Highly accurate log skew normal

Natarajan Meghanathan et al. (Eds) : NeTCoM, CSIT, GRAPH-HOC, SPTM - 2014
pp. 41–52, 2014. © CS & IT-CSCP 2014 DOI : 10.5121/csit.2014.41304
HIGHLY ACCURATE LOG SKEW NORMAL
APPROXIMATION TO THE SUM OF
CORRELATED LOGNORMALS
Marwane Ben Hcine1
and Ridha Bouallegue2
¹‫²׳‬Innovation of Communicant and Cooperative Mobiles Laboratory,
INNOV’COM
Sup’Com, Higher School of Communication
Univesity of Carthage
Ariana, Tunisia
¹marwen.benhcine@supcom.tn
2
ridha.bouallegue@supcom.rnu.tn
ABSTRACT
Several methods have been proposed to approximate the sum of correlated lognormal RVs.
However the accuracy of each method relies highly on the region of the resulting distribution
being examined, and the individual lognormal parameters, i.e., mean and variance. There is no
such method which can provide the needed accuracy for all cases. This paper propose a
universal yet very simple approximation method for the sum of correlated lognormals based on
log skew normal approximation. The main contribution on this work is to propose an analytical
method for log skew normal parameters estimation. The proposed method provides highly
accurate approximation to the sum of correlated lognormal distributions over the whole range
of dB spreads for any correlation coefficient. Simulation results show that our method
outperforms all previously proposed methods and provides an accuracy within 0.01 dB for all
cases.
KEYWORDS
Correlated Lognormal Sum, Log Skew Normal, Interference, Outage Probability
1. INTRODUCTION
Multipath with lognormal statistics is important in many areas of communication systems. With
the emergence of new technologies (3G, LTE, WiMAX, Cognitive Radio), accurate interference
computation becomes more and more crucial for outage probabilities prediction, interference
mitigation techniques evaluation and frequency reuse scheme selection. For a given practical
case, Signal-to-Interference-plus-Noise (SINR) Ratio prediction relies on the approximation of
the sum of correlated lognormal RVs. Looking in the literature; several methods have been
proposed in order to approximate the sum of correlated lognormal RVs. Since numerical methods
require a time-consuming numerical integration, which is not adequate for practical cases, we
consider only analytical approximation methods. Ref [1] gives an extension of the widely used
iterative method known as Schwartz and Yeh (SY) method [2]. Some others resources uses an
extended version of Fenton and Wilkinson methods [3-4]. These methods are based on the fact
that the sum of dependent lognormal distribution can be approximated by another lognormal

42 Computer Science & Information Technology (CS & IT)
distribution. The non-validity of this assumption at distribution tails, as we will show later, is the
main raison for its fail to provide a consistent approximation to the sum of correlated lognormal
distributions over the whole range of dB spreads. Furthermore, the accuracy of each method
depends highly on the region of the resulting distribution being examined. For example, Schwartz
and Yeh (SY) based methods provide acceptable accuracy in low-precision region of the
Cumulative Distribution Function (CDF) (i.e., 0.01–0.99) and the Fenton–Wilkinson (FW)
method offers high accuracy in the high-value region of the CDF (i.e., 0.9–0.9999). Both
methods break down for high values of standard deviations. Ref [5] propose an alternative
method based on Log Shifted Gamma (LSG) approximation to the sum of dependent lognormal
RVs. LSG parameters estimation is based on moments computation using Schwartz and Yeh
method. Although, LSG exhibits an acceptable accuracy, it does not provide good accuracy at the
lower region.
In this paper, we propose a very highly accurate yet simple method to approximate the sum of
correlated lognormal RVs based on Log Skew Normal distribution (LSN). LSN approximation
has been proposed in [6] as a highly accurate approximation method for the sum of independent
lognormal distributions, Furthermore a modified LSN approximation method is proposed in [7].
However, LSN parameters estimation relies on a time-consuming Monte Carlo simulation. The
main contribution on this work is to provide a simple analytical method for LSN parameters
estimation without the need for a time-consuming Monte Carlo simulation or curve fitting
approximation. Our analytical fitting method is based on moments and tails slope matching for
both distributions. This work can be seen as extension to the correlated case for our work done in
[8].
The rest of the paper is organized as follows: In section 2, a brief description of the lognormal
distribution and sum of correlated lognormal distributions is given. In section 3, we introduce the
Log Skew Normal distribution and its parameters. In section 4, we use moments and tails slope
matching method to estimate LSN distribution parameters. In section 5, we provide comparisons
with well-known approximation methods (i.e. Schwartz and Yeh, Fenton–Wilkinson, LSG) based
on simulation results.
The conclusion remarks are given in Section 6.
2. SUM OF CORRELATED LOGNORMAL RVS
Given X, a Gaussian RV with mean Xµ and variance 2
Xσ , then X
L e= is a lognormal RV with (PDF):
[ ]2
2
1 1
exp( ln(l) ) 0
2(l, , ) 2
0
X
XL X X X
l
f l
otherwise
µ
σµ σ π σ
 
