Error Rates of M-PAM and M-QAM in Generalized

Fading and Generalized Gaussian Noise Environments

Item Type Article

Authors Soury, Hamza; Yilmaz, Ferkan; Alouini, Mohamed-Slim

Eprint version Post-print

DOI 10.1109/LCOMM.2013.081913.131409

Publisher Institute of Electrical and Electronics Engineers (IEEE)

Journal IEEE Communications Letters

Rights (c) 2013 IEEE. Personal use of this material is permitted.

Permission from IEEE must be obtained for all other users,
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
components of this work in other works.

Download date 21/08/2020 18:15:14

Link to Item http://hdl.handle.net/10754/294163

 1

Error Rates of M-PAM and M-QAM in Generalized

Fading and Generalized Gaussian Noise
Environments
Hamza Soury, Student Member, IEEE, Ferkan Yilmaz, Member, IEEE,
and Mohamed-Slim Alouini, Fellow, IEEE

Abstract
This letter investigates the average symbol error probability (ASEP) of pulse amplitude modulation and
quadrature amplitude modulation coherent signaling over flat fading channels subject to additive white generalized
Gaussian noise. The new ASEP results are derived in a generic closed-form in terms of the Fox H function and the
bivariate Fox H function for the extended generalized-K fading case. The utility of this new general closed-form is
that it includes some special fading distributions, like the Generalized-K, Nakagami-m, and Rayleigh fading and
special noise distributions such as Gaussian and Laplacian. Some of these special cases are also treated and are
shown to yield simplified results.

Index Terms
Symbol error probability, pulse amplitude modulation, quadrature amplitude modulation, additive generalized
Gaussian noise, Laplacian noise, generalized-K fading, and Nakagami-m fading.

I. I NTRODUCTION
In pulse amplitude modulation (PAM), the information is conveyed by the amplitude of the pulse, and
from two independent pulse amplitude modulation (PAM) signals (in-phase and quadrature PAM), one can
easily generate any general order, and rectangular quadrature amplitude modulation (QAM) constellation.
The average symbol error probability of PAM and QAM was extensively studied in the presence of a
Gaussian noise and flat fading [1]–[5]. Although the Gaussian noise is widely used, the actual additive
noise may deviate from it, so a more general noise model is sometimes needed. In fact, it has been widely
reported that the generalized Gaussian distribution (GGD) can model different type of noise (see Table
I), for example the Gaussian, Laplacian, and uniform distributions are just a special cases of the GGD
[6], [7].
In the derivation of the ASEP of QAM in the standard Gaussian case, the challenge consists of solving
an integral involving the product of two traditional Gaussian Q functions [1]–[4]. Here, in the presence
of generalized Gaussian noise, the problem is more complicated because we have to derive an integral
of the product of two generalized Gaussian Q functions [8]. This integral appears when we consider the
ASEP of QAM over Extended Generalized-K (EGK) fading [9] and subject to additive white generalized
Gaussian noise (AWGGN) [8]. In the PAM case, the problem is less complex than the QAM because
we need to evaluate an integral involving one generalized Gaussian Q function which is discussed in [8].
The EGK distribution has five parameters and includes most of the well-known fading distributions in the
literature as either special or limit cases as shown in [9, Table 1] and Table II. Our goal in this letter to
extend the analysis in [8] and [10], which were limited to binary coherent signalling and square M -QAM,
respectively, and obtain a generic expression for the ASEP for M -PAM and M -QAM constellations and
then to simplify these expressions as much as possible for some special cases of interest corresponding
to different values of the fading and noise parameters.
H. Soury, F. Yilmaz, and M.-S. Alouini are with King Abdullah University of Science and Technology (KAUST), Thuwal, Makkah
Province, Kingdom of Saudi Arabia (e-mail: soury.hamza, ferkan.yilmaz, slim.alouini}@kaust.edu.sa).
 2

