Advances in Data Analysis and Classification (2019) 13:1083–1103

https://doi.org/10.1007/s11634-019-00360-z

REGULAR ARTICLE

A Kendall correlation coefficient between functional data

Dalia Valencia1 · Rosa E. Lillo2 · Juan Romo1

Received: 26 May 2018 / Revised: 12 May 2019 / Accepted: 17 May 2019 /

Published online: 25 May 2019
© Springer-Verlag GmbH Germany, part of Springer Nature 2019

Abstract
Measuring dependence is a very important tool to analyze pairs of functional data.
The coefficients currently available to quantify association between two sets of curves
show a non robust behavior under the presence of outliers. We propose a new robust
numerical measure of association for bivariate functional data. We extend in this paper
Kendall coefficient for finite dimensional observations to the functional setting. We
also study its statistical properties. An extensive simulation study shows the good
behavior of this new measure for different types of functional data. Moreover, we
apply it to establish association for real data, including microarrays time series in
genetics.

Keywords Concordance · Dependence · Functional data · Kendall’s tau

Mathematics Subject Classification 62-07 · 62G35 · 62G09

1 Introduction

Many processes currently used in different fields of science and research lead to
random observations that can be analyzed as curves. We can also find a large amount
of data for which it would be more appropriate to use some interpolation techniques
and consider them as functional data. This approach turns out to be essential when
data have been observed at different time intervals. Examples of functional data are

B Dalia Valencia
djvalega@gmail.com
Rosa E. Lillo
rosaelvira.lillo@uc3m.es
Juan Romo
juan.romo@uc3m.es

1 Department of Statistics, Universidad Carlos III de Madrid, Calle Madrid 126, Madrid, Spain
2 Department of Statistics, UC3M-Santander Big Data Institute, Universidad Carlos III de Madrid,
Madrid, Spain

123
1084 D. Valencia et al.

found in areas such as meteorology, where, for example, the ozone level measured
during a day is a curve; finance, where, for example, an asset price takes values at very
close time instants, and medicine, where the observed gene expressions over time can
also be considered as realizations of random curves.
Several multivariate methods have been extended to functional data. Multivariate
techniques such as regression (Cardot et al. 1999; He et al. 2000), analysis of variance
(Cuevas et al. 2004; Delicado 2007), principal components (Pezulli and Silverman
1993), generalized linear model (Escabias et al. 2004) and depth (López-Pintado and
Romo 2007, 2009) have already been extended to a functional context. Other useful
methodologies can be found in Ramsay and Silverman (2005). However, there are
still some concepts that have not been fully explored for functional data: measures of
association and dependence structures between curves, for example.
Leurgans et al. (1993) considered the canonical correlation between two sets of
curves. This technique provides a pair of functions called canonical variates and the
sample correlation among these variates leads to the canonical correlation between the
two sets of curves. He et al. (2000) propose an alternative way of finding the canonical
correlation through the extension of multivariate analysis ideas. Opgen-Rhein and
Strimmer (2006) proposed an estimator of the dynamic correlation that provides a
measure of similarity between pairs of functional observations. It is based on the
concept of dynamical correlation introduced by Dubin and Müller (2005) to analyze
a nonparametric method to quantify the covariation of components of multivariate
longitudinal observations.
In this paper, we extend the Kendall τ correlation coefficient to the functional
framework. Kendall’s τ allows dependence to be measured in the bivariate case through
the definition of concordance, which is based on the idea of order. Since there exists
no total order among functions, we use preorders that allow us to sort the functional
observations and count the concordant and discordant pairs of a bivariate sample of
curves. Once a preorder is introduced, the functional τ coefficient can be defined in a
way similar to the bivariate τ coefficient. We show that it fulfils natural properties for a
dependence measure and also establish the consistency of the sample version. Finally,
we illustrate with simulated and real data the performance of this new dependence
measure as well as its robustness, which is a basic characteristic of the Kendall τ in
its bivariate version. We analyze a data set corresponding to a microarray time series
from a human T-cell experiment with 58 genes, 10 time points and 44 replications.
We obtain the functional τ for each pair of genes and construct the partial correlation
matrix to compare the gene network resulting from functional τ with those from
dynamical correlation.
This paper is organized as follows. In Sect. 2, the functional τ is defined extending
the concept of concordance for bivariate random variables. Section 3 is devoted to
some properties of this correlation coefficient and to studying convergence results.
A summary of the classic techniques, simulation results and sensitivity analysis are
given in Sect. 4. Section 5 contains a study of dependence using the genes data set.
In Sect. 6, we present a robustness empirical study. Finally, in Sect. 7, we outline the
main conclusions of this paper.

123
A Kendall correlation coefficient between functional data 1085

2 Functional Kendall correlation coefficient

Kendall (1938) introduced his correlation coefficient τ as a measure for dependence

between two random variables X , Y , based on the ranks of sampled observations
of X and Y . It makes use of the idea of concordance: two random variables are
concordant if large (small) values of one are related to large (small) values of the
other. When large (small) values of one are related to small (large) values of the
other, the random variables are discordant. More formally, let (x1 , y1 ) and (x2 , y2 )
be two observations of a random vector (X , Y ). We say that (x1 , y1 ) and (x2 , y2 ) are
concordant if (x1 − x2 )(y1 − y2 ) > 0 and discordant if (x1 − x2 )(y1 − y2 ) < 0. This
means that they are concordant if either x1 < x2 and y1 < y2 or x2 < x1 and y2 < y1 ;
in other cases with strict inequality, the observations are discordant.
Kendall’s correlation coefficient is defined as the difference between the probabilities
of concordance and discordance in two different realizations (X 1 , Y1 ), (X 2 , Y2 ) of
(X , Y ),

τ = P{(X 1 − X 2 )(Y1 − Y2 ) > 0} − P{(X 1 − X 2 )(Y1 − Y2 ) < 0}.

The above expression can be also written as

τ = 2[P{X 1 < X 2 , Y1 < Y2 } + P{X 2 < X 1 , Y2 < Y1 }] − 1.

If (x1 , y1 ), (x2 , y2 ) . . . (xn , yn ) is a sample from (X , Y ), the sample coefficient is

S
τ̂ =   ,
n
2

where S = cp − dp is the difference between the number of concordant pairs (cp)

and the number of discordant pairs (dp).
The aim of this paper is to present a functional version of this correlation coefficient
where (X , Y ) are two random processes X = {X (t)|t ∈ I } and Y = {Y (t)|t ∈ I } that
are observed continuously or at disrete time points. Let f and g be in the space C(I )
of real continuous functions on the compact interval I . First, we need to introduce
relationships allowing the comparison between curves. A natural choice is the usual
order, i. e., f  g ⇔ f (t) ≤ g(t), for all t ∈ I . It fulfills the partial order conditions;
however, most functions are not comparable with this order. To avoid this difficulty,
we waive the antisymmetry condition and use preorders instead of orders.