− − > 
= 
 
 
(1)
ln(l)
( )X
X
µ
φ
σ
−
=
Where (x)φ is the standard normal cumulative distribution function (cdf).
Usually X represents power variation measured in dB. Considering XdB
with mean dBµ and
variance 2
dBσ , the corresponding lognormal RV 1 0
10
dB
d B
X
X
L eζ
= = has the following pdf:
[ ]
2
2
1 1
exp( 10log(l) ) 0
22(l, , )
0
dB
dBdBL dB dB
l
lf
otherwise
µ
σπ σµ σ
 
− − > 
=  
 
 
(2)

Computer Science & Information Technology (CS & IT) 43
Where
X
d B
X
d B
µ
µ
ξ
σ
σ
ξ
=
=
and ln(10)
10
ξ =
The first two central moments of L may be written as:
2
2 2
/2
2 2
( 1)
m e e
D e e e
µ σ
µ σ σ
=
= −
(3)
Correlated Lognormals sum distribution corresponds to the sum of dependent lognormal RVs, i.e.
1 1
n
N N
X
n
i i
L e
= =
Λ = =∑ ∑ (4)
We define as a strictly positive random vector such that the vector
with log( )j jX L= , 1 j N≤ < has an n-dimensional normal distribution with
mean vector 1 2( , ... )Nµ µ µ µ= and covariance matrix M with (i,j) ( , )i jM Cov X X= ,
1 ,1i N j N≤ < ≤ < . is called an n-dimensional log-normal vector with parameters andM.
( , )i jCov L L may be expressed as [9, eq. 44.35]:
2 21
( )
( , )2
( , ) (e 1)
i j i j M i j
i jCov L L e
µ µ σ σ+ + +
= − (5)
The first two central moments of Λ may be written as:
2
/ 2
1 1
m i i
N N
i
i i
m e eµ σ
= =
= =∑ ∑ (6)
2 2
2
1, 1
1
( )
( , )2
1, 1
( , )
(e 1)
i j i j
N
i j
i j
N
M i j
i j
D Cov L L
e
µ µ σ σ
= =
+ + +
= =
=
= −
∑
∑
(7)
3. LOG SKEW NORMAL DISTRIBUTION
The standard skew normal distribution was firstly introduced in [10] and was independently
proposed and systematically investigated by Azzalini [11]. The random variable X is said to
have a scalar ( , , )SN λ ε ω distribution if its density is given by:
2
(x; , , ) ( ) ( )X
x x
f
ε ε
λ ε ω ϕ φ λ
ω ω ω
− −
= (8)
Where
2
/2
e
(x)
2
x
ϕ
π
−
= , (x ) ( ) d
x
φ ϕ ζ ζ
− ∞
= ∫
With λ is the shape parameter which determines the skewness, ε and ω represent the usual
location and scale parameters andϕ , φ denote, respectively, the pdf and the cdf of a standard
Gaussian RV.