II. ASEP OF M -PAM AND M -QAM

A typical single input single output communication system is considered in this letter. The transmitted
signal S is mapped according to an M -PAM or M -QAM constellation. In fact, the QAM constellation is
formed by two independent quadrature M -ary PAM signals, where the in-phase and quadrature signals are
MI -ary PAM and MQ -ary PAM, respectively, and M = MI MQ . Next, S is multiplied by a channel fading
envelope H that has a generalized flat fading characteristics and the resulting signal is then corrupted at the
receiver end by an AWGGN N with zero mean, variance σ 2 , and shaping parameter α (i.e. R = H S +N ).
In [1] (respectively [5]), the probability of error of the system above is well studied for the Gaussian
noise case, and the resulting symbol error probability (SEP) was shown to be given by [1, Eq. (10)] for
the QAM constellation (respectively [5, Eq. (8.3.5)] for the PAM constellation).
Since the generalized generalized Gaussian distribution and the Gaussian distribution have the same
symmetry properties, the SEP of the M -PAM signalling over an AWGGN channel can be easily written
as ( ) ( )
1 d
PP AM (E) = 2 1 − Qα , (1)
M σ
where d is the decision distance. Using the same symmetry properties, the SEP of the M -QAM can be
written as ( ) ( )
1 1
PQAM (E) = 2 1 − Qα (AI ) + 2 1 − Qα (AQ )
( )M(I ) MQ
(2)
1 1
−4 1 − 1− Qα (AI ) Qα (AQ ) ,
MI MQ
d
where AI = dσI , AQ = σQ , dI and dQ are the in-phase and quadrature decision distances, respectively. In
(1) and (2), Qα (.) is the generalized Gaussian Q function, defined defined, for x ≥ 0, in [8, Eq. (A1)] as
∫ ∞
αΛ0
Qα (x) = exp (−Λα0 tα ) dt, (3)
2Γ(1/α) x
√
where Γ(.) is the Gamma function and Λ0 = Γ(3/α)/Γ(1/α), the absolute (|t|α ) is omitted from [8,
Eq. (A1)] because we deal with positive parameters (x ≥ 0). The results in (1) and (2) can be found
directly.
Before proceeding to the proof we have to compute the Bit error probability (BEP) of the BPSK
constellation which is given in [8] as ( )
d
PBP SK = Qα , (4)
σ
where d is the decision distance and σ 2 is the noise variance.
It appears from [8] that the decision region is based on the euclidian distance between the received signal
and the constellation symbols. We focus on computing the SEP of an M -PAM constellation. Consider
an M -PAM constellation with decision distance d. The probability of error for the PAM is divided into
two terms, one for the outermost points (2 points) and the other for the remaining M − 2 points. The
first term is similar for both points and is equal to the error between one outermost point ( d and
) its closest
1
neighbor which is equal to the error of BPSK with decision distance d so Pouter = M Qα σ . In the other
hand, each of the remaining M − 2 points has two neighbors from both( sided, ) so the probability of error
2 d
is twice the probability of error between two symbols Pinter = M Qα σ . Hence to total probability of
error of the M -PAM is given by
( ) ( ) ( ) ( )
1 d 1 d 1 d
PP AM (E) = 2 Qα + 2(M − 2) Qα =2 1− Qα . (5)
M σ M σ M σ
| {z } | {z }
Two outermost points The remaining M − 2 points
 3

Back to QAM model, the SEPs of the in-phase MI -PAM and the quadrature MQ -PAM are given by
( ) ( )
1 dI
PI = 2 1 − Qα , (6)
MI σ
( ) ( )
1 dQ
PQ = 2 1 − Qα . (7)
MQ σ
A correct reception appears when we have correct reception in the independent in-phase signals and the
quadrature phase signals. Then the probability of correct symbol reception for the rectangular M -QAM
system is
Pc = (1 − PI )(1 − PQ ),
and the SEP is given by
PS = 1 − P c = 1 − (1 − PI )(1 − PQ ) = PI + PQ − PI PQ . (8)
Using (6) and (7) in (8), we get
( ) ( ) ( )( )
1 1 1 1
PS = 2 1 − Qα (AI ) + 2 1 − Qα (AQ ) − 4 1 − 1− Qα (AI ) Qα (AQ ) .
MI MQ MI MQ
In the presence of a slow fading channel, the ASEP is given by averaging the conditional SEPs (1) and
(2) over the probability density function (PDF) of the fading power γ = H2 , pγ (γ), yielding
( ) ( )
1 d
PrP AM (e) = 2 1 − I , (9)
M σ
( ) ( ) ( )( )
1 1 1 1
PrQAM (e) = 2 1 − I(AI ) + 2 1 − I(AQ ) − 4 1 − 1− J , (10)
MI MQ MI MQ
where we define
∫ ∞
√
I(x) = Qα ( γx) pγ (γ) dγ (11)
∫0 ∞
√ √
J = Qα ( γAI ) Qα ( γAQ ) pγ (γ) dγ, (12)
0
d
with x taking the values AI , AQ , or σ
in (11).