Definition 1 Let f and g be in C(I ). Then, we consider two alternatives.

f m g ≡ max f (t) ≤ max g(t). (1)

t∈I t∈I
 b
f i g ≡ (g(t) − f (t))dt ≥ 0. (2)
a

123
1086 D. Valencia et al.

It follows easily that, for constant functions defined in the same compact interval
I , both preorders are equivalent to the usual ordering on the real line. Given any
preorder definition among functions, we may define the concordance concept between
functions.

Definition 2 (Functional concordance) Let  be a preorder between functions, and

let ≺ address the case without considering ties. Two pairs of functions ( f 1 , g1 ) and
( f 2 , g2 ) are called concordant if either f 1 ≺ f 2 and g1 ≺ g2 or f 2 ≺ f 1 and g2 ≺ g1 ;
otherwise they are called discordant.

Definition 2 allows us to extend Kendall’s correlation coefficient to the functional

case, as described in the next Definition.

Definition 3 (Functional τ ) If (X 1 , Y1 ), (X 2 , Y2 ) are copies of a bivariate stochastic

process {(X (t), Y (t)) : t ∈ I }, the population version of this dependence measure is

τ = 2[P{X 1 ≺ X 2 , Y1 ≺ Y2 } + P{X 2 ≺ X 1 , Y2 ≺ Y1 }] − 1. (3)

Consider a sample (x1 , y1 ), . . . , (xn , yn ) of a two-dimensional random process

(X , Y ) = {(X (t), Y (t)) : t ∈ I } within the compact interval I , with X , Y ∈ C(I ).
Then Kendall’s extended correlation coefficient is estimated by the empirical version:
 −1 
n
n  
τ̂ = 2I (xi ≺ x j and yi ≺ y j ) + 2I (x j ≺ xi and y j ≺ yi ) − 1.
2
i< j
(4)
Some of the asymptotic properties of the traditional Kendall τ coefficient arise from
the fact that it can be expressed as a U -statistic. To obtain an asymptotic result in the
functional field, which will be stated in Theorem 2, we need the definition of U B-
statistics, U -statistics taking values in a Banach space. We also need some convergence
results for this kind of statistics. These concepts can be defined as follows:

Definition 4 (U B-Statistics. Borovskikh 1996, page 5) Let B be a real separable

Banach space with a norm · and let B ∗ be the dual space of real-valued linear
functions on B. Denote by x ∗ (x) the value of functional x ∗ ∈ B ∗ at x ∈ B. Let
X 1 , . . . , X n be independent random variables taking values in the measurable space
(X , X), where X is a σ -algebra, and all with identical distribution P. Consider a
Bochner integrable symmetric function (kernel) Φ : X m → B of m variables given
on X m and taking values in B . Then, a U -statistic is
 −1 
n
Un = Φ{(X i1 , . . . X im )}. (5)
m
1≤i 1 <···<i m ≤n

It is clear that Un ∈ B. Hence, the U -statistic (5) with a B-valued kernel Φ is called a
U B-statistic. In particular, if B = R, it is called a U R-statistic and if B = H , where
H is a real separable Hilbert space, it is called a U H -statistic.

123
A Kendall correlation coefficient between functional data 1087

Let P = {P} be a class a probability distribution on (X , X). By θ : P → θ (P),

we denote a functional given on P and taking values in B where
 
θ (P) = ... Φ(x1 , . . .x m)P(d x1 ) . . . P(d xm )

If E Φ < ∞, then Un is an unbiased estimate of the B-valued element θ (P) =

EΦ(X 1 , . . . X m )

The following theorem provides an asymptotic result, which will be very useful in
what follows.

Theorem 1 (Borovskikh 1996, page 73) Assume that the B-valued kernel Φ is such
that E Φ < ∞. Then,

Un → θ a.s n → ∞,

and

E||Un − θ || → 0.

Now, consider (X 1 , Y1 ), . . . , (X n , Yn ) to be independent copies of the bivariate

stochastic process (X , Y ) with identical distribution P and whose realizations or paths
are pairs of functions that take values in the measurable space (C[a, b] × C[a, b], X).
Then, the functional τ̂ given in (4) can be expressed as a U B-statistic,
 −1 
n
Un = Φ{(X i1 , Yi1 ), (X i2 , Yi2 )}, (6)
2
1≤i 1 <i 2 ≤n

where Φ : C 2 [a, b] × C 2 [a, b] → R is a Bochner integrable symmetric function

according to Definition 1.3.11 in Schwabik and Guoju (2005), and given by

Φ[(xi , yi ), (x j , y j )] = 2I (xi ≺ x j , yi ≺ y j ) + 2I (x j ≺ xi , y j ≺ yi ) − 1,

where I denotes the indicator function.

3 Properties of functional 

We analyze in this Section some desirable properties of τ as a dependence measure.

Scarsini (1984) studies the measures of concordance in terms of copulas and proposes
a set of axioms that a concordance measure for ordered pairs of continuous random
variables should fulfill. The extension of these axioms to the multivariate case was
studied in Taylor (2007, 2008). The following list gives the properties of the functional
τ . Some of them come from the axioms proposed by Scarsini (1984). Other properties

123
1088 D. Valencia et al.

are a natural extension of the well known properties of the bivariate τ itself (Kendall
1938).
Let (X (t), Y (t)) be a bivariate stochastic process. Then,
1. τ (X (t), Y (t)) = τ (Y (t), X (t)).
2. −1 ≤ τ (X (t), Y (t)) ≤ 1.
3. τ (−X (t), Y (t)) = −τ (X (t), Y (t)).
4. τ (X (t), g(X (t))) = 1, for any monotone increasing function g.
5. τ (X (t), g(X (t))) = −1, for any monotone decreasing g.
6. If X (t) and Y (t) are stochastically independent, then τ (X (t), Y (t)) = 0.
7. The correlation coefficient functional is invariant under strictly increasing and
continuous transformations of the functional variables,

τ [α(X (t)), β(Y (t))] = τ (X (t), Y (t)),

where α and β are strictly increasing functions.

Note that τ with the preorder (1) verifies 1, 2, 4, 6 and 7, and τ with the preorder
(2) verifies 1, 2, 3, 6 but 4, 5 and 7 just for affine transformations. The proof of the
properties are given in the Appendix of the Supplementary Material.
The consistency of functional τ̂ is established in the next theorem.

