, which is a generalization of Mallows’ Cp [62], also yields a linear penalty,
namely penAIC,L2 , defined as follows.

26
Algorithm 5 Algorithm BotUp
Input: signal {yt }Tt=1 , cost function c(·), stopping criterion, grid size δ > 2.
Initialize L ← {δ, 2δ, . . . , (bT /δc − 1) δ}. . Estimated breakpoints.
repeat
k ← |L|. . Number of breakpoints
t0 ← 0 and tk+1 ← T . Dummy variables.
Denote by ti (i = 1, . . . , k) the elements (in ascending order) of L, ie L = {t1 , . . . , tk }.
Initialize G a (k − 1)-long array. . list of gains
for i = 1, . . . , k − 1 do
G[i − 1] ← c(yti−1 ..ti+1 ) − [c(yti−1 ..ti ) + c(yti ..ti+1 )] .
end for
bi ← argmini G[i]
Remove tbi+1 from L.
until stopping criterion is met.
Output: set L of estimated breakpoint indexes.

Penalty 4 (penAIC,L2 ). The penalty function penAIC,L2 is given by

penAIC,L2 (T ) := σ 2 |T |. (31)

where σ is the standard deviation.

6.2. Fused lasso

For the special case where the cost function is cL2 , a faster alternative to penl0 can be used. To
that end, the l0 penalty is relaxed to a l1 penalty [18, 48]. The resulting penalty function, denoted
penl1 , is defined as follows.
Penalty 5 (penl1 ). The penalty function penl1 is given by

|T |
X
penl1 (T ) := β ȳtk−1 ..tk − ȳtk ..tk+1 1
(32)
k=1

where β > 0 is the smoothing parameter, the tk are the elements of T and ȳtk−1 ..tk is the empirical
mean of sub-signal ytk−1 ..tk .
This relaxation strategy (from l0 to l1 ) is shared with many developments in machine learning, for
instance sparse regression, compressive sensing, sparse PCA, dictionary learning [83], where penl1 is
also referred to as the fused lasso penalty. In numerical analysis and image denoising, it is also known as
the total variation regularizer [13, 18, 48]. Thanks to this relaxation, the optimization of the penalized
sum of costs (1) in (P2) is transformed into a convex optimization problem, which can be solved
efficiently using Lars (for “least absolute shrinkage and selection operator”) [18, 48]. The resulting
complexity is of this order of O(T log T ) [83, 141]. From a theoretical standpoint, under the mean-shift
model (piecewise constant signal with Gaussian white noise), the estimated change point fractions are
asymptotically consistent [48]. This result is demonstrated for an appropriately converging sequence
of values of β. This consistency property is obtained even though classical assumptions from the Lasso
regression framework (such as the irrepresentable condition) are not satisfied [48]. In the literature,
penl1 , combined with cL2 , is applied on DNA sequences [16, 18], speech signals [12] and climatological
data [142].

6.3. Complex penalties

Several other penalty functions can be found in the literature. However they are more complex, in
the sense that the optimization of the penalized sum of cost is not tractable. In practice, the solution

27
 Change point detection

Cost function Search method Constraint

Known K? Unknown K?

Penalty l0 Penalty l1 Other methods

penl0 , penBIC , Stopping

penBIC,L2 , penl1 criterion,
penBIC,L2 penLeb ,
penmBIC

Figure 11: Typology of the constraints (on the number of change points) described in Section 6.

is found by computing the optimal segmentations with K change points, with K = 1, 2, . . . , Kmax for a
sufficiently large Kmax , and returning the one that minimizes the penalized sum of costs. When possi-
ble, the penalty can also be approximated by a linear penalty, in which case, Pelt can be used. In this
section, we describe two examples of complex penalties. Both originate from theoretical considerations,
under the univariate mean-shift model, with the cost function cL2 . The first example is the modified
BIC criterion (mBIC) [143], which consists in maximizing the asymptotic posterior probability of the
data. The resulting penalty function, denoted penmBIC , depends on the number and repartition of
the change point indexes: intuitively, it favours evenly spaced change points.
Penalty 6 (penmBIC ). The penalty function penmBIC is given by

|T |+1
X tk+1 − tk
penmBIC (T ) := 3|T | log T + log( ) (33)
T
k=0

where the tk are the elements of T .

In [144], a model selection procedure leads to another complex penalty function, namely penLeb . Upon
using this penalty function, the penalized sum of costs satisfied a so-called oracle inequality, which
holds in a non-asymptotic setting, contrary to the other penalties previously described.
Penalty 7 (penLeb ). The cost function penLeb is given by

|T | + 1 2 |T | + 1
penLeb (T ) := σ (a1 log + a2 ) (34)
T T
where a1 > 0 and a2 > 0 are positive parameters and σ 2 is the noise variance.