III. AVERAGE SEP OVER EGK FADING

We consider an EGK fading environment [9] for which γ follows an EGK PDF [9, Eq. (2)], with
average fading power γ = E[γ], fading severity figure m ≥ 0, fading shaping factor ξ ≥ 0, shadowing
severity figure ms ≥ 21 , and shadowing shaping factor ξs ≥ 0
( )mξ ( ( )mξ )
ξ βs β ξ β s β ξ
pγ (γ) = γ mξ−1 Γ ms − m , 0, γξ, , (13)
Γ(ms )Γ(m) γ ξs γ ξs
where β = Γ(m + 1/ξ)/Γ(m), βs = Γ(ms + 1/ξs )/Γ(ms ), and Γ(·, ·, ·, ·) is the extended incomplete
Gamma function defined in [11, Eq. (6.2)] as
∫ ∞
Γ(a, x, b, c) = ra−1 exp (−r − br−c )dr, (14)
x

where a, b, c ∈ C and x ∈ R +

As I and J are difficult to manipulate using the classical expressions of the EGK distribution and the
GGD, we use for alternative expressions of Qα (.) as given in (3) and pγ (γ) as defined in [9, Eq. (2)], in
order to derive closed-from expressions for I and J and as a result for the ASEP. More specifically, as
 4

was done in [8], we first express (3) and [9, Eq. (2)] in terms of the Fox H function (FHF) H.. .. [.] [12, Eq.
(1.1.1)], whose MATHEMATICA implementation is given in [13, Appendix]. As such Qα (.) and pγ (γ)
are expressed in terms of the FHF in [8, Eq. (A.5)] and [8, Eq. (8)] respectively
[  ]
1 α
 (1, 1)
Qα (x) = H 2,0
Λ |x|  1
α
, (15)
2Γ(1/α) 1,2 0 ( α , 1), (0, 1)
[  ]

1 2,0 βs β 
pγ (γ) = H γ , (16)
Γ(ms )Γ(m)γ 0,2 γ (ms , ξ1s ), (m, 1ξ )
The FHF is defined in [12, Eq. (1.1.1)] as
∏m ∏n
[  ] ∫ Γ(bj + βj s) Γ(1 − ai − αi s)
 (ai , αi )1,p 1
z 
j=1 i=1
Hm,n = z −s ds, (17)
p,q
(bj , βj )1,q 2πi C ∏
p ∏q
Γ(ai + αi s) Γ(1 − bj − βj s)
j=n+1 j=m+1

In (11), I(x) includes an integral involving the product of two FHFs which is similar to what the
authors studied in [8]. Following similar steps to the ones adopted in [8], the identity [12, Eq. (2.8.4)]
was very helpful to find a closed form of I(.)
∫ ∞ [  ] [  ]

σ (a , α )  (ci , γi )1,P
t η−1 m,n
Hp,q zt 
i i 1,p M,N
HP,Q wt   dt
(bj , βj )1,q (dj , δj )1,Q
0
[  ]
−η m+N,n+M

−σ  (ai , αi )1,n , (1 − dj − ηδj , σδj )1,Q , (ai , αi )n+1,p
= w Hp+Q,q+P zw  . (18)
(bj , βj )1,m , (1 − cj − ηγj , σγj )1,P , (bj , βj )m+1,q
and a closed form for I(.) in terms of the FHF is easily found as
[  ]
2,2

βs β  (1− α , α ),(1, α )
1 2 2
H2,3 x2 Λ2 γ  (m , 1 ),(m, 1 ),(0, 2 )
0 s ξ ξ α
I(x) =
s
. (19)
αΓ(1/α)Γ(ms )Γ(m)
Now substituting (15) and (16) in (12), J is defined as an integral involving the product of three FHFs.
Using the identity [14, Eq. (2.3)]
∫ ∞ [  ] [  ] [  ]
 (ai , αi )1,p 
h  (ci , ri )1,p2

k  (ei , Ei )1,p3
λ−1 m,0
x Hp,q ax   m2 ,n2
Hp2 ,q2 βx  m3 ,n3
Hp3 ,q3 δx  dx
0 (bj , βj )1,q (dj , δj )1,q2 (fj , Fj )1,q3
  