Theorem 2 Let (x1 , y1 ), . . . , (xn , yn ) be a sample of independent and identical func-

tional observations from (X , Y ). Then,

τ̂n → τ a.s. as n → ∞

for the two preorders considered in Definition 1.

It is easy to check that the function

Φ[(xi , yi ), (x j , y j )] = 2I (xi ≺ x j , yi ≺ y j ) + 2I (x j ≺ xi , y j ≺ yi ) − 1,

which represents the kernel for the τ̂n belongs to the interval [−1, 3]. Then, the func-
tional τ̂ , given in (4) and expressed as the U B-statistic (6), has associated a kernel Φ
such that E Φ is finite. Therefore, from Theorem 1, we have that, if Φ is such that
E Φ < ∞, then the U B-statistic will converge almost surely to the parameter τ .
Observe that Theorem 2 is valid in general for any well-defined preorder ().
To illustrate how the functional τ̂ works in simulated functional samples with
different kinds of dependence, we provide some examples. From now on, τ̂1 , τ̂2 denote
the functional τ̂ when (1) and (2) preorders are considered, respectively. Consider five
joint realizations of the processes X (t) = t 2 + Z 1 and Y (t) = −(t + Z 2 )2 − 8t + Z 2 ,
where (Z 1 , Z 2 ) follow a bivariate standard normal distribution with correlation σ12
representing the random part of the processes. The bivariate functional sample shown
in Fig. 1 was generated with a high value of σ12 close to 1. Each pair of curves is
represented by the same color. In this first case, the ordering for the preorder (1) in
the first group is (red > cyan > green > blue > magenta), and for the second group

123
A Kendall correlation coefficient between functional data 1089

7 2
First
0 First
6 Second
Second
Third
−2 Third
5 Fourth
Fourth
Fifth
−4 Fifth
4
−6
3
−8
2
−10

1 −12

0 −14
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 1 τ̂1 = 0.6 τ̂2 = 0.4

12 8
First 7 First
10 Second Second
6
Third Third
5
8 fourth fourth
fifth 4 fifth

6 3
2
4
1
0
2
−1
0 −2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Fig. 2 τ̂1 = −0.8 τ̂2 = −0.8

it is (cyan > green > red > blue > magenta). In both panels, the cyan and green
curves are in the same relative position with respect to the other curves. The blue and
magenta curves are also in the same position in both groups. In this case τ̂1 = 0.6.
For the ordering with preorder (2), in the first group it is (red > cyan > green > blue
> magenta), and for the second group it is (green > cyan > red > blue > magenta).
In both panels, blue and magenta curves are in the same position in the two groups.
At the same time, the remainder of the curves are almost completely ordered in the
opposite way and τ̂2 = 0.4. Figure 2 shows another example where both coefficients
take the same value.

4 Empirical results and comparisons

In this Section, we illustrate the performance of the functional τ with both preorders
given in Eqs. 1 and 2, as well as its behavior with respect to other dependence mea-
sures already introduced in the literature. We briefly describe two of them (dynamical
correlation and canonical correlation) and, in order to compare our results with these
dependence measures, we carry out a simulation study.

123
1090 D. Valencia et al.

A commonly used technique to find the correlation between two groups of func-
tions is the dynamical correlation, which is a measure of similarity between two curves.
Dubin and Müller (2005) introduced the dynamical correlation as the following infor-
mal idea: “if both trajectories tend to be mostly on the same side of their time average
(a constant) then the dynamical correlation is positive; if the opposite occurs, then
dynamical correlation is negative”. Opgen-Rhein and Strimmer (2006) proposed an
estimator for the dynamical correlation considering functional data instead of lon-
gitudinal data. We will use the estimator of the dynamical correlation proposed in
Opgen-Rhein and Strimmer (2006), which is a slightly revised version of the dynam-
ical correlation introduced in Dubin and Müller (2005):

1  s
n
ρ̂d = xi (t), yis (t) ,
n−1
i=1

where

x c (t)
x s (t) = ,
n
1
n−1 i=1 xic (t), xic (t)

and x c (t) are functions centered in space and time simultaneously, i.e.,

1
n
x (t) = x(t) − x̄(t), 1,
c
wher e x̄(t) = xi (t),
n
i=1

and · is the usual inner product for functions x(t), y(t) = I x(t)y(t)dt. As we
can see, ρ̂d is an estimator of the population dynamical correlation

ρd = E X S (t), Y S (t) ,

which can been seen as an average of individual correlations.

Another well-known technique to measure functional dependence is the canonical
correlation, which was defined in Leurgans et al. (1993). This procedure seeks to
investigate which modes of variability in the two sets of curves are most associated
with one another. This analysis provides a pair of functions called canonical variates

(ξ(s), η(s))

such that ξ X i and ηYi are well correlated with one another and the sample cor-
relation between these variables will be what in Leurgans et al. (1993) was called the
canonical correlation between the two variables or groups. In a formal way, consider
n observed pairs of curves (xi (t), yi (t)) with t in the same finite interval I and all
integrals are taken over I (see Ramsay and Silverman 2005, page 204). Given canon-
ical variates ξ and η, the canonical correlation was defined by Leurgans et al. (1993)
as the sample squared correlation of ξ xi and ηyi , i.e.,

123
A Kendall correlation coefficient between functional data 1091

Table 1 Steps for calculating τ̂1,2

1. Input: F, G matrices that contain the two sets of functions

2. Compare the functions in each matrix using the maximum preorder or integral preorder
3. Obtain the number of concordant and discordant pairs
4. Calculate τ̂1,2 = 2∗(concor dat pair s−discor dant pair s)
n∗(n−1) (n = total functions)
5. Output: τ̂1,2

  2
cov ξ xi , ηyi
ρ̂c (ξ, η) =       ,
var ξ xi + λ||D 2 ξ ||2 var ηyi + λ||D 2 η||2

 2 2
where λ is a positive smoothing parameter and ||D 2 f ||2 = D f , that is, the inte-
grated squared curvature of f that quantifies its roughness. Having a pair of canonical
variables with fairly smooth weight functions and correlations that are not excessively
low is necessarily a good choice for the smoothing parameter. This parameter can be
chosen subjectively, but can also be selected through a cross-validation score if an
automatic procedure is required.
As usual in FDA, calculations related to functional data are performed by observing
the data in certain points of the functions. Hence, it is necessary before introducing
the simulations that are going to be shown in the paper to define the finite dimensional
version of these coefficients and the preorders given in Definition (1). Denote by
t1 , . . . , td the time points from I where the n sampled functions xi = {xi (t)|t ∈
I }, i = 1, . . . , n have been observed. Then,
– x1 (t) m x2 (t) ⇔ max(x1 (t1 ), . . . , x1 (td )) ≤ max(x2 (t1 ), . . . , x2 (td )).
−t1 n−1 −t1
– x1 (t) i x2 (t) ⇔ td2d [x1 (t1 )+x1 (td )+2 i=2 x1 (ti )] ≤ td2d [x2 (t1 )+x2 (td )+
n−1
2 i=2 x2 (ti )].
In the same way we can order the functions yi (t). The last expression corresponds
to the composite trapezoidal rule of numerical integration, which has been used for
calculating the values of the integrals. Then, the coefficients are calculated as
 −1 
n
n  
τ̂ = 2I (xi ≺ x j and yi ≺ y j ) + 2I (x j ≺ xi and y j ≺ yi ) − 1.
2
i< j

