1

Chapter I
The Poisson Process

1. Three Ways To Define The Poisson Process

A stochastic process (N (t))t≥0 is said to be a counting process if N (t) counts the total
number of ’events’ that have occurred up to time t. Hence, it must satisfy:
(i) N (t) ≥ 0 for all t ≥ 0.
(ii) N (t) is integer-valued.
(iii) If s < t, then N (s) ≤ N (t).
def
(iv) For s < t, the increment N ((s, t]) = N (t) − N (s) equals the number of events that have
occurred in the interval (s, t].
A counting process is said to have independent increments if the numbers of events that
occur in disjoint time intervals are independent, that is, the family (N (Ik ))1≤k≤n consists
of independent random variables whenever I1 , ..., In forms a collection of pairwise disjoint
intervals. In particular, N (s) is independent of N (s + t) − N (s) for all s, t ≥ 0.
A counting process is said to have stationary increments if the distribution of the number
of events that occur in any interval of time depends only on the length of the time interval.
In other words, the process has stationary increments if the number of events in the interval
(s, s + t], i.e. N ((s, s + t]) has the same distribution as N ((0, t]) for all s, t ≥ 0.
One of the most important types of counting processes is the Poisson process, which can
be defined in various ways.

Definition 1.1. [The Axiomatic Way]. A counting process (N (t))t≥0 is said to be

a Poisson process with rate (or intensity) λ, λ > 0, if:
(PP1) N (0) = 0.
(PP2) The process has independent increments.
(PP3) The number of events in any time interval of length t is Poisson distributed with mean
d
λt. That is, N ((s, t]) = Poi(λt) for all s, t ≥ 0:

(λt)n
P(N ((s, t]) = n) = e−λt , n ∈ N0 .
n!
If λ = 1, then (N (t))t≥0 is also called standard Poisson process.

Note that condition (PP3) implies that (N (t))t≥0 has stationary increments and also that

EN (t) = λt, t ≥ 0,
 2

which explains why λ is called the rate of the process.

In order to determine if an arbitrary counting process is actually a Poisson process, the
conditions (PP1–3) must be shown. Condition (PP1), which simply states that the counting
of events begins at time t = 0, and condition (PP2) can usually be verified directly from our
knowledge of the process. However, it is not at all clear how we could determine validity of
condition (PP3), and for this reason an equivalent definition of a Poisson process would be
useful.
A function f : R → R is said to be o(h) (for h → 0), if

f (h)
lim = 0.
h→0 h

Definition 1.2. [By Infinitesimal Description]. A counting process (N (t))t≥0 is

said to be a Poisson process with rate λ, λ > 0, if:
(PP1) N (0) = 0.
(PP4) The process has stationary and independent increments.
(PP5) P(N (h) = 1) = λh + o(h).
(PP6) P(N (h) ≥ 2) = o(h).

That the processes defined by 1.1 form a subclass of those defined by 1.2 is easily assessed,
but a proof of the reverse inclusion requires some work which we postpone to the end of this
section. However, the essence of the proof is disclosed by the following heuristic argument
based upon the Poisson limit theorem which states that

lim B(n, θn )({k}) = Poi(θ)({k}), k ∈ N0 ,

n→∞

whenever θ, θ1 , θ2 , ... are positive numbers such that nθn → θ, as n → ∞ (+ [1, Satz 29.4]).
d
Plainly, we must only argue that (PP1) and (PP4–6) ensure N (t) = Poi(λt) for all t > 0.
To see this subdivide the interval [0, t] into k equal parts where k is very large. Note that, by
(PP6), the probability of having two or more events in any subinterval goes to 0 as k → ∞.
This follows from

P(2 or more events in any subinterval)

k
X
≤ P(2 or more events in the ith subinterval)
i=1
 
t o(t/k)
= ko = t → 0
k t/k

as k → ∞. Hence, N (t) will (with probability going to 1) just equal the number of subintervals
in which an event occurs. However, by (PP4) this number will have a binomial distribution
with parameters k and pk = λt/k + o(t/k). By letting k → ∞, we thus see that N (t) will have
 3

a Poisson distribution with mean equal to

    
t t o(t/k)
lim k λ + o = λt + lim t = λt.
k→∞ k k k→∞ t/k

The astute reader will have noticed the possibility that the previous two definitions may
only be wishful thinking, in other words, that processes satisfying (PP1–6) do not exist. It
is indeed the merit of our third constructive definition of a Poisson process that it settles the
question of existence in an affirmative way.

Definition 1.3. [The Constructive Way]. A counting process (N (t))t≥0 is said to

be a Poisson process with rate λ, λ > 0, if
X
N (t) = 1(0,t] (Tn ), t ≥ 0, (1.1)
n≥1

for a sequence (Tn )n≥1 having i.i.d. increments Y1 , Y2 , ..., say, with an Exp(λ)-distribution.
The Tn are called jump or arrival epochs and the Yn interarrival or sojourn times associated
with (N (t))t≥0 .

It is clear that any counting process (N (t))t≥0 is completely determined by its associated
sequence of jump epochs (Tn )n≥1 via (1.1). Hence, the equivalence of Definitions 1.1 and
d
1.3 follows if one can show that in the case of i.i.d. Y1 , Y2 , ... with Y1 = Exp(λ), and thus
d
Tn = Γ(n, λ) for each n ≥ 1, the conditions (PP1–3) are satisfied. While (PP1) holds trivially
true, we note for (PP3) that

P(N (t) = n) = P(Tn ≤ t < Tn+1 )

= P(Tn ≤ t < Tn + Yn+1 )
Z t
= P(Yn+1 > t − s) Γ(n, λ)(ds)
0
Z t
λn sn−1 −λs
= e−λ(t−s) e ds
0 (n − 1)!
Z t
λn
= e−λt sn−1 ds
(n − 1)! 0
(λt)n −λt
= e
n!
d
for each t > 0 and n ∈ N (⇒ P(N (t) = 0) = e−λt ). This shows N (t) = Poi(λt). Finally,
it remains to argue that (N (t))t≥0 has independent increments (condition (PP2)). The key
to this is provided by the following lemma hinging on the lack of memory property of the
exponential distribution. We state it without proof here.