(1 − bj − λβj ; hβj , kβj )1,q 

 (1 − ai − λαi ; hαi , kαi )1,p  
  β δ
 (c , r )  
= a−λ H0,m;m 2 ,n2 ;m3 ,n3
 i i 1,p2
 h, k, (20)
p,q;p2 ,q2 ;p3 ,q3
 (dj , δj )1,q2 a a 
  
(ei , Ei )1,p3 
(fj , Fj )1,q3 
J can be expressed in terms of the FHF of two variables [14], also known as the Bivariate Fox H-function
..;..;.. [., .] and whose MATLAB implementation is outlined in [15]. Hence a closed-form for J
(BFHF) H..;..;..
is found as
1
J = 2
2αΓ(1/α) Γ(ms )Γ(m)
  
(1 − α1 ; α2 , 1), (1; α2 , 1)
 (0; α2 , 1)  
  ( )α
  βs β AQ 
H0,2;2,0;2,0
2,1;0,2;1,2 
  , . (21)
(m , 1
), (m, 1
) A 2
Λ 2
γ A 
 s ξs ξ  I 0 I 
 (1, 1)  

1
( α , 1), (0, 1) 
 5

In conclusion, a closed-form of the ASEP of QAM (respectively PAM) is obtained using (19) and (21)
in (10) (respectively (19) in (9)). This closed-form is a general expression of the ASEP of rectangular
(MI ̸= MQ ), square (MI = MQ ) QAM, and PAM in arbitrary EGK fading subject to AWGGN, and holds
as such for a considerable range of noise and fading parameters.

IV. S PECIAL C ASES OF N OISE AND FADING

A. EGK Fading with Additive Laplacian Noise
The first special case of the generalized Gaussian noise appears when α equals to 1 (i.e. the noise is
Laplacian). Using the properties and the special cases of the FHF [12], and BFHF functions [14], the
general ASEP expression in (10) and (9) can be simplified by re-writing I(.) and J as
[  ]

1 β s β  (1, 2)
I(x) = H2,1  (22)
Γ(ms )Γ(m) 1,2 2x2 γ  (ms , ξ1s ), (m, 1ξ )
  
(1; 2, 1) 

  
  
1 
0,1;2,0;1,0 
 βs β A Q 
J = H  , . (23)
2Γ(ms )Γ(m) 1,0;0,2;0,1 
1 1  
 (ms , ξs
), (m, )
ξ  2A2
I γ A I 
  

(0, 1) 
In fact, J can be expressed in terms of I(.) for the Laplacian noise (LN). Indeed using the definition of
BHFH [14, Eq. (1.1)], we re-write J as
∫ ∫ ( )−s ( )−t
1 1 βs β AQ
J = Γ(−2s − t)Γ(ms + s/ξs )Γ(m + s/ξ)Γ(t) dt ds.
2Γ(ms )Γ(m) (2πi)2 L1 L2 2A2I γ AI
(24)
Using the inverse Mellin transform of the extended incomplete Gamma function Γ(., .; .; .) [11, Eq. (6.29)],
and its special case presented in [11, Eq. (6.42)], we get
∫ ( )−t
1 AQ AQ
Γ(−2s − t)Γ(t) dt = Γ(−2s, 0; ; −1)
2πi L2 AI AI
( )2s
AQ
= Γ(−2s) 1 + . (25)
AI
Replacing (25) in (24)
∫ ( )−s
1 1 βs β
J = Γ(ms + s/ξs )Γ(m + s/ξ)Γ(−2s) ds
2Γ(ms )Γ(m) 2πi L1 2 (AI + AQ )2 γ
[  ]

1 2,1 βs β  (1, 2)
= H 
2Γ(ms )Γ(m) 1,2 2(AI + AQ )2 γ  (ms , ξ1s ), (m, 1ξ )
1
= I(AI + AQ ). (26)
2
Hence, in the presence of LN, the ASEP is defined in terms of I(.) only. The expression (22) contains
also the FHF but with a lower rank than the general expressions given in (19), which means that it can be
computed with a reduced complexity of evaluation. Since the LN is an interesting case of study, it is also
of interest to study other special fading cases in conjunction with the LN. For instance, considering the
generalized Nakagami-m (GNM) fading by setting ξs = 1 and ms → ∞ in (22), I(.) further simplifies to
[  ]
 (1, 2)
1 β 
I(x) = H1,1  . (27)
Γ(m) 1,1 2x2 γ  (m, 1ξ )
 6

