Chapter 3

Poisson process

3.1 The exponential distribution

A continuous random variable X ∼ Exponential(λ ) is said to have an exponential

distribution with parameter λ (> 0), if its p.d.f. and its c.d.f. are given by
(
λ e−λ x , x ≥ 0,
f (x) =
0, x < 0,

and (
1 − e−λ x , x ≥ 0,
F (x) =
0, x < 0,
respectively. Then,
1 2 1
, E X 2 = 2 , Var [X] = 2 .
 
E [X] =
λ λ λ
The moment generating function (m.g.f.) φ (t) is given by
  Z ∞ λ
φ (t) = E etX = etx λ e−λ x dx = , t < λ.
0 λ −t
Definition 3.1. (Memoryless property) A random variable X is said to be memo-
ryless if
P (X > s + t | X > t) = P (X > s) for all s, t ≥ 0.

For example, let X represent the lifetime of some instrument. The probability that
it lives for at least s + t hours given that it has survived t hours is the same as the
initial probability that it lives for at least s hours. In other words, if the instrument is
alive at time t, then the distribution of the remaining amount of lifetime is the same
as the original lifetime distribution, i.e. the instrument does not remember that it has
already been in use for a time t.

45
46 3 Poisson process

Property 3.1. Exponentially distributed random variable is the only memoryless

continuous random variable, which has a monotonically increasing c.d.f.

Proof. Let F(x) = P(X > x). Then, the Memoryless property is equivalent to

F(s + t) = F(s)F(t) (3.1)

for all s,t ≥ 0. Suppose that ξ ≥ 0 is any rational number. Then, ξ = m/n for some
integers m ≥ 0 and n > 0. By (3.1), it is easy to see that

F(ξ ) = F(1/n + 1/n + · · · + 1/n) = [F(1/n)]m ,

F(1) = F(1/n + 1/n + · · · + 1/n) = [F(1/n)]n .

Therefore, it follows that F(ξ ) = [F(1)]ξ . For any real number x, we can find two
sequences of rational numbers {ξ1n } and {ξ2n } such that ξ1n ≤ x ≤ ξ2n , ξ1n → x,
and ξ2n → x. Note that

[F(1)]ξ2n = F(ξ2n ) ≤ F(x) ≤ F(ξ1n ) = [F(1)]ξ1n .

Taking the limiting in both sides, we can obtain that F(x) = [F(1)]x . Finally, the
result holds by taking λ = − log F(1). tu

The memoryless property can be justified by the hazard rate function.

Definition 3.2. The hazard rate function r (t) is defined by

f (t)
r (t) = for t ≥ 0.
1 − F (t)

To interpret r (t), suppose that an item with lifetime X has survived for t hours.
We consider the probability that it does not survive for an additional time dt.

P (X ∈ (t,t + dt) , X > t)

P (X ∈ (t,t + dt) | X > t) =
P (X > t)
P (X ∈ (t,t + dt))
=
P (X > t)
f (t) dt
≈ = r (t) dt.
1 − F (t)

That is, r (t) represents the conditional probability density that a t-year-old item will
fail.
Suppose now the lifetime distribution is exponential. Then, by the memoryless
property, it follows that the distribution of remaining life for a t-year-old item is the
same as for a new item. Hence r (t) should be a constant. This can be checked by

f (t) λ e−λt
r (t) = = −λt = λ .
1 − F (t) e
3.1 The exponential distribution 47

The hazard rate function r (t) also uniquely determines the c.d.f. F (one-to-one
mapping):
d
dt F (t)
r (t) =
1 − F (t)
Z t
⇐⇒ log (1 − F (t)) = − r (t) dt + k
0
 Zt 
k
⇐⇒ 1 − F (t) = e exp − r (t) dt .
0

Letting t = 0 shows that k = 0 and thus

 Zt 
F (t) = 1 − exp − r (t) dt .
0

Example 3.1. Consider a post office that is run by two clerks. Suppose that when
Mr. Smith enters the system he discovers that Mr. Jones is being served by one of
the clerks and Mr. Brown by the other. Suppose also that Mr. Smith is told that his
service will begin as soon as either Jones or Brown leaves. If the amount of time that
a clerk spends with a customer is exponentially distributed with mean 1/λ , what is
the probability that, of the three customers, Mr. Smith is the last to leave the post
office?
Solution. The answer is obtained by this reasoning: Consider the time at which Mr.
Smith first finds a free clerk. At this point either Mr. Jones or Mr. Brown would have
just left and the other one would still be in service. However, by the lack of memory
of the exponential, it follows that the amount of time that the other man (either Jones
or Brown) would still have to spend in the post office is exponentially distributed
with mean 1/λ . That is, it is the same as if he were just starting his service at this
point. Hence, by symmetry, the probability that he finishes before Smith must equal
1
2. t u
Example 3.2. Suppose that the amount of time that a lightbulb works before burning
itself out is exponentially distributed with mean ten hours. Suppose that a person
enters a room in which a lightbulb is burning. If this person desires to work for five
hours, then what is the probability that she will be able to complete her work without
the bulb burning out? What can be said about this probability when the distribution
is not exponential?
Solution. Since the bulb is burning when the person enters the room it follows, by the
memoryless property of the exponential, that its remaining lifetime is exponential
with mean ten. Hence the desired probability is