7. Summary table

This literature review is summarized in Table 2. When applicable, each publication is associated
with a search method (such as Opt, Pelt, BinSeg or Win); this is a rough categorization rather than an

28
exact implementation. Note that Pelt (introduced in 2012) is sometimes associated with publications
prior to 2012. It is because some linear penalties [62, 143] were introduced long before Pelt was, and
authors then resorted to quadratic (at best) algorithms. Nowadays, the same results can be obtained
faster with Pelt. A guide of computational complexity is also provided. Quadratic methods are
the slowest and have only one star while linear methods are given three stars. Algorithms for which
the number of change points is an explicit input parameter work under the “known K” assumption.
Algorithms that can be used even if the number of change points is unknown work under the “unknown
K” assumption. (Certain methods can accommodate both situations.)

29
 Publication Search method Cost function Known K Scalability (wrt T ) Package Additional information
Yes No
Sen and Srivastava (1975), Vostrikova (1981) BinSeg cL 3 - HHH 3
2
Yao (1988) Opt cL - 3 HII - Bayesian information criterion (BIC)
2
Basseville and Nikiforov (1993) Opt ci.i.d. , cL - - HHH - single change point
2
Bai (1994), Bai and Perron (2003) Opt clinear,L - - HHI - single change point
2
Bai (1995) Opt clinear,L - - HHI - single change point
1
Lavielle (1998) Opt cAR 3 - HII -
Bai (2000) Opt cAR 3 - HII -
Birgé and Massart (2001), Birgé and Massart (2007) Opt cL - 3 HII - model selection
2
Bai and Perron (2003) Opt cL 3 - HII -
2
Olshen et al. (2004), Venkatraman and Olshen (2007) BinSeg cL 3 3 HHH 3
2
Lebarbier (2005) Opt cL - 3 HII - model selection
2
Desobry et al. (2005) Win ckernel - 3 HHH - dissimilarity measure (one-class SVM), see Remark 4
Harchaoui and Cappé (2007) Opt ckernel , crbf 3 - HII -

Zhang and Siegmund (2007) Pelt cL - 3 HHI - modified BIC

2
Harchaoui et al. (2009) Win - 3 3 HHH - dissimilarity measure (Fisher discriminant), see Remark 4
Lévy-Leduc and Roueff (2009), Lung-Yut-Fong et al. (2012) Win crank 3 3 HHH 3 dissimilarity measure (rank-based), see Remark 4
Bai (2010) Opt cL , c Σ - - HHI - single change point
2
Vert and Bleakley (2010) Fused Lasso cL - 3 HHH - Tikhonov regularization
2
Harchaoui and Lévy-Leduc (2010) Fused Lasso cL - 3 HHH - total variation regression (penl )
2 1
Arlot et al. (2012) Opt ckernel , crbf 3 3 HII -

Killick et al. (2012) Pelt any c(·) - 3 HHI 3

Angelosante and Giannakis (2012) Fused Lasso cAR - 3 HHH - Tikhonov regularization
Liu et al. (2013) Win - - 3 HHH - dissimilarity measure (density ratio), see Remark 4
Hocking et al. (2013) Pelt cL - 3 HHI - supervised method to learn a penalty level (penl )
2 0
Fryzlewicz (2014) BinSeg cL 3 3 HHH 3 univariate signal
2
Lajugie et al. (2014) Opt cM 3 - HII - supervised method to learn a suitable metric
Frick et al. (2014) BinSeg ci.i.d. 3 3 HHH 3 exponential distributions family
Lung-Yut-Fong et al. (2015) Opt crank 3 - HII 3
Garreau and Arlot (2017) Pelt ckernel , crbf 3 3 HII -

Haynes et al. (2017) Pelt any c(·) - 3 HHI -

Chakar et al. (2017) Pelt cAR 3 3 HII 3

Table 2: Summary table of literature review.

 Input signal Evalutation
Change detection
Simulated signal Display

Search method Cost function Constraint

User’s signal Metrics

Figure 12: Schematic view of the ruptures package.

8. Presentation of the Python package

Most of the approaches presented in this article are included in a Python scientific library for
multiple change point detection in multivariate signals called ruptures [37]. The ruptures library is
written in pure Python and available on Mac OS X, Linux and Windows platforms. Source code is
available from [37] under the BSD license and deployed with a complete documentation that includes
installation instructions and explanations with code snippets on advance use.
A schematic view is displayed on Figure 12. Each block of this diagram is described in the following
brief overview of ruptures’ features.