Re-writing the expression of the FHF, using some changes of variable, and the identity [11, Eq. (6.29)],
a new expression of I(.) can be obtained as
( )mξ ( ( )mξ )
ξ β β
I(x) = Γ 2mξ, 0; ; −2ξ . (28)
Γ(m) 2x2 γ 2x2 γ
B. Rayleigh Fading with Additive Laplacian Noise
Keeping an additive LN, the Rayleigh fading case is obtained by setting m = 1 and ξ = 1 in (28).
From (28) and with a simple manipulation of Γ(., .; .; .), it can be shown that I is equivalent to
( )
1 1
I(x) = Γ 2, 0; 2 ; −2
2x2 γ 2x γ
∫ ∞ 2
1 −t− t2
= te 2x γ dt
2x2 γ 0
x2 γ ∫ ∞ (√ ) √
e 2 −u 2 t 2x2 γ
= √ √ 2x2 γu − x γ e
2
du where u = √ +
2x2 γ x2 γ
2
2x2 γ 2
∫ ∞ 2 √ ∫
2x2 γ ∞ −u2
x γ
x2 γ
−u2 e 2
= e 2 √ ue du − √ e du
x2 γ 2 x2 γ

( √ )
2 2
1 x√ γx2
= − 2πγe 2 Q x γ (29)
2 2
which is a very simple expression involving only the standard Gaussian Q function (i.e. Q(x) = Q2 (x)).
To the best of the author’s knowledge, the expressions of the ASEP with additive LN and Rayleigh fading
in (29) is a new expression that has not been reported previously.

C. Generalized K (GK) Fading with Additive Gaussian Noise

The last two special cases will focus on the classical Gaussian noise with different fading distributions.
The first example deals with the performance of M -PAM and M -QAM over a GK fading subjected to
an additive white gaussian noise (AWGN) channel. This case is obtained by setting α = 2, ξ = 1, and
ξs = 1. With these settings, the main integrals in (19) and (21) reduce to
[  ]
1 2,2 2ms m 
 ( 1 , 1)
I(x) = √ G 2
(30)
2 πΓ(ms )Γ(m) 2,3 x2 γ  (ms , m, 0)
1
J =
4πΓ(ms )Γ(m)
  
( 12 , 1) 

 (0)  
  2m m ( A )2 
  s Q 
G0,2;2,0;2,0
2,1;0,2;1,2  (m , m)  , , (31)
 s  AI γ
2
AI 
 (1)  
1
( 2 , 0) 

where I is expressed in terms of the standard Meijer G-function (MGF) [12, Eq. (2.9.1)] which is a built
in function in MATHEMATICA, and J is expressed in terms of the bivariate Meijer G-function (BMGF)
whose MATHEMATICA implementation is given in [16, Table II].
It is necessary to mention that in these special cases, we are not just setting the parameters in the
general expression (19) and (21). Rather, in each case, we offered simplified closed-form expressions of
the main parts of the ASEP with reduced computational complexity. In fact, from the definition of the
FHF [12, Eq.(1.1.1)], the complexity of evaluating the Mellin integral is based on the rank of the FHF,
which is reduced in these special cases.
 7

D. Rayleigh Fading with Additive Gaussian Noise

The AWGN was studied in the second special case and an expression of the ASEP of M -PAM and
M -QAM over GK fading with AWGN was derived and can be obtained by (30) and (31). The Rayleigh
fading is a special case of the GK distribution by simply setting m = 1 and ms → ∞. In this case, the
main integrals, I and J in (30) and (31), respectively reduce to
[  ]
1 2  ( 21 , 1)
I(x) = √ G2,2 2  1,2
(32)
2 π x γ (1, 0)
 1  
( 2 , 1) 

 (0)  
1 0,2;1,0;2,0   2 ( A )2 
  Q 
J = G2,1;0,1;1,2   2 , . (33)
4π  (1)  AI γ AI 
 (1)  