The CDF of the skew normal distribution can be easily derived as:
(x; , , ) ( ) 2T( , )X
x x
F
ε ε
λ ε ω φ λ
ω ω
− −
= − (9)
Where function T(x, )λ is Owen’s T function expressed as:
2 2
2
0
1
exp x (1 t )
1 2
T(x, )
2 (1 t )
dt
λ
λ
π
 
− + 
 =
+∫
(10)
A fast and accurate calculation of Owen’s T function is provided in [12].
Similar to the relation between normal and lognormal distributions, given a skew normal RV X
then 10
10
dB
dB
X
X
L eζ
= = is a log skew normal distribution. The cdf and pdf of L can be easily derived
as:
10log(l) 10log(l)2
( ) ( ) 0
(l; , , )
0
dB dB
db db dbL dB db
l
lf
otherwise
ε ε
ϕ φ λ
ξω ω ωλ ε ω
− − 
> 
= 
 
 
(11)
10log(l) 10log(l)
( ) 2T( , ) 0
(l; , , )
0
dB dB
db dbL dB db
l
F
otherwise
ε ε
φ λ
ω ωλ ε ω
− − 
− > 
=  
 
 
(12)
Where d B
d B
ε
ε
ξ
ω
ω
ξ
=
=
and ln(10)
10
ξ =
The Moment Generating Function (MGF) of the skew normal distribution may be written as [11]:
2
/ 2
2
M (t) E
2 ( t),
1
tX
X
t
e
e
λ
φ β β
λ
 =  
= =
+
(13)
Thus the first two central moments of L are:
2
/ 2
2 ( )e eε ω
ζ φ β ω= (14)
2 2
2 2 2
2 ( (2 ) 2 ( ))e e eε ω ω
ϖ φ βω φ βω= −
4. VALIDITY OF LOGNORMAL ASSUMPTION FOR THE SUM OF
LOGNORMAL RVS AT DISTRIBUTION TAILS
Several approximation methods for the sum of correlated lognormal RVs is based on the fact that
this sum can be approximated, at least as a first order, by another lognormal distribution. On the
one hand, Szyszkowicz and Yanikomeroglu [14] have published a limit theorem that states that
the distribution of a sum of identically distributed equally and positively correlated lognormal
RVs converges, in distribution, to a lognormal distribution as N becomes large. This limit

theorem is extended in [15] to the sum of correlated lognormal RVs having a particular
correlation structure.
On the other hand, some recent results [16, Theorem 1. and 3.] show that the sum of lognormal
RVs exhibits a different behaviour at the upper and lower tails even in the case of identically
distributed lognormal RVs. Although the lognormal distribution have a symmetric behaviours in
both tails, this is not in contradiction with results proven in [14-15] since convergence is proved
in distribution, i.e., convergence at every point x not in the limit behaviour.
This explain why some lognormal based methods provide a good accuracy only in the lower tail
(e.g. Schwartz and Yeh), where some other methods provide an acceptable accuracy in the upper
tail (e.g. Fenton–Wilkinson). This asymmetry of the behaviours of the sum of lognormal RVs at
the lower and upper tail, motivates us to use the Log Skew Normal distribution as it represents the
asymmetric version of lognormal distribution. So, we expect that LSN approximation provide the
needed accuracy over the whole region including both tails of the sum of correlated lognormal
RVs distribution.
5. LOG SKEW NORMAL PARAMETERS DERIVATION
5.1. Tails properties of sum of Correlated lognormal RVs
Let be an N-dimensional log-normal vector with parameters µ andM . Let 1
B M −
= the inverse
of the covariance matrix.
To study tails behavior of sum of correlated lognormal RVs, it is convenient to work on
lognormal probability scale [13], i.e., under the transformation G:
(15)
We note that under this transformation, the lognormal distribution is mapped onto a linear
equation.
We define iB as row sum of B :
1
( , )
N
i
k
B B i k
=
= ∑ , 1 i N≤ ≤ (16)
We let:
We define such that:

Since variables iL are exchangeable, we can assume for the covariance matrix B, with no loss of
generality, that with .
Let such that:
(17)
So that, we may write:
(18)
Now, we set as
(19)
Assuming that for every
(20)
Where satisfies and otherwise.
In [16, Theorem 3.], Gulisashvili and Tankov proved that the slope of the right tail of the SLN cdf
on lognormal probability scale is equal to when assumption (20) is valid.
(21)
Considering the left tail slope, they proved that the slope of the left tail of the SLN cdf on
lognormal probability scale is equal to [16, Theorem 1.].
(22)
In general, we can assume that 0iB ≠ for 1 i N≤ ≤ , so that and tails slope can be
expressed as:
(23)
(24)