• Search methods Our package includes the main algorithms from the literature, namely dy-
namic programming, detection with a l0 constraint, binary segmentation, bottom-up segmenta-
tion and window-based segmentation. This choice is the result of a trade-off between exhaus-
tiveness and adaptiveness. Rather than providing as many methods as possible, only algorithms
which have been used in several different settings are included. In particular, numerous “mean-
shift only” detection procedures were not considered. Implemented algorithms have sensible
default parameters that can be changed easily through the functions’ interface.
• Cost functions Cost functions are related to the type of change to detect. Within ruptures,
one has access to parametric cost functions that can detect shifts in standard statistical quan-
tities (mean, scale, linear relationship between dimensions, autoregressive coefficients, etc.) and
non-parametric cost functions (kernel-based or Mahalanobis-type metric) that can, for instance,
detect distribution changes [30, 96].
• Constraints All methods can be used whether the number of change points is known or not. In
particular, ruptures implements change point detection under a cost budget and with a linear
penalty term [17, 111].
• Evaluation Evaluation metrics are available to quantitatively compare segmentations, as well
as a display module to visually inspect algorithms’ performances.
• Input Change point detection can be performed on any univariate or multivariate signal that
fits into a Numpy array. A few standard non-stationary signal generators are included.
• Consistent interface and modularity Discrete optimization methods and cost functions are
the two main ingredients of change point detection. Practically, each is related to a specific
object in the code, making the code highly modular: available optimization methods and cost
functions can be connected and composed. An appreciable by-product of this approach is that
a new contribution, provided its interface follows a few guidelines, can be integrated seamlessly
into ruptures.

31
 • Scalability Data exploration often requires to run several times the same methods with different
sets of parameters. To that end, a cache is implemented to keep intermediate results in memory,
so that the computational cost of running the same algorithm several times on the same signal
is greatly reduced. We also add the possibility for a user with speed constraints to sub-sample
their signals and set a minimum distance between change points.

9. Conclusion

In this article, we have reviewed numerous methods to perform change point detection, organized
within a common framework. Precisely, all methods are described as a collection of three elements: a
cost function, a search method and a constraint on the number of changes to detect. This approach is
intended to facilitate prototyping of change point detection methods: for a given segmentation task,
one can pick among the described elements to design an algorithm that fits its use-case. Most detection
procedures described above are available within the Python language from the package ruptures [37],
which is the most comprehensive change point detection library. Its consistent interface and modularity
allow painless comparison between methods and easy integration of new contributions. In addition, a
thorough documentation is available for novice users. Thanks to the rich Python ecosystem, ruptures
can be used in coordination with numerous other scientific libraries .

32
References
[1] E. S. Page. Continuous inspection schemes. Biometrika, 41:100–105, 1954.

[2] E. S. Page. A test for a change in a parameter occurring at an unknown point. Biometrika, 42:523–527, 1955.

[3] M. Basseville and I. Nikiforov. Detection of abrupt changes: theory and application, volume 104. Prentice Hall
Englewood Cliffs, 1993.

[4] B. E. Brodsky and B. S. Darkhovsky. Nonparametric methods in change point problems. Springer Netherlands,
1993.

[5] M. Csörgö and L. Horváth. Limit theorems in change-point analysis. Chichester, New York, 1997.

[6] Jie Chen and Arjun K Gupta. Parametric statistical change point analysis: With applications to genetics,
medicine, and finance. Springer Science & Business Media, 2011.

[7] M. Lavielle and G. Teyssière. Adaptive detection of multiple change-points in asset price volatility. In Long-
Memory in Economics, pages 129–156. Springer Verlag, Berlin, Germany, 2007.

[8] V. Jandhyala, S. Fotopoulos, I. Macneill, and P. Liu. Inference for single and multiple change-points in time series.
Journal of Time Series Analysis, 34(4):423–446, 2013.

[9] K. Haynes, I. A. Eckley, and P. Fearnhead. Computationally efficient changepoint detection for a range of penalties.
Journal of Computational and Graphical Statistics, 26(1):134–143, 2017.

[10] F. Desobry, M. Davy, and C. Doncarli. An online kernel change detection algorithm. IEEE Transactions on Signal
Processing, 53(8):2961–2974, 2005.

[11] Z. Harchaoui, F. Vallet, A. Lung-Yut-Fong, and O. Cappé. A regularized kernel-based approach to unsuper-
vised audio segmentation. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), pages 1665–1668, Taipei, Taiwan, 2009.

[12] D. Angelosante and G. B. Giannakis. Group lassoing change-points piece-constant AR processes. EURASIP
Journal on Advances in Signal Processing, 70, 2012.

[13] N. Seichepine, S. Essid, C. Fevotte, and O. Cappé. Piecewise constant nonnegative matrix factorization. In
Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages
6721–6725, Florence, Italy, 2014.

[14] J. Bai and P. Perron. Estimating and testing linear models with multiple structural changes. Econometrica, 66
(1):47–78, 1998.

[15] K. Frick, A. Munk, and H. Sieling. Multiscale change point inference. Journal of the Royal Statistical Society.
Series B: Statistical Methodology, 76(3):495–580, 2014.

[16] T. Hocking, G. Schleiermacher, I. Janoueix-Lerosey, V. Boeva, J. Cappo, O. Delattre, F. Bach, and J.-P. Vert.
Learning smoothing models of copy number profiles using breakpoint annotations. BMC Bioinformatics, 14(1):
164, 2013.