( 1 , 0) 
2

The expressions in (32) and (33) involve reduced rank MGF and BMGF functions in comparison with
(30) and (31). However, it is interesting to note that these formulas are numerically equivalent to
√
x2 γ
I(x) = 2 − 2 2+x
1 1
[√ γ 2 ( √ )
2

AI γ −1 AI 2+A2I γ
J = 14 − 2π
1
tan (34)
√ ( √ AQ )]AI γ
2+A2I γ 2

2
A γ 2
2+AQ γ
+ 2+AQ2 γ tan−1 AQI
A
A2 γ
,
Q Q

where tan−1 (.) is the inverse tangent function, in agreement with the known expressions of Beaulieu [1,
Eq. (5)] of the ASEP of M -QAM over Rayleigh fading and subject to AWGN.
V. N UMERICAL R ESULTS AND C ONCLUDING R EMARKS
Let us consider an 8 × 4-QAM system, as presented in [1], and investigate the system performance as
a function of the signal-to-noise ratio (SNR) for different type of noise (i.e. different values of α), and
fading (i.e. different set of values of m, ms , ξ, and ξs ), proceeding of PAM performance is less complex
than the QAM so we focus in this part on the QAM performance. Let us define the in-phase-to-quadrature
decision distance ratio τ = dQ /dI = AQ /AI . In this case the average total energy per symbol ET can
be defined as ET = 21 2
γd2I + 52 γd2Q = 0.5(21 + 5τ 2 )γd2I . According to this definition, the SNR has the
following expression EσT2 = 0.5(21 + 5τ 2 )γA2I . For instance, τ is fixed (τ = (21/5)1/5 ), so the average
energies of the in-phase and quadrature signals are equal.
The first studied case is a comparison between the Gaussian and Laplacian noises in conjunction with
Nakagami-m fading. Note that the Nakgami-m fading is obtained by setting the fading shaping factor
ξ = 1 in (28) for the LN. Fig. 1 shows the ASEP as a function of the average SNR per symbol EσT2 for
both cases Gaussian and Laplacian and for four values of the fading parameter, namely m = 12 , 2, 4, ∞.
A comparison between the analytical results, presented in this letter, and a computer-based Monte-Carlo
simulations results are presented. The simulation results match perfectly the analytical results derived
in this letter. A general look at the figure shows us that the performance of the system is improved by
increasing the fading parameter m, as expected. In addition, for lower SNR (i.e. SNR< 15dB), we note
that the ASEP of the LN is better than the ASEP of the Gaussian noise. The situation reverse for high
SNR and low amount of fading (m ≥ 2) since the Gaussian noise yields better results than LN. However
for high amount of fading, such as for example m = 1/2, the performance in LN is better than in the
Gaussian noise case.
In the second numerical example, we compare the Rayleigh fading and the Nakagami-4 fading. For
instance, in Fig. 2, we draw the ASEP in function of the SNR per QAM symbol for different values of α.
Note again that the simulation results match perfectly the analytical results obtained from (19) and (21).
 8

In Rayleigh fading case, it is worth mentioning that the system has better performance by decreasing
the noise parameter α, which confirm the result found in the previous example when the LN had better
performance than the Gaussian noise for high amount of fading. However, the situation becomes different
for lower amount of fading (i.e. when m = 4), and we get two regions. For low SNR the ASEP decreases
with the noise parameter, and for high SNR it gets better by increasing α.
A new parameter appears on the energy expression, which is τ , and in this numerical example we want
to see the effect of τ on the system performance. Therefore Fig. 3 draws the ASEP of the system described
above as a function of the SNR and for different values of the in-phase-to-quadrature decision distance
ratio. It is clear that the best case is when the in-phase and quadrature distance are equal (i.e. τ = 1) for
both cases of noise. For equal energy between the in-phase and quadrature signal (i.e. τ = (21/5)1/2 ),
the system looses in performance but in small amount (about 1 dB in SNR for large SNRs). However,
the loss is more important when the quadrature signal has 21/5 times the average energy of the in-phase
signal (i.e. τ = 21/5), since it incurs a loss of about 4 dB in SNR for large SNRs, relative to the case
where τ = 1.

