2.1. Population versions of PRVCC
Let X and Y be two random variables following a
continuous bivariate distribution. Denote by F
X
and F
Y
the marginal distributions of X and Y, respectively. Then,
according to Pearson [30], one of the population versions
of PRVCC can be defined, in modern notation, by
r
P
H
Y; XðÞ9 corr F
X
XðÞ; Y
½
¼
C½F
X
ðXÞ; Y
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
V½F
X
ðXÞ
p ffiffiffiffiffiffiffiffiffiffiffi
VðYÞ
p
: ð2Þ
Swapping X and Y in (2) gives the other version
r
P
H
X; YðÞ9 corr F
Y
YðÞ; X
½
¼
C½F
Y
ðYÞ; X
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
V½F
Y
ðYÞ
p ffiffiffiffiffiffiffiffiffiffiffi
VðXÞ
p
:
Pearsonhadshownthatwhen(X,Y) follows a bivariate normal
distribution with parent correlation
ρ 9 corrðX; YÞ [30],
r
P
H
Y; XðÞ¼r
P
H
X; YðÞ¼
ffiffiffiffi
3
π
r
ρ:
2.2. Sample versions of PRVCC
Let fðX
i
; Y
i
Þg
n
i ¼ 1
denote n independent and identically
distributed (i.i.d.) data pairs drawn from a continuous
bivariate population. Rearranging fX
i
g
n
i ¼ 1
in ascending
order, we get a new sequence, X
ð1Þ
o ⋯ o X
ðnÞ
, which is
termed the order statistics of X [33–35]. Suppose that X
j
is
at the kth position in the sorted sequence fX
ðiÞ
g
n
i ¼ 1
. The
number 1 r kr n is termed the rank of X
j
and is denoted
by P
j
. Similarly we can also define the rank of Y
j
which is
denoted by Q
j
. Denote by ξ the arithmetic mean of n data
points f
ξ
i
g
n
i ¼ 1
. Then, based on (2), the sample version of
PRVCC with respect to fðX
i
; Y
i
Þg
n
i ¼ 1
can be constructed as
the product-moment correlation coefficient between
fY
i
g
n
i ¼ 1
and the empirical cdf of fX
i
g
n
i ¼ 1
, f
^
F
X
ðX
i
Þg
n
i ¼ 1
, being
r
H
Y
i
; X
i
ðÞ9
P
n
i ¼ 1
h
^
F
X
ðX
i
Þ
^
F
X
ðX
i
Þ
i
ðY
i
Y Þ
P
n
i ¼ 1
2
4
^
F
X
ðX
i
Þ
^
F
X
ðX
i
Þ
3
5
2
P
n
i ¼ 1
ðY
i
Y Þ
2
)
1=2
8
>
<
>
:
ð3Þ
which, upon substitution of the relationship
^
F
X
ðX
i
Þ¼P
i
=n
with some straightforward algebra (see Appendix A), yields
r
H
Y
i
; X
i
ðÞ¼
3n
nþ1
1=2
1
nðn1Þ
P
n
i ¼ 1
2P
i
1nðÞY
i
1
n1
P
n
i ¼ 1
ðY
i
Y Þ
2
1=2
: ð4Þ
In a similar manner, r
H
ðX
i
; Y
i
Þ can also be obtained by
replacing P
i
and Y
i
with Q
i
and X
i
respectively in (4).
Note that in general r
H
ðX
i
; Y
i
Þar
H
ðY
i
; X
i
Þ. The choice
between r
H
ðX
i
; Y
i
Þ and r
H
ðY
i
; X
i
Þ depends on different roles
played by X and Y under Scenarios 0–2 mentioned in the
previous section.
2.3. General properties
From (3) and (4), it follows that PRVCC possesses the
following general properties:
1. r
H
ðY
i
; X
i
Þ and r
H
ðX
i
; Y
i
ÞA ½1; 1 for all (X,Y);
2. r
H
ðY
i
; X
i
Þ¼r
H
ðY
i
; Ψ ðX
i
ÞÞ for any strict monotone
increasing function
ΨðÞ;
3. r
H
ðX
i
; Y
i
Þ¼r
H
ðX
i
; Ψ ðY
i
ÞÞ for any strict monotone
increasing function
ΨðÞ;
4. the expectations of r
H
ðY
i
; X
i
Þ and r
H
ðX
i
; Y
i
Þ equal zero if
X and Y are independent;
5. r
H
ð†; ‡Þ¼r
H
ð†; ‡Þ¼r
H
ð†; ‡Þ¼r
H
ð†; ‡Þ for
both r
H
ðY
i
; X
i
Þ and r
H
ðX
i
; Y
i
Þ;
6. r
H
ðY
i
; X
i
Þ and r
H
ðX
i
; Y
i
Þ are scale and shift invariant with
respect to both X and Y;
7. r
H
ðY
i
; X
i
Þ belongs to the family of Daniels's generalized
correlation coefficients [19], that is
r
H
Y
i
; X
i
ðÞ¼
P
n
i ¼ 1
P
n
j ¼ 1
a
ij
b
ij
P
n
i ¼ 1
P
n
j ¼ 1
a
2
ij
P
n
i ¼ 1
P
n
j ¼ 1
b
2
ij
1=2
ð5Þ
where a
ij
¼P
i
P
j
and b
ij
¼Y
i
Y
j
;
8. both r
H
ðY
i
; X
i
Þ and r
H
ðX
i
; Y
i
Þ converge to normal dis-
tributions under some mild conditions when the sam-
ple size n-1.
Note that, although not difficult to verify (see Appendix
B), the result (5) is very important. It not only reveals the
close relationship of PRVCC with PPMCC, SR and KT, the
three most commonly used correlation coefficients in the
literature, but also leads to the last property directly from
the argument of Daniels [19]. Moreover, since (5) is
interpretable as a function of U-statistics, the asymptotic
normality of PRVCC can also be established from
Hoeffding's theory (Theorem 7.5 in [36]).
3. Asymptotic properties of PRVCC in bivariate normal
samples
In this section we derive the closed form formulas
concerning the expectation and variance of r
H
ðY
i
; X
i
Þ in the
normal cases (Scenario 0). For notational compactness, the
arguments of r
H
ðY
i
; X
i
Þ will be omitted in the following
discussion unless ambiguity occurs. We first formulate
Lemmas (1) and (2) which are mandatory for establishing
the major result of Theorem 1 in this work.
3.1. Auxiliary results
Lemma 1. Assume that the quadruple ðW
1
; W
2
; W
3
; W
4
Þ
follows a quadrivariate normal distribution with EðW
r
Þ¼0,
VðW
r
Þ¼σ
2
r
, and ϱ
rs
¼corr W
r
; W
s
ðÞfor r; s ¼ 1; …; 4. Let
HðtÞ¼1 for t 4 0 and HðtÞ¼0 for t r 0. Then
I 9 E HðW
1
ÞW
2
W
3
W
4
ð6Þ
¼
σ
2
σ
3
σ
4
ffiffiffiffiffiffi
2
π
p
ϱ
12
ϱ
34
þϱ
13
ϱ
24
þϱ
14
ϱ
23
ϱ
12
ϱ
13
ϱ
14
: ð7Þ
Proof. See Appendix C □.
Lemma 2. Let fðX
i
; Y
i
Þg
n
i ¼ 1
be n i.i.d. sample pairs drawn
from a standard bivariate normal population N ð0; 0; 1; 1;
ρÞ.
W. Xu et al. / Signal Processing 119 (2016) 190–202192