P (remaining lifetime > 5) = 1 − F (5) = e−5λ = e−1/2 .

However, if the lifetime distribution F is not exponential, then the relevant proba-
bility is
48 3 Poisson process

1 − F (t + 5)
P (lifetime > t + 5 | lifetime > t) = ,
1 − F (t)
where t is the amount of time that the bulb had been in use prior to the person
entering the room. That is, if the distribution is not exponential then additional in-
formation is needed (namely, t) before the desired probability can be calculated.
In fact, it is for this reason, namely, that the distribution of the remaining lifetime
is independent of the amount of time that the object has already survived, that the
assumption of an exponential distribution is so often made. t u

3.2 The Poisson process

Definition 3.3. (Counting process) A stochastic process {N (t) ,t ≥ 0} is said to be

a counting process if N (t) represents the total number of “events” that occur by time
t. Thus, N (t) must satisfy:
1. N (t) ≥ 0
2. N (t) is integer-valued.
3. If s < t, then N (s) ≤ N (t).
4. For s < t, N (t) − N (s) equals the number of events that occur in the interval (s,t].

Some examples of counting processes:

1. N (t) = # of phone calls in the time interval (0,t].
2. N (t) = # of claims in an insurance company in the time interval (0,t].
3. N (t) = # of people who were born in the time interval (0,t].

Definition 3.4. (Independent increments) — # of events that occur in disjoint time

intervals are independent.

Remark 3.1. In other words, N(t4 ) − N(t3 ) is independent of N(t2 ) − N(t1 ) for any
t1 < t2 ≤ t3 < t4 .

Assumption of independent increments is:

• reasonable for Example 1, as phone calls are made independently by individuals.
• not reasonable for Example 3, as the number of people who are alive at t will
affect the number of people who are born in the time interval (t,t + s).

Definition 3.5. Stationary increments — the distribution of the number of events

that occur in any interval of time depends only on the length of the time interval.

Remark 3.2. In other words, N(t2 + s) − N(t1 + s) has the same distribution as
N(t2 ) − N(t1 ) for any t1 < t2 and s > 0.

Definition 3.6. (Poisson process) The counting process {N (t) ,t ≥ 0} is said to be

a Poisson process having rate λ for λ > 0 if
3.2 The Poisson process 49

1. N (0) = 0.
2. The process has independent increments.
3. The number of events in any interval of length t follows Poisson(λt). That is, for
all s, t ≥ 0

(λt)n
P (N (t + s) − N (s) = n) = e−λt , n = 0, 1, . . .
n!
This means that the Poisson process has stationary increments and also that
E [N (t)] = λt.

From the definition, we know that λ = E(N(t))/t, and hence λ , which refers to
# of arrivals per time unit, is called rate (or intensity) of a Poisson process.

Property 3.2. If N(t) is a Poisson process with rate λ , then for any u ≥ 0,
h i −u
E e−uN(t) = e(λt)(e −1) .

(This is the Laplace transform of N(t)).

Proof. Note that

∞
xi
ex = ∑ .
i=0 i!

The proof is left as an exercise. t

u

Usually, conditions 1-2 are easy to be verified, but condition 3 is not. Hence, an
alternative definition of Poisson process is useful.

Definition 3.7. (An alternative definition of the Poisson process) The counting
process {N (t) ;t ≥ 0} is said to be a Poisson process with rate λ , λ > 0 if
1. N (0) = 0.
2. The process has stationary and independent increments.
3. P (N (h) = 1) = λ h + o (h).
4. P (N (h) ≥ 2) = o (h).

Remark 3.3. The function f (·) is said to be o (h) if

f (h)
lim = 0.
h→0 h
Remark 3.4. Condition 3 in Definition 3.7 means that an event occurs in a short time
interval with probability almost proportional to the length of it.
Condition 4 in Definition 3.7 means that there is almost no chance to have more
than one event occurring in a short time interval.

Theorem 3.1. Definitions 3.6 and 3.7 are equivalent.