Lemma 1.4. If (Tn )n≥1 has independent increments which are exponentially distributed
 4

with parameter λ > 0, then the sequence

def
Z (t) = (TN (t)+1 − t, TN (t)+2 , TN (t)+3 , ...)

is independent of (N (t), T1 , ..., TN (t) ) and distributed as Z (0) = (Tn )n≥1 for every t ≥ 0.

Now, since (N (s + t) − N (s))t≥0 = H(Z (s)) for some measurable function H and all
s ≥ 0, Lemma 1.4 implies the independence of N (s) and (N (s + t) − N (s))t≥0 , in particular
of N (s) and N (t2 ) − N (t1 ) for all 0 ≤ s ≤ t1 ≤ t2 . The reader is invited to complete this
argument to conclude (PP2).
Let us further note here that, given a counting process satisfying (PP1–3), the distribution
of the first jump epoch T1 follows immediately from

P(T1 > t) = P(N (t) = 0) = e−λt

for all t > 0.

Finally, we will show now that Definition 1.2 does indeed imply Definition 1.1.

Proof of ”1.2 ⇒ 1.1”. Assuming (PP1) and (PP4–6), the task is to verify (PP3), i.e.
d
N (t) = Poi(λt) for each t > 0. Put
def
Pn (t) = P(N (t) = n)

and start by considering P0 (t). We derive a differential equation for P0 (t) in the following
manner: For t ≥ 0 and h > 0, we have

P0 (t + h) = P(N (t + h) = 0)
= P(N (t) = 0, N (t + h) − N (t) = 0)
= P(N (t) = 0) P(N (t + h) − N (t) = 0) (1.2)
= P0 (t)P0 (h)
= P0 (t)[1 − λh + o(h)]

where the final three equations follow from (PP4) and the fact that (PP5) and (PP6) give
P0 (h) = P(N (h) = 0) = 1 − λh + o(h). Notice that the latter together with P0 (t + h) =
P0 (t)P0 (h) ensures P0 (t) > 0 for all t > 0. Replacing t with t − h in (1.2), we also have

P0 (t) = P0 (t − h)[1 − λh + o(h)]. (1.3)

It follows that P0 (t) is continuous, as P0 (t ± h) → P0 (t) for h → 0. But (1.2) and (1.3) further
yield
P0 (t + h) − P0 (t) o(h)
= −λP0 (t) +
h h
as well as
P0 (t − h) − P0 (t) o(h)
= −λP0 (t − h) + .
−h h
 5

Again, by letting h → 0 and using the continuity of P0 (t), we infer

P00 (t) = −λP0 (t)

or
P00 (t)
= −λ,
P0 (t)
which implies, by integration,
log P0 (t) = −λt + c
or
P0 (t) = Ke−λt .
Since P0 (0) = P(N (0) = 0) = 1, we arrive at

P0 (t) = e−λt , t ≥ 0. (1.4)

Turning to the case n ≥ 1, we begin by noting that

Pn (t + h) = P(N (t + h) = n)
= P(N (t) = n, N (t + h) − N (t) = 0)
+ P(N (t) = n − 1, N (t + h) − N (t) = 1)
+ P(N (t + h) = n, N (t + h) − N (t) ≥ 2).

By (PP6), the last term in the above is o(h); hence, by using (PP4) and (PP5), we obtain

Pn (t + h) = Pn (t)P0 (h) + Pn−1 (t)P1 (h) + o(h)

(1.5)
= (1 − λh)Pn (t) + λhPn−1 (t) + o(h).

This and the same identity, but with t replaced by t − h, shows the continuity of Pn (t) by an
inductive argument. Rewriting (1.5) as

Pn (t + h) − Pn (t) o(h)
= −λPn (t) + λPn−1 (t) +
h h
and further using the corresponding equation with t − h in place of t, i.e.
Pn (t − h) − Pn (t) o(h)
= −λPn (t − h) + λPn−1 (t − h) + ,
−h h
we obtain upon letting h tend to 0

Pn0 (t) = −λPn (t) + λPn−1 (t)

or, equivalently,
eλt [Pn0 (t) + λPn (t)] = eλt λPn−1 (t).
Hence,
d λt
[e Pn (t)] = eλt λPn−1 (t). (1.6)
dt
 6

Now use mathematical induction over n, the hypothesis being Pn−1 (t) = e−λt (λt)n−1 /(n − 1)!,
to infer from (1.6)
d λt λ(λt)n−1
[e Pn (t)] =
dt (n − 1)!
implying that
(λt)n
eλt Pn (t) = + c.
n!
Finally, since Pn (0) = P(N (0) = n) = 0, we arrive at the desired conclusion

(λt)n
Pn (t) = e−λt , t ≥ 0.
n!
This completes the proof of ”1.2 ⇒ 1.1”. ♦

2. Conditional Distribution Of The Jump Epochs

Suppose we are told that exactly one event of a Poisson process has taken place by time
t, and we are asked to determine the distribution of the time at which the event occured. Since
a Poisson process possesses stationary and independent increments, it seems reasonable that
each interval in [0,t] of equal length should have the same probability of containing the event.
In other words, the time of the event should be uniformly distributed over [0, t]. This is easily
checked since, for s ≤ t,

P(T1 ≤ s, N (t) = 1)
P(T1 ≤ s|N (t) = 1) =
P(N (t) = 1)
P(1 event in (0, s], 0 events in (s, t])
=
P(N (t) = 1)
P(1 event in (0, s]) P(0 events in (s, t])
=
P(N (t) = 1)
λse−λs e−λ(t−s)
=
λte−λt
s
= .
t
This result may be generalized, but before doing so we need to introduce the concept of order
statistics.
Let Y1 , ..., Yn be n random variables. We say that (Y(1) , ..., Y(n) ) is the order statistic
corresponding to (Y1 , ..., Yn ) if Y(k) is the kth smallest among Y1 , ..., Yn , k = 1, ..., n. If the Yi ’s
are i.i.d. continuous random variables with probability density f , then the joint density of the
order statistics is given by
n
Y
f(·) (y1 , ..., yn ) = n! f (yi ) 1S (y1 , ..., yn ), (2.1)
i=1

def
where S = {(s1 , ..., sn ) ∈ Rn : s1 < s2 < ... < sn }. The above follows because
 7

(i) (Y(1) , ..., Y(n) ) will equal (y1 , ..., yn ) ∈ S if (Y1 , ..., Yn ) is equal to any of the n! permutations
of (y1 , ..., yn ) and
(ii) the probability density of (Y1 , ..., Yn ) at (yi1 , ..., yin ) equals f (yi1 )f (yi2 ) · ... · f (yin ) =
Qn
i=1 f (yi ) when (i1 , ..., in ) is a permutation of (1, ..., n).

