Sum of Lognormals

SUMS OF LOGNORMALS
Daniel Dufresne
Centre for Actuarial Studies
University of Melbourne
Abstract
The problem of finding the distribution of sums of lognormally distributed random variables
is discussed. References going back to the 1930s are given, as well as some possible solutions.
A formula for the characteristic function of one lognormal is stated, and then the moments and
distribution of the logarithm of sums of lognormals are considered.
1. Introduction
Finance: In financial mathematics, the most popular model for a stocks price is the lognormal distribution: if P is the stock price, then log P has a normal distribution.
Suppose there are two such stock prices, P1 and P2 , and that they both have a lognormal distribution.
What is the probability that the sum of the prices will be greater than y? Making the simplifying assumption
that the stocks are stochastically independent, then this probability is (for y > 0):
Z
`1 +`2 >y
d`1 d`2
[(log `1 ) 1 ]2
[(log `2 ) 2 ]2
exp
.
21 2 `1 `2
212
222
This integral can be evaluated numerically, but it would seem that nothing can be done explicitly for the
distribution of the sum of lognormals. This is one of the most surprising facts of elementary probability theory:
almost nothing is known of the distribution of the sum of lognormals, although the lognormal distribution
is a simple transformation of the very pleasant normal distribution.
Option pricing: Same problem with Asian and basket options, which involve sums of two or more
lognormals.
Actuarial science: Individual claims are often well represented by a lognormal distribution; what is the
distribution of total claims?
Engineering: The oldest and widest literature on the sum of lognormals is in engineering. Amplitudes
of signals are modelled as lognormals. In telecommunications, engineers talk of the power sum, or the
logarithm of a sum of signals. This is of importance in wireless systems.
Other applications: The lognormal distribution has been used in many other fields: in economics, finance,
reliability, biology, ecology, atmospheric sciences, geology. Even to model the duration of marriage (Aitchison
& Brown, 1957).
Historical Summary
Weber (1834). Said to be one of the first studies of the properties of the lognormal distributions (cited
by Limpert et al. (2001)).
Dixon, J.T. (1932, unpublished, cited by Marlow (1967)): method for approximating the sum of lognormals
Wilkinson, R.I. (1934, Bell Telephone Labs, unpublished, cited by Marlow (1967)): possibly the first to
use the lognormal approximation:
X
if S =
eNj then log S N (, 2 ).
1
No mathematical justification was apparently given.

Fenton, L.F. (1960). Takes up Wilkinsons (1934) idea of a lognormal approximation for sums of lognormals based on moment-matching, henceforth called the Fenton-Wilkinson approximation. (Later used
in pricing Asian and basket options.)
Mitchell, R.L. (1968). Applies a modified Gram-Charlier series to approximate sums of lognormals.
Pursued by Schleher (1977) and others. Similar formulas (Edgeworth series) used in finance by Jarrow
& Rudd (1982) and Turnbull & Wakeman (1991). Resulting series has not been shown to converge.
Leipnik, R.B. (1991). Derives an exact integral expression for the characteristic function of the lognormal
distribution.
Dufresne, D. (2004). Proves that in the limit as volatilities (j ) tend to 0 a sum of lognormals tends
both to a lognormal and to a normal, depending on the normalization used.
Wu et al. (2005). Sum of lognormals approximated by a single lognormal, based on two approximate
values of Laplace transform. Used in wireless systems.
The only exact results known to date are, as far as I know:
1. Convolution integrals such as:
Z
fL1 +L2 (w) =
12 (log `1 )2 12 (log(w`)2 )2
d`
2
2
e 21
21 2 `(w `)
(or (n 1)-fold integrals for the sum of n lognormals). The case n = 2 is not bad numerically, but n 3 is
a problem.
2. Series for E(eN1 + + eNk )r .
3. Series for E log(eN1 + + eNk ) and E[log(eN1 + + eNk )]2 .

4. Epression for the characteristic function of a single lognormal (Leipnik, 1991).
2. The characteristic function of the lognormal distribution
If N N(0, 1), then L = eN Lognormal(0, 2 ). Define
t
H(t) = E eie
t N
= E eie
t R.
Then:
t N
H 0 (t) = E iet+N eie

2
t+ 2
= ie
2
t+ 2
= ie
2
t+ 2
H (t) = ie
E eN
2
2
t N
eie
t (N +)
E eie
(Cameron-Martin)
H(t + ).
This is a delay-dierential equation, and Leipnik (1991) uses it as a starting point to find an expression
for the characteristic function of the lognormal distribution. His derivation uses de Bruijns method for
delay-dierential equations. After some algebra, Leipniks result is:
if L Lognormal(0, 2 ), then (u R {0}, 0 < c < 1):
EeiuL =
1
2
c+i1
dz e
2
2
z 2 z(log u+i/2)
ci1
sin(z)(z).
(1)
then
Proceeding dierently, this author gets a slightly dierent expression: if L,2 Lognormal(, 2 ),
E eiuL,2 =
1
2i
c+i1
dz e(log |u|+ 2 sgn(u))z+

i
2 z2
2
(z).
(2)
ci1
Here c > 0 is arbitrary, and the integration path is to the right of the origin. It has not been possible to
reconcile those two expressions. Numerically, (2) does agree with the direct integral
Z 1
dy
1
exp iuy 2 ( log y)2

