Sankhyā : The Indian Journal of Statistics

2002, Volume 64, Series A, Pt 1, pp 156-166

BIVARIATE EXPONENTIAL DISTRIBUTIONS

USING LINEAR STRUCTURES

By SRIKANTH K. IYER
Indian Institute of Technology, Kanpur
D. MANJUNATH
Indian Institute of Technology, Bombay
and
R. MANIVASAKAN
Indian Institute of Technology, Bombay

SUMMARY. We derive bivariate exponential distributions using independent auxil-

iary random variables. We develop separate models for positive and negative correlations
between the exponentially distributed variates. To obtain a positive correlation, we define
a linear relation between the variates X and Y of the form Y = aX + Z where a is a posi-
tive constant and Z is independent of X. To obtain exponential marginals for X and Y we
show that Z is a product of a Bernoulli and an Exponential random variables. To obtain
negative correlations, we define X = aP + V and Y = bQ + W where either P and Q or V
and W or both are antithetic random variables. For the case of positive correlations, we
also characterize a bivariate Poisson process generated by using the bivariate exponential
as the interarrival distribution.
Our bivariate exponetial model is useful in introducing dependence between the inter-
arrivals and service times in a queueing model and in the failure process in multicomponent
systems. The primary advantage of our model is that the resulting queueing and reliability
analysis remains mathematically tractable because the Laplace Transform of the joint dis-
tribution is a ratio of polynomials of s. Further, the variates can be very easily generated
for computer simulation.

1. Introduction

The single variable exponential distribution has long been the favorite
of analysts working with queueing systems, life testing and reliability models.

Paper received August 2000; revised March 2001.

AMS (1991) subject classification. Primary 62E10, 62H05; secondary 62N05, 60K25,
62H20.
Key words and phrases. Bivariate exponential distribution, joint distribution, Laplace
transform, correlation.
 bivariate exponential distributions 157

This is because its memorylessness property lends itself to easy analy-

sis. To provide a flavor of real life models, many distributions, like Gamma,
hyper exponential, and Coxian, have been derived from the exponential dis-
tribution. Whenever there are two or more variables in a system model, in
many analyses they are assumed to be independent. Consider, for example,
the interarrival and the service times in a queuing system which are assumed
independent in queuing theory while it can be shown that this is not true in
most realistic situations particularly in packet communication networks. As
another example consider the failure analysis of a multicomponent system
in which the lifetime of one component influences the lifetime of the other
component in which case it becomes necessary to introduce a dependence
between the two variables. To continue to use analytically amenable dis-
tributions we will need to assume them to be marginally exponential and
introduce multivariate exponential distributions that can model the depen-
dence between the variables. However, unlike the normal distribution, there
is no unique natural extension of the univariate exponential distribution and
many models have been proposed. In this paper we consider a family of bi-
variate distributions that are marginally exponential. Mathematical analysis
remains fairly straightforward with this family of distributions in spite of the
dependence that is introduced among the variates. This is not true of many
bivariate exponential distributions in literature. Further, these variables are
very easy to simulate on a computer.
The most well known of the bivariate exponential distributions was de-
rived by Marshall and Olkin(1967a,b) by considering shock models. If X and
Y are the lifetimes of two components of a system, then the distributions of
X and Y are exponential with parameters λ1 +λ12 and λ2 +λ12 and the joint
distribution of X and Y is F (x, y) = 1 − exp(−λ1 x − λ2 y − λ12 max(x, y)).
This distribution is not absolutely continuous but satisfies the lack of mem-
ory property (LMP),
P [X > t1 + s1 , Y > t2 + s2 /X > s1 , Y > s2 ] = P [X > t1 , Y > t2 ]. (1)
In fact it has been shown that the above is the only bivariate distribution
that has exponential marginals and which satisfies LMP. Properties of this
Marshall Olkin model are described in detail in Barlow and Proschan(1975).
Block and Basu (1974) consider a bivariate distribution whose marginals are
mixtures of exponentials and having an absolutely continuous joint distribu-
tion. Arnold and Strauss (1991) characterize bivariate distributions where
the conditional distributions are in prescribed exponential families. A sur-
vey of various bivariate and multivariate exponential models can be found
in Basu (1988).
158 srikanth k. iyer, d. manjunath, r. manivasakan