By treating densities as if they were probabilities, we then indeed obtain

X
P(Y(1) = y1 , ..., Y(n) = yn ) = P(Y1 = yi1 , ..., Yn = yin )
(i1 ,...,in )
X
= f (yi1 ) · ... · f (yin )
(i1 ,...,in )
n
Y
= n! f (yi ), (y1 , ..., yn ) ∈ S,
i=1

where summation is over all permutations (i1 , ..., in ) of (1, ..., n).
If the Yi , i = 1, ..., n, are uniformly distributed over (0, t), then it follows from the above
that the joint density function of the order statistics is given by

n!
f(·) (y1 , ..., yn ) = 1S (y1 , ..., yn ). (2.2)
tn
We are now ready for the following useful theorem.

Theorem 2.1. Given that N (t) = n, the n jump epochs T1 , ..., Tn have the same dis-
tribution as the order statistics corresponding to n independent random variables uniformly
distributed on the interval (0, t).

Proof. We shall compute the conditional density function of T1 , ..., Tn given that N (t) =
n. So let 0 < t1 < ... < tn < tn+1 = t and let hi be small enough so that ti + hi < ti+1 for
i = 1, ..., n. Now,

P(ti < Ti ≤ ti + hi , i = 1, ..., n|N (t) = n)

P(exactly 1 event in (ti , ti + hi ], i = 1, ..., n, no events elsewhere in (0, t])
=
P(N (t) = n)
λh1 e−λh1 · ... · λhn e−λhn e−λ(t−h1 −h2 −...−hn )
=
e−λt (λt)n /n!
n!
= n h1 · h2 · ... · hn .
t
Consequently,
P(ti < Ti ≤ ti + hi , i = 1, ..., n|N (t) = n) n!
= n,
h1 · h2 · ... · hn t
and by letting hi → 0, we obtain that the conditional density of T1 , ..., Tn given that N (t) = n
is
n!
f(·) (t1 , ..., tn ) = , 0 < t1 < ... < tn ,
tn
 8

which completes the proof. ♦

Before turning to an example, let us point out that the above result suggests the following
efficient way of simulating a Poisson process on a time interval [0, t]:
(i) Generate a random number N having a Poisson distribution with mean λt.
(ii) If N = n ≥ 1, then generate n random numbers U1 , ..., Un with a uniform distribution on
(0, 1) and choose
def
(T1 , ..., Tn ) = (tU(1) , ..., tU(n) ) (2.3)
as the arrival times in (0, t).

Example 2.2. Suppose that travelers arrive at a train depot in accordance with a
Poisson process with rate λ. If the train departs at time t, let us compute the expected sum
PN (t)
of the waiting times of travelers arriving in (0, t). That is, we want E[ i=1 (t − Ti )] where Ti
is the arrival time of the ith traveler. Conditioning on N (t) yields
" N (t)  # " n  #
X  X 
E (t − Ti )N (t) = n = E (t − Ti )N (t) = n
i=1 i=1
" n  #
X 
= nt − E Ti N (t) = n .

i=1

Now if we let U1 , ..., Un be independent random variables with a uniform distribution on (0, 1),
then
" n  # " n #
X  X
E Ti N (t) = n = E
 tU(i) (by Theorem 2.1, + (2.3))
i=1 i=1
" n # n n
!
X X X
= tE Ui since U(i) = Ui
i=1 i=1 i=1
nt
= .
2
Hence,
" N (t)  #
X  nt nt
E (t − Ti )N (t) = n = nt −
 =
i=1
2 2
and " N (t) #
X t λt2
E (t − Ti ) = EN (t) = . ♠
i=1
2 2

Tagging. As an important application of Theorem 2.1 suppose that each event of a

Poisson process with rate λ is classified (”tagged”) as being either a type I or type II event,
and suppose that the probability of an event being classified as type I depends on the time at
which it occurs. Specifically, suppose that if an event occurs at time s, then, independently of
 9

all else, it is classified as being a type I event with probability P (s) and a type II event with
probability 1 − P (s). By using Theorem 2.1 we can prove the following proposition.

Proposition 2.3. If Ni (t) represents the number of type i events that occur by time
t, i = 1, 2, then N1 (t) and N2 (t) are independent Poisson random variables having respective
means λpt and λ(1 − p)t, where
Z t
1
def
p = P (s) ds.
t 0

Proof. As ususal, denote by T1 , T2 , ... the successive jump epochs of (N (t))t≥0 and let
I1 , I2 , ... be Bernoulli variables such that Ik equals 1 or 0 depending on whether the kth occur-
ring event is classified as a type I or type II event. In accordance with the above description
the Ik are conditionally independent given (Tn )n≥1 , and

P(Ik = 1|T1 = t1 , T2 = t2 , ...) = P(Ik = 1|Tk = tk ) = P (tk )

for 0 < t1 < t2 < ... Fix any n ≥ 1 and let (Tk∗ , Ik∗ )1≤k≤n be a random shuffle of (Tk , Ik )1≤k≤n .
Then
P(I1∗ = i1 ,..., In∗ = in |T1∗ = t1 , ..., Tn∗ = tn )
1 X
= P(Iπ(1) = i1 , ..., Iπ(n) = in |Tπ(1) = t1 , ..., Tπ(n) = tn )
n! π
n
1 XY (2.4)
= P(Iπ(k) = ik |Tπ(k) = tk )
n! π
k=1
n
Y
= [P (tk )ik (1 − P (tk ))1−ik ]
k=1

for any i1 , ..., in ∈ {0, 1}, where the summation is over all permutations π of 1, ...n. So the Ik∗ ,
k = 1, ..., n, are conditionally independent given T1∗ , ..., Tn∗ , and