Since Rd is a Banach space, the convergence of the τ̂n,d → τd a.s. as n → ∞,

follows directly by from a proof similar to Theorem 2. From now on, subscript d is
deleted from the notation.
The algorithms to calculate the coefficients were implemented in MATLAB using
the preorders defined for the finite dimensional case. Table 1 summarizes the steps
necessary to obtain them.
Once we have defined the dependence measures that will be used to com-
pare the performance of our coefficient, we show through simulation exercises the
behavior of the measure introduced in this paper and those chosen to compare it.

123
1092 D. Valencia et al.

The data are simulated in the following way. Consider the bivariate stochastic process
(X (t), Y (t)) = [ f 1 (t, Z 1 ), f 2 (t, Z 2 )], where (Z 1 , Z 2 ) represents the random part of
the process, a bivariate standard normal distribution with correlation σ12 . We consider
a different structure for the functions f i , i = 1, 2 as well as different values for σ12 . In
each case, 50 realizations of the process (X (t), Y (t)) are generated, where the paths
are discretized taking d = 50 points over the interval [0, 1] and calculating the mea-
sures of dependence previously mentioned. This procedure is carried out 100 times
and the results reported refer to the average and deviation over the 100 setups.
Table 2 presents the average of the measures τ̂1 and τ̂2 as well as ρ̂c and ρ̂d , which
denote the canonical correlation and dynamical correlation, respectively. The value in
brackets reports the standard deviation of the measures considered. We also include,
in each case, the value of the correlation σ12 . We can see that the coefficients τ̂1 and
τ̂2 in some cases take different values, which is a consequence of the preorders not
sorting the data in the same way. In the case of processes in which one of them is an
increasing transformation of the other, both coefficients take value 1, which confirms
the perfect dependence between the processes considered. However, this fact does not
occur in the measures used for comparison, see for example rows 3 and 4 in Table 2.
Indeed, the value of ρ̂d in row 4 does not reflect the true dependence between those
processes, which is positive and perfect. Observe that a similar conclusion can be
drawn when the dependence is perfect but negative as may be seen in row 5. There,
only our coefficients were able to capture the negative perfect dependence. Note also
that in the independent case (row 11), our coefficients reflect this fact better than the
other measures. Finally, the standard deviation of τ̂2 in most cases is the smallest
among the other measures.
We can see that the canonical correlation ρ̂c is always positive, which means that it
does not capture the direction of the dependence. This is because it seeks variability
in the two sets of curves that maximize the sample correlation between the pairs
of canonical variates. Dynamical correlation ρ̂d reflects the average of individual
similarities rather than considering the set of curves as a sample of the same population.
This makes the dynamical correlation capture changes at an individual performance
level, while Kendall’s coefficient detects changes at a more general level, since it is
inspired in the bivariate coefficient, which is based on evaluating the relationship of
the two sets of data.
Thus, the functional τ̂ is appropriate to indicate how related two functional variables
are, regardless of the shape of their realizations. This coefficient measures the joint
tendency of the variables to have increasing or decreasing behavior.
In the bivariate case, the decision of which coefficient to use depends on several
factors, such as the type of measurement scale in which each variable is expressed,
the nature of the distributions (continuous or discrete) and if the dependence sought
is linear or nonlinear. Also, each coefficient is designed so that the dependence that it
is able to capture measures different aspects such as concordance and discordance in
the data or the ordering relation of a group of data with respect to another, etc. Having
several coefficients to measure dependency from different points of view is therefore
natural in the bivariate case, as happens in the functional case. The final user of these
measures will make her decisions based on the joint information offered by the pull
of coefficients available to him. In the case of functional data, as in the multivariate

123
 Table 2 Dependence measures in simulated data

X (t) = f 1 (t, Z 1 ) Y (t) = f 2 (t, Z 2 ) σ12 τ̂¯1 τ̂¯2 ρ̂¯ c ρ̂¯ d

1 (t + Z 1 )3 + (t + Z 1 )2 + 3(t + Z 1 ) (t + Z 2 )2 + 78 (t + Z 2 ) − 10 0.8 0.4861 0.4874 0.7448 0.7098

(0.0657) (0.0711) (0.0898) (0.1139)
2 sin(t + Z 1 ) cos(t + Z 2 ) − 0.7 0.3084 0.2774 0.5367 0.3605
(0.0923) (0.0835) (0.1004) (0.11)
3 (t + Z 1 )2 (t + Z 1 )4 1 1 1 0.9566 0.922
(0) (0) (0.0118) (0.0125)
4 (t + Z 1 )2 + 7(t + Z 1 ) + 2 ((t + Z 2 )2 + 7(t + Z 2 ) + 2)3 1 1 1 0.9989 0.7779
(0) (0) (0) (0.0347)
5 (t + Z 1 )2 + 7(t + Z 1 ) + 2 1 − ((t + Z 2 )2 + 7(t + Z 2 ) + 2)3 1 −1 −1 0.999 − 0.78
(0) (0) (0.0009) (0.0275)
A Kendall correlation coefficient between functional data

6 exp(t + Z 1 ) (t + Z 2 )3 + (t + Z 2 )2 + 3(t + Z 2 ) 0.6 0.4047 0.4138 0.5098 0.5682

(0.0811) (0.0751) (0.1431) (0.1301)
7 exp(t + Z 1 )2 cos(t + Z 2 ) − 0.8 0.3097 0.2982 0.3101 0.0408
(0.0922) (0.1035) (0.07) (0.1458)
8 sin(t + Z 1 ) (t + Z 2 )2 0.4 0.1080 0.1059 0.3382 0.1647
(0.1035) (0.1021) (0.1132) (0.0916)
9 (t + Z 1 )2 + 9(t + Z 1 ) − 5 cos(3t + Z 2 ) 1 − 0.7198 − 0.9476 0.9334 − 0.7244
(0.0853) (0.0358) (0.0458) (0.0562)
10 exp(t 2 + Z 1 ) (t + Z 2 )2 − 8t + Z 2 0.9 0.3621 0.5991 0.8544 0.4620
(0.1078) (0.0706) (0.0485) (0.1215)
11 exp(t + Z 1 ) sin(t + Z 2 ) 0 − 0.0076 0.0087 0.1438 0.0560
(0.1004) (0.0883) (0.0861) (0.1275)

123
1093
1094 D. Valencia et al.