5.2. Tail properties of Skew Log Normal
In [17], it has been showed that the rate of decay of the right tail probability of a skew normal
distribution is equal to that of a normal variate, while the left tail probability decays to zero faster.
This result has been confirmed in [18]. Based on that, it is easy to show that the rate of decay of
the right tail probability of a log skew normal distribution is equal to that of a lognormal variate.
Under the transformation G, skew lognormal distribution has a linear asymptote in the upper limit
with slope
(25)
In the lower limit, it has no linear asymptote, but does have a limiting slope
(26)
These results are proved in [8, Appendix A]. Therefore, it will be possible to match the tail slopes
of the LSN with those of the sum of correlated lognormal RVs distribution in order to find LSN
optimal parameters.
5.3. Moments and lower tail slope matching
In order to derive log skew normal optimal parameters, we proceed by matching the two central
moments of both distributions. Furthermore, use we lower slope tail match. By simulation, we
point out that upper slope tail match is valid only for the sum of high number of lognormals RVs.
However we still need it to find an optimal starting guess solution to the generated nonlinear
equation. Thus we define as solution the following nonlinear equation:
(27)
Such nonlinear equation can be solved using different mathematical utility (e.g. fsolve in matlab).
Using upper slope tail match we derive a starting solution guess 0λ to (23) in order to converge
rapidly (only few iterations are needed):
(28)
Optimal location and scale parameters, are obtained according to

(29)
6. SIMULATION RESULTS
In this section, we propose to validate our approximation method and compare it with other
widely used approximation methods. The comparison of the Complementary Cumulative
Distribution Function (CDF) of the sum of N ( 2,8,20N = ) correlated lognormal RVs [ ]P λΛ > of
Monte Carlo simulations with LSN approximation and lognormal based methods for low value of
standard deviation 3 dBσ = with 0dBµ = and 0.7ρ = are shown in Fig.1. Although these methods
are known by its accuracy in the lower range of standard deviation, it obvious that LSN
approximation outperforms them. We note that fluctuation at the tail of sum of lognormal RVs
distribution is due to Monte Carlo simulation, since we consider 7
10 samples at every turn. We
can see that LSN approximation results are identical to Monte Carlo simulation results.
5 10 15 20 25 30 35
10
-16
10
-14
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
Complementary CDF of the Sum of N correlated Lognormal RVs Λ with µ=0,σ=3dB ρ=0.7
γ(dB)
P(Λ>λ)
Wilkinson
Ext. Sch. and Yeh
Log Skew Normal
Simulation
N=2
N=20
N=8
Figure 1. Complementary CDF of the sum of N correlated lognormal RVs with 0 dBµ = , 3 d Bσ = and
0.7ρ =
Fig.2 and Fig.3 show the complementary CDF of the sum of N correlated lognormal RVs for
higher values of standard deviation 6, 9 dBσ = and two values of correlation
coefficients 0.9, 0.3ρ = .We consider the Log Shifted Gamma approximation for comparison
purposes. We consider the Log Shifted Gamma approximation for comparison purposes. We can
see that LSN approximation highly outperforms other methods especially at the CDF right tail
( 2
1 10CDF −
− < ). Furthermore, LSN approximation give exact Monte Carlo simulation results even
for low range of the complementary CDF of the sum of correlated lognormal RVs
( 6
1 10CDF −
− < ).