2
2y
0
computed using Mathematica, for many values of and 2 . However, in all those cases formula (1) gave
vastly incorrect numbers. For instance, if = 0, = 1, then the direct formula yields (using Mathematica)
0.340301 + 0.507190i
to six decimal places. Formula (2) yields the same, 0.340301+0.507190i. Formula (1) yields 13.2399+19.8899i.
Observe that the latter violates the condition that the characteristic function has norm at most equal to 1.
However, this is not a rigorous proof that (1) is incorrect, because I could have used incorrect Mathematica
code, nor does itprove that (2) is correct.
Holgate (1989) finds approximations for the characteristic function of the lognormal
3. Moments of the log of the sum of lognormals
The logonormal approximation naturally leads to the problem of finding the first two moments of the
logarithm of such a sum. Apart from the direct multiple integral there is no simple, explicit expression
for those moments. Series are known in some simple cases. What follows is in the literature (see Crow &
Shimizu, 1988 for references). Consider two independent normal variables Xj N(0, 2 ), j = 1, 2, and let
Y1 = min(X1 , X2 ), Y2 = max(X1 , X2 ). Then
eX1 + eX2 = eY2 (1 + eY1 Y2 )
and so
E[log(eX1 + eX2 )]k = E[Y2 + log(1 + eY1 Y2 )]k ,
k = 1, 2, . . . .
The appearance of the order statitics of the normal vector yields some simplification, especially for the first
moment. In that case the problem reduces to computing two single integrals. The first one has a well-known
explicit expression, while the second integral may be expressed as a series. Letting N N(0, 1), one finds:
EY2 = EY1 ;
2
E(Y2 Y1 ) = 2EY2 = E|X1 X2 | = E( 2|N |) = ,
and thus EY2 = / . Next,

E log(1 + eY1 Y2 ) = E log(1 + e|X1 X2 | ) = E log(1 + e
2|N |
),
which can be expanded using the Taylor expansion about 0 of log(1 + z), yielding
E log(1 + eY1 Y2 ) =
where (x) =
1
1
X
X
(1)n+1 n2|N |
(1)n+1 n2 2
Ee
= 2
e
(n 2),
n
n
n=1
n=1
y2
dy
e 2 . Finally,
2
1
X
(1)n+1 n2 2
E log(eX1 + eX2 ) = + 2
e
(n 2).
n
n=1
A similar formula is also known for the second moment. This series converges slowly. This is because
2 2
1
1
1

en (n 2)
+
2 n 2 (n 2)3
as n 1 (Feller, 1968, p.193). It also converges faster for larger . For instance, the relative error of the
ten-term truncated series is 6% when = .01, while it is .05% when = 3. Convergence is significantly
improved using Richardsons extrapolation.
4. Density of the sum of two lognormals
It is possible to find series for the density of the logarithm of the sum of two lognormals. Each term is
a polynomial times the normal density times the normal distribution function.
Example. Suppose X1 , X2 N(0, 1) are independent, and let Y = log(eX1 + eX2 ). Figures 1 to 4 compare the exact density of Y (found by numerical integration using Mathematica) with n-term approximations.
The truncated series is very well behaved, and gets closer to the exact density as n increases.
References
Aitchison, J., and Brown, J.A.C. (1957). The Lognormal Distribution: With Special Reference to Its Use
in Economics. Cambridge University Press.
Crow, E.L., and Shimizu, K. (1988). Lognormal Distributions: Theory and Applications. New York,
Marcel Dekker.
Dufresne, D. (2004). The log-normal approximation in financial and other applications. Adv. Appl. Prob.
36: 747-773.
Feller, W. (1968). An Introduction to Probability Theory and its Applications, Vol.1. (Third Edition.)
Wiley, New York.
Fenton, L.F. (1960). The sum of log-normal probability distibutions in scattered transmission systems.
IRE Trans. Commun. Systems 8: 57-67.
Holgate, P. (1989). The lognormal characterisitic function. Communications in Statistics Theory and
Methods 18: 4539-4548.
Jarrow, R., and Rudd, A. (1982). Approximate option valuation for arbitrary stochastic processes. J. of
Financial Economics 10: 347-369.
Leipnik, R.B. (1991). On lognormal random variables: I The characteristic function. J. Australian
Math. Soc. Ser. B 32: 327-347.
Limpert, E., Stahel, W.A., and Abbt, M. (2001). Log-normal distributions across the sciences: Keys and
clues. Bioscience 51: 341-352.
Mitchell, R.L. (1968). Permanence of the log-normal distribution. J. Optical Society of America. 58:
1267-1272.
Marlow, N.A. (1967). A normal limit theorem for power sums of independent random variables. Bell
System Technical J. 46: 2081-2089.
Schleher (1977). Generalized Gram-Charlier series with applications to sums of log-normal variates.
IEEE Trans. Inform. Theory: 275-280.
Turnbull, S., and Wakeman, L. (1991). A quick algorithm for pricing European average options. Journal
of Financial and Quantitative Analysis 26: 377-389.
Wu, J, Mehta, N., and Zhang, J. (2005). A flexible lognormal sum approximation. Proceedings of IEEE
Global Telecommunications Conference GLOBECOM 2005 6: 3413-3417.
4
0.5
0.4
0.3
0.2
0.1
!2
Figure 1. Exact density of log(eX1 + eX2 ) and 1-term approximation (dotted line)
0.5
0.4
0.3
0.2
0.1
!2
0.6
0.5
0.4
0.3
0.2
0.1
!2
0.5
0.4
0.3
0.2
0.1
!2

Sum of Lognormals

Uploaded by

Copyright:

Available Formats

Sum of Lognormals

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sum of Lognormals

Uploaded by

Copyright:

Available Formats

SUMS OF LOGNORMALS

No mathematical justification was apparently given.

fL1 +L2 (w) =

3. Series for E log(eN1 + + eNk ) and E[log(eN1 + + eNk )]2 .

H 0 (t) = E iet+N eie

dz e(log |u|+ 2 sgn(u))z+

exp iuy 2 ( log y)2

and thus EY2 = / . Next,

You might also like