50 3 Poisson process

Proof. First, we show that Definition 3.7 implies Definition 3.6. As conditions 1
and 2 of Definition 3.7 are the same as those in Definition 3.6, what we need to do
is to prove that for any t, N (t) follows Poisson(λt).
To start, fix u ≥ 0 and let
h i
gu (t) = E e−uN(t)

be the Laplace transform of N(t). In view of Property 3.2, we have to show that
−u −1
gu (t) = e(λt)(e ).

By independent and stationary increments, we have

h i
gu (t + h) = E e−uN(t+h)
h i
= E e−uN(t) e−u(N(t+h)−N(t))
h i h i
= E e−uN(t) E e−u(N(t+h)−N(t))
h i
= gu (t) E e−uN(h) .

Now, conditions 3-4 of Definition 3.7 imply that

P (N (h) = 0) = 1 − λ h + o (h) .

So we have
h i ∞
E e−uN(h) = P (N (h) = 0) + e−u P (N (h) = 1) + ∑ e−u j P (N (h) = j)
j=2
−u
= 1 − λ h + o (h) + e (λ h + o (h)) + o (h)
−u
= 1 − λ h + e λ h + o (h) .

Therefore, we have

gu (t + h) = gu (t) 1 − λ h + e−u λ h + o (h)

implying that

gu (t + h) − gu (t)
o (h)
= gu (t) λ e−u − 1 + .
h h
Letting h → 0 gives the derivative of gu (t) at t

g0u (t) = gu (t) λ e−u − 1

or equivalently,
3.3 Properties of Poisson process 51

g0u (t) d
= log (gu (t)) = λ e−u − 1 .


gu (t) dt
Integrating and using gu (0) = 1 gives

log (gu (t)) = λt e−u − 1

−u
⇐⇒ gu (t) = eλt (e −1) .

Next, we show that Definition 3.6 implies Definition 3.7. By condition 3 of Def-
inition 3.6,
 
P (N (h) = 1) = e−λ h λ h = λ h + λ h e−λ h − 1 = λ h + o (h) ,

and
∞
e−λ h (λ h)n
P (N (h) ≥ 2) = ∑
n=2 n!
= 1 − e−λ h − λ he−λ h
= 1 − (1 − λ h + o (h)) − (λ h + o (h))
= o (h) .

This completes the proof. t

u

3.3 Properties of Poisson process

Proposition 3.1. (Backcasting) If N(t) is a Poisson process with rate λ , the condi-
tional distribution of N(s) given N(t) = n, for s < t, is binomial with parameters n
and p = s/t.

Proof. Consider the conditional p.d.f. of N(s) given N(t) = n:

P(N(s) = k, N(t) = n)
P(N(s) = k | N(t) = n) =
P(N(t) = n)
P(N(s) = k, N(t) − N(s) = n − k)
=
P(N(t) = n)
P(N(s) = k)P(N(t) − N(s) = n − k)
=
P(N(t) = n)
k n−k
e−λ s (λk!s) e−λ (t−s) (λ (t−s))
(n−k)!
= n
e−λt (λt)
     n! 
n s k s n−k
= 1−
k t t
52 3 Poisson process

for k = 0, 1, · · · , n. t
u
The above proposition allows us to backcast the process at unobserved time
points.
Example 3.3. Suppose that the number of goals in a soccer match follows the Pois-
son process with rate 2.7 per 90 mins match. If at the end the outcome of the match
is 3:2, what is the probability that the majority of these goals were scored in the first
half of the match?
Solution. Let N(t) =# of goals scored in t mins. Then, by Proposition 3.1, given that
N(90) = 5, the conditional distribution of N(45) is binomial with n = 5 and p = 0.5.
Hence,
   
5 3 2 5
P(N(45) > 2 | N(90) = 5) = (0.5) (0.5) + (0.5)4 (0.5)1
3 4
 
5
+ (0.5)5 (0.5)0
5
= 0.5.

t
u
Proposition 3.2. (Superposition) Let {N j (t);t ≥ 0}, j = 1, 2, · · · , m, be m indepen-
dent Poisson processes each with corresponding rate λ j . Let

N(t) = N1 (t) + N2 (t) + · · · + Nm (t).

Then, {N(t);t ≥ 0} is a Poisson process with rate λ = ∑mj=1 λ j .

Example 3.4. Consider two typos of insurance claims: car and home. Suppose that
(i) # of car insurance claims: Poisson process with rate 20 per day;
(ii) # of home insurance claims: Poisson process with rate 5 per day;
(iii) The two processes are independent.
Calculate P(more than 800 claims in a 30-day period).
Solution. Let N(t) =# of claims received until day t. Then, by Proposition 3.2,
{N(t);t ≥ 0} is a Poisson process with rate 25 per day. Hence, N(30) follows
Poisson(25 × 30) and

E(N(30)) = Var(N(30)) = 25 × 30 = 750.