To further verify the accuracy of our method in the left tail as well as the right tail of CDF of the
sum of correlated lognormal RVs, Fig. 4 and Fig. 5 show respectively the CDF [ ]P λΛ < of the
sum of 6 correlated lognormal RVs with 0dBµ = 0.7ρ =
10 15 20 25 30 35 40 45
10
-14
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
γ(dB)
P(Λ>λ)
Wilkinson
Extended Sch. and Yeh
Log Shifted Gamma
Log Skew Normal
Simulation
N=8
N=20
Figure 2. Complementary CDF of the sum of N correlated lognormal RVs with 0 dBµ = , 6 d Bσ = and
0.9ρ =
10 20 30 40 50 60 70
10
-16
10
-14
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
γ(dB)
P(Λ>λ)
Ext. Sch. and Yeh
Log Shifted Gamma
Log Skew Normal
Simulation
N=10
N=30
N=60
Figure 3. Complementary CDF of the sum of N correlated lognormal RVs with
0 dBµ = 9 d Bσ = 0.3ρ =
and the complementary CDF of the sum of 20 correlated lognormal RVs with 0dBµ = 0.3ρ = for
different standard deviation values. It is obvious that LSN approximation provide exact Monte
Carlo simulation results at the left tail as well as the right tail. We notice that accuracy does not
depend on standard deviation values as LSN performs well for the highest values of standard
deviation of the shadowing.
To point out the effect of correlation coefficient value on the accuracy of the proposed
approximation, we consider the CDF of the sum of 12 correlated lognormal RVs for different
combinations of standard deviation and correlation coefficient values with 0dBµ = (Fig. 6). One
can see that LSN approximation efficiency does not depend on correlation coefficient or standard
deviation values. So that, LSN approximation provides same results as Monte Carlo simulations
for all cases.

-70 -60 -50 -40 -30 -20 -10 0 10
10
-20
10
-15
10
-10
10
-5
10
0
CDF of the Sum of 6 correlated Lognormal RVs Λ with µ=0, ρ=0.7
γ(dB)
P(Λ<λ)
Log Skew Normal
Simulation
σ=6dB σ=3dBσ=9dBσ=12dB
σ=15dB
σ=20dB
Figure 4. CDF of the sum of 6 correlated lognormal RVs with 0 dBµ = , 0.7ρ =
10 20 30 40 50 60 70 80 90 100 110
10
-16
10
-14
10
-12
10
-10
10
-8
10
-6
10
-4
10
-2
10
0
Complementary CDF of the Sum of 20 correlated Lognormal RVs Λ with µ=0, ρ=0.3
γ(dB)
P(Λ>λ)
Log Skew Normal
Simulation
σ=3dB σ=6dB σ=9dB
σ=12dB σ=15dB
σ=20dB
Figure 5. Complementary CDF of the sum of 20 correlated lognormal RVs with 0 dBµ = , 0.3ρ =
-60 -50 -40 -30 -20 -10 0 10
10
-20
10
-15
10
-10
10
-5
10
0
CDF of the Sum of 12 correlated Lognormal RVs Λ with µ=0
γ(dB)
P(Λ<λ)
Log Skew Normal
Simulation
ρ=0.8,σ=6dB
ρ=0.4,σ=15dB
ρ=0.1,σ=6dB
ρ=0.4,σ=6dB
ρ=0.1,σ=10dB
ρ=0.4,σ=10dB
ρ=0.8,σ=10dB
ρ=0.1,σ=15dB
ρ=0.8,σ=15dB
Figure 6. CDF of the sum of 12 correlated lognormal RVs with different correlation coefficients ρ ,
0 dBµ =