P(Ik∗ = 1|T1∗ = t1 , ..., Tn∗ = tn ) = P(Ik∗ = 1|Tk∗ = tk ) = P (tk )

for each k = 1, ..., n. We now prove that I1∗ , ..., In∗ conditioned upon N (t) = n are i.i.d. with
P(Ik∗ = 1|N (t) = n) = p. By Theorem 2.1, T1∗ , ..., Tn∗ conditioned upon N (t) = n are i.i.d. with
a uniform distribution on (0, t). By combining this with (2.4), we infer for i1 , ..., in ∈ {0, 1}

P(I1∗ = i1 ,..., In∗ = in |N (t) = n)

= P(I1 = i1 , ..., In = in |N (t) = n)
Z
= P(I1∗ = i1 , ..., In∗ = in |T1∗ = t1 , ..., Tn∗ = tn )
(0,t)n
× P(T1∗ ∈ dt1 , ..., Tn∗ ∈ dtn |N (t) = n)
n
1 t
Z Z tY
P (tj )ij (1 − P (tj ))1−ij dtn . . . dt1
 
= n ...
t 0 0 j=1
 10
" Z #s " #n−s
1 t 1 t
Z
def
= P (s) ds 1− P (s) ds (s = i1 + ... + in )
t 0 t 0
= ps (1 − p)n−s

which proves the above claim. But this result in combination with the observation that N1 (t) =
PN (t) PN (t) ∗
k=1 Ik = k=1 Ik implies

P(N1 (t) = m,N2 (t) = n|N (t) = m + n)

= P(N1 (t) = m|N (t) = m + n)
N (t)  !
X
∗

= P Ik = mN (t) = m + n
k=1
 
m+n m
= p (1 − p)n
m

for all m, n ∈ N0 , and so

(λt)m+n
 
m+n m
P(N1 (t) = m, N2 (t) = n) = p (1 − p)n e−λt
m (m + n)!
m m
  
−λpt (λpt) −λ(1−p)t (λ(1 − p)t)
= e e
m! m!

which completes the proof. ♦

The importance of the above proposition, which we will extend to the counting processes
(N1 (t))t≥0 and (N2 (t))t≥0 in Theorem 3.4, is illustrated by the following example.

Example 2.4. The Infinite Server Poisson Queue. Suppose that customers arrive
at a service station in accordance with a Poisson process with rate λ. Upon arrival the customer
is immediately served by one of an infinite number of possible servers, and the service times
are assumed to be independent with a common distribution G.
To compute the joint distribution of the number of customers that have completed their
service and the number of customers that are in service (queue length) at t, call an entering
customer a type I customer if it completes its service by time t and a type II customer if it
does not complete service by time t. Now, if the customer enters at time s, s ≤ t, then it will
be a type I customer if its service time is less than t − s, and since the service time distribution
is G, the probability of this will be G(t − s). Hence,

P (s) = G(t − s), s ≤ t,

and from Proposition 2.3 we obtain that the distribution of N1 (t) – the number of customers
that have completed service by time t – is Poisson with mean
Z t Z t
EN1 (t) = λ G(t − s) ds = λ G(y) dy.
0 0
 11

Similarly, N2 (t), the number of customers being served at time t, is Poisson distributed with
mean Z t
EN2 (t) = λ G(y) dy.
0

Further N1 (t) and N2 (t) are independent. ♠

Example 2.5. Suppose that a device is subject to shocks that occur in accordance with
a Poisson process having rate λ. The ith shock gives rise to a damage Di . The Di , i ≥ 1,
are assumed to be i.i.d. and also to be independent of (N (t))t≥0 , where N (t) is the number of
shock in [0, t]. The damage due to a shock is assumed to decrease exponentially in time. That
is, if a shock causes an initial damage D, then a time t later its damage is De−αt .
If we suppose that the damages are additive, then D(t), the damage at t, can be expressed
as
N (t)
X
D(t) = Di e−α(t−Ti ) ,
i=1

where Ti represents the arrival time of the ith shock. We can determine ED(t) as follows:
" N (t)  #
X
Di e−α(t−Ti ) N (t) = n

E[D(t)|N (t) = n] = E
i=1
" n  #
X
−α(t−Ti ) 

= E Di e N (t) = n
i=1
n
X
= E[Di e−α(t−Ti ) |N (t) = n]
i=1
n
X
E e−α(t−Ti ) |N (t) = n
 
= ED
i=1
" n  #
X
−α(t−Ti ) 

= ED E e N (t) = n
i=1
" n  #
X
−αt αTi 

= ED e E e N (t) = n .
i=1

Now, letting U1 , ..., Un be once again be i.i.d. uniform variables on (0, 1), we obtain by another
appeal to Theorem 2.1
" n  # " n #
X X
αTi  αtU(i)

E e N (t) = n = E e
i=1 i=1
" n
#
X
αtUi
= E e
i=1
Z 1
n αt
= n eαtx dx = (e − 1).
0 αt
 12

Hence,
N (t)
E[D(t)|N (t)] = (1 − e−αt ) ED
αt
and, taking expectations,
λ ED
ED(t) = (1 − e−αt ). ♠
α

Remark. Another approach to obtaining ED(t) is to break up the interval (0, t] into
nonoverlapping intervals of length h and then add the contribution at time t of shocks orig-
inating in these intervals. More specifically, let h be given and define Xi as the sum of the
def
damages at time t of all shocks arriving in the interval Ii = (ih, (i + 1)h], i = 0, 1, ..., [t/h],
where [a] denotes the largest integer less than or equal to a. Then we have the representation
[t/h]
X
D(t) = Xi ,
i=0

and so
[t/h]
X
ED(t) = EXi .
i=0
To compute EXi condition on whether or not a shock arrives in the interval Ii . This yields
(recalling Definition 1.2)
[t/h]
X
ED(t) = (λh E[De−α(t−Li ) ] + o(h)],
i=0

where Li is the arrival time of the shock in the interval Ii . Hence,

" [t/h] #  
X
−α(t−Li ) t
ED(t) = λ ED E he + o(h). (2.5)
i=0
h

But since Li ∈ Ii , it follows upon letting h → 0 that

[t/h] t
1 − e−αt
X Z
−α(t−Li )
he → e−α(t−y) dy =
i=0 0 α

and thus from (2.5) upon letting h → 0

λ ED
ED(t) = (1 − e−αt ).
α
It is worth noting that the above is a more rigorous version of the following argument:
Since the shock occurs in the interval (y, y + dy] with probability λdy and since its damage at
time t will equal e−α(t−y) times its initial damage, it follows that the expected damage at t
from shocks originating in (y, y + dy] is