[17] R. Maidstone, T. Hocking, G. Rigaill, and P. Fearnhead. On optimal multiple changepoint algorithms for large
data. Statistics and Computing, 27(2):519–533, 2017.

[18] J.-P. Vert and K. Bleakley. Fast detection of multiple change-points shared by many signals using group LARS.
In Advances in Neural Information Processing Systems 23 (NIPS 2010), volume 1, pages 2343–2351, Vancouver,
Canada, 2010.

[19] F. Picard, S. Robin, M. Lavielle, C. Vaisse, and J.-J. Daudin. A statistical approach for array CGH data analysis.
BMC Bioinformatics, 6(1):27, 2005.

[20] Y. Guédon. Exploring the latent segmentation space for the assessment of multiple change-point models. Compu-
tational Statistics, 28(6):2641–2678, 2013.

[21] S. Chakar, É. Lebarbier, C. Levy-Leduc, and S. Robin. A robust approach for estimating change-points in the
mean of an AR(1) process. Bernouilli Society for Mathematical Statistics and Probability, 23(2):1408–1447, 2017.

[22] L. Oudre, R. Barrois-Müller, T. Moreau, C. Truong, R. Dadashi, T. Grégory, D. Ricard, N. Vayatis, C. De Waele,
A. Yelnik, and P.-P. Vidal. Détection automatique des pas à partir de capteurs inertiels pour la quantification de
la marche en consultation. Neurophysiologie Clinique/Clinical Neurophysiology, 45(4-5):394, 2015.

33
[23] J. Audiffren, R. Barrois-Müller, C. Provost, É. Chiarovano, L. Oudre, T. Moreau, C. Truong, A. Yelnik, N. Vayatis,
P.-P. Vidal, C. De Waele, S. Buffat, and D. Ricard. Évaluation de l’équilibre et prédiction des risques de chutes
en utilisant une Wii board balance. Neurophysiologie Clinique/Clinical Neurophysiology, 45(4-5):403, 2015.

[24] S. Liu, A. Wright, and M. Hauskrecht. Change-point detection method for clinical decision support system rule
monitoring. Artificial Intelligence In Medicine, 91:49–56, 2018.

[25] R. Maidstone. Efficient Analysis of Complex Changepoint Models. page 34, 2013.

[26] Jan Verbesselt, Rob Hyndman, Glenn Newnham, and Darius Culvenor. Detecting trend and seasonal changes in
satellite images time series. Remote Sensing of Environment, (114):106–115, 2010.

[27] J. Reeves, J. Chen, X. L. Wang, R. Lund, and Q. Q. Lu. A review and comparison of changepoint detection
techniques for climate data. Journal of Applied Meteorology and Climatology, 46(6):900–915, 2007. ISSN 15588424.
doi: 10.1175/JAM2493.1.

[28] C. Lévy-Leduc and F. Roueff. Detection and localization of change-points in high-dimensional network traffic
data. The Annals of Applied Statistics, 3(2):637–662, 2009.

[29] A. Lung-Yut-Fong, C. Lévy-Leduc, and O. Cappé. Distributed detection/localization of change-points in high-

dimensional network traffic data. Statistics and Computing, 22(2):485–496, 2012.

[30] R. Lajugie, F. Bach, and S. Arlot. Large-margin metric learning for constrained partitioning problems. In
Proceedings of the 31st International Conference on Machine Learning (ICML), pages 297–395, Beijing, China,
2014.

[31] T. Hocking, G. Rigaill, and G. Bourque. PeakSeg: constrained optimal segmentation and supervised penalty
learning for peak detection in count data. In Proceedings of the International Conference on Machine Learning
(ICML), pages 324–332, Lille, France, 2015.

[32] R. Barrois-Müller, D. Ricard, L. Oudre, L. Tlili, C. Provost, A. Vienne, P.-P. Vidal, S. Buffat, and A. Yelnik.
Étude observationnelle du demi-tour à l’aide de capteurs inertiels chez les sujets victimes d’AVC et relation avec
le risque de chute. Neurophysiologie Clinique/Clinical Neurophysiology, 46(4):244, 2016.

[33] R. Barrois-Müller, T. Gregory, L. Oudre, T. Moreau, C. Truong, A. Aram Pulini, A. Vienne, C. Labourdette,
N. Vayatis, S. Buffat, A. Yelnik, C. de Waele, S. Laporte, P.-P. Vidal, and D. Ricard. An automated recording
method in clinical consultation to rate the limp in lower limb osteoarthritis. PLoS One, 11(10):e0164975, 2016.

[34] L. Oudre, R. Barrois-Müller, T. Moreau, C. Truong, A. Vienne-Jumeau, D. Ricard, N. Vayatis, and P.-P. Vidal.
Template-Based Step Detection with Inertial Measurement Units. Sensors, 18(11), 2018.

