Acta Sci. Math. (Szeged) 18 (1957), pp. 2328.

ON THE COMPOUND POISSON DISTRIBUTION

Andr
as Prekopa (Budapest)
Received: March 1, 1957
A probability distribution is called a compound Poisson distribution if its characteristic
function can be represented in the form



 0
iux
iux
(e 1) dM (x) +
(e 1) dN (x) ,
(1)
(u) = exp iu +

where is a constant, M (x) and N (x) are defined on the intervals (, 0) and (0, ),
respectively, both are monotone non-decreasing, M () = N () = 0, further the integrals


0

x dM (x),

x dN (x)

exist. We shall prove that under certain conditions we obtain (1) as a limit distribution
of double sequences of independent and infinitesimal random variables and apply this
theorem to stochastic processes with independent increments.
Theorem 1 Let n1 , n2 , . . . , nkn (n = 1, 2, . . .) be a double sequence of random variables. Suppose that the random variables in each row are independent, they are infinitesimal, i.e. for every > 0
lim max P(|nk | > ) = 0,
n 1kkn

finally, there exists a finite-valued, non-negative random variable such that

kn


|nk |

(n = 1, 2, . . .)

k=1

with probability 1. (This last condition means that the sums of the absolute values of
the sample summands are uniformly bounded.) Suppose, moreover, that the sequence of
probability distributions of the variables
n = n1 + n2 + + nkn
converges to a limiting distribution. Then this is a compound Poisson distribution.
1

Proof. Let us define the functions f + (x), f (x) as follows:



x if x 0,
0 if x 0,
+

f (x) =
f (x) =
0 if x < 0,
x if x < 0.
Clearly, f + (x)f (x) 0 and
n+

kn


f + (nk )

k=1
kn


n =

(n = 1, 2, . . .),

f (nk )

(n = 1, 2, . . .),

k=1

with probability 1. Hence it follows that for every K > 0 the relations
P(n+ K) P( K)
P(n

(n = 1, 2, . . .),

K) P( K)

(n = 1, 2, . . .)

hold. This imply that the distributions of the sequences n+ and n are compact sets. Let
Fn+ (x) and Fn (x) denote the distribution functions of the variables n+ and n , respectively. Let us choose a sequence of integers n1 , n2 , . . . for which
lim Fn+i (x) = F + (x),

(2)

lim Fni (x) = F (x),

(where F + (x) and F (x) are distribution functions) at every point of continuity of the latters. Let be a positive number such that the functions F + (x) and F (x) are continuous
at and , respectively. Since the random variables in the double sequences
f + (n1 ), f + (n2 ), . . . , f + (nkn ),
f (n1 ), f (n2 ), . . . , f (nkn )
are infinitesimal and independent in each row, moreover the relations (2) hold, we conclude
+

(x) = P(f + (nk ) < x), Fnk

(x) = P(f (nk ) < x), then the sequences
that if Fnk
kni 

k=1 0<x<

kni 


x dFn+i k (x),

k=1

<x<0

x dFni k (x)

are convergent (see [1] 25, Theorem 4, Remark). This implies that
lim

kni 

k=1

0<x<

2

x dFn+i k (x)

= lim

kni 

k=1

<x<0

2
x dFni k (x) = 0.

Thus if
+
nk (t) =

+
eitx dFnk
(x),

nk (t) =

then from the inequality






itx


(e 1) dG(x) |t|


eitx dFnk
(x),


|x| dG(x) + 2

|x|<

dG(x),
|x|

valid for every distribution function G(x) and every > 0, it follows (using Theorem 4 of
[1] 25) that
kni
kni


+
2
2
|ni k (t) 1| = lim
|
lim
ni k (t) 1| = 0.
i

k=1

k=1

Hence the conditions of Theorem 2 of [2] are fulfilled and thus the variables n+i and ni
are asymptotically independent, i.e.
(3)

lim P(n+i < x, ni < y) = F + (x)F (y).

Let F (x) denote the limiting distribution of the random variables n . Since n = n+ + n ,
we get from (3)
(4)

F (x) = F + (x) F (x).

evys formula
The laws F (x), F + (x), F (x) are infinitely divisible. In L






0

iux
iux
2 u2
iux
iux
+
e 1
dM (x) +
e 1
dN (x)
iu
2
1 + x2
1 + x2

0
there correspond to F (x), F + (x) and F (x) constants and functions, which we denote by
 , 1 , 2 ; 2 , 12 , 22 ; M (x), M1 (x), M2 (x); N (x), N1 (x), N2 (x), respectively. According
to (4)
 = 1 + 2 ,

2 = 12 + 22 ,

M (x) = M1 (x) + M2 (x),

N (x) = N1 (x) + N2 (x).

