On the Multivariate Gamma-Gamma (ΓΓ) Distribution with Arbitrary

Correlation and Applications in Wireless Communications

Zhang, J., Matthaiou, M., Karagiannidis, G. K., & Dai, L. (2016). On the Multivariate Gamma-Gamma (ΓΓ)
Distribution with Arbitrary Correlation and Applications in Wireless Communications. IEEE Transactions on
Vehicular Technology, 65(5), 3834-3840. https://doi.org/10.1109/TVT.2015.2438192

Published in:
IEEE Transactions on Vehicular Technology

Document Version:
Peer reviewed version

Queen's University Belfast - Research Portal:

Link to publication record in Queen's University Belfast Research Portal

Publisher rights
©2015 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future
media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or
redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

General rights
Copyright for the publications made accessible via the Queen's University Belfast Research Portal is retained by the author(s) and / or other
copyright owners and it is a condition of accessing these publications that users recognise and abide by the legal requirements associated
with these rights.

Take down policy

The Research Portal is Queen's institutional repository that provides access to Queen's research output. Every effort has been made to
ensure that content in the Research Portal does not infringe any person's rights, or applicable UK laws. If you discover content in the
Research Portal that you believe breaches copyright or violates any law, please contact openaccess@qub.ac.uk.

Download date:20. Nov. 2022


On the Multivariate Gamma-Gamma (ΓΓ) Distribution
with Arbitrary Correlation and Applications in
Wireless Communications
Jiayi Zhang, Member, IEEE, Michail Matthaiou, Senior Member, IEEE, George K. Karagiannidis, Fellow, IEEE,
and Linglong Dai, Senior Member, IEEE
arXiv:1505.06685v1 [cs.IT] 25 May 2015
Abstract—The statistical properties of the multivariate Gamma-
Gamma (ΓΓ) distribution with arbitrary correlation have remained
unknown. In this paper, we provide analytical expressions for the
joint probability density function (PDF), cumulative distribution
function (CDF) and moment generation function of the multivariate
ΓΓ distribution with arbitrary correlation. Furthermore, we present
novel approximating expressions for the PDF and CDF of the sum
of ΓΓ random variables with arbitrary correlation. Based on this
statistical analysis, we investigate the performance of radio frequency
and optical wireless communication systems. It is noteworthy that
the presented expressions include several previous results in the
literature as special cases.
Index Terms—Diversity receivers, free-space optical communica-
tions, multivariate ΓΓ distribution.
I. I NTRODUCTION
In practical wireless communications, the electromagnetic sig-
nal experiences composite small-scale fading and shadowing
simultaneously. Over the past decades, several works have fo-
cused on the performance analysis of wireless systems over
composite fading channels [1], [3], [4]. On a different note,
diversity techniques are widely used to reduce the effects of
multipath fading. If the antennas are sufficiently separated, it is
reasonable to assume that signals received by different antennas
are independent. However, this assumption is intimately crude for
systems with closely spaced antennas, such as mobile phones. In
general, correlation between channels results in a degradation of
the diversity gain. Therefore, it is important to quantify rigorously
the effects of correlation in real-life practical scenarios.
Recently, the Gamma-Gamma (ΓΓ) distribution has been pro-
posed as a general and mathematically tractable composite fading
model [2], [5], [6]. The ΓΓ distribution is equivalent to the
Copyright (c) 2015 IEEE. Personal use of this material is permitted. However,
permission to use this material for any other purposes must be obtained from the
IEEE by sending a request to pubs-permissions@ieee.org.
This work was supported by the National Key Basic Research Program of China
(No. 2013CB329203), National Natural Science Foundation of China (Grant No.
61201185), National High Technology Research and Development Program of
China (Grant No. 2014AA01A704), and China Postdoctoral Science Foundation
(No. 2014M560081).
J. Zhang and L. Dai are with the Department of Electronic Engineering as
well as Tsinghua National Laboratory of Information Science and Technology
(TNList), Tsinghua University, Beijing 100084, P. R. China (e-mail: {jiayizhang,
daill}@tsinghua.edu.cn).
M. Matthaiou is with the School of Electronics, Electrical Engineering and
Computer Science, Queen’s University Belfast, Belfast, BT3 9DT, U.K. and with
the Department of Signals and Systems, Chalmers University of Technology,
Gothenburg SE-412 96, Sweden (e-mail: m.matthaiou@qub.ac.uk).
G. K. Karagiannidis is with the Department of Electrical and Computer
Engineering, Khalifa University, PO Box 127788, Abu Dhabi, UAE and with
the Department of Electrical and Computer Engineering, Aristotle University of
Thessaloniki, 54 124, Thessaloniki, Greece (e-mail: geokarag@ieee.org).
1
squared generalized-K distribution and includes the squared K
distribution and double-Rayleigh distribution as special cases.
In radio frequency (RF) communications, the performance of
diversity systems over ΓΓ fading channels has been analyzed in
[5], [6]. In particular, [5] investigated the outage probability (OP)
of selection combining (SC) receivers, as well as the average bit
error rate (BER) of maximal ratio combining (MRC) receivers
operating over exponentially correlated distributed fading chan-
nels. Furthermore, a simple approximation for the distribution of
the sum of independent distributed ΓΓ random variables (RVs)
was provided in [6]. However, all these works are based on the
assumption of independent or exponentially correlated ΓΓ RVs.
Furthermore, the ΓΓ distribution has been used in the performance
investigation of free-space optical (FSO) links over atmospheric
turbulence conditions [6]–[9]. These previous results are not
theoretically attractive when the underlying FSO links are not
independently distributed. To the best of the authors’ knowledge,
a statistical analysis of the multivariate ΓΓ distribution with
arbitrary correlation is still not available in the literature.
Motivated by the above discussion, this paper makes the
following specific contributions:
• Using the Green’s matrix approximation [10], we derived
new analytical expressions for the probability density func-
tion (PDF), cumulative distribution function (CDF) and mo-
ment generation function (MGF) of the joint ΓΓ distribution
with arbitrary correlation. We point out that the presented re-
sults encompass several previously known results as special
cases (e.g., those in [2], [5], [11]).
• Efficient approximations to the PDF and CDF of the sum of
arbitrarily correlated ΓΓ RVs are derived for integer values
of fading severity m. The resulting expressions can be easily
and quickly evaluated by using a recursive formula.
• Furthermore, in order to reveal the importance of the
proposed statistical formulations, we study the perfor-
mance of SC and MRC receivers over arbitrarily correlated
generalized-K fading channels.
• Finally, in the context of FSO communication systems with
spatial diversity, we analyze the BER performance when
strong turbulence channels is assumed. The derived expres-
sion extends the results on independent and exponentially
correlated ΓΓ fading channels in [5].
Notations: We use upper and lower case boldface to denote
matrices and vectors, respectively. The expectation is given by
E (·). The norm of a vector is given by || · ||, while the matrix
†
determinant by | · |. The Hermitian operation is defined as (·) .
The set of integer numbers is expressed via Z. Finally, let X i,j
denote the (i, j)-th element of the matrix X .
 2