Table 3 Sensitivity to sample

Sample size Model 1 Model 1 Model 2 Model 2
size
τ̂¯1 τ̂¯2 τ̂¯1 τ̂¯2

25 0.4035 0.4017 0.2809 0.3014

(0.1285) (0.1129) (0.1475) (0.1429)
50 0.4044 0.4190 0.3084 0.2774
(0.0719) (0.0724) (0.0923) (0.0835)
100 0.4130 0.4047 0.2882 0.2945
(0.0575) (0.0495) (0.0600) (0.0636)
150 0.4093 0.4094 0.2999 0.2880
(0.0394) (0.0485) (0.0517) (0.0489)
1000 0.4077 0.4096 0.2903 0.2945
(0.0162) (0.0185) (0.0219) (0.0196)

case, this diversity of coefficients increases because there is no total order of the data,
but partial orders that preclude a canonical definition of the dependency measures.
This is exactly what happens for example in the definition of depth measurements for
functional or multivariate data that, depending on each “idea” of order, originates a
different measure.
As we can see, τ̂ depends on the sample size n and on the number of points d
used to discretize the functions. In order to assess the stability of the functional τ̂ with
respect to (n, d), we perform two sensitivity analysis, using the following two pairs
of stochastic processes:

– Model 1: X (t) = exp(t + Z 1 ) and Y (t) = (t + Z 2 )3 + (t + Z 2 )2 + 3(t + Z 2 ) with

σ12 = 0.6.
– Model 2: X (t) = sin(t + Z 1 ) and Y (t) = cos(t + Z 2 ) with σ12 = − 0.7.

The first analysis has as its objective to evaluate the sensibility with respect to the
sample size n. In this case, we move n = 25, 50, 100, 150 and 1000 with d = 50
(the number of points to discretize the functions) fixed. This procedure is repeated
100 times and we report their average and standard deviation. Table 3 shows that the
changes in τ̂¯1 , τ̂¯2 are negligible and they are quite stable with respect to the sample
size.
Now, the same scheme is applied to d. Fix n = 50, and consider d =
25, 50, 100, 150 and 1000 points. This procedure is repeated 100 times. Table 4 illus-
trates the sensitivity with respect to d. It is noteworthy that the coefficients present
good stability with respect to the number of points taken to discretize the functions.
We also carry out the sensitivity analysis for other models, but we do not report them
in this work, since we obtain the same conclusions.
We have also smoothed the curves of Table 2 with cubic spline and take the smooth-
ing parameter λ = 0.1, 0.5, and 1. The results are given in Table 5. In general, the
coefficient τ̂2 is more stable than τ̂1 when we vary the smoothing parameter as also
happened when the data are not smoothed.

123
A Kendall correlation coefficient between functional data 1095

Table 4 Sensitivity to the number of points in the discretization

Number of points Model 1 Model 1 Model 2 Model 2

τ̂¯1 τ̂¯2 τ̂¯1 τ̂¯2

25 0.4056 0.4163 0.3017 0.2822

(0.0761) (0.0770) (0.0902) (0.0843)
50 0.4108 0.4071 0.2834 0.3052
(0.0788) (0.0769) (0.0929) (0.0917)
100 0.4071 0.3911 0.2761 0.2786
(0.0769) (0.0772) (0.0837) (0.0829)
150 0.4163 0.4109 0.2983 0.2801
(0.0770) (0.0804) (0.0919) (0.0825)
1000 0.4108 0.4092 0.3005 0.2978
(0.0788) (0.0822) (0.0839) (0.0834)

Table 5 Sensitivity to smoothing parameter (λ)

λ τ̂1 τ̂2
0.1 0.5 1 0.1 0.5 1

1 0.4759 0.4759 0.4759 0.4694 0.4694 0.4694

2 0.1527 0.1559 0.1869 0.2180 0.2180 0.2180
3 1 1 1 1 1 1
4 0.9788 0.9804 1 1 1 1
5 − 0.1086 − 0.1363 −1 −1 −1 −1
6 0.2735 0.2735 0.2735 0.2735 0.2735 0.2735
7 0.2457 0.2490 0.2180 0.3045 0.3045 0.3045
8 0.2816 0.280 0.2963 0.2816 0.3061 0.3061
9 − 0.5135 − 0.4857 − 0.7502 − 0.9478 − 0.9478 − 0.9478
10 0.5282 0.5282 0.5282 0.6473 0.6473 0.6473
11 0.073 0.0057 0.0743 0.0792 0.0792 0.0792

As already mentioned, the functional Kendall’s τ coefficient depends on the pre-

order used. We have observed that when the sets of curves do not cross each other,
either of the preorders would work properly and they would be very similar. How-
ever, when the function bundle presents crosses and different shapes, it seems more
advisable to use the preorder of the integral (coefficient τ̂2 ).
To finish the simulation study, we have obtained the bootstrap intervals to capture the
uncertainty related to the estimation of coefficient τˆ2 . For this procedure we have used
the standard nonparametric bootstrap interval. Table 6 shows five pairs of processes
that have been re-sampled 500 times and provides the estimated coefficient τˆ2 and the
95% confidence intervals. As can be seen, the intervals contain values very close to
the one obtained with the coefficient.

123
1096 D. Valencia et al.

Table 6 Confidence intervals

Stochastic process 1 Stochastic process 2 τ2 Confidence interval

(t + Z 1 )2 + 7(t + Z 1 ) + 2 ((t + Z 2 )2 + 7(t + Z 2 ) + 2)3 1 (0.9698, 0.9869)

exp(t + Z 1 ) sin(t + Z 2 ) 0.0087 (−0.1755, 0.2278)
sin(t + Z 1 ) (t + Z 2 )2 0.1059 (0.0433, 0.3673)
sin(t + Z 1 ) cos(t + Z 2 ) 0.2774 (0.0980, 0.4735)
exp(t + Z 1 ) (t + Z 2 )3 + (t + Z 2 )2 + 3(t + Z 2 ) 0.4138 (0.2302, 0.5771)

