Sum of independent exponentials

Lemma 1. Let (Xi )i=1...n , n ≥ 2, be independent exponential random variables with pairwise distinct
respective parameters λi . Then the density of their sum is
" n # n
Y X e−λj x
(1) fX1 +X2 +···+Xn (x) = λi Qn , x > 0.
i=1 j=1 (λk − λj )
k6=j
k=1

Remark. I once (in 2005, to be more precise) thought this stuff would be part of some research-related
arguments, but I ended up not using it. Later on I realized it’s actually Problem 12 of Chapter I in
Feller: An Introduction to Probability Theory and its Applications, Volume II. And recently I have
read about it, together with further references, in “Notes on the sum and maximum of independent
exponentially distributed random variables with different scale parameters” by Markus Bibinger under
http://arxiv.org/abs/1307.3945.

Proof. First we compute the convolutions needed in the proof.

Zx
−ax −bx e(a−b)x − 1 e−bx − e−ax
e ∗e = e−a(x−u) e−bu du = e−ax = .
a−b a−b
0

For n = 2,

e−λ2 x − e−λ1 x
 −λ1 x
e−λ2 x

e
fX1 +X2 (x) = fX1 (x) ∗ fX2 (x) = λ1 λ2 = λ1 λ2 + ,
λ1 − λ2 λ2 − λ1 λ1 − λ2

in accordance to (1). Now inductively, fix n ≥ 3, and assume the statement is true for n − 1. Then
"n−1 # n−1
Y X e−λj x
fX1 +X2 +···+Xn (x) = fX1 +X2 +···+Xn−1 (x) ∗ fXn (x) = λi n−1
∗ fXn (x)
Q
i=1 j=1 (λk − λj )
k6=j
k=1
" n
# n−1 " n # "n−1 n−1
#
Y X e−λn x − e−λj x Y X e−λj x X e−λn x
= λi n−1
= λi n
Q − n
Q .
(λk − λj ) j=1 (λk − λj )
Q
i=1 j=1 (λj − λn ) (λk − λj ) i=1 j=1
k6=j k6=j k6=j
k=1 k=1 k=1

The proof is done as soon as we show that the coefficient of e−λn x fits the coefficients seen in the sum
of (1), i.e.
n−1
X 1 1
(2) − n
Q = n−1
(λk − λj )
Q
j=1 (λk − λn )
k6=j k=1
k=1

1
or, equivalently,
n
X 1
n
Q = 0.
j=1 (λk − λj )
k6=j
k=1

To this order, we write

n
Q
(λk − λl )
n n k6=l6=j
X 1 X k, l=1
n
Q = Qn
j=1 (λk − λj ) j=1 (λk − λl )
k6=j k6=l
k=1 k, l=1

which is zero if and only if

n
X n
Y
(λk − λl )
j=1 k6=l6=j
k, l=1

is zero. We transform the latter in the following display. The nontrivial steps are changing orders of
λ’s and thus signs in the factors of the products.
n
X n
Y n
X n
Y n
Y
(λk − λl ) = (λk − λl ) (λk − λl )
j=1 k6=l6=j j=1 j6=k6=l6=j k=j6=l
k, l=1 k, l=1 k, l=1
Xn Y n Yn n
Y
2
=± (λk − λl ) (λk − λl ) (λk − λl )
j=1 j6=k>l6=j k=j>l k=j<l
k, l=1 k, l=1 k, l=1
n
X Y n Y n Y n
=± (λk − λl )2 (λk − λl ) (λk − λl ) (−1)n−j =
j=1 j6=k>l6=j j=k>l k>l=j
k, l=1 k, l=1 k, l=1
Yn n
X n
Y
=± (λk − λl ) (λk − λl ) (−1)n−j ,
k>l j=1 j6=k>l6=j
k, l=1 k, l=1

which is zero if and only if

n
X n
Y
(3) (λk − λl ) (−1)j
j=1 j6=k>l6=j
k, l=1

is zero. Notice that the product here is a Vandermonde determinant of the form
 
1 λ
 1 λ21 ··· λ1n−2 
λ22 λ2n−2 
 
1 λ 2 ···

. .. .. .. .. 
 .. . . . . 

 
n−2 ,
1 λj−1
 λ2j−1 ··· λj−1 
n−2 
λ2j+1
 
1 λj+1 ··· λj+1 

. .. .. .. .. 
 .. . . . . 

 
1 λ n λ2n ··· n−2 
λn

2
and hence (3) is nothing but the expansion of the determinant
 
1
 1 λ1 λ21 ··· λ1n−2 
λ22 λ2n−2 
 
1 1 λ2 ···

. .. .. .. .. .. 
 .. . . . . . 

 
1 1 λn λ2n ··· λn−2 
n

w.r.t. its second column. As this determinant is zero, so is (3) and thus (2) is proven.