II. M ULTIVARIATE DISTRIBUTION WITH ARBITRARY equation C = Σ, i.e., ζi ϑi = Σi,i = 1 and ζi ϑj = Σi,j = ρi,j .
CORRELATION The accuracy of the Green’s matrix approximation technique will
A. Multivariate Nakagami-m Distribution with Arbitrary Corre- be verified in Section V. In any case, its computational com-
lation plexity, determined by the dimension of the correlation matrix, is
given by O(N 2 ).
Let r be a Nakagami-m RV, whose PDF is given by [10]
 2
2r2m−1 r B. Multivariate Gamma Distribution with Arbitrary Correlation
fr (r) = exp − , r ≥ 0, (1)
Γ (m) Ωm Ω
It is well known that a Nakagami-m RV is the square root
where Γ(·) represents the Gamma function [15, Eq. (8.310.1)], of a Gamma RV. Furthermore, it has been numerically validated
Ω , E(r2 )/m with E(r2 ) being the average power, and m ≥ 1/2 in [14] that the correlation coefficients of Gamma RVs can be
denotes the fading severity. Note that (1) is another representation taken identical to the corresponding coefficients of Nakagami-
of the classical formula for the single Nakagami-m PDF. Without m RVs with satisfactory accuracy. Performing N RV trans-
loss of generality, let u be a Gamma RV which represents formations onto (3), the joint PDF of multiple Gamma RVs,
shadowing and it is assumed in the following that E(u2 ) = 1. ω , [ω1 , ω2 , . . . , ωN ] with ωi = ri2 = ||xi ||2 for i = 1, . . . , N ,
Then, the PDF of u can be expressed as [7] can be expressed as
β β β−1 −βu P
N pn,n ωn
fu (u) = u e , u, β ≥ 0, (2) m − n=1 Ω ∞
Γ (β) W| e
|W X m+iN −1 −1
ω) =
fω (ω ω1m+i1 −1 ωN
where β is a channel parameter related to the effective number of Γ (m) i1 ,··· ,iN −1 =0
discrete scatterers. Let y1 , y2 , . . . , y2m be N -dimensional column P1
N− " #
−N m−2 ij N−
Y1 N−1
Y N−1
Y |pn,n+1 |2in
vectors, which are independent of each other and normally m+i +i −1
×Ω j=1
ωj j−1 j . (5)
distributed with zero means and arbitrary correlation matrix Σ n=1 j=2 n=1
in !Γ (m + in )
which is defined as Σ i,j = 1 for i = j, and Σ i,j = ρi,j for
i 6= j, where 0 ≤ ρi,j < 1 is the correlation coefficient and To derive (5), we have used the infinite series representations
1 ≤ i, j ≤ N . Here we assume that all vectors have the same of Iv (·) [15, Eq. (8.445.1)]. For the special case of exponential
correlation matrix Σ . correlation, where the correlation matrix is defined as Σ i,j = 1
Now, let x1 , x2 , . . . , xN be 2m-dimensional column vectors, for i = j and Σ i,j = ρ|i−j| for i 6= j, (5) reduces to [5, Eq. (2)].
with xn composed of the n-th components of the yi and the
norm of xn being rn , i.e., rn , ||xn ||. Then, ri is a Nakagami- C. Multivariate ΓΓ Distribution with Arbitrary Correlation
m RV. We recall that the joint PDF of N Nakagami-m RVs, The multivariate ΓΓ distribution can be derived from the
r , [r1 , r2 , . . . , rN ], with arbitrary correlation and identically product of two multivariate Gamma RVs. Let us assume that
distributed is given by [10, Eq. (2)] [12, Eq. (9)], as follows the Gamma RVs in one multivariate Gamma distribution are
"
2 N −1 correlated but they are independent from the Gamma RVs of the
W |m 2N r1m−1 rN
|W m pN,N rN Y 1−m
fr (rr ) = e− Ω |pn,n+1 | rn other multivariate Gamma distribution.1 Moreover, let u and ω
ΩN +m−1 Γ (m) n=1 be two independent Gamma RVs with PDFs fu (u) and fω (ω),
 #
p
− n,n
2
rn 2 |pn,n+1 | respectively. The ΓΓ RV is denoted by z = uω.
×e Ω Im−1 rn rn+1 , (3)
Ω Theorem 1. The joint PDF of the multivariate ΓΓ distribution
where W denotes the inverse of Σ with elements pi,j , and Iv (·) with arbitrary correlation is given by
NP
is the modified Bessel function of the first kind with order v [15, m ∞   N (m+β) +
−1
ij YN µ
Eq. (8.406)]. 2N |W
W| X β 2
j=1 zj j
fz (zz ) = η
[Γ (β)]N Γ (m) i1 ,...,iN −1 =0 Ω p j
j=1 j,j
Remark. Note that, although the parameter 2m seems to be ! N −1 " #
r 2i
restricted as a positive integer, (3) can be used for any positive βpj,j Y |pn,n+1 | n
values of m not less than 1/2. This is because it satisfies the × K2ηj 2 zj , (6)
Ω n=1
in !Γ (m + in )
necessary and sufficient conditions to be a joint distribution
function, as it was emphasized in [10]. m+α +β m+α −β
where µj , 2
j
− 1, ηj , 2
j
with αj = ij for j = 1,
The joint PDF expression (3) requires the matrix W to have the αj = iN −1 for j = N , and αj = ij−1 +ij for j = 2, 3, . . . , N −1.
tridiagonal property. However, in the general case, the inverse of Also, Kv (·) denotes the modified Bessel function of the second
Σ is not a tridiagonal matrix. Therefore, we need to approximate kind with order v [15, Eq. (8.407.1)].
Σ with a Green’s matrix C [13], which has the form Proof: The multivariate ΓΓ distribution can be obtained as
 