In this paper we present a bivariate exponential distribution that derives

its primary motivation from packet communication networks but can be used
in other models. Consider a high speed packet communication link that can
be modeled as a single server First Come-First Served (FCFS) queue. Let X
be the length of a packet that has just joined the output queue of a network
node at its tail and Y be the time till the arrival of the next packet. In
most networks, a correlation can be established between X and Y described
above. The nature of the correlation would depend on the kind of application
generating the packets. For example in bulk transfer applications like ftp,
the correlation will be negative. In some interactive applications like telnet,
the correlations could be positive. In the next section we define three models
- one for positive correlation and two for negative correlation in which the
marginals are exponential, and the Laplace-Steiltjes transform (LST) of the
joint distribution is a ratio of polynomials. In section 3 we consider a joint
counting process (N, M ) in which time between the increments of the M
process is X and that between N is Y and obtain the mass function for the
joint counting process.
2. The Bivariate Exponential
The bivariate models that we develop are based on Exponential Autore-
gressive sequences of Gaver and Lewis (1980). While borrowing freely from
the ideas in the Gaver and Lewis (1980) paper, we would like to point out
the differences as well. The major difference is that while Gaver and Lewis
construct an autoregressive AR(1) sequence Xn such that Xn and Xn+1 are
correlated and the Xn have identically distributed exponential marginals,
our interest is to construct a bivariate exponential distribution. The pri-
mary purpose of the models introduced by Gaver and Lewis is to introduce
a correlation between the successive interarrivals or service times in a queue.
Our main interest is to introduce a cross correlation between the service time
and the next interarrival. Further, in the model proposed for introducing a
negative cross correlation, we do not take recourse to (a fictitious) antithetic
queue.
2.1 Positive cross correlation. Consider two random variables X and Y
whose marginals are exponential with parameters λx and λy respectively.
Let
Y = aX + Z (2)
where a > 0 is a constant and Z is a random variable independent of X. The
constant a can be thought of as a scaling factor. Let us first characterize
Z. Let X (s), Y(s) and Z(s) be the LST of the distributions of X, Y and Z
 bivariate exponential distributions 159

respectively. From equation (2), we see that

Y(s) = X (as)Z(s)

From this we obtain

Y(s) λy λx + as
Z(s) = =
X (as) s + λy λx

Define λy /λx = ρ. After some algebra, for 0 ≤ aρ ≤ 1.0, we get

λx + a(s + λy − λy ) λy
Z(s) = ρ = (1 − ρa) + (ρa) (3)
λy + s s + λy

It is easy to see that (3) corresponds to the LST of the distribution of a

random variable that is a product of two independent random variables - a
Bernoulli random variable with mean (1 − aρ) and an exponential random
variable with parameter λy . Thus the distribution of Z will be aρ + (1 −
aρ)(1 − e−λy z ), z ≥ 0. The LST of the joint distribution is easily obtained
as follows.
   
XY(s1 , s2 ) = E e−s1 X−s2 Y = E e−(s1 +s2 a)X )E(e−s2 Z
" #
λx λy

= (1 − aρ) + (aρ) (4)
λx + s1 + s2 a s2 + λy

Let rxy denote the correlation between X and Y . E(XY ), and hence rxy is
obtained as follows.

E(XY ) = E(X(aX + Z)) = aE(X 2 ) + E(X)E(Z)

2 1 1 − aρ
= a 2 + (5)
λx λx λy
1 1−aρ
a λ22 + λx λy − 1
λy λx
x
rxy = 1 = aρ (6)
λy λx

To see the application of this model to a multicomponent reliability sys-

tem, assume for the moment that a = 1. That is, Y = X + Z. Consider a
system protected by two safety devices D1 and D2 that are subject to shock
from the same source(s). The marginal lifetime distributions of D1 and D2
are exponential with parameters λx and λy respectively. The shocks are first
absorbed by device D1 . As soon as the device D1 fails, the device D2 takes
over. So, the lifetime Y of D2 is the lifetime X of D1 plus a random amount
160 srikanth k. iyer, d. manjunath, r. manivasakan

Z. From (2) we see that Z = 0 or Y = X with probability ρ. Thus ρ is

the probability that the shock is strong enough to destroy both the devices
simultaneoulsy. If the device D2 survives the shock that destroys D1 , then
it lasts a lifetime that is exponentially distributed with parameter λy .
As a second application, consider a packet communication network with
packet transmission times denoted by X and having an exponential distribu-
tion. Let Y be the packet interarrival time which also have an exponential
distribution. In most networks, packet transmissions are best modelled as
burst processes in which during the burst, consecutive packets have strongly
correlated arrival times, typically a constant or a constant multiple of the
packet length, and bursts arrive independently. If we use the above model
to model such packet processes, the burst length, number of packets in the
burst, will be geometrically distributed with mean 1/(aρ) and the time from
the end of one burst till the beginning of the next one will be exponen-
tially distributed with parameter λy . During the burst, the time at which
the “next packet” arrives will be a constant multiple of the current packet
length.
Apart from modeling correlations in network traffic the models consid-
ered in this paper can be used to study transmission control schemes over
a network. A tansmitter sends packets after a delay which is distributed as
a constant multiple of the length of the previous packet (transmission time)
plus a random amount. Packet sizes are synonymous with the service times.
Using the linear structure and the fact that the marginals have exponential
distributions, queueing analysis can be carried out. This type of control will
help in increasing bandwith utilisation while reducing congestion.