[35] C. Truong, L. Oudre, and N. Vayatis. Segmentation de signaux physiologiques par optimisation globale. In
Proceedings of the Groupe de Recherche et d’Etudes en Traitement du Signal et des Images (GRETSI), Lyon,
France, 2015.

[36] R. Barrois-Müller, L. Oudre, T. Moreau, C. Truong, N. Vayatis, S. Buffat, A. Yelnik, C. de Waele, T. Gregory,
S. Laporte, P. P. Vidal, and D. Ricard. Quantify osteoarthritis gait at the doctor’s office: a simple pelvis accelerom-
eter based method independent from footwear and aging. Computer Methods in Biomechanics and Biomedical
Engineering, 18 Suppl 1:1880–1881, 2015.

[37] Charles Truong. ruptures: change point detection in python, 2018. URL http://ctruong.perso.math.cnrs.fr/
ruptures. [Online].

[38] L.R. Rabiner. A tutorial on hidden Markov models and selected applications in speech recognition. Proceedings
of the IEEE, 77(2):257–286, 1989.

[39] S. I. M. Ko, T. T. L. Chong, and P. Ghosh. Dirichlet process hidden Markov multiple change-point model. Bayesian
Analysis, 10(2):275–296, 2015.

[40] A. F. Martı́nez and R. H. Mena. On a Nonparametric Change Point Detection Model in Markovian Regimes.
Bayesian Analysis, 9(4):823–858, 2014.

[41] D. Barry and J. A. Hartigan. Product partition models for change point problems. The Annals of Statistics, 20
(1):260–279, 1992.

[42] D. Barry and J. A. Hartigan. A bayesian analysis for change point problems. Journal of the American Statistical
Association, 88(421):309–319, 1993.

34
[43] S. Aminikhanghahi and D. J. Cook. A survey of methods for time series change point detection. Knowledge and
information systems, 51(2):339–367, 2017.

[44] Y. S. Niu, N. Hao, and H. Zhang. Multiple change-point detection: a selective overview. Statistica Sciences, 31
(4):611–623, 2016.

[45] J. Bai and P. Perron. Multiple structural change models: a simulation analysis. Journal of Applied Econometrics,
18:1–22, 2003.

[46] S. Chakar, É. Lebarbier, C. Levy-Leduc, and S. Robin. AR1seg: segmentation of an autoregressive Gaussian
process of order 1, 2014. URL https://cran.r-project.org/package=AR1seg.

[47] L. Boysen, A. Kempe, V. Liebscher, A. Munk, and O. Wittich. Consistencies and rates of convergence of jump-
penalized least squares estimators. The Annals of Statistics, 37(1):157–183, 2009.

[48] Z. Harchaoui and C. Lévy-Leduc. Multiple Change-Point Estimation With a Total Variation Penalty. Journal of
the American Statistical Association, 105(492):1480–1493, 2010.

[49] M. Lavielle. Detection of multiples changes in a sequence of dependant variables. Stochastic Processes and their
Applications, 83(1):79–102, 1999.

[50] F. Pein, H. Sieling, and A. Munk. Heterogeneous change point inference. Journal of the Royal Statistical Society.
Series B (Statistical Methodology), 79(4):1207–1227, 2017.

[51] H. Keshavarz, C. Scott, and X. Nguyen. Optimal change point detection in Gaussian processes. Journal of
Statistical Planning and Inference, 193:151–178, 2018.

[52] M. Lavielle and É. Moulines. Least-squares estimation of an unknown number of shifts in a time series. Journal
of Time Series Analysis, 21(1):33–59, 2000.

[53] A. Sen and M. S. Srivastava. On tests for detecting change in mean. The Annals of Statistics, 3(1):98–108, 1975.

[54] P. R. Krishnaiah. Review about estimation of change points. Handbook of Statistics, 7:375–402, 1988.

[55] A. Aue and L. Horvàth. Structural breaks in time series. Journal of Time Series Analysis, 34:1–16, 2012.

[56] P Fearnhead. Exact and efficient Bayesian inference for multiple changepoint problems. Statistics and Computing,
16(2):203–213, 2006.

[57] T. Górecki, L. Horváth, and P. Kokoszka. Change point detection in heteroscedastic time series. Econometrics
and Statistics, 7:63–88, 2018.

[58] Y.-X. Fu and R. N. Curnow. Maximum likelihood estimation of multiple change points. Biometrika, 77(3):563–573,
1990.

[59] H. He and T. S. Severini. Asymptotic properties of maximum likelihood estimators in models with multiple change
points. Bernoulli, 16(3):759–779, 2010.

[60] H Chernoff and S Zacks. Estimating the Current Mean of a Normal Distribution which is Subjected to Changes
in Time. The Annals of Mathematical Statistics, 35(3):999–1018, 1964.

[61] G. Lorden. Procedures for reacting to a change in distribution. The Annals of Mathematical Statistics, 42(6):
1897–1908, 1971.

[62] C. L. Mallows. Some comments on Cp. Technometrics, 15(4):661–675, 1973.