5 Gene data

Existing relationships among genes contain broad information on the structure and
functioning of living beings. Accordingly, the interaction between genes allows us
to understand many life phenomena. These interactions give rise to the construction
of genetic networks. By studying the structural properties of such networks, much
more information may be extracted in order to understand the complex functioning
of living organisms. Different statistical methodologies have been used to estimate
genetic networks, such as graphical models which represent stochastic conditional
dependence between the investigated variables. Graphical Gaussian models and the
Bayesian network are examples of simple graphical models (see, e.g., Whittaker 1990),
but their drawback is that these methods are based on the assumption of identically
and independently distributed variables. Opgen-Rhein and Strimmer (2006) studied
the graphical Gaussian models from the perspective of functional data, where these
two assumptions are not necessary.
Opgen-Rhein and Strimmer (2006) considered the gene expression as a functional
observation, rather than describing the individual time points separately. They built the
networks in the following way: the network nodes are the genes and the correlations
are the connectivity strengths assigned to the edges of the network. They consider the
dynamical correlation introduced in Sect. 4. However, they do not use the dynamical
correlation itself because it represents only marginal dependencies, including indirect
interactions between two variables, since it contains information on the relationships
of each variable with the rest. They use the concept of partial correlation, which
describes the correlation between any two variables i and j, conditioned on all the
other variables, which is the correlation between two variables when the effect of the
other is eliminated. Therefore, if the variables are linearly and conditionally associated,
the partial correlation coefficient is different from zero.
The partial correlation matrix is constructed as follows: Let P = (ρkl ) be the
correlation coefficients, and let Ω be the inverse relationships

Ω = P−1 = (wi j );

then the partial correlations are given by

−wkl
ρ̃kl = √ and P̃ = (ρ̃kl ).
wkk wll

123
A Kendall correlation coefficient between functional data 1097

To test the significance of these correlations and decide which are significant edges,
they employ a large-scale simultaneous hypothesis testing, the “local fdr”, which is
an empirical Bayes estimator of the false discovery rate proposed by Efron (2004)
and (2005). This method computes the posterior probability for an edge to be present
or absent in the gene network. An important question in the use of this method is
whether we can identify a small percentage of interesting cases that deserve further
investigation. In this study, these cases will be the edges present in the network.
We propose a new form of finding connectivity strengths (edges) using the func-
tional τ̂2 and applying the “local fdr” to investigate valid relations.
In order to compare the results, we use the same pre-processed data as in the paper
by Opgen-Rhein and Strimmer (2006). The data set characterizes the response of a
human T-cell line (Jirkat) to a treatment with PMA and ioconomin. After preprocessing
the time course data, 58 genes measured across 10 time points with 44 replications are
provided, (see Rangel et al. 2004). Data were smoothed with linear spline, taking four
basis functions and a smoothing parameter λ = 0.00001. Table 7 shows the correlation
coefficients including the canonical correlation ρ̂c and dynamical correlation ρ̂d for
some pairs of genes. Note how the correlations vary depending on the coefficient used,
which was considered when we analyzed simulated data in Sect. 4.
In order to compare our results with those obtained by Opgen-Rhein and Strimmer
(2006), we calculate the partial correlation matrix from the correlation matrix found
with the functional τ̂2 and we use the “local fdr” algorithm in GeneNet packages,
available in library R-software, to find whether significant edges are present or absent
in our network, with the same cut-off = 0.2 used for calculating the network with
dynamical correlation.
Figures 3 and 4 show the network proposed by Opgen-Rhein and Strimmer (2006)
and our proposed network, respectively. The network calculated with partial dynamical
correlation contains 15 nodes and 9 edges, whereas the network calculated with partial
functional τ̂2 contains 22 nodes and 12 edges. The common nodes in both networks
are CASP8, SOD1, MAPK9, CDC2 and CCNA.
The advantage of using functional τ̂2 instead of the dynamical correlation studied
in Opgen-Rhein and Strimmer (2006) is that our coefficient identifies relationships
between the variables based on the relative ordering among realizations in each group.
And it is not only based on the shape of individual realizations; our coefficient also
takes into account the temporal evolution of each gene, so it is able to identify additional
and different relationships than those given by the dynamical correlation.
Tables 8 and 9 show the results of partial correlation with dynamical correlation
and partial correlation with functional τ̂2 , respectively, which were found through the
“local fdr” algorithm. In addition, we can see the p value for each of the correlations
as well as the nodes included in the networks.
Finally, to explore the relationship between the dynamical correlation and the func-
tional τ̂2 , we make a regression analysis between the partial dynamical correlation
and partial functional τ̂2 for T-cell data. We obtain a R 2 = 0.0634, which is low and
indicates a weak relationship.

123
1098 D. Valencia et al.

Table 7 Gene data

GEN 1 GEN 2 τ̂1 τ̂2 ρ̂c ρ̂d

RB1 CCNG1 − 0.3425 − 0.3996 0.8296 − 0.3266

TRAF5 CLU − 0.3975 − 0.3383 0.7322 − 0.2461
MAPK9 SIVA 0.3298 0.3890 0.9031 0.4665
EDG9 ZNFN1A1 − 0.1839 − 0.3858 0.9081 − 0.011
IL4R MAP2K4 0.2656 0.2706 0.9063 0.4193
JUND LCK − 0.2146 − 0.2114 0.9311 − 0.4443
SCYA2 PPSGKA1 − 0.1522 − 0.2622 0.6055 − 0.1518
ITGAM CTNNB1 0.0962 0.0317 0.8491 0.2373
SMN1 CASP8 − 0.0338 − 0.1755 0.9311 − 0.7743
E2F4 PCNA 0.3869 0.4989 0.9394 0.6312
CCNC PDE4B − 0.3087 − 0.5687 0.8562 − 0.5738
IL16 APC − 0.2474 − 0.3192 0.7916 − 0.1763
ID3 SLA − 0.4027 − 0.4334 0.8905 − 0.7363
CDK4 EGR1 0.1734 − 0.2421 0.9605 0.2091
TCF12 MCL1 0.3467 0.2960 0.9610 0.8361
CDC2 SOD1 0.0486 0.4080 0.9749 0.4871
CCNA2 PIG3 − 0.4017 − 0.4820 0.9361 − 0.3394
IRAK1 SKIIP − 0.0560 − 0.1871 0.5658 0.1197
MYD88 CASP4 0.4778 0.4376 0.9266 0.2225
TCF8 API2 − 0.0063 − 0.1966 0.9292 0.5261
GATA3 RBL2 0.3467 0.4038 0.9352 0.5604
C3X1 IFNAR1 0.2653 0.3805 0.8923 0.6694
FYB IL2R6 − 0.0782 0.5254 0.9301 0.3324
CSF2RA MPO − 0.4588 − 0.4778 0.9048 0.0831
API1 CYP19 − 0.3245 0.1036 0.9116 0.1227
CIR CASP7 − 0.2220 − 0.3827 0.8003 − 0.2234
MAP3K8 JUNB − 0.3044 − 0.4630 0.8913 − 0.6764
IL3RA NFKBIA − 0.4165 − 0.3848 0.7861 − 0.1457
LAT AKT1 − 0.3404 − 0.1649 0.8210 − 0.0764
RB1 MAPK9 0.5328 0.6964 0.9767 0.7740
RB1 CASP4 − 0.4567 − 0.4207 0.9672 − 0.4748
TRAF5 LCK 0.3647 0.5856 0.8970 0.4583
TRAF5 ITGAM − 0.4820 − 0.5941 0.9494 − 0.6519
TRAF5 CTNNB1 0.4397 0.5920 0.8145 0.2573
TRAF5 CSF2RA − 0.5116 − 0.6342 0.9318 − 0.6458
EDG9 C3X1 0.5370 0.7030 0.9626 0.6056
ZNFN1A1 CASP8 − 0.2611 − 0.63 0.9467 − 0.4740
IL4R ITGAM 0.4926 0.5856 0.9611 0.8036
MAP2K4 IL16 0.1078 0.1015 0.6217 0.0634
JUND SMN1 − 0.5846 − 0.4419 0.9528 − 0.6019