2.2 Negative cross correlation. To obtain a negative cross correlation

between X and Y , we take a different but similar approach than the one in
the previous section. Consider the following equation

X = aP + V Y = bQ + W (7)

Here a and b are non negative constants, P and V are independent of each
other and so are Q and W . Further, from the previous section, if P and
Q are exponential with parameters λp and λq respectively, and V and W
are like Z of equation (2), the marginal distributions of X and Y will be
exponential. To induce a negative correlation between X and Y , we could
make either P and Q or V and W antithetic random variables and choose the
other two to make the marginals of X and Y exponential. We will consider
both these options separately.
 bivariate exponential distributions 161

Model 1. First, consider P and Q to be antithetic exponential ran-

dom variables. Further let V and W be independent. Following Gaver and
Lewis (1980), we define P and Q through U which is a uniformly distributed
random variable in (0, 1), as follows
1
P =− ln(U ) Q = − λ1q ln(1 − U ) (8)
λp
Under this definition of P and Q, V is the product of a Bernoulli variable
with mean (1 − a λλxp ) and an exponential random variable with parameter
λ
λx . Similarly, W is the product of a Bernoulli variable with mean (1 − b λyq )
and an exponential random variable with parameter λy . Also, a and b are
λ
to be chosen such that 0 ≤ a λλxp ≤ 1 and 0 ≤ b λyq ≤ 1. It is easy to see that
cov(X, Y ) = E(XY )−E(X)E(Y ) = abcov(P, Q). Therefore, the correlation
coefficient between X and Y , rxy , is given by
!
abλx λy π2 
aλx
 
bλy

rxy = 1− for 0 ≤ λp , λq ≤ 1 and a, b ≥ 0 (9)
λp λ q 6
 2

From equation 9 it follows that 1 − π6 ≤ rxy ≤ 0.
To simplify parameter selection, we could make λq = λp and b = a.
The LST of the joint distribution of X and Y XY(s1 .s2 ), is obtained as
follows.

XY(s1 , s2 ) = E(e−s1 X−s2 Y ) = E(e−s1 V −s2 W )E(e−as1 P −bs2 Q )

" ! !#
aλx λx aλx
= 1− +
λp s1 + λx λp
" ! !#
bλy λy bλy
1− + ×
λq s2 + λy λq
Z 1
uas1 /λp (1 − u)bs2 /λq du (10)
0

Note that the integral on the RHS of equation 10 corresponds to the beta
function.
To minimize the number of parameters to choose, we could have W = V,
instead of assuming them to be independent. In this case the cov(X, Y ) =
abcov(P, Q) + σv2 where σv2 is the variance of V . Also, since P and Q are
exponential marginals, we could consider this model to define a 4-variate
exponential in which case we could choose not to make P and Q antithetic.
In the latter case, cov(X, Y ) = abσv2 .
162 srikanth k. iyer, d. manjunath, r. manivasakan

Model 2. In the second model to obtain a negative correlation between

the exponential marginals of X and Y we make V and W antithetic. Once
again we follow Gaver and Lewis (1980), and define V and W as functions
of a random variable U that is uniform in (0, 1) as follows
( (
V =
0   if U ≤ c W =
0   if U ≥ d (11)
− λ1v ln 1−U
1−c if U > c − λ1w ln Ud if U < d

Here 0 ≤ c, d ≤ 1. Note that the random variables V and W as defined above

are of the same form as the random variable Z of equation 2. In fact V =
F1−1 (U ) and W = F2−1 (1 − U ), where F1 and F2 are distribution functions of
the same form as the distribution of Z but with different parameter values.
Thus we see that for X and Y to be marginally exponential with parameters
λx and λy respectively, λv = λx , λw = λy , and P and Q are independent
exponential random variables with λp and λq such that, c = aλ λp and 1 − d =
x

bλy
λq .
Under this model, cov(X, Y ) = cov(V, W ) and rxy , the correlation be-
tween X and Y is given by
( R  
d
ln 1−u u


c 1−c ln d du − (1 − c)d if c < d
rxy = (12)
−(1 − c)d if d ≤ c

The LST of the joint distribution of X and Y will be given by

XY(s1 , s2 ) = E(e−s1 X−s2 Y ) = E(e−as1 P −bs2 Q )E(e−s1 V −s2 W )