If 2 > 0, then at least one of 12 and 22 is positive too. This is, however, impossible,
since F + (x) = 0 if x 0 and F (x) = 1 if x > 0.
We have therefore only to prove that the integrals
 1
 0
x dM (x),
x dN (x)
1

exist. We prove the existence of the second integral, the existence of the first one can be
proved similarly. We know that if 1 is a point of continuity of N (x), then


(5)
0

xd

kni

k=1

Fni k (x)

converges ([1], 25, Theorem 4) hence it is bounded. If


0

x dN (x) =

then we can choose such a number (0 < < 1 ) that

 1
x dN (x) > L,
(6)

where L is the upper bound of the terms in the sequence (5) and N (x) is continuous at the
point . But we know from the limiting distribution theorems (cf. [1] 25, Theorem 4)
that
kni

(Fni k (x) 1) = N (x)
(x > 0)
lim
i

k=1

at every point of continuity of N (x) whence


(7)

lim

xd

kni



Fni k (x) =

k=1

x dN (x).

Obviously (6) and (7) contain a contradiction.

vys formula the terms
Let us separate in Le

 0
x
x
dM (x),
iu
dN (x)
iu
2
1 + x2
1 + x
0
and unite them with i  u, then we obtain the required form of the limiting distribution.
Thus our theorem is completely proved.

In the sequel we apply our result to the theory of stochastic processes with independent
increments. We say that a stochastic process with independent increments t is weakly
continuous if for every > 0
P(|t+h t | > ) 0
when h 0, uniformly in t. We suppose that P(0 = 0) = 1.
Theorem 2 Let us suppose that the stochastic process with independent increments t
is weakly continuous and its sample functions are of bounded variation with probability 1
in every finite time interval. If (u, t) is the characteristic function of the random variable
t then it has the form



 0
(eiux 1) dM (x, t) +
(eiux 1) dN (x, t) ,
(8)
(u, t) = exp i(t)u +

where (t) is a continuous function of bounded variation in every finite time interval,
M (x, t) and N (x, t) are continuous functions of the variable t and the integrals



x dM (x, t),

1
0

x dN (x, t)

exist for every t.

Proof. According to our suppositions the double sequence of independent random
variables
t , 2t t , . . . , t n1 t
n

satisfies all the conditions of Theorem 1. Moreover, for every n

t =

n

k=1

k t k1 t ,
n

hence we have only to prove the assertion regarding the functions (t), M (x, t), N (x, t).
The continuity in t of these functions follows at once from the weak continuity of the
process t and the convergence theorems of infinitely divisible distributions (see e.g. [1]
Chapter 3).
Now we show that for every T > 0 (t) is of bounded variation in the interval 0 t T .
Let us consider the sequence of subdivisions


k
k1
(n)
T, n T
(k = 1, 2, . . . , 2n ; n = 1, 2, . . .)
Ik =
2n
2
of the interval [0, T ] and let us denote the distribution function of the random variable
(n)
by F (x, Ik ). We know from the limiting distribution theorems that
kn T k1
n T
2

  0
 



k1
k
k1

T +
x d M x, n T M x, n T
2n
2
2







 
k
k1
(N )
x d N x, n T N x, n T
= lim
x dF (x, Ij )
N

2
2
0
(N)
(n) |x|<

k
T
2n

+

Ij

Ik

(cf. [1] 25, Theorem 4), hence



 
2n  



 k T k 1 T 


n
n
2
2

k=1


x dN (x, T )
2 


x dM (x, T )

(9)

+ lim

k=1

(N )

|x|<

|x| dF (x, Ik

).

The boundedness of the sequence on the right-hand side of (9) is a consequence of the fact
that the non-decreasing sequence
2 



N

k=1

k
2N



k1 T 
2N

converges with probability 1, and of Theorem 4 of [1] 25. Since (T ) is continuous, this
implies that it is of bounded variation. Thus Theorem 2 is proved.


References
[1] Gnedenko, B. V. and A. N. Kolmogorov (1949). Predeln
ue raszpredelenija dlja
szumm nezaviszim
uh szlucsajn
uh velicsin, MoszkvaLeningrad.
[2] Pr
ekopa, A. and A. R
enyi (1956). On the independence in the limit of sums depending on the same sequence of independent random variables, Acta Math. Acad. Sci.
Hung., 7, 319326.