R EFERENCES
[1] N. Beaulieu, “A useful integral for wireless communication theory and its application to rectangular signaling constellation error rates,”
IEEE Trans. Commun., vol. 54, no. 5, pp. 802–805, May 2006.
[2] G. Karagiannidis, “On the symbol error probability of general order rectangular QAM in Nakagami-m fading,” IEEE Commun. Lett.,
vol. 10, no. 11, pp. 745–747, November 2006.
[3] R. Mallik, “Average of product of two Gaussian Q-functions and its application to performance analysis in Nakagami fading,” IEEE
Trans. Commun., vol. 56, no. 8, pp. 1289–1299, August 2008.
[4] H. Suraweera and J. Armstrong, “A simple and accurate approximation to the SEP of rectangular QAM in arbitrary Nakagami-m fading
channels,” IEEE Communications Letters, vol. 11, no. 5, pp. 426–428, May 2007.
[5] J. G. Proakis and M. Salehi, Communication Systems Engineering, 2nd ed. Upper Saddle River, NJ, USA: Prentice-Hall, August
2001.
[6] R. Viswanathan and A. Ansari, “Distributed detection of a signal in generalized Gaussian noise,” IEEE Transcations on Acoustics,
Speech and Signal Processing, vol. 37, no. 5, pp. 775–778, May 1989.
[7] S. Zahabi and A. Tadaion, “Local spectrum sensing in non-Gaussian noise,” in Proc. of the IEEE 17th International Conference on
Telecommunication (ICT’2010), Doha, Qatar, April 2010, pp. 843–847.
[8] H. Soury, F. Yilmaz, and M.-S. Alouini, “Average bit error probability of binary coherent signaling over generalized fading channels
subject to additive generalized Gaussian noise,” IEEE Commun. Lett., vol. 16, no. 6, pp. 785–788, June 2012.
[9] F. Yilmaz and M.-S. Alouini, “A new simple model for composite fading channels: Second order statistics and channel capacity,” in
Proc. of the IEEE 7th International Symposium on Wireless Communication Systems (ISWCS’2010), York, UK, Sept. 2010, pp. 676–680.
[10] H. Soury, F. Yilmaz, and M.-S. Alouini, “Exact symbol error probability of square M-QAM signaling over generalized fading channels
subject to additive generalized Gaussian noise,” in Proc. of the IEEE International Symposium on Information Theory (ISIT’2013),
Istanbul, Turkey, July 2013.
[11] M. A. Chaudhry and S. M. Zubair, On a Class of Incomplete Gamma Function with Applications. Boca Raton-London-Ney York-
Washington, D.C.: Chapman & Hall/CRC, 2002.
[12] A. Kilbas and M. Saigo, H-Transforms : Theory and Applications (Analytical Method and Special Function), 1st ed. CRC Press,
2004.
[13] F. Yilmaz and M.-S. Alouini, “Product of the powers of generalized Nakagami-m variates and performance of cascaded fading channels,”
in Proc. of the IEEE Global Telecommunication Conference (GLOBECOM’2009), Honolulu, Hawai, USA, Dec 2009, pp. 1–8.
[14] P. Mittal and K. Gupta, “An integral involving generalized function of two variables,” in Indian Acad. Sci., 1972, pp. 117–123.
[15] K. Peppas, “A new formula for the average bit error probability of dual-hop amplify-and-forward relaying systems over generalized
shadowed fading channels,” IEEE Wireless Commun. Lett., vol. 1, no. 2, pp. 85–88, Apr 2012.
[16] I. Ansari, S. Al-Ahmadi, F. Yilmaz, M.-S. Alouini, and H. Yanikomeroglu, “A new formula for the BER of binary modulations with
dual-branch selection over generalized-K composite fading channels,” IEEE Trans. Commun., vol. 59, pp. 2654–2658, Oct 2011.
[17] E. Stacy, “A generalization of the gamma distribution,” The Annals of Mathematical Statistics, vol. 3, no. 33, pp. 1187–1192, Sep.
1962.
[18] V. Aalo, T. Piboongungon, and C.-D. Iskander, “Bit-error rate of binary digital modulation schemes in generalized gamma fading
channels,” IEEE Communications Letters, vol. 9, no. 2, pp. 139–141, Feb. 2005.
[19] H. Suzuki, “A statistical model for urban radio propogation,” IEEE Transactions on Communications, vol. 25, no. 7, pp. 673–680, July
1977.
[20] M. H. Ismail and M. M. Matalgah, “Outage probability in multiple access systems with weibull-faded lognormal-shadowed
communication links,” in in Proc of the IEEE 62nd Vehicular Technology Conference (VTC-2005-Fall), vol. 4, Dallas, TX, Sept.
2005, pp. 2129–2133.
 9