[63] J. Chen and A. K. Gupta. Parametric Statistical Change Point Analysis: With Applications to Genetics, Medicine,
and Finance. 2011.

[64] G. Hébrail, B. Hugueney, Y. Lechevallier, and F. Rossi. Exploratory analysis of functional data via clustering and
optimal segmentation. Neurocomputing, 73(7-9):1125–1141, 2010.

[65] S. Chib. Estimation and comparison of multiple change-point models. Journal of Econometrics, 86(2):221–241,
1998.

[66] M. Lavielle. Using penalized contrasts for the change-point problem. Signal Processing, 85(8):1501–1510, 2005.

[67] J. Bai. Least squares estimation of a shift in linear processes. Journal of Time Series Analysis, 15(5):453–472,
1994.

35
[68] J. Bai. Least absolute deviation of a shift. Econometric Theory, 11(3):403–436, 1995.

[69] J. Bai. Testing for parameter constancy in linear regressions: an empirical distribution function approach. Econo-
metrica, 64(3):597–622, 1996.

[70] J. Bai. Vector autoregressive models with structural changes in regression coefficients and in variance???covariance
matrices. Annals of Economics and Finance, 1(2):301–336, 2000.

[71] J. Bai. Estimation of a change-point in multiple regression models. Review of Economic and Statistics, 79(4):
551–563, 1997.

[72] Z. Qu and P. Perron. Estimating and testing structural changes in multivariate regressions. Econometrica, 75(2):
459–502, 2007.

[73] J. Bai. Estimation of multiple-regime regressions with least absolutes deviation. Journal of Statistical Planning
and Inference, 74:103–134, 1998.

[74] J. Bai. Likelihood ratio tests for multiple structural changes. Journal of Econometrics, 91(2):299–323, 1999.

[75] J. Bai and P. Perron. Critical values for multiple structural change tests. Econometrics Journal, 6(1):72–78, 2003.

[76] P. Perron. Dealing with structural breaks. Palgrave handbook of econometrics, 1(2):278–352, 2006.

[77] J. Bai. Common breaks in means and variances for panel data. Journal of Econometrics, 157:78–92, 2010.

[78] J. Bai, R. L. Lumsdaine, and J. H. Stock. Testing for and dating common breaks in multivariate time series. The
Review of Economic Studies, 65(3):395–432, 1998.

[79] P. Perron and Z. Qu. Estimating restricted structural change models. Journal of Econometrics, 134(2):373–399,
2006.

[80] B. M. Doyle and J. Faust. Breaks in the variability and comovement of G-7 economic growth. The Review of
Economics and Statistics, 87(4):721–740, 2005.

[81] C. F. H. Nam, J. A. D. Aston, and A. M. Johansen. Quantifying the uncertainty in change points. Journal of
Time Series Analysis, 33:807–823, 2012.

[82] P. C. Mahalanobis. On the generalised distance in statistics. Proceedings of the National Institute of Sciences of
India, 2(1):49–55, 1936.

[83] T. Hastie, R. Tibshirani, and J. Friedman. The elements of statistical learning, volume 1. 2009.

[84] E. P. Xing, M. I. Jordan, and S. J. Russell. Distance metric learning, with application to clustering with side-
Information. In Advances in Neural Information Processing Systems 21 (NIPS 2003), pages 521–528, 2003.

[85] J. V. Davis, B. Kulis, P. Jain, S. Sra, and I. S. Dhillon. Information-theoretic metric learning. In Proceedings of
the 24th International Conference on Machine Learning (ICML), pages 209–216, Corvalis, Oregon, USA, 2007.

[86] J. H. J. Einmahl and I. W. McKeague. Empirical likelihood based hypothesis testing. Bernoulli, 9(2):267–290,
2003.

[87] C. Zou, G. Yin, F. Long, and Z. Wang. Nonparametric maximum likelihood approach to multiple change-point
problems. The Annals of Statistics, 42(3):970–1002, 2014.

[88] J. Zhang. Powerful two-sample tests based on the likelihood ratio. Technometrics, 48(1):95–103, 2006.

[89] K. Haynes, P. Fearnhead, and I. A. Eckley. A computationally efficient nonparametric approach for changepoint
detection. Statistics and Computing, 27:1293–1305, 2017.

[90] S. Clemencon, M. Depecker, and N. Vayatis. AUC optimization and the two-sample problem. In Advances in
Neural Information Processing Systems 22 (NIPS 2009), pages 360–368, Vancouver, Canada, 2009.

[91] J. H. Friedman and L. C. Rafsky. Multivariate Generalizations of Wald-Wolfowitz and Smirnov two-sample tests.
The Annals of Statistics, 7(4):697–717, 1979.

[92] A. Lung-Yut-Fong, C. Lévy-Leduc, and O. Cappé. Homogeneity and change-point detection tests for multivariate
data using rank statistics. Journal de la Société Française de Statistique, 156(4):133–162, 2015.