λ dy ED e−α(t−y) ,
 13

and so Z t
λ ED
ED(t) = λ ED e−α(t−y) dy = (1 − e−αt ). ♠
0 α

2.1. The M/G/1 Busy Period

Consider the queueing system, known as M/G/1, in which customers arrive in accordance
with a Poisson process with rate λ. Upon arrival they either enter service if the server is free
or else they join the queue. The successive service times are independent and identically
distributed according to G, and are also independent of the arrival process. When an arrival
finds the server free, we say that a busy period begins. It ends when there are no longer any
customers in the system. We would like to compute the distribution of the length of a busy
period.
Suppose that a busy period has just begun at some time, which we shall designate as time
0. Let Tk denote the time until k additional customers have arrived. Thus Tk has a Gamma
d
distribution with parameters k, λ [Tk = Γ(k, λ)]. Also let X1 , X2 , ... denote the sequence of
def
service times and put Sk = X1 + ... + Xk . Now the busy period will last a time t and will
consist of n services if, and only if,
(i) Tk ≤ Sk , k = 1, ..., n − 1.
(ii) Sn = t.
(iii) There are n − 1 arrivals in (0, t).
Equation (i) is necessary for, if Tk > Sk , then the kth arrival after the initial customer will
find the system empty of customers and thus the busy period would have ended prior to k + 1
(and thus prior to n) services. The reasoning behind (ii) and (iii) is straightforward and left
to the reader.
Hence, reasoning heuristically (by treating densities as if they were probabilities) we see
from the above that

P(busy period is of length t and consists of n services)

= P(Sn = t, n − 1 arrivals in (0, t), Tk ≤ Sk for k = 1, ..., n − 1)
(2.6)
= P(Tk ≤ Sk for k = 1, ..., n − 1|n − 1 arrivals in (0, t), Sn = t)
× P(n − 1 arrivals in (0, t), Sn = t).

Now the arrival process is independent of the service times and thus

(λt)n−1
P(n − 1 arrivals in (0, t), Sn = t) = e−λt Gn (dt), (2.7)
(n − 1)!

where Gn is the n-fold convolution of G with itself (the distribution of Sn ). In addition, we have
from Theorem 2.1 that, given n − 1 arrivals in (0, t), the ordered arrival times are distributed
as the ordered values of a set of n − 1 independent uniform(0, t) random variables. Hence,
 14

using this fact along with (2.6) and (2.7) yields

P(busy period is of length t and consists of n services)

(λt)n−1 (2.8)
= e−λt Gn (dt) P(Tk∗ ≤ Sk , k = 1, ..., n − 1|Sn = t),
(n − 1)!

where T1∗ , ..., Tn−1

∗
are independent of X1 , ..., Xn and represent the ordered values of a set of
n − 1 uniform(0, t) random variables.
To compute the remaining probability in (2.8) we need some lemmata. Given i.i.d.
uniform(0, 1) random variables U1 , ..., Un , we shall denote by (Uk:1 , ..., Uk:k ) the order statistic
based on the first k of these variables (1 ≤ k ≤ n). Lemma 2.6 is elementary and its proof is
left as an exercise.

Lemma 2.6. Let S1 , ..., Sn be the partial sums of n i.i.d. nonnegative random variables
X1 , ..., Xn . Then
k
E[Sk |Sn = s] = s, k = 1, ..., n.
n

Lemma 2.7. Let S1 , ..., Sn be as in Lemma 2.6 and U1 , ..., Un be the i.i.d. uniform(0, 1)
random variables that are also independent of S1 , ..., Sn . Then

 1 − s,

if 0 < s ≤ t,
P(Sk ≤ tUn:k for k = 1, ..., n|Sn = s) = t (2.9)
 0, otherwise

for each t > 0.

Proof. The proof is by induction on n. For n = 1 we must compute P(S1 ≤ tU1 |S1 = s).
But  
s s
P(S1 ≤ tU1 |S1 = s) = P U1 ≥ = 1− .
t t
So assume the lemma be true if n is replaced by n − 1 and consider the n case. Since the
result is obvious for s > t, suppose that s ≤ t. To make use of the induction hypothesis we will
compute the left-hand side of (2.9) by conditioning on the values of Sn−1 and Un:n and then
using the quite intuitive fact that

P((Un:1 , ..., Un:n−1 ) ∈ ·|Un:n = u) = P((uUn−1:1 , ..., uUn−1:n−1 ) ∈ ·).

Doing so, we have for y ≤ s

P(Sk ≤ tUn:k , k = 1, ..., n|Sn−1 = y, Un:n = u, Sn = s)

= P(Sk ≤ tUn:k , k = 1, ..., n|Sn−1 = y, Un:n = u, Xn = s − y)
P(Sk ≤ tuUn−1:k , k = 1, ..., n − 1|Sn−1 = y), if s ≤ tu
(
=
0, otherwise.
 15
 y
 1 − , if s ≤ tu,
= tu (induction hypothesis)
0, otherwise.