ζ1 ϑ1 ζ1 ϑ2 . . . ζ1 ϑN Z ∞ Z ∞YN  
 ζ1 ϑ2 ζ2 ϑ2 . . . ζ2 ϑN  −1 zn
  fz (zz ) = ... ωn fun ω ) dω
fω (ω ω , (7)
C = .. .. .. , (4) 0 0 n=1
ωn
 . . . ... 
ζ1 ϑN ζ2 ϑN . . . ζN ϑN where z , [z1 , z2 , . . . , zN ], with zn = un ωn . Moreover, fun (·) is
the PDF of un following a Gamma distribution with parameter β
where ζi and ϑi , are two sequences of real numbers. In order
to determine C , nonlinear methods (e.g., Levenberg Marquardt, 1 This assumption is reasonable, since in most practical wireless channels the
quasi Newton, and conjugate gradient) may be used to solve the effects of fast fading and shadowing are usually independent from each other.
 3

given in (2), while u , [u1 , u2 , . . . , uN ] is a vector containing N the multivariate ΓΓ distribution is given by
independent and identically distributed (i.i.d.) Gamma-distributed Z ∞ Z ∞ N
!
X
RVs with the same parameter β. By substituting (2) and (5) into Mz (ss) = ... exp − sn zn fz (zz ) dzz . (10)
(7) and using [15, Eq. (3.471.9)], (6) can be derived. 0 0 n=1
Following a similar procedure as in deriving (8), we simply
To the best of the authors’ knowledge, (6) is new and has not
express the exponential function in terms of a Meijer’s-G function
been presented in previous literature. Although (6) is given in
[16, Eq. (11)] and utilize the relation [16, Eq. (21)]. Then, the
terms of infinite series, only a finite number of terms is needed
joint MGF expression in (9) is derived.
to get a satisfactory accuracy (e.g., smaller than 10−6 ) for all
considered cases (a detailed discussion is provided in Section V).
D. Special Cases
Corollary 1. The joint CDF of the multivariate ΓΓ distribution Now we present some PDF and CDF expressions for special
z with arbitrary correlation is given by cases, e.g., bivariate arbitrarily correlated or exponentially corre-
∞ N
lated ΓΓ RVs. Note that these expressions give the link to previous
W |m
|W X Y −m−α results and prove the correctness of our generalized formulas. For
Fz (zz ) = N
pj,j j
[Γ (β)] Γ (m) i1 ,...,iN −1 =0 j=1 bivariate correlated ΓΓ RVs, where z = [z1 , z2 ], the joint PDF
   NY−1
"
2in
# and CDF reduce to [2, Eq. (4)] and [2, Eq. (10)], respectively.
2,1 βpj,j 1 |p |
 n,n+1 For the special case of the exponentially correlated multivariate
× G1,3 zj  , (8)
Ω m+αj , β, 0 in !Γ (m+in ) ΓΓ distribution, (6) simplifies to [5, Eq. (5)]. Furthermore, when
n=1
we consider independent ΓΓ RVs, the joint PDF, CDF and
where G[·] is a Meijer’s-G function [15, Eq. (9.301)]. MGF simplify to the product of multivariate independent ΓΓ
distributions [11, Eqs. (2), (3) and (4)], respectively. We finally
Proof: We start with the definition of the CDF as point out that for exponentially correlated multivariate ΓΓ RVs,
P1
  N (m2+β) +N− the joint CDF formula (8) reduces to [5, Eq. (7)] and the joint
∞ ij
N W |m
2 |W X β j=1 MGF expression (9) reduces to [5, Eq. (10)].
Fz (zz ) = N
[Γ (β)] Γ (m) i1 ,··· ,iN−1 =0 Ω
" # N Z r ! III. S UM OF ΓΓ RV S WITH A RBITRARY C ORRELATION
N−1
Y |pj,j+1 |2ij Y zj zjµj βpj,j
× K2ηj 2 zj dzz . A. PDF and CDF of the Sum of ΓΓ RVs
n=1
in !Γ (m + in ) j=1 0 pηj,jj Ω
The sum of N ΓΓ RVs is defined as
By transforming the Bessel function Kv (·) into a Meijer’s-G N
X N
X
function through [16, Eq. (14)] and using [16, Eq. (26)], we can S, zn = u n ωn , (11)
conclude the proof with [15, Eq. (9.31.5)]. n=1 n=1