[93] F. Wilcoxon. Individual comparisons by ranking methods. Biometrics Bulletin, 1(6):80–83, 1945.

36
 [94] E. L. Lehman and J. P. Romano. Testing Statistical Hypotheses, volume 101. springer, 3 edition, 2006.

[95] M. G. Kendall. Rank correlation methods. Charles Griffin, London, England, 1970.

[96] Z. Harchaoui and O. Cappé. Retrospective mutiple change-point estimation with kernels. In Proceedings of the
IEEE/SP Workshop on Statistical Signal Processing, pages 768–772, Madison, Wisconsin, USA, 2007.

[97] B. Schölkopf and A. J. Smola. Learning with kernels. MIT Press, Cambridge, USA, 2002.

[98] A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola. A kernel two-sample test. Journal of
Machine Learning Research (JMLR), 13:723–773, 2012.

[99] A. Celisse, G. Marot, M. Pierre-Jean, and G. Rigaill. New efficient algorithms for multiple change-point detection
with reproducing kernels. Computational Statistics and Data Analysis, 128:200–220, 2018.

[100] B. K. Sriperumbudur, A. Gretton, K. Fukumizu, G. Lanckriet, and B. Schölkopf. Injective Hilbert space embed-
dings of probability measures. In Proceedings of the 21st Conference on Learning Theory (COLT), pages 9–12,
Helsinki, Finland, 2008.

[101] J. Shawe-Taylor and N. Cristianini. Kernel Methods for Pattern Analysis. Cambridge university press, 2004.

[102] D. Garreau and S. Arlot. Consistent change-point detection with kernels. arXiv preprint arXiv:1612.04740v3,
pages 1–41, 2017.

[103] S. Arlot, A. Celisse, and Z. Harchaoui. Kernel change-point detection. arXiv preprint arXiv:1202.3878, pages
1–26, 2012.

[104] J. Cabrieto, F. Tuerlinckx, P. Kuppens, F. H. Wilhelm, M. Liedlgruber, and E. Ceulemans. Capturing correlation
changes by applying kernel change point detection on the running correlations. Information Sciences, 447:117–139,
2018.

[105] S. M. Kay and A. V. Oppenheim. Fundamentals of Statistical Signal Processing, Volume II: Detection Theory.
Prentice Hall, 1993.

[106] J. Bai and P. Perron. Computation and analysis of multiple structural change models. Journal of Applied
Econometrics, 18(1):1–22, 2003.

[107] R. Bellman. On a routing problem. Quaterly of Applied Mathematics, 16(1):87–90, 1955.

[108] M. Lavielle. Optimal segmentation of random processes. IEEE Transactions on Signal Processing, 46(5):1365–
1373, 1998.

[109] G. Rigaill. A pruned dynamic programming algorithm to recover the best segmentations with 1 to K max change-
points. Journal de la Société Française de Statistique, 156(4):180–205, 2015.

[110] B. Hugueney, G. Hébrail, Y. Lechevallier, and F. Rossi. Simultaneous clustering and segmentation for functional
data. In Proceedings of 16th European Symposium on Artificial Neural Networks (ESANN), pages 281–286, Bruges,
Belgium, 2009.

[111] R. Killick, P. Fearnhead, and I. A. Eckley. Optimal detection of changepoints with a linear computational cost.
Journal of the American Statistical Association, 107(500):1590–1598, 2012.

[112] Jie Chen and Arjun K. Gupta. Parametric Statistical Change Point Analysis. Birkhäuser Boston, 2011. doi:
10.1007/978-0-8176-4801-5.

[113] J. Chen and A. K. Gupta. Testing and locating variance changepoints with application to stock prices. Journal
of the American Statistical Association, 92(438):739–747, 1997.

[114] H. Vullings, M. Verhaegen, and H. Verbruggen. ECG segmentation using time-warping. In Lecture notes in
computer science, pages 275–286. Springer, 1997.

[115] B. E. Brodsky, B. S. Darkhovsky, A. Y. Kaplan, and S. L. Shishkin. A nonparametric method for the segmentation
of the EEG. Computer Methods and Programs in Biomedicine, 60(2):93–106, 1999.

[116] R. Esteller, G. Vachtsevanos, J. Echauz, and B. Litt. A Comparison of waveform fractal dimension algorithms.
IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications, 48(2):177–183, 2001.

[117] K. Karagiannaki, A. Panousopoulou, and P. Tsakalides. An online feature selection architecture for Human Activity
Recognition. In Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), pages 2522–2526, New Orleans, LA, USA, 2017.

37
[118] S. Adak. Time-dependent spectral analysis of nonstationary time series. Journal of the American Statistical
Association, 93(444):1488–1501, 1998.

[119] S. S. Chen and P. S. Gopalakrishnan. Speaker, environment and channel change detection and clustering via the
bayesian information criterion. In Proceedings of the DARPA Broadcast News Transcription and Understanding
Workshop, page 8, Landsdowne, VA, 1998.