R d  1−u  λs1x
λs2

λp λq u
 y du if c < d,
= λp +as1 λq +bs2 hc 1−c d i
λp λq λy

λp +as1 λq +bs2 d λy +s 2
+ (c − d) + (1 − c) λxλ+s
x
1
if d ≤ c.
(13)

The magnitude of negative correlations from Model 2 can exceed π 2 /6 − 1

and hence can be larger than that of Model 1.
Just like with Model 1, we could choose Q = P, instead of assuming
them to be independent and minimize the number of parameters. In this
case we have cov(X, Y ) = cov(V, W ) + abσp2 . Further, just like in Model 1,
we could consider this model to describe a trivariate exponential in which
case we could also choose not to make V and W antithetic. In the latter
case, cov(X, Y ) = abσp2 .

Model 3. In this model we can make both P and Q and V and W

antithetic with P and Q being independent of V and W . In this case,
 bivariate exponential distributions 163

cov(X, Y ) = abcov(P, Q) + cov(V, W ). Also, just like in model 1, we could

treat this to be a model for a 4-variate exponential.
3. Joint Counting Process
As an application we consider a bivariate counting process (Mt , Nt ). Let
the interarrival times between the jumps of M and N be distributed as X
and Y respectively, which in turn are each marginally exponential and the
relation between them is defined as in equation 2. It is easy to see that Mt
and Nt are marginally Poisson. Let (Xi , Yi , Zi ), i = 1, 2, . . . , be an i.i.d.
sequence of random vectors with the same joint distribution as (X, Y, Z) of
section 2.1. Observe that ni=1 Zi has the same distribution as B × G(B, λy )
P

where B is binomial with parameter (n, p), where p = 1−aρ, and conditional
on B, G is a Erlang (Gamma) with parameters, B and λy . Define gj (z) as
the density of a gamma random variable with parameters λy and j,
λy (λy z)j−1 −λy z
gj (z) = e , z > 0.
(j − 1)!
In evaluating fm,n = Pr[Mt = m, Nt = n], m, n = 0, 1, 2, . . ., the joint prob-
ability mass function for (Mt , Nt ), we have to consider three cases.

Case 1: a > 1. In this case fm,n is non zero only if m ≥ n, since

Yi = aXi + Zi ≥ aXi > Xi . So we restrict ourselves to the case when m ≥ n.
Thus, we have
m n
" #
X X
Pr[Mt ≥ m, Nt ≥ n] = Pr Xi ≤ t, Yi ≤ t
i=1 i=1
m n
" #
n t X X
= (1 − p) Pr Xi ≤ t, Xi ≤
i=1 i=1
a
n
!
X n
+ pj (1 − p)n−j
j=1
j
m n
" #
Z t X t−z X
Pr Xi ≤ t, Xi ≤ gj (z)dz
0 i=1 i=1
a

m n
" #
X X t−z h i
Pr Xi ≤ t, Xi ≤ = Pr Mt ≥ m, M( t−z ) ≥ n
i=1 i=1
a a

m−1
X h i h i h i
= Pr M( t−z ) = k Pr M(t− t−z ) ≥ m − k + Pr M( t−z ) ≥ m (14)
a a a
k=n
164 srikanth k. iyer, d. manjunath, r. manivasakan

In the above we have used the homogeneity and independent incre-

ment properties of the Poisson process. Substitute for P r[Ms = k] =
exp(−λx s)(λx s)k /k! in the above expression and simplify using

Pr[Mt = m, Nt = n] = Pr[Mt ≥ m, Nt ≥ n] − Pr[Mt ≥ m + 1, Nt ≥ n] −

−Pr[Mt ≥ m, Nt ≥ n + 1]
+Pr[Mt ≥ m + 1, Nt ≥ n + 1]

to get, for m ≥ n,

n
n  m−n !
λx t t
 
n
X n
fm,n = (1 − p) λx t− + pj (1 − p)n−j ×
a a j=1
j
#
e−λx t
Z t n  m−n
t−z t−z

× λx λx t− gj (z)dz (15)
0 a a n!(m − n)!