where un are i.i.d. Gamma RVs with parameters (β, 1/β) whose
Note that for (m + αj − β) ∈ / Z and by using [17, Eq.
PDF expression was given in (2), and ωn are non-identical and
(07.34.26.0004.01)], (8) can be written in terms of the more
correlated Gamma RVs with parameters (mn , Ωn ) and joint PDF
familiar generalized hypergeometric function 1 F2 (·) as
" # given in[18, Eq. (11)]. According to the approach presented in
∞ N−1
W |m
|W X Y |pn,n+1 |2in [6], we can rewrite (11) as
Fz (zz ) = N
[Γ (β)] Γ (m) i1 ,··· ,iN−1 =0 n=1 in !Γ (m+in ) N
1 X
N
X N−1 N
1 X X
S= un ωn + (ui −uj ) (ωi −ωj ), (12)
N  m+αj N n=1 n=1 N i=1 j=i+1
−m−αj Γ (β −m−αj ) βpj,j
Y
× pj,j zj | {z } | {z }
j=1
m+αj Ω Ŝ ε
 
βpj,j where Ŝ is the approximation of the sum and ε is the approxi-
× 1 F2 m+αj ; 1−β +m+αj , 1+m+αj ; zj
Ω mation error, respectively. The validity of this approximation has
 ! been investigated by Kolmogorov-Smirnov (KS) goodness-of-fit
Γ (m+αj −β) βpj,j
+  −β 1 F2 β; 1−m−αj +β, 1+β; zj . statistical test in [6]. Then, Ŝ can be expressed as the product of
βp Ω N
β Ωj,j zj P
two sums of Gamma RVs, i.e. Ŝ = S1 S2 , where S1 , N1 un
n=1
N
P
Corollary 2. The joint MGF of the ΓΓ distribution z with and S2 , ωn . It is well known that the sum of i.i.d. Gamma
n=1
arbitrary correlation matrix is given by RVs un remains Gamma distributed with parameters (N β, 1/β)
∞ N
W |m
|W X Y −m−αj [19, Eq. (7)]. After variate transformation, we can conjecture that
Mz (ss) = N
pj,j
[Γ (β)] Γ (m) i1 ,...,iN −1 =0 j=1 S1 is Gamma distributed with parameters (N β, 1/N β).
   NY −1
" # Following the similar method presented in [18], let δn be the
2,1 βpj,j 
 1 |pn,n+1 |2in set of N distinct eigenvalues of the N × N covariance matrix
× G1,2 , (9)
Ωsj  m + αj , β n=1 in !Γ (m + in ) given by Ky = E(yy† ), where

where s , [s1 , s2 , . . . , sN ].  Ωi /2,
p
if i = j and k = l
E(yi,k yj,l ) = ρi,j Ωi Ωj /2, if i 6= j and k = l

Proof: Based on the PDF expression (6), the joint MGF of 0, otherwise.
 4

We assume that δn has algebraic multiplicity νn , where νn /2 ∈ 1

Z. Then, we can derive a closed-form PDF expression for S2 as
0.9
mi
N X
X

fS2 (z) = ΞN i, j, {mn }N N
n=1 , {Ωn }n=1 fu (z; j, Ωi ) ,
0.8

i=1 j=1 0.7 N = 2, m1 = m2 = 1,

Ω1 = Ω2 = 1
where fu (z; j, Ωi ) is given in (5), mn = νn /2 should be integers, 0.6

CDF FŜ (z)

Ωn = 4δn /νn , and 0.5
N = 3, m1 = m2 = m3 = 1
N  −j Ω1 = Ω2 = Ω3 = 1

1 X mn 1 1 0.4
ΞN i, mi −k, {mn}N n=1 , {Ω } N
n n=1 = − N = 2, m1 = m2 = 1, Ω1 = Ω2 = 1
k j,n=1 Ωjn Ωi Ωn 0.3

n6=i 0.2


× ΞN i, mi − k + j, {mn }N N
n=1 , {Ωn }n=1 , (13) 0.1 Simulation
Approximation
with k = 1, . . . , mi − 1 and 0
0 5 10 15 20
N  −mj 10log10(z) (dB)

Ωm i Y 1 1
ΞN i, mi , {mn }N N
n=1 , {Ωn }n=1 = QN
i
mt
− .
t=1 Ωt
Ωj Ωi Fig. 1. Analytical and simulated CDF of the sum of ΓΓ RVs (ρ = 0.1 and
j6=i
β = 1).
With the PDF expressions of S1 and S2 , we can derive the PDF
of Ŝ as2 statistic T is defined as the maximal difference between the
mi
q approximate CDFs. The critical value Tmax is given
N X
X simulated and


