A Flexible Lognormal Sum Approximation
A Flexible Lognormal Sum Approximation
A Flexible Lognormal Sum Approximation
http://www.merl.com
TR2005-099
November 2005
Abstract
A simple and novel method is presented to approximate the distribution of the sum of independent, but not necessarily identical, lognormal random variables, by the lognormal distribution.
It is shown that matching a short GaussHermite approximation of the moment generating function of the lognormal sum with that of the lognormal distribution leads to an accurate lognormal
sum approximation. The advantage of the proposed method over the ones in the literature, such
as the Fenton-Wilkinson method, Schwartz-Yeh method, and the recently proposed BeaulieuXie method, is that it provides the parametric flexibility to handle the inevitable trade-off that
needs to be made in approximating different regions of the probability distribution function. The
accuracy is verified using extensive simulations based on a cellular layout.
IEEE Globecom 2005
This work may not be copied or reproduced in whole or in part for any commercial purpose. Permission to copy in whole or in part
without payment of fee is granted for nonprofit educational and research purposes provided that all such whole or partial copies include
the following: a notice that such copying is by permission of Mitsubishi Electric Research Laboratories, Inc.; an acknowledgment of
the authors and individual contributions to the work; and all applicable portions of the copyright notice. Copying, reproduction, or
republishing for any other purpose shall require a license with payment of fee to Mitsubishi Electric Research Laboratories, Inc. All
rights reserved.
c Mitsubishi Electric Research Laboratories, Inc., 2005
Copyright
201 Broadway, Cambridge, Massachusetts 02139
MERLCoverPageSide2
Abstract A simple and novel method is presented to approximate the distribution of the sum of independent, but
not necessarily identical, lognormal random variables, by the
lognormal distribution. It is shown that matching a short GaussHermite approximation of the moment generating function of
the lognormal sum with that of the lognormal distribution leads
to an accurate lognormal sum approximation. The advantage of
the proposed method over the ones in the literature, such as
the Fenton-Wilkinson method, Schwartz-Yeh method, and the
recently proposed Beaulieu-Xie method, is that it provides the
parametric flexibility to handle the inevitable trade-off that needs
to be made in approximating different regions of the probability
distribution function. The accuracy is verified using extensive
simulations based on a cellular layout.
I. I NTRODUCTION
The lognormal distribution arises in several wireless systems
such as cellular mobile communication systems [1, Chp. 3] and
ultra wide band transmission [2]. For example, it models the
attenuation due to shadowing in wireless channels. Therefore,
one often encounters the sum of lognormal random variables
(RV) in analyzing wireless system performance. Given the
importance of the lognormal sum distribution, considerable
efforts have been devoted to analyze its statistical properties.
While exact closed-form expressions for the lognormal sum
probability density function (pdf) are unknown, several analytical approximation methods exist in the literature [3][8].
The methods proposed in the literature can be classified into
two broad categories. The methods by Fenton-Wilkinson [3],
Schwartz-Yeh [4], and Beaulieu-Xie [6] approximate the lognormal sum by a single lognormal RV, and provide different
recipes for determining the parameters of the lognormal pdf.
The proven permanence of the lognormal pdf when the number
of summands approaches infinity lends credence to these methods [5], [9]. The methods by Farley [1], [4], Ben Slimane [7],
and Schleher [8] instead compute a compound distribution or
specify it implicitly. For example, the first two methods derive
strict lower bounds of the cumulative distribution function
(CDF), while the last one partitions the lognormal RVs range
into three segments, with each segment being approximated
by a different lognormal RV.
Beaulieu et al. [6], [10] have studied in detail the accuracy of several of the above methods, and shown that each
method has its own advantages and disadvantages; none is
unquestionably better than the others. Farleys method and,
more generally, the formulae derived in [7] are strict bounds
The authors are with the Dept. of Engineering Science, Sonoma State
University, Rohnert Park, CA 94928, USA, and the Mitsubishi Electric
Research Labs (MERL), 201 Broadway, Cambridge, MA 02139, USA (Email:
[email protected],{mehta, jzhang}@merl.com).
This work was done when the author was at MERL.
(x X )2 pX (x)dx =
i=1 0
K Z
X
i=1
i=1 i
and Y = 10 log10
i=1 Xi , respectively. While the match
is exact for K = 2, an iterative technique needs to be used for
K > 2. The parameters Y and Y are evaluated numerically.
i=1
The simplicity of the F-W method arises from the fact that
the mean and variance of a sum of independent RVs can be
written directly as the sum of the mean and variance of the
individual RVs. The MGF of the sum of independent RVs also
possesses this desirable property that it can be written directly
in terms of the MGFs of the individual RVs as follows:
(PK
i=1
Xi
(s) =
)
K
Y
i=1
Xi (s),
(Re(s) 0).
(5)
X (s) =
exp
dx, (6a)
2Y2
xY 2
0
"
!#
N
X
wn
2Y an +Y
exp s exp
+ RN , (6b)
=
n=1
where Y and Y are the mean and standard deviation of the
Gaussian RV Y = 10 log10 X. Eqn. (6b) is the Gauss-Hermite
series expansion of the MGF function, N is the Hermite
integration order, = 10/ loge 10 is a scaling constant, and
RN is a remainder term. The weights, wi , and the abscissas,
ai , are tabulated in [14, Tbl. 25.10] for N 20. From (6b),
we can define the Gauss-Hermite representation of the MGF,
b , by removing R as follows:
X
N
!#
"
N
X wn
2an +
b
exp s exp
. (7)
X (s; , ) =
n=1
X
PK
The lognormal sum i=1 Xi can now be approximated by
a lognormal RV X = 100.1Y , where Y N (Y , Y2 ), by
PK
matching the MGF of X with the MGF of i=1 Xi at two
different values of s: s1 and s2 . This sets up a system of two
independent equations to calculate Y and Y2 , as follows:
"
!#
N
X
wn
2Y an + Y
exp sm exp
n=1
=
K
Y
i=1
b X (sm ; Y , Y ),
i
i
Note that the right hand side of the above two equations is a
constant number. These non-linear equations in Y and Y can
be readily solved numerically using standard functions such as
fsolve in Matlab and NSolve in Mathematica.
2 A formula for the MGF, in the form of an infinite series, was derived by
Naus [15] for the special case of the sum of two independent and identically
distributed lognormal RVs.
10
10
|w(x)|
10
10
10
FW (mean: x)
FW (variance: x2)
SY |log (x)|
10
10
MGF exp(0.1x)
MGF exp(5x)
5
10
0.5
1
x
1.5
Fig. 1. Weight functions employed by F-W, S-Y, and the proposed MGFbased method
1
GaussHermite approximation to MGF (s)
N=2
N=4
N = 12
N = 20
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 2
10
10
10
10
10
3 The
pathloss factor
variance.
d
k
18
2
11
17
BS 0
3
MS
12
16
15
13
14
10
10
10
CDF
10
Proposed
FW
SY
Simulation
10
10
10
10
10
10
V. C ONCLUSIONS
0
10
10
10
10
MGF
FW
SY
Simulation
10
10
10
10
10
10
10
CDF
10
10
10
= 12 dB
Proposed
SY
Simulation
= 4 dB
10
10
10
10
10
10
Fig. 6.
CDF of co-channel interference from both first and second-tier
interferers (K = 18) for = 4 dB and = 12 dB
1
0.9
0.8
Outage Probability
0.7
0.6
0.5
= 12 dB
= 4 dB
0.4
0.3
0.2
0.1
0
0
Fig. 7.
Proposed
SY
FW
Simulation
0.2
0.4
0.6
0.8
MSBS distance as a fraction of cell radius