Hence, for s ≤ tu,

Sn−1
P(Sk ≤ tUn:k , k = 1, ..., n|Sn−1 , Un:n = u, Sn = s) = 1 −
tu
and thus, for s ≤ tu,

P(Sk ≤ tUn:k , k = 1, ..., n|Un:n = u, Sn = s)

  
Sn−1 
= E 1− U n:n = u, Sn = s
tu 
  
Sn−1 
= E 1− Sn = s
tu 
1
= 1− E[Sn−1 |Sn = s]
tu
1 n−1
= 1− s,
tu n
where we have made use of Lemma 2.6 in the above. Taking expectations once more yields

P(Sk ≤ tUn:k , k = 1, ..., n|Sn = s)

 
s n−1 (2.10)
Z
= 1− P(Un:n ∈ du).
(s/t,1) tu n

Now the distribution function of Un:n = max1≤i≤n Ui is given by

P(Un:n ≤ u) = P(Ui ≤ u, i = 1, ..., n) = un , 0 < u < 1,

and its density is therefore

fn (u) = nun−1 1(0,1) (u).
In (2.10), this leads to
 
s n−1
Z
1− P(Un:n ∈ du)
(s/t,1) tu n
Z 1  
s n−1
= 1− nun−1 du
s/t tu n
 n   n−1 
s s s s
= 1− − 1− = 1−
t t t t
and the proof is complete. ♦

Lemma 2.8. Under the same conditions as in Lemma 2.7 it holds true that
1
P(Sk ≤ tUn−1:k for k = 1, ..., n − 1|Sn = t) =
n
for all t > 0.
 16

Proof. To compute the above probability we will make use of Lemma 2.7 by conditioning
on Sn−1 . Then

P(Sk ≤ tUn−1:k for k = 1, ..., n − 1|Sn−1 = s, Sn = t)

= P(Sk ≤ tUn−1:k for k = 1, ..., n − 1|Sn−1 = s)
 s
 1 − , if 0 < s ≤ t,
= t
0, otherwise


for 0 ≤ s ≤ t. Hence, as Sn−1 ≤ Sn , we have that

P(Sk ≤ tUn−1:k for k = 1, ..., n − 1|Sn = t)

  
Sn−1 
= E 1− Sn = t
t 
n−1
= 1− (by Lemma 2.6)
n

which proves the result. ♦

Returning to the joint distribution of the length of a busy period and the number of
customers served, we must, from (2.8), compute

P(tUn−1:k ≤ Sk for k = 1, ..., n − 1|Sn = t),

as (T1∗ , ..., Tn−1

∗
) = (tUn−1:1 , ..., tUn−1:n−1 ) for some i.i.d. uniform(0,1) variables U1 , ..., Un−1 .
d
Now, since 1 − U = U for any uniform(0,1) variable U , it follows that

d
(1 − Un−1:n−1 , ..., 1 − Un−1:1 ) = (Un−1:1 , ..., Un−1:n−1 ).

Hence, upon replacing Un−1:k by 1 − Un−1:n−k throughout, 1 ≤ k ≤ n − 1, we obtain

P(tUn−1:k ≤ Sk for k = 1, ..., n − 1|Sn = t)

= P(t − tUn−1:n−k ≤ Sk for k = 1, ..., n − 1|Sn = t)
= P(t − tUn−1:n−k ≤ t − (Sn − Sk ) for k = 1, ..., n − 1|Sn = t)
= P(tUn−1:n−k ≥ Sn − Sk for k = 1, ..., n − 1|Sn = t)
= P(tUn−1:n−k ≥ Sn−k for k = 1, ..., n − 1|Sn = t)
= P(tUn−1:k ≥ Sk for k = 1, ..., n − 1|Sn = t)

where the next-to-last equality follows because the conditional laws of (X1 , ..., Xn ) and (Xn , ...,
X1 ) given Sn = t coincide, and so any probability statement involving the Xi ’s and further
random variables independent of X1 , ..., Xn remains valid if X1 is replaced by Xn , X2 by
 17

Xn−1 ,...,Xk by Xn−k , ...,Xn by X1 . Hence, we see that

P(tUn−1:k ≤ Sk for k = 1, ..., n − 1|Sn = t)

= P(tUn−1:k ≥ Sk for k = 1, ..., n − 1|Sn = t)
1
= (from Lemma 2.8).
n
Now, from (2.6), if we let
def
B(t, n) = P(busy period is of length ≤ t, n customers served in a busy period),

then
d (λt)n−1
B(t, n) = e−λt Gn (dt)
dt n!
or Z t
(λt)n−1
B(t, n) = e−λt Gn (dt).
0 n!
def P
The distribution function of the length of a busy period, call it B(t) = n≥1 B(t, n), is then
given by
t
XZ (λt)n−1
B(t) = e−λt Gn (dt). ♠
0 n!
n≥1

3. The Nonhomogeneous Poisson Process

We will now generalize the Poisson process by allowing the arrival rate to be time-depen-
dent. Again we will provide a number of definitions that focus on different characterizing as-
pects of this type of process.

Definition 3.1. [The Axiomatic Way] A counting process (N (t))t≥0 is said to be a

nonstationary or nonhomogeneous Poisson process with rate (or intensity) function λ(t), t ≥ 0,
if:
(NPP1) N (0) = 0.
(NPP2) The process has independent increments.
(NPP3) The number of events in any time interval (s, t] is Poisson distributed with mean
Rt d def R t
s
λ(x) dx, i.e. N ((s, t]) = Poi(m(t) − m(s)), where m(t) = 0 λ(x) dx is the cumu-
lative rate function.

Plainly, condition (NPP3) states that (N (t))t≥0 does not have stationary increments
unless λ(t) ≡ λ for some λ > 0. It should further be understood from this condition that the
Rt
rate function λ(t) is supposed to be nonnegative and locally integrable, i.e. 0 λ(x) dx < ∞ for
all t > 0. Note that the function
Z t
m(t) = λ(x) dx, t≥0
0
 18

defines a locally finite measure ν on [0, ∞) via

def
ν((s, t]) = m(t) − m(s), 0 ≤ s < t < ∞,

which is usually called the intensity measure of the process. In the homogeneous case λ(t) ≡
λ > 0 it obviously equals λ times Lebesgue measure on [0, ∞).
Our second definition of a nonhomogeneous Poisson process provides an infinitesimal de-
scription and as such must impose the additional condition that the rate function be continuous.
It is therefore more restrictive than the previous one.

Definition 3.2. [By Infinitesimal Description]. A counting process (N (t))t≥0 is

said to be a nonhomogeneous Poisson process with continuous rate function λ(t), t ≥ 0, if:
(NPP1) N (0) = 0.
(NPP2) The process has independent increments.
(NPP4) P(N (t + h) − N (t) = 1) = λ(t)h + o(h).
(NPP5) P(N (t + h) − N (t) ≥ 2) = o(h).

As in the homogeneous case, it is straightforward to assess that processes defined by 3.1

(with continuous rate function) form a subclass of those defined by 3.2. For the converse, more
has to be done but follows similar lines as in the homogeneous case.