123
A Kendall correlation coefficient between functional data 1099

Fig. 3 Gene dependence network using dynamical correlation

Fig. 4 Gene dependence network using functional τ̂2

6 Robustness

As commented in the Introduction, we now analyze the robustness of our coefficients

τ̂1 and τ̂2 and compare them with the results obtained with the dynamical and canonical
correlation (ρ̂d and ρ̂c , respectively). We contaminate the dataset with outliers, defining
a functional outlier as in Febrero et al. (2008): a “curve [that] has been generated by
a stochastic process with a different distribution than the rest of curves, which are
assumed to be identically distributed”. Given this definition, we use three types of
outliers: shape outliers, magnitude outliers and shape–magnitude outliers.

123
1100 D. Valencia et al.

Table 8 Partial correlation with dynamical correlation

Correlation node1 node2 pval prob

0.5196 JUND EGRI 4.549e−09 0.9821

0.3971 CDC2 CCNA2 1.490e−05 0.9821
0.3888 API2 N FK BI A 2.325e−05 0.9821
0.3817 CASP8 IFNAR1 3.365e−05 0.9778
0.3749 IL16 EGRI 4.755e−05 0.9317
− 0.3543 MAPK9 SLA 0.2917e−04 0.9317
0.3503 IL16 SOD1 1.560e−04 0.9317
0.3477 IL2RG API1 1.759e−04 0.9079
0.3414 MCL1 API2 2.337e−04 0.8790

Table 9 Partial correlation with functional τ̂2

Correlation node1 node2 pval prob

−0.3235 PPS6KA1 FYB 2.286e−05 0.9599

0.3029 IRAK1 MPO 7.744e−05 0.9599
0.3019 SMN1 CC N C 8.202e−05 0.9599
0.2990 RB1 MAP3K8 9.678e−05 0.9400
0.2932 RB1 MAPK9 1.336e−04 0.9287
− 0.2842 ITGAM SOD1 2.184e−04 0.9287
− 0.2839 CDC2 CYP19 2.211e−04 0.8543
− 0.2687 IL4R C3X1 4.880e−04 0.8543
− 0.2680 GATA3 C3X1 5.059e−04 0.8543
0.2628 CASP8 PIG3 6.554e−04 0.8543
0.2627 CTNNB1 SKIIP 6.586e−04 0.8543
0.2600 TCF12 CCNA2 7.488e−04 0.8543

We generate 50 curves for the previously studied processes. (Recall that σ12 is the
correlation between the normal random variables Z 1 and Z 2 .)

X (t) = exp(t + Z 1 ), and Y (t) = (t + Z 2 )3 + (t + Z 2 )2 + 3(t + Z 2 ), σ12 = 0.6,

and the types of outliers to be considered are:

– Shape outliers. Changing the argument, t to (1 − t).
– Magnitude outliers. Adding a constant to the original process, X (t) to X (t) + k.
In our case we will use k = 60.
– Shape–magnitude outliers. Changing the argument and adding a constant to the
original function, X (t) to X (1 − t) + k.

123
A Kendall correlation coefficient between functional data 1101

Table 10 Contamination with shape outliers

Contaminated groups Type of outliers No. outl τ̂1 τ̂2 ρ̂d ρ̂c

None None 0 0.454 0.454 0.549 0.544

X (t) Shape 1 0 0 0.0231 0.0007
X (t) Shape 2 0 0 0.0242 0.0669
X (t) Shape 3 0 0 0.0244 0.1292
X (t) Shape 4 0 0 0.0245 0.1284
X (t), Y (t) same position Shape 1 0 0 0 0.2122
X (t), Y (t) same position Shape 2 0 0 0 0.4137
X (t), Y (t) same position Shape 3 0 0 0 0.2707
X (t), Y (t) same position Shape 4 0 0 0 0.27
X (t), Y (t) different position Shape 1 0 0 0.0296 0
X (t), Y (t) different position Shape 2 0 0 0.0301 0.0698
X (t), Y (t) different position Shape 3 0 0 0.0303 0.1446
X (t), Y (t) different position Shape 4 0 0 0.0305 0.1393
Bold values indicate the greatest variation

Table 11 Contamination with magnitude outliers

Contaminated groups Type of outliers No. outl τ̂1 τ̂2 ρ̂d ρ̂c

None None 0 0.454 0.454 0.549 0.544

X (t) Magnitude 1 0.0033 0.0033 0.096 0.002
X (t) Magnitude 2 0.0016 0 0.009 0.043
X (t) Magnitude 3 0.008 0.008 0.17 0.18
X (t) Magnitude 4 0.026 0.026 0.095 0.126
X (t), Y (t) same position Magnitude 1 0.008 0.009 0.16 0.34
X (t), Y (t) same position Magnitude 2 0.0131 0.0147 0.2757 0.4022
X (t), Y (t) same position Magnitude 3 0.0163 0.0196 0.3346 0.4239
X (t), Y (t) same position Magnitude 4 0.0343 0.0375 0.3419 0.4292
X (t), Y (t) different position Magnitude 1 0.0196 0.0245 0.1786 0.0079
X (t), Y (t) different position Magnitude 2 0.0212 0.0261 0.1766 0.0384
X (t), Y (t) different position Magnitude 3 0.0131 0.0196 0.1135 0.1652
X (t), Y (t) different position Magnitude 4 0.1192 0.1274 0.2091 0.1076
Bold values indicate the greatest variation

We use different ways to contaminate the data:

1. Contaminating a group.
2. Contaminating two groups in the same position, i.e, the pair (xi , yi ) is replaced
by an outlier.
3. Contaminating two groups in different positions, i.e, the pair (xi , y j ) where i = j.
Each measure is calculated before contaminating the data (row 1). Once data have
been contaminated with outliers from different types, we report the relative variation

123
1102 D. Valencia et al.