fŜ (z) = ΞN i, j, {mn }N n=1 , {Ω n } N
n=1 f γ (z; N β, j, jΩ i ) , as Tmax ≈ − 2v 1
ln α2 , where α = 5% is the significance level
i=1 j=1 4
(14) and v = 10 is the number of random samples of RVs [20]. Then,
Tmax = 0.0136. The hypothesis H0 is not always accepted with
where fγ (·) is the PDF of a ΓΓ RV and is given by [11, Eq. (2)] 95% significance and the accuracy of the proposed approximation
N β+j N β+j r ! depends on the combinations of the parameters ρ, mn , Ωn ,
2(N β) 2 z 2 −1 N βz
fγ (z; N β, j, jΩi ) = N β+j KN β−j 2 . β and N . For example, the hypothesis H0 is accepted with
Γ (j) Γ (N β) (Ωi ) 2 Ωi T = 0.0118 < Tmax when considering exponential correlation
(15) with the following typical values, m1 = m2 = 1, Ω1 = Ω2 = 1,
Finally, the corresponding CDF is given by β = 1, ρ = 0.5, and N = 2. However, when we set ρ = 0.1 and
keep other parameters the same, the hypothesis H0 is rejected
N X mi
X
1 with T = 0.0198 > Tmax .
FŜ (z) = ΞN i, j, {mn }N n=1 , {Ωn }n=1
N

i=1 j=1
Γ (j) Γ (N β)
   IV. A PPLICATIONS IN W IRELESS C OMMUNICATIONS
N β  1
× G2,11,3 z . (16) A. Diversity Receivers
Ωi  N β, j, 0
In this section, the performance of various classes of multi-
Note that the approximating PDF and CDF expressions (14) and
channel diversity receivers, operating over arbitrarily correlated
(16) are both in analytical form, since they involve finite nested
generalized-K fading channels, is analyzed in terms of OP and
summations. The weights coefficients can be easily and quickly
BER. For each receive antenna, the transmitted signal is passing
evaluated by using the recursive formula (13).
through the fading channel and perturbed by complex additive
white Gaussian noise (AWGN) with zero mean and variance N0 .
B. Approximation Accuracy The instantaneous signal-to-noise ratio (SNR) of the baseband
In order to evaluate the accuracy of our approximating result, received symbol in the nth diversity branch is given by λn =
2
the simulated and analytical (16) CDF curves of the sum of |hn | Es /N0 , where Es is the energy of the transmitted complex
exponentially correlated ΓΓ RVs are plotted in Fig. 1. Note symbol and hn is the complex gain of the nth generalized-
that the analytical CDF expression can be only obtained by K fading channel. By assuming that each fading channel has
calculating the eigenvalues and algebraic multiplicity for each identical parameters, the average SNR  per  branch is the same
specific covariance matrix. From Fig. 1, it is clear that the and can be expressed as λ̄n = E |hn |2 Es /N0 = λ̄, where
analytical results approximate the exact ones with good accuracy.  
2
Moreover, note that the accuracy of the analytical CDF expression E |h n | = mΩ. Furthermore, since shadowing occurs in large
depends on the values of parameters ρ, N , β, mn , and Ωn . As a geographical areas, we assume that the shadowing parameters on
general comment, higher values of N and smaller values of Ωn each antenna are same, i.e., βn = β, and the AWGN is uncorre-
yield better accuracy. lated among the diversity channels [21]. With λn = |hn |2 λ̄/mΩ,
Furthermore, we use a KS goodness-of-fit statistical test to the joint PDF and CDF of λ , [λ1 , λ2 , . . . , λN ]
can be ob-
N
validate the accuracy of the approximation results. The KS test tained as fλ (λ λ ) = fz mΩ λ̄
λ1 , mΩ
λ̄
λ2 . . . , mΩ
λ̄
λN / λ̄/mΩ


and Fλ (λ λ) = Fz mΩ λλ̄1 , mΩ λλ̄2 , . . . , mΩ λλ̄N , respectively.
2 Since the PDF expression involves the distinct eigenvalues of the covariance
1) Selection Combining: The output SNR of SC diversity
matrix Ky , it is not straightforward to derive similar expressions for the case of
exponentially correlated ΓΓ RVs. Yet, the authors in [6] have given the PDF and receivers is the highest instantaneous SNR among the multiple
CDF expressions for the case of independent ΓΓ RVs. branches, i.e., λsc = max{λ1 , λ2 , . . . , λN }. An outage event
 5

happens when the SNRs of all branches fall below a given For the BER of BPSK in the high-SNR regime, we invoke the
threshold λth . From (8), the OP is obtained as seminal parametrization in terms of diversity order Gd and coding
∞ N−1
" # gain Gc . By keeping only the dominant term of (21), the BER of
W |m
|W X Y 2i
|pn,n+1 | n MRC receivers over arbitrarily correlated generalized-K fading
Pout (λth ) = N
[Γ (β)] Γ (m) i1 ,...,iN −1 =0 n=1 in !Γ (m+in ) channels is given by
 N −G

PbMRC,∞ = (Gc × Ω) d + o Ω−Gd ,
 Y
2,1 βmpj,j λth 
 1 −m−αj (22)
× G1,3  m+αj , β, 0 pj,j . (17)
λ̄ j=1 where
N
The OP for small values of λth is typically interesting in X
communication systems. We now focus on the high-SNR regime. Gd , τj (23)
j=1
Based on [15, Eq. (9.303)] and [17, Eq. (07.22.02.0001.01)], the
∞ N
high-SNR outage can be derived from (17) and for (m+αj −β) ∈ / W |m
|W X Y −m−α
Z as Gc , N
pj,j j Γ (tj −τj ) Γ (τj )
4 [Γ (β)] Γ (m) i1 ,...,iN−1=0 j=1
∞ N
∞ W |m
|W X −m−αj Γ (tj − τj )
Y
 τ N −1 " #! −1
Pout (λth ) = N
pj,j 3βpj,j j Y |pn,n+1 | n
2i τj