0
10
m = 1/2
−1
10

−2 m=2
Average Symbol Error Probability

10

−3
10

−4
10 Laplacian Noise
Gaussian Noise
−5
10

m=4
−6
10

−7 m=∞
10

−8
10
10 15 20 25 30 35
ET
Average SNR σ2 (dB)

Fig. 1. ASEP of 8×4-QAM over Nakagami-m fading subject to Laplacian noise and Gaussian noise. The markers denote simulation results
while the lines represent analytical results.
 10

0
10

Rayleigh
−1
Average Symbol Error Probability

10

−2
10 α=10
α=2
α=1/2
α=1/3
−3
10

Nakagami − 4

−4
10
10 12 14 16 18 20 22 24 26 28
ET
Average SNR σ2 (dB)

Fig. 2. ASEP of 8×4-QAM over Rayleigh and Nakagami-4 fading subject to generalized Gaussian noise. The markers denote simulation
results while the lines represent analytical results.
 11

0
10

−1
10
Average Symbol Error Probability

−2
10

τ = 21/5
−3
10
τ = (21/5)1/2
τ=1

−4
10

−5
10
10 12 14 16 18 20 22 24 26 28 30
ET
Average SNR σ2 (dB)

Fig. 3. ASEP of 8×4-QAM over Nakagami-4 fading subject to Laplacian and Gaussian noise. The markers denote simulation results, the
solid lines represent the Laplacian noise case, and the dashed lines represent Gaussian noise case.
 12

TABLE I
S OME S PECIAL C ASES OF G ENERALIZED G AUSSIAN D ISTRIBUTION

Noise Distribution α
Impulsive 0
Gamma 0.5
Laplacian 1
Gaussian 2
Uniform ∞
 13

TABLE II
S OME S PECIAL C ASES OF T HE EGK D ISTRIBUTION

Envelope Distribution m ξ ms ξs
1 1 ∞ 1
Rayleigh
∞ 1 1 1
3/2 1 ∞ 1
Maxwell
∞ 1 3/2 1
1/2 1 ∞ 1
Half-Normal
∞ 1 1/2 1
1 1/2 ∞ 1
Exponential
∞ 1 1 1/2
1 ξ ∞ 1
Weibull
∞ 1 1 ξs
m 1 ∞ 1
Nakagami-m
∞ 1 ms 1
m ξ ∞ 1
GNM [13]
∞ 1 ms ξs
m 1/2 ∞ 1
Gamma
∞ 1 ms 1/2
m ξ/2 ∞ 1
Generalized Gamma [17], [18]
∞ 1 ms xis /2
∞ 0 ∞ 1
Lognormal
∞ 1 ∞ 1
1/2 1 1 1
Half-Normal-Exponential
1 1 1/2 1
1/2 1 ms 1
Half-Normal-Gamma
1 1 1/2 1
∞ 0 ms ξs
GNM-Lognormal
m ξ ∞ 0
∞ 0 1 1
Suzuki [19]
1 1 ∞ 0
Rayleigh-Exponential 1 1 1 1
∞ 0 3/2 1
Maxwell-Lognormal
3/2 1 ∞ 0
1 1 3/2 1
Maxwell-Exponential
3/2 1 1 1
m 1 3/2 1
Maxwell-Gamma
3/2 1 ms 1
∞ 0 1 ξs
Weibull-Lognormal [20]
1 ξ ∞ 0
1 1 1 ξs
Weibull-Exponential
1 ξ 1 1
Weibull-Weibull 1 ξ 1 ξs
m 1 1 ξs
Weibull-Gamma
1 ξ ms 1
∞ 0 ms 1
Nakagami-Lognormal
m 1 ∞ 0
m 1 1 1
K-Distribution
1 1 ms 1
Generalized-K m 1 ms 1
1 1 ms ξs
GNM-Exponential
m ξ 1 1
m ξ 1 ξs
GNM-Weibull
1 ξ ms ξs
m 1 ms ξs
GNM-Gamma
m ξ ms 1