7. CONCLUSIONS
In this paper, we proposed to use the Log Skew Normal distribution in order to approximate the
sum of correlated lognormal RV distribution. Our fitting method uses moment and tails slope
matching technique to derive LSN distribution parameters. LSN provides identical results to
Monte Carlo simulations results and then outperforms other methods for all cases. Such
approximation can be effectively used for accurate and fast computation of interference and
outage probabilities in cellular networks.
REFERENCES
[1] A. Safak, “Statistical analysis of the power sum of multiple correlated log-normal components”, IEEE
Trans. Veh. Tech., vol. 42, pp. 58–61, Feb. 1993.
[2] S.Schwartz and Y.S. Yeh, “On the distribution function and moments of power sums with log-normal
components”, Bell System Tech. J., Vol. 61, pp. 1441–1462, Sept. 1982.
[3] M. Pratesi , F. Santucci , F. Graziosi and M. Ruggieri "Outage analysis in mobile radio systems with
generically correlated lognormal interferers", IEEE Trans. Commun., vol. 48, no. 3, pp.381 -385
2000.
[4] A. Safak and M. Safak "Moments of the sum of correlated log-normal random variables", Proc.
IEEE 44th Vehicular Technology Conf., vol. 1, pp.140 -144 1994.
[5] C. L. Joshua Lam,Tho Le-Ngoc " Outage Probability with Correlated Lognormal Interferers using
Log Shifted Gamma Approximation” , Wireless Personal Communications, Volume 41, Issue 2, pp
179-192, April 2007.
[6] Z. Wu, X. Li, R. Husnay, V. Chakravarthy, B. Wang, and Z. Wu. A novel highly accurate log skew
normal approximation method to lognormal sum distributions. In Proc. of IEEE WCNC 2009.
[7] X. Li, Z. Wu, V. D. Chakravarthy, and Z. Wu, "A low complexity approximation to lognormal sum
distributions via transformed log skew normal distribution," IEEE Transactions on Vehicular
Technology, vol. 60, pp. 4040-4045, Oct. 2011.
[8] M. Benhcine, R. Bouallegue, “Fitting the Log Skew Normal to the Sum of Independent Lognormals
Distribution”, accepted in the sixth International Conference on Networks and Communications
(NetCom-2014) .
[9] Samuel Kotz, N. Balakrishnan, Norman L. Johnson "Continuous Multivariate Distributions", Volume
1, Second Edition, John Wiley & Sons, avril 2004.
[10] O’Hagan A. and Leonard TBayes estimation subject to uncertainty about parameter constraints,
Biometrika, 63, 201–202, 1976.
[11] Azzalini A, A class of distributions which includes the normal ones, Scand. J. Statist., 12, 171–178,
1985.
[12] M. Patefield, “Fast and accurate calculation of Owen’s t function,” J. Statist. Softw., vol. 5, no. 5, pp.
1–25, 2000.
[13] N. C. Beaulieu, Q. Xie, “Minimax Approximation to Lognormal Sum Distributions”, IEEE VTC
Spring, Vol. 2, pp. 1061-1065, April 2003.
[14] S. S. Szyszkowicz and H. Yanikomeroglu "Limit theorem on the sum of identically distributed
equally and positively correlated joint lognormals", IEEE Trans. Commun., vol. 57, no. 12,
pp.3538 -3542 2009
[15] N. C. Beaulieu, "An extended limit theorem for correlated lognormal sums," IEEE Trans. Commun.,
vol. 60, no. 1, pp. 23-26, Jan. 2012
[16] Archil Gulisashvili, Peter Tankov, “Tail behavior of sums and differences of log-normal random
variables”, ARXIV 09/2013.
[17] Antonella Capitanio, “On the approximation of the tail probability of the scalar skew-normal
distribution”, in METRON (2010).
[18] W. Hürlimann, “Tail Approximation of the Skew-Normal by the Skew-Normal Laplace: Application
to Owen’s T Function and the Bivariate Normal Distribution”, Journal of Statistical and Econometric
Methods, vol. 2, no.1, 2013, 1-12, Scienpress Ltd, 2013.

AUTHORS
Marwane Ben Hcine was born in Kébili, Tunisia, on January 02, 1985. He
graduated in Telecommunications Engineering, from The Tunisian Polytechnic
School (TPS), July 2008. In June 2010, he received the master’s degree of research
in communication systems of the Higher School of Communication of Tunisia
(Sup’Com). Currently he is a Ph.D. student at the Higher School of
Communication of Tunisia. His research interests are network design and
dimensioning for LTE and beyond Technologies.
Pr. Ridha BOUALLEGUE was born in Tunis, Tunisia. He received the M.S degree
in Telecommunications in 1990, the Ph.D. degree in telecommunications in 1994, and
the HDR degree in Telecommunications in 2003, all from National School of
engineering of Tunis (ENIT), Tunisia. Director and founder of National Engineering
School of Sousse in 2005. Director of the School of Technology and Computer
Science in 2010. Currently, Prof. Ridha Bouallegue is the director of Innovation of
COMmunicant and COoperative Mobiles Laboratory, INNOV’COM Sup’COM,
Higher School of Communication. His current research interests include mobile and cooperative
communications, Access technique, intelligent signal processing, CDMA, MIMO, OFDM and UWB
systems.

Highly accurate log skew normal

More Related Content

What's hot (18)

Similar to Highly accurate log skew normal (20)

Recently uploaded (20)

Highly accurate log skew normal