[Γ (β)] Γ (m) i1 ,...,iN −1 =0 j=1 τj × . (24)

 N  4 n=1
in !Γ (m + in )
 τ N−1 " # P
λth j Y |pn,n+1 |2in τj
Based on the aforementioned expressions, we can notice that
× βmpj,j +o λth
 j=1  , (18)
λ̄ n=1
i n !Γ (m+i n ) the diversity order does not depend on the correlation between
fading channels, but depends on the shape parameter of small-
where τj , min(m + αj , β) and tj , max(m + αj , β). (18) scale fading and scale parameter of Gamma shadowing. Note
indicates that using more antennas at the receiver affects the high- that, for large values of β, the diversity order Gd is limited by
SNR slope, since the high-SNR slope grows linearly with N . the shape parameter of small-scale fading.
2) Maximal-Ratio Combining: The MRC receiver can achieve
optimal performance in terms of maximizing the SNR at the
output of the combiner [1]. The instantaneous SNR per symbol of B. Free-Space Optical Communications
PN
an N -branch MRC receiver is given by λMRC = n=1 λn . Then, We consider a diversity FSO system with N apertures and
the approximating expression for the OP of MRC receivers is one beam, where the underlying sub-channels between pairs of
given by transmit-receive apertures are correlated. Furthermore, we assume
mi
N X
X
that the optimal combining (OC)3 scheme is deployed. Let Pn be
Pout (λth ) ≈ ΞN i, j, {mn }N N
n=1 , {Ωn }n=1 the fading coefficient of the n-th sub-channel, n = 1, 2, . . . , N ,
i=1 j=1 that follows the ΓΓ distribution with arbitrary correlation. The
× Fγ (λth ; N β, j, jΩi ) , (19) instantaneous electrical SNR at the n-th receive aperture can be
defined as λn = (ηPn )2 /N0 and the average electrical SNR
where Fγ (·) is given in (16). For (N β − j) ∈
/ Z, the OP in the 2
is λ̄n = |ηE(Pn )| /N0 , where η is the optical-to-electrical
high-SNR regime is given by conversion coefficient. Without loss of generality, we normalize
mi
N X
X
the average irradiance as E(Pn ) = 1 and assume that the average
∞
Pout (λth ) = ΞN i, j, {mn }N N
n=1 , {Ωn }n=1 electrical SNRs at different receive apertures are equal, i.e.,
i=1 j=1 λ̄n = λ̄. The variance of the noise in each receiver is N0 /2N
 c
Γ (ej − cj ) λth j and the average BER is defined as [22, Eq. (17)]
× Nβ + o (λth cj ) , (20) v
cj Ωi Z ∞ Z ∞
 uN 
η uX
where cj , min(N β, j) and ej , max(N β, j). It is clear that PbOC = ··· Q √ t P 2  fp (pp) dpp, (25)
0 0 2N N0 n=1 n
the diversity order depends on the values of the parameters N , β
and mi . For the case of MRC, mi in (20) should be integer.
where Q(·) denotes the Gaussian Q-function [1, Eq. (4.1)], p ,
We now consider the average BER of the MRC receiver for
[P1 , P2 , . . . , PN ], and fp (pp) is the joint PDF of the vector p
noncoherent binary frequency-shift keying (NBFSK). Based on
obtained with the aid of (6).
[1, Eq. (9.254)], the average BER expression is given as PbMRC =
In strong atmospheric channels, the fading parameters β and
0.5MλMRC (g), where g = 0.5 for NBFSK. Futhermore, the MGF
m can be respectively expressed as [8, Eqs. (5)-(6)]
of the output SNR can be obtained from (9) as MλMRC (ss) =
   −1
Mλ (s, s, . . . , s).
Following a similar method of [5], the approximate BER of   0.49σ22  
β = exp   , (26)
7/6  − 1
MRC receivers for binary phase-shift keying (BPSK) modulation 12/5
1 + 0.18d2 + 0.56σ2
is given by
     −5/6  −1
MRC 1 1 4 0.51σ 2
1 + 0.69σ
12/5
Pb ≈ MλMRC (1) + MλMRC . (21)   2 2  
12 4 3 m = exp   5/6  − 1 , (27)
12/5
The approximate BER (21) provides a very useful metric of 1 + 0.9d2 + 0.62d2 σ2
communication systems. Furthermore, (21) can reduce to [5, Eq.
(15)] for exponentially correlated multivariate ΓΓ RVs. 3 In FSO communications, the term OC is used instead of MRC.
 6

0 0
10 10
Simulation Results Simulation Results
High SNR Approximation High SNR Approximation
Analytical Results Analytical Results
−2 −2
10 10

Average Bit Error Rate

Outage Probability

−4 −4
10 10

β=0.5, m=2
−6 −6
10 10

β = 1.0931, 7.9115, 75.1155 β=1.0931, m=3

−8 −8
10 10

β=1.0931, m=2
−10 −10
10 10
0 5 10 15 20 25 0 5 10 15 20 25
Inverse Normalised Outage Threshold (dB) Average SNR Per Brance (dB)