The calculations are similar for the case a ≤ 1. We only write the expression
for fm,n in the remaining two cases.
Case 2: a ≤ 1, m > n. In this case, it is easy to see that

n
! n
X n
Z t 
t−z
fm,n =  pj (1 − p)n−j λx
j=1
j t(1−a) a
#
e−λx t
m−n
t−z
 
λx t− gj (z)dz × (16)
a n!(m − n)!

Case 3: a ≤ 1, m ≤ n. As before, we will just write down the

corresponding equations for m < n and m = n respectively. If m < n then,
n t !
e−λx a
n−m
t
 
n m
X n j
fm,n = (1 − p) (λx t) λx −t + p (1 − p)n−j
a m!(n − m)! j=1 j
t−z
e−λx a
t(1−a) n−m
t−z
Z  
m
× (λx t) λx −t gj (z)dz (17)
0 a m!(n − m)!
If m = n then,
 
n
! n
X n
Z t  
t−z
fm,n = (1 − p)n (λx t)n + pj (1 − p)n−j λx gj (z)dz 
j=1
j t(1−a) a
t−z
−λx at n
!
e n j t(1−a) e−λx
Z
X a
× + p (1 − p)n−j (λx t)n gj (z) dz (18)
n! j=1
j 0 n!
 bivariate exponential distributions 165

The joint counting processes for the models with negative cross-correlation
between the interarrivals of the processes M and N do not yield com-
pact formulas as above. In this case it is easier to write an expression for
P r[Mt ≥ m, Nt ≥ n] than in the positive correlation case. This is due to the
fact that the interarrivals Xi and Yi satisfy two different linear equations.
The difficulty however stems from the link between X and Y via antithetic
variables which makes subsequent calculations very messy. However, numer-
ical simulation still remains straightforward.

4. Discussion and Conclusion

We have presented bivariate distributions that have exponential marginal
distributions and a positive or a negative correlation between the variates.
When the correlation is positive, a linear relation between the variates can
be used. To obtain a negative correlation between the variates, four more
variables, two of which are exponential, are used and the concept of antithetic
variables has been used.
A major advantage of our description of X and Y according to equa-
tions (2) and (7) is that they lend themselves to easy simulation. Consider
for example generating a sequence (Xn , Yn ) in which the Xn and Yn are
marginally exponential with a positive cross correlation. We use equation 2
and generate Xn as iid exponential random variables and to obtain the cor-
responding Yn , we need to generate a Bernoulli variable, say Bn with mean
(1 − aρ), and another exponential random variable Zn with mean λy and
obtain Yn = aXn + Bn Zn . Negatively correlated X and Y can be similarly
generated using equation 7.
We wish to emphasize that the distributions derived here are different
from the bivariate exponential distribution of Marshall and Olkin (1967a,b)
and other bivariate exponential models described in Basu (1988). In par-
ticular, the bivariate exponential described here does not satisfy the lack of
memory property even though the marginals are exponentials.
Acknowledgement. The research of the first and second authors was
supported in part by DST grant No. III.5(20)/98-ET.

References

Arnold B.C. and Strauss D.J. (1991). Bivariate distributions with conditionals in
prescribed exponential families. J. Roy. Stat. Soc. Series B, 53, 365-375.
Barlow, R.E. and Proschan, F. (1975). Statistical Theory of Reliability and Life
Testing: Probability Models. Hall. Rinehart and Winston, New York.
166 srikanth k. iyer, d. manjunath, r. manivasakan

Basu A.P. (1988). Multivariate exponential distributions and their applications in re-
liability. In: Krishnaiah, P.R. and Rao, C.R., eds., Handbook of Statistics, Vol. 7.
Elsevier Science Publishers B.V., 467-477.
Block, H.W. and Basu, A.P. (1974). A continuous bivariate exponential extension, J.
Amer. Statist. Assoc., 69, 1031-1037.
Gaver, D.P. and Lewis, P.A.W (1980). First order autoregressive gamma sequences
and point processes, Adv. Appl. Probab., 12, 727-745.
Marshall, A.W. and Olkin, I. (1967a). A multivariate exponential distribution, J.
Amer. Statist. Assoc., 62, 30-44.
Marshall, A.W. and Olkin, I. (1967b). A generalized bivariate exponential distribu-
tion. J. Appl. Probab., 4, 291-302.

D. Manjunath, R. Manivasakan Srikanth K. Iyer

Department of Electrical Engg. Department of Mathematics
Indian Institute of Technology Indian Institute of Technology
Mumbai, 400076, India Kanpur, UP 208016 India
E-mail: [email protected] E-mail: [email protected]