Proof of ”3.2 ⇒ 3.1” when λ(t) is continuous. Assuming (NPP1,2) and (NPP4,5),
d
the task is to verify (NPP3), i.e. N (s + t) − N (s) = Poi(m(s + t) − m(s)) for each s ≥ 0 and
t > 0. Fix s and define

def
Pn (t) = P(N (s + t) − N (s) = n), n ∈ N0 ,

so that
[m(s + t) − m(s)]n
Pn (t) = e−(m(s+t)−m(s)) , n ∈ N0 (3.1)
n!
must be verified. Start by considering P0 (t) for which a differential equation can be derived in
the following manner: We leave it to the reader as an exercise to show that

P(N (s + t) − N (s) = 0) > 0 for all 0 ≤ s < t < ∞.

For h > 0, we infer with the help of (NPP2) and (NPP4,5) that

P0 (t + h) = P(N (s + t + h) − N (s) = 0)
= P(N (s + t) − N (s) = 0, N (s + t + h) − N (s + t) = 0)
(3.2)
= P(N (s + t) − N (s) = 0) P(N (s + t + h) − N (s + t) = 0)
= P0 (t)[1 − λ(s + t)h + o(h)]
 19

and thereupon
P0 (t + h) − P0 (t)
lim = −λ(s + t)P0 (t).
h↓0 h
Replacing s with s − h in (3.2), we also have

P0 (t) = P0 (t − h)[1 − λ(s + t − h)h + o(h)] (3.3)

and thus see that P0 (t − h) → P0 (t) as h ↓ 0. By combining this with the continuity of λ(t),
we further infer from (3.3) that

P0 (t − h) − P0 (t) o(h)
lim = lim −λ(s + t − h)P0 (t − h) + = −λ(s + t)P0 (t).
h↓0 −h h↓0 h

Consequently, P0 (t) is differentiable and satisfies

P00 (t) = −λ(s + t)P0 (t)

or (recalling that P0 (t) > 0 for all t ≥ 0)

Z t
log P0 (t) = − λ(s + u) du
0

or
P0 (t) = e−(m(s+t)−m(s)) .

The remainder of the verification of (3.1) follows similarly and is left as an exercise. ♦

A particularly quick way of introducing the nonhomogeneous Poisson process, which at

the same time settles the question of its existence, is by changing the time scale of a standard
Poisson process.

Definition 3.3. [The Constructive Way: Time Change]. A counting process

(N (t))t≥0 is said to be a nonhomogeneous Poisson process with rate function λ(t), t ≥ 0, if

N (t) = N
b (m(t)), t ≥ 0, (3.4)

for a standard Poisson process (N

b (t))t≥0 .

The reader will readily check that Definition 3.3 implies Definition 3.1. Conversely,
if (N (t))t≥0 is a nonhomogeneous Poisson process with cumulative rate function m(t), the
standard Poisson process (Nb (t))t≥0 in (3.4) can be obtained via a time change based on the
pseudo-inverse m−1 (t) of m(t), defined as

def
m−1 (t) = inf{s ≥ 0 : m(s) ≥ t}, t ≥ 0. (3.5)
 20

The continuity of m(t) implies m(m−1 (t)) = t, while m−1 (m(t)) = tmin with tmin being the
minimal s such that m(s) = m(t). Since m−1 (t) is nondecreasing and

d
b ((s, t]) = N ((m−1 (s), m−1 (t)]) =
N Poi(m(m−1 (t)) − m(m−1 (s))) = Poi(t − s)

b (t) def
for all 0 ≤ s < t < ∞, we infer that N = N (m−1 (t)), t ≥ 0, constitutes a standard Poisson
process, and it obviously also satisfies (3.4).

Tagging revisited. Suppose we are given a homogeneous Poisson process (N (t))t≥0

with rate λ > 0 and furthermore a rate function λ(t), t ≥ 0, which is bounded by λ. An arrival
when occurring from (N (t))t≥0 at time s is classified (tagged) as a type I event with probability
p(s) = λ(s)/λ and as a type II event with probability 1 − p(s). Let (N1 (t))t≥0 and (N2 (t))t≥0
denote the resulting counting processes of type I and type II events. By Proposition 2.3, N1 (t)
and N2 (t) are independent Poisson variables with means m(t) and λt − m(t), respectively, for
each t ≥ 0. The now natural question whether the whole processes (N1 (t))t≥0 and (N2 (t))t≥0
are independent Poisson processes is affirmatively answered by the following result.

Theorem 3.4. In the situation just described the following assertions hold:
(a) (N1 (t))t≥0 and (N2 (t))t≥0 are nonhomogeneous Poisson processes with rate functions λ(t)
and λ − λ(t), respectively.
(b) (N1 (t))t≥0 and (N2 (t))t≥0 are independent.

Proof. We restrict ourselves to some intuitive explanations. Given n ≥ 2 pairwise

disjoint time intervals, arrivals within these intervals from the proposal process (N (t))t≥0 occur
independently. Moreover, since each classification result depends on the proposal process only
through the particular pertinent arrival that is to be classified, we see that the number of
type I as well as type II events within these time intervals are also independent. But this in
combination with what has been inferred from Proposition 2.1 easily shows the theorem. ♦

Remarks. (a) If λ(t) is merely locally bounded the previous result is still valid in the
following local sense: For any fixed T > 0, let λ = λT > 0 be such that sup0≤t≤T λ(t) ≤ λ.
Then Theorem 3.4 remains true when sampling from a homogeneous Poisson process with rate
λ and using the same classification procedure but restricted to the time interval [0, T ].
(b) One may also generalize Theorem 3.4 — and Proposition 2.1 as well — by allow-
ing more than two types of events. More precisely, let m ≥ 2 and λ1 (t), ..., λm (t) be rate
Pm
functions such that j=1 λj (t) ≡ λ > 0. Any event of a given rate λ Poisson process is clas-
sified as a type k event with probability λk (t)/λ when occurring at time t. Let (Nk (t))t≥0
be the resulting counting process of type k events for k = 1, ..., m. Then the conclusion is
that (N1 (t))t≥0 , ..., (Nm (t))t≥0 are independent nonhomogeneous Poisson processes with rate
functions λ1 (t), ..., λm (t), respectively.
 21

Example 3.5. The Output Process of an Infinite Server Poisson Queue. It

turns out that the output process of the M/G/∞ queue — that is of the infinite server queue
having Poisson arrivals and general service distribution G (+ Example 2.4) — is a nonhomo-
geneous Poisson process having intensity function λ(t) = λG(t). To prove this we shall argue
that
R s+t
(1) the number of departures in (s, s + t] is Poisson distributed with mean λ s
G(x) dx,
and
(2) the numbers of departures in disjoint time intervals are independent.
To prove statement (1), call an arrival type I if it departs in the interval (s, s + t]. Then an
arrival at x will be type I with probability


 G(s + t − y) − G(s − y), if y < s,

def
p(y) = G(s + t − y), if s < y ≤ s + t,


0, if y > s + t.