Fig. 2. Analytical, simulated and high-SNR approximation OP of SC diversity Fig. 3. Average BER of triple SC diversity receivers over linearly correlated
receivers with three branches over exponentially correlated generalized-K fading generalized-K fading channels against the average SNR per branch λ̄.
channels against the inverse normalized outage threshold λ̄/λth (ρ = 0.25 and
m = 2).
0
p 10
where σ22 = 0.492Cn2 k 7/6 L11/6 and d = kD2 /4L with L being
the distance between transmitter and receiver, k is the optical 10
−1

wavenumber, D is the aperture diameter and Cn2 = 1.7 × 10−14 N = 2, ρ = 0.5, 0.1

is the refractive-index structure parameter. −2

10
The integral in (25) is very difficult to be evaluated in closed- Outage Probability
form, so we use a simple and accurate exponential approximation −3
10
for Q(·) [5, Eq. (14)] to derive an approximate result. Then, (25)
can be written as −4
! 10 N = 3, ρ = 0.5, 0.1
Z ∞ Z ∞ N
OC 1 η2 X 2
Pb ≈ ··· exp − P
0 0 12 4N N0 n=1 n 10
−5

N
!! Simulation Results
1 η2 X 2 Analytical Results
+ exp − P fp (pp) dpp. (28) 10
−6

4 3N N0 n=1 n 0 5 10 15 20
Inverse Normalised Outage Threshold (dB)

Substituting (6) into (28) and using [16, Eqs. (11), (14)] and [16, Fig. 4. OP of MRC diversity receivers over exponentially correlated generalized-
Eq. (21)], we derive the average BER as K fading channels against the inverse normalized outage threshold λ̄/λth (β = 1,
r ! r ! mn = m = 1).
OC 1 λ̄ 1 λ̄
Pb ≈ Λ m, β, + Λ m, β, , (29)
12 4N 4 3N frequent heavy shadowing, Karasawa shadowing and infrequent
where light shadowing, respectively [5]. It is clear that for exponentially
∞ N−1
" # correlated generalized-K fading channels, the analytical results
W |m
2N (m+β−2) |W X Y |2pn,n+1|2in agree closely with the exact OP results obtained via simulations.
Λ (m, β, x) , N
[πΓ (β)] Γ (m) i1 ,...,iN−1 =0 n=1 in !Γ (m+in) Furthermore, the high-SNR approximations are sufficiently tight
N
"   2
 # and become exact, even at moderate SNR values. As expected, the
Y 4xΩ  1−(m+αj ) , 2−(m+αj ) , 1−β , 2−β
. OP that corresponds to light shadowing (β = 75.1155) is larger
−m−αj 1,4
× pj,j G4,1  2 2 2 2
j=1
βp j,j  0 than the OP that corresponds to heavy shadowing (β = 1.0931).
Moreover, this trend is more pronounced for smaller values of β.
Note that for the special case of exponentially correlated
Figure 3 presents the simulated, analytical (21) and high-
atmospheric turbulence, (29) coincides with a previously known
SNR asymptotic results (22) for the BER of triple SC diversity
expression [5, Eq. (19)].
receivers over linearly correlated generalized-K fading channels.
The linearly correlated model is a more general case than the ex-
V. N UMERICAL R ESULTS ponential one with a correlation Toeplitz structure matrix defined
A. RF Communication Systems as Σ i,j = 1 for i = j and Σ i,j = Σ j,i = ρi,j for i 6= j. We note
In our simulation, without loss of generality, it is assumed that as the scale and shadowing coefficients increase, the average
here that I1 = I2 = . . . = IN −1 and every branch has the BER decreases. For all scenarios, the analytical results agree with
same average SNR λ̄. For the OP of three branches with SC the Monte-Carlo ones. Likewise, the diversity order and coding
receivers, the Monte-Carlo simulation results, analytical results gains are accurately predicted. It is clear that the diversity order
(17) and high-SNR approximation (18) are compared in Fig. depends on the shadowing parameters.
2. We again assume exponential correlation. The shadowing The simulated and approximate OP of MRC receivers over
parameter β corresponds to 1.0931, 7.9115 and 75.1155 for exponentially correlated generalized-K fading channels are inves-
 7

10
−1 and correlated coefficients ρ = 0.2, 0.5, 0.7, respectively. Upon
Simulation results using (26) and (27), the respective values for β and m can be
Analytical results
−2
determined. From Fig. 6, it is clear that the approximate results
10
of cases under consideration closely agree with the ones obtained
−3
via Monte-Carlo simulations. Note that the average BER increases
10
by increasing ρ and L or decreasing λ̄.
Average BER

−4
10 R EFERENCES
[1] M. K. Simon and M.-S. Alouini, Digital Communication over Fading
10
−5
Channels. Jone Wiley & Sons, 2005.
β = 1.0931, 7.9115, 75.1155 [2] P. S. Bithas, N. C. Sagias, and P. T. Mathiopoulos, “The bivariate
generalized-K (KG ) distribution and its application to diversity receivers,”
−6
10 IEEE Trans. Commun., vol. 57, no. 9, pp. 2655–2662, Sep. 2009.
[3] J. Zhang, M. Matthaiou, Z. Tan, and H. Wang, “Performance analysis of dig-
ital communication systems over composite η-µ/gamma fading channels,”
−7
10 IEEE Trans. Veh. Technol., vol. 61, no. 7, pp. 3114–3124, Sep. 2012.
0 1 2 3 4 5 6 7 8 9 10
Average SNR Per Branch (dB) [4] J. F. Paris, “Statistical characterization of κ-µ shadowed fading,” IEEE
Trans. Veh. Technol., vol. 63, no. 2, pp. 518–526, Feb. 2014.
Fig. 5. Bit error rate of quadruple branch MRC diversity receivers over [5] K. P. Peppas, G. C. Alexandropoulos, C. K. Datsikas, and F. I. Lazarakis,
exponentially correlated generalized-K fading channels against the average SNR “Multivariate gamma-gamma distribution with exponential correlation and
per branch λ̄ (m = 2, ρ = 0.3). its applications in radio frequency and optical wireless communications,”
IET Microw. Antennas Propag., vol. 5, no. 3, pp. 364–371, Mar. 2011.
[6] N. D. Chatzidiamantis and G. K. Karagiannidis, “On the distribution of the
0 sum of gamma-gamma variates and applications in RF and optical wireless
10
communications,” IEEE Trans. Commun., vol. 59, no. 5, pp. 1298–1308,
May 2011.
−1
10 [7] M. Uysal, S. M. Navidpour, and J. Li, “Error rate performance of coded
ρ = 0.2, 0.5, 0.7
free-space optical links over strong turbulence channels,” IEEE Commun.
r = 5000 m Lett., vol. 8, no. 10, pp. 635–637, Oct. 2004.
−2
10 [8] J. Zhang, L. Dai, Y. Zhang, and Z. Wang, “Unified performance analysis of
mixed radio frequency/free-space optical dual-hop transmission systems,” J.
Average BER