Table 12 Contamination with shape–magnitude outliers

Contaminated groups Type of outliers No. outl τ̂1 τ̂2 ρ̂d ρ̂c

None None 0 0.454 0.454 0.549 0.544

X (t) Shape-magnit 1 0.003 0.004 0.09 0.0008
X (t) Shape-magnit 2 0.001 0 0.006 0.028
X (t) Shape-magnit 3 0.008 0.008 0.15 0.18
X (t) Shape-magnit 4 0.02 0.02 0.079 0.11
X (t), Y (t) same position Shape-magnit 1 0.008 0.009 0.16 0.41
X (t), Y (t) same position Shape-magnit 2 0.013 0.014 0.27 0.43
X (t), Y (t) same position Shape-magnit 3 0.016 0.019 0.33 0.41
X (t), Y (t) same position Shape-magnit 4 0.034 0.037 0.34 0.41
X (t), Y (t) different position Shape-magnit 1 0.019 0.024 0.18 0.002
X (t), Y (t) different position Shape-magnit 2 0.021 0.026 0.18 0.04
X (t), Y (t) different position Shape-magnit 3 0.013 0.019 0.12 0.19
X (t), Y (t) different position Shape-magnit 4 0.119 0.127 0.22 0.11
Bold values indicate the greatest variation

of the association measure with respect to its value in the uncontaminated data set. We
compare our results with those obtained by the dynamical correlation and canonical
correlation. We can see that functional τ̂1 and τ̂2 coefficients are not affected by the
presence of shape outliers, while the dynamical correlation and canonical correla-
tion coefficients are sensitive to them. For magnitude outliers and shape–magnitude
outliers, our coefficients present small variations unlike the other coefficients, which
present variations up to 40 percent of the original value. The results are given in
Tables 10, 11 and 12, where the values in bold are those that provide the largest vari-
ation in each of the cases. We can see that the functional τ̂1 as well as the functional
τ̂2 do not present a significant variation, while ρ̂d and ρ̂c give the largest variations in
almost all cases.

7 Conclusions

We have introduced a new numerical dependence measure between two sets of func-
tional data. Our technique is a natural extension of the Kendall τ coefficient when the
data are curves. In order to build this new coefficient, we have also introduced the
concordance concept between pairs of functional data. We have presented examples
of applications showing the usefulness of the new coefficients introduced for both
simulated and real data.
We have compared the performance of our measure with other coefficients, such
as dynamical correlations and canonical correlations. The coefficients presented here
allow us to identify the global dependence between two groups of functional data
regardless of the shape of their realizations. These coefficients are presented as a new
alternative to find dependence between functional data. In addition, this coefficient’s
implementation is straightforward.

123
A Kendall correlation coefficient between functional data 1103

An interesting example with real data is studied. The data set corresponds to a
microarray time series from a human T-cell experiment. We obtain the partial func-
tional τ̂2 for each pair of genes and construct a gene network.
We also study the sensitivity of our coefficients and conclude that these coefficients
present good stability with respect to sample size and the number of points taken
to discretize the functions. However the coefficient defined with the preorder of the
maximun presents less stability when the data are smoothed. In terms of robustness,
our coefficients can be considered quite stable in the presence of functional outliers
in comparison with the measures used as a benchmark.
As one of the referees suggested, it would be very interesting to carry out a deeper
inference study on these new coefficients. In this work, a first approximation with the
boostrap intervals is presented, but an analysis based on hypothesis testing will be the
subject of future work.

References
Borovskikh Y (1996) U-statistics in Banach space. VSP BV, Oud-Beijerland
Cardot H, Ferraty F, Sarda P (1999) Functional linear model. Stat Probab Lett 45:11–22
Cuevas A, Febrero M, Fraiman R (2004) An ANOVA test for functional data. Comput Stat Data Anal
47:111–122
Delicado P (2007) Functional k-sample problem when data are density functions. Comput Stat 22:391–410
Dubin JA, Müller HG (2005) Dynamical correlation for multivariate longitudinal data. J Am Stat Assoc
100:872–881
Efron B (2004) Large-scale simultaneous hypothesis testing: the choice of a null hypothesis. J Am Stat
Assoc 99:96–104
Efron B (2005) Local false discovery rates. Technical report, Department of Statistics, Stanford University
Escabias M, Aguilera A, Valderrama M (2004) Principal components estimation of functional logistic
regression: discussion of two different approaches. J Non Parametr Stat 16(3–4):365–384
Febrero M, Galeano P, González-Manteiga W (2008) Outlier detection in functional data by depth measures,
with application to identify abnormal N O x levels. Envirometrics 19:331–345
He G, Müller HG, Wang JL (2000) Extending correlation and regression from multivariate to functional
data. In: Puri ML (ed) Asymptotics in statistics and probability. VSP, Leiden, pp 197–210
Kendall M (1938) A new measure of rank correlation. Biometrika Trust 30(1/2):81–93
Leurgans SE, Moyeed RA, Silverman BW (1993) Canonical correlation analysis when data are curves. J R
Stat Soc B 55:725–740
López-Pintado S, Romo J (2007) Depth-based inference for functional data. Comput Stat Data Anal
51:4957–4968
López-Pintado S, Romo J (2009) On the concept of depth for functional data. J Am Stat Assoc 104:718–734
Opgen-Rhein R, Strimmer K (2006) Inferring gene dependency networks from genomic longitudinal data:
a functional data approach. REVSTAT 4(1):53–65
Pezulli S, Silverman B (1993) Some properties of smoothed components analysis for functional data.
Comput Stat 8:1–16
Ramsay JO, Silverman BW (2005) Functional data analysis, 2nd edn. Springer, New York
Rangel C, Angus J, Ghahramani Z et al (2004) Modelling T-cell activation using gene expression profiling
and state-space models. Bioinformatics 20:1361–1372
Scarsini M (1984) On measure of concordance. Stochastica 8(3):201–218
Schwabik S, Guoju Y (2005) Topics in Banach space integration. World Scientific Publishing, Singapore
Taylor MD (2007) Multivariate measures of concordance. Ann Inst Stat Math 59:789–806
Taylor MD (2008) Some properties of multivariate measures of concordance. arXiv:0808.3105 [math.PR]
Whittaker J (1990) Graphical models in applied multivariate statistics. Wiley, New York

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

123