Hence, from Proposition 2.1 the number of such departures will be Poisson distributed with
mean
Z s+t Z s

Z s+t
λ p(y) dy = λ G(s + t − y) − G(s − y) dy + λ G(s + t − y) dy
0 0 s
Z s+t
= λ G(y) dy.
s

To prove statement (2), let I1 , ..., In , n ≥ 2, be pairwise disjoint time intervals of the form (a, b]
and call an arrival type k if it departs in Ik for k = 1, ..., n, and call it type n + 1 otherwise.
Again, from Proposition 2.1 (and Remark (b) above), it follows that the number of departures
in I1 , ..., In are mutually independent Poisson variables.
Using statements (1) and (2) it is clear that the output (departure) process (N (t))t≥0 ,
say, satisfies (NPP1-3) with λ(t) = λG(t), t ≥ 0. ♠

Example 3.6. Record Values. Let X1 , X2 , ... denote a sequence of i.i.d. nonnegative
random variables with distribution function F , density function f and hazard rate function
def 0
λ(t) = f (t)/F (t), where F = 1 − F . Recall that λ(t) = −(log F )0 (t) = −F (t)/F (t) and
therefore !
Z t
F (t) = exp − λ(s) ds , t > 0.
0

def
We say that a record occurs at time n if Xn > max(X1 , ..., Xn−1 ), where X0 = 0. In this case
Xn is called a record value. We claim that, if N (t) denotes the number of record values less
than or equal to t, then (N (t))t≥0 is a nonhomogeneous Poisson process with rate function
λ(t), t ≥ 0.
 22
def
In order to prove this claim we first give a formal definition of N (t). Define T0 = 0 and
then recursively
def
Tn = inf{k > Tn−1 : Xk > XTn−1 }, n ≥ 1.
Plainly, T1 , T2 , ... are the times at which records occur, called record epochs, and XT1 , XT2 , ...
are the record values. Now

def
X
N (t) = 1(0,t] (XTn ), t ≥ 0. (3.6)
n≥1

In the following, we first consider the case where the Xn have a standard exponential
def
distribution, i.e. F (t) = e−t or λ(t) = 1 for all t ≥ 0. We claim that the Dn = XTn − XTn−1 ,
n ≥ 1, are also i.i.d. standard exponentials and verify this by an induction on n:
For n = 0 it suffices to note that D1 = XT1 − XT0 = X1 . Assume now that D1 , ..., Dn are
i.i.d. standard exponentials (inductive hypothesis). We must verify that Dn+1 is independent of
def
D1 , ..., Dn and also a standard exponential. Put τ = Tn+1 − Tn = inf{k ≥ 1 : XTn +k > XTn }.
Lemma 3.6 below shows that the sequence XTn +1 , XTn +2 , ... forms a copy of X1 , X2 , ... and
is further independent of Tn and D1 , ..., Dn . By using this fact, we infer for any k ≥ 1 and
x1 , ..., xn , t > 0 (treating densities like probabilities)

P(Dn+1 = t, τ = k, D1 = x1 , ..., Dn = xn )
= P(XTn +k − XTn = t, τ = k, D1 = x1 , ..., Dn = xn )
def
= P(XTn +k = t + sn , τ = k, D1 = x1 , ..., Dn = xn ) [sn = x1 + ... + xn ]
= P(XTn +k = t + sn , XTn +j ≤ sn , 1 ≤ j < k, D1 = x1 , ..., Dn = xn )
= e−t−sn (1 − e−sn )k−1 P(D1 = x1 , ..., Dn = xn )
= e−t−sn (1 − e−sn )k−1 P(D1 = x1 ) · ... · P(Dn = xn ) [inductive hypothesis]

and thereby upon summation over k

P(Dn+1 = t, D1 = x1 , ..., Dn = xn )
X
= P(Dn+1 = t, τ = k, D1 = x1 , ..., Dn = xn )
k≥1
X
= e−t−sn P(D1 = x1 ) · ... · P(Dn = xn ) (1 − e−sn )k−1
k≥1
−t
= e P(D1 = x1 ) · ... · P(Dn = xn ).

This clearly proves the assertion.

Having shown that the XTn are sums of i.i.d. standard exponentials, we directly conclude
from Definition 1.3 that (N (t))t≥0 forms a standard Poisson process because, by (3.6), the XTn
are the jump epochs of this counting process.
Turning to the general situation, we first note that, since F is continuous, the sequence
F (X1 ), F (X2 ), ... consists of i.i.d. uniform(0,1) random variables. Indeed, for t ∈ (0, 1) put
 23
def
u = max{u : F (u) = t} which exists by the continuity of F . Then

P(F (Xn ) ≤ t) = P(Xn ≤ u) = F (u) = t

shows that Xn has the asserted distribution. Next put

Z Xn
def
Yn = − log F (Xn ) = λ(s) ds = m(Xn )
0

for n ≥ 1 and check that these variables are i.i.d. standard exponentials. By the strict mono-
tonicity of the transformation t 7→ − log(1 − t), the Tn are also the record epochs for the
sequence (Yn )n≥1 . Consequently, the counting process

b (t) def
X
N = 1(0,t] (YTn ), t≥0
n≥1

constitutes a standard Poisson process by what has been proved in the first step. Finally, as
X X
N (t) = 1(0,− log F (t)] (YTn ) = 1(0,m(t)] (YTn ) = N
b (m(t))
n≥1 n≥1

for all t ≥ 0, we arrive at the asserted result by an appeal to Definition 3.3. ♠