[120] E. Keogh, S. Chu, D. Hart, and M. Pazzani. Segmenting time series: a survey and novel approach. Data Mining
in Time Series Databases, 57(1):1–22, 2004.

[121] Z. Harchaoui, F. Bach, and É. Moulines. Kernel change-point analysis. In Advances in Neural Information
Processing Systems 21 (NIPS 2008), pages 609–616, Vancouver, Canada, 2008.

[122] S. Liu, M. Yamada, N. Collier, and M. Sugiyama. Change-point detection in time-series data by relative density-
ratio estimation. Neural Networks, 43:72–83, 2013.

[123] D. Kifer, S. Ben-David, and J. Gehrke. Detecting change in data streams. In Proceedings of the Thirtieth
International Conference on Very Large Data Bases (VLDB) - Volume 30, pages 180–191, Toronto, Canada,
2004.

[124] R. Prescott Adams and D. J. C. MacKay. Bayesian Online Changepoint Detection. Technical report, 2007.

[125] A. B. Olshen, E. S. Venkatraman, R. Lucito, and M. Wigler. Circular binary segmentation for the analysis of
array-based DNA copy number data. Biostatistics, 5(4):557–572, 2004.

[126] J. Bai. Estimating multiple breaks one at a time. Econometric Theory, 13(3):315–352, 1997.

[127] Piotr Fryzlewicz. Wild binary segmentation for multiple change-point detection. Annals of Statistics, 42(6):
2243–2281, 2014.

[128] J. Himberg, K. Korpiaho, H. Mannila, J. Tikanmaki, and H. T. Toivonen. Time series segmentation for context
recognition in mobile devices. In Proceedings of the IEEE International Conference on Data Mining (ICDM),
pages 203–210, 2001.

[129] Y. S. Niu and H. Zhang. The screening and ranking algorithm to detect DNA copy number variations. The Annals
of Applied Statistics, 6(3):1306–1326, 2012.

[130] W. R. Lai, M. D. Johnson, R. Kucherlapati, and P. J. Park. Comparative analysis of algorithms for identifying
amplifications and deletions in array CGh data. Bioinformatics, 21(19):3763–3770, 2005.

[131] H. Willenbrock and J. Fridlyand. A comparison study: applying segmentation to array CGH data for downstream
analyses. Bioinformatics, 21(22):4084–4091, 2005.

[132] E. S. Venkatraman and A. B. Olshen. A faster circular binary segmentation algorithm for the analysis of array
CGH data. Bioinformatics, 23(6):657–663, 2007.

[133] E. Keogh, S. Chu, D. Hart, and M. Pazzani. An online algorithm for segmenting time series. In Proceedings of
the IEEE International Conference on Data Mining (ICDM), pages 289–296, 2001.

[134] Y.-C. Yao. Estimating the number of change-points via Schwarz’ criterion. Statistics and Probability Letters, 6
(3):181–189, 1988.

[135] Y.-C. Yao and S. T. Au. Least-squares estimation of a step function. Sankhy??: The Indian Journal of Statistics,
Series A, 51(3):370–381, 1989.

[136] S. Arlot and A. Celisse. A survey of cross-validation procedures for model selection. Statistical Surveys, 4:40–79,
2010.

[137] L. Birgé and P. Massart. Minimal penalties for Gaussian model selection. Probability Theory and Related Fields,
138(1):33–73, 2007.

[138] T. Hocking, G Rigaill, J.-P. Vert, and F. Bach. Learning sparse penalties for change-point detection using max
margin interval regression. In Proceedings of the International Conference on Machine Learning (ICML), pages
172–180, Atlanta, USA, 2013.

[139] C. Truong, L. Oudre, and N. Vayatis. Penalty learning for changepoint detection. In Proceedings of the European
Signal Processing Conference (EUSIPCO), Kos, Greece, 2017.

[140] G. Schwarz. Estimating the dimension of a model. The Annals of Statistics, 6(2):461–464, 1978.

38
[141] R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society. Series B
(Methodological), 58(1):267–288, 1996.

[142] J.-J. Jeon, J. Hyun Sung, and E.-S. Chung. Abrupt change point detection of annual maximum precipitation using
fused lasso. Journal of Hydrology, 538:831–841, 2016.

[143] N. R. Zhang and D. O. Siegmund. A modified Bayes information criterion with applications to the analysis of
comparative genomic hybridization data. Biometrics, 63(1):22–32, 2007.

[144] É. Lebarbier. Detecting multiple change-points in the mean of gaussian process by model selection. Signal
Processing, 85(4):717–736, 2005.

[145] L. Y. Vostrikova. Detecting disorder in multidimensional random processes. Soviet Math. Dokl., 24:55–59, 1981.

[146] L. Birgé and P. Massart. Gaussian model selection. Journal of the European Mathematical Society, 3(3):203–268,
2001.

39