Lightw. Technol., vol. 33, no. 11, pp. 2286–2293, Jun. 2015.
−3
10 [9] J. Zhang, L. Dai, Y. Han, Y. Zhang, and Z. Wang, “On the ergodic capacity
of MIMO free-space optical systems over turbulence channels,” IEEE J. Sel.
Areas Commun., to appear, 2015.
−4
10 [10] G. K. Karagiannidis, D. A. Zogas, and S. A. Kotsopoulos, “An efficient
approach to multivariate Nakagami-m distribution using Green’s matrix
−5 ρ = 0.2, 0.5, 0.7 approximation,” IEEE Trans. Wireless Commun., vol. 2, no. 5, pp. 883–889,
10 r = 3000 m Sep. 2003.
Simulation Results [11] P. S. Bithas, N. C. Sagias, P. T. Mathiopoulos, G. K. Karagiannidis, and A. A.
−6
Approximation Results Rontogiannis, “On the performance analysis of digital communications over
10
0 5 10 15 20 25 30 generalized-K fading channels,” IEEE Commun. Lett., vol. 10, no. 5, pp.
Average Electrical SNR (dB) 353–355, May 2006.
[12] G. C. Alexandropoulos, N. C. Sagias, F. I. Lazarakis, and K. Berberidis,
Fig. 6. Average BER of space diversity FSO system with three apertures and “New results for the multivariate Nakagami-m fading model with arbitrary
one beam over strong turbulence fading channels against the average electrical correlation matrix and applications,” IEEE Trans. Wireless Commun., vol. 8,
SNR λ̄. no. 1, pp. 245–255, Jan. 2009.
[13] R. Nabben, “On Green’s matrices of trees,” SIAM J. Matrix Anal. & Appl.,
vol. 22, no. 4, pp. 1014–1026, 2001.
tigated in Fig. 4. It is obvious that the proposed approximation is [14] K. Zhang, Z. Song, and Y. L. Guan, “Simulation of Nakagami fading
a lower bound of the simulated results. However, the difference is channels with arbitrary cross-correlation and fading parameters,” IEEE
not greater than 2 dB in all cases. Furthermore, the approximation Trans. Wireless Commun., vol. 3, no. 5, pp. 1463–1468, Sep. 2004.
[15] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and Products,
results are more accurate with the increased correlation coefficient 7th ed. San Diego, CA: Academic Press, 2007.
ρ and the decreased value of N . [16] V. S. Adamchik and O. I. Marichev, “The algorithm for calculating integrals
We present the simulated and approximate BER perfor- of hypergeometric type functions and its realization in REDUCE system,”
in Proc. Intern. Conf. Symbolic Algebraic Computation, 1990, pp. 212–224.
mance of quadruple MRC receivers over exponentially correlated [17] Wolfram, “The wolfram functions site,” Available:
generalized-K fading channels in Fig. 5. It is clear that the BER http://functions.wolfram.com.
decreases as β increases. Approximate BER curves are close [18] G. K. Karagiannidis, N. C. Sagias, and T. A. Tsiftsis, “Closed-form statistics
for the sum of squared Nakagami-m variates and its applications,” IEEE
to the simulated results and can be used as an upper bound. Trans. Commun., vol. 54, no. 8, pp. 1353–1359, Aug. 2006.
Moreover, the gap between two curves becomes smaller as the [19] E. Al-Hussaini and A. Al-Bassiouni, “Performance of MRC diversity
SNR increasing. systems for the detection of signals with Nakagami fading,” IEEE Trans.
Commun., vol. 33, no. 12, pp. 1315–1319, Dec. 1985.
[20] A. Papoulis and S. U. Pillai, Probability, random variables, and stochastic
B. FSO Systems processes. Tata McGraw-Hill Education, 2002.
[21] P. S. Bithas, P. T. Mathiopoulos, and S. A. Kotsopoulos, “Diversity reception
In Fig. 6, the simulated and approximate (29) error performance over generalized-K (KG ) fading channels,” IEEE Trans. Wireless Commun.,
of space diversity FSO links with N = 3 receive apertures vol. 6, no. 12, pp. 4238–4243, Dec. 2007.
[22] T. A. Tsiftsis, H. G. Sandalidis, G. K. Karagiannidis, and M. Uysal, “Optical
employing OC over exponentially correlated atmospheric turbu- wireless links with spatial diversity over strong atmospheric turbulence
lence channels, is depicted. We consider the case of the optical channels,” IEEE Trans. Wireless Commun., vol. 8, no. 2, pp. 951–957, Feb.
wavenumber k = 0.405 × 107 , link distances L = 3000, 5000m 2009.