Recall the normal approximation with continuity correction of the Poisson dis-
tribution:
P(X ≤ x) ≈ P(Y ≤ x + 1/2),
where X follows Poisson(λ ), and Y follows N(λ , λ ).
By normal approximation, we have
 
800.5 − 750
P(N(30) > 800) = 1 − P(N(30) ≤ 800) ≈ 1 − Φ √ ≈ 0.0329.
750
3.4 Interarrival and waiting time distributions 53

Proposition 3.3. (Decomposition) Let {N(t);t ≥ 0} be a Poisson process with rate

λ . Suppose that each event can be independently classified as type j with probability
p j , for j = 1, 2, · · · , m. Let

N j (t) = # of type j events in [0,t].

Then, {N j (t);t ≥ 0}, j = 1, 2, · · · , m, are independent Poisson processes each with

rate λ j = λ p j .

Example 3.5. Customers arrive at a clothing shop according to a Poisson process

with rate 20 per hour. Suppose that the probability that a customer buys a suit is
p = 0.3. Determine
(i) the expected number of suits sold during an eight-hour business day;
(ii) the probability that 10 or more suits are sold in a period of one hour.

Solution. Let N1 (t) =# of suits sold in t hours. Then, by Proposition 3.3, {N1 (t);t ≥
0} is a Poisson process with rate λ1 , where λ1 = λ p = 20 × 0.3 = 6.
(i) Since N1 (8) follows Poisson(8λ1 ),

E(N1 (8)) = 8 × 6 = 48.

(ii) Since N1 (1) follows Poisson(λ1 ),

9
6k
P(N1 (1) ≥ 10) = 1 − ∑ e−6 = 0.0426.
k=0 k!

t
u

Example 3.6. If immigrants to area A arrive at a Poisson rate of ten per week, and
1
if each immigrant is of English descent with probability 12 , then what is the proba-
bility that no people of English descent will emigrate to area A during the month of
February?

Solution. Let N1 (t) =# of English descent immigrants made in t weeks. By Propo-

sition 3.3, {N1 (t);t ≥ 0} is a Poisson process with rate λ1 , where λ1 = λ p =
10 × 1/12 = 5/6. Then, the desired probability is

P(N1 (4) = 0) = e−4λ1 = e−10/3 = 0.0357,

since N1 (4) follows Poisson(4λ1 ). t

u

3.4 Interarrival and waiting time distributions

Definition 3.8. (Interarrival time) Consider a Poisson process,

T1 — the time of the first event
54 3 Poisson process

Tn — the elapsed time between the (n − 1)th and the nth event
The sequence {Tn ; n = 1, 2, . . .} is called the sequence of interarrival times.

Say, if T1 = 5 and T2 = 10, then the first event of the Poisson process would have
occurred at time 5 and the second at time 15.

Theorem 3.2. {Tn ; n = 1, 2, . . .} is a sequence of i.i.d. Exponential(λ ) random vari-

ables.

Proof. Note that

{T1 > t} ⇔ N (t) = 0.
Then,
P (T1 > t) = P (N (t) = 0) = e−λt .
Hence, T1 follows Exponential(λ ).
Moreover,

P (T2 > t | T1 = s) = P (no events in (s, s + t] | T1 = s)

= P (no events in (s, s + t]) (independent increments)
= e−λt (stationary increments).

So, T2 also follows Exponential(λ ), and T2 is independent of T1 . Repeating the same

argument yields our proposition. Generally,

P (Tn > t | T1 = s1 , . . . , Tn−1 = sn−1 )

!
 n−1   n−1    n−1 

= P N t + ∑ si − N ∑ si = 0N ∑ si = n − 1
i=1 i=1 i=1
−λt
= P (N (t) = 0) = e .

This completes the proof. t

u

Definition 3.9. (Arrival time or waiting time) Consider a Poisson process,

Sn — the arrival time of the nth event, also called the waiting time until the nth
event, is defined by
n
Sn = ∑ Ti , n ≥ 1.
i=1

The sequence {Sn ; n = 1, 2, . . .} is called the sequence of arrive times.

By using the m.g.f., it is easy to see that Sn follows a gamma distribution with
parameters n and λ (denoted as Gamma(n, λ )). Its p.d.f. is given by

(λt)n−1
fSn (t) = λ e−λt , t ≥ 0,
(n − 1)!

and its m.g.f. is

 3.4 Interarrival and waiting time distributions 55

(λ )n
E etSn =
 
, t < λ.
(λ − t)n
Example 3.7. Suppose that people immigrate into a territory at a Poisson rate λ = 1
per day.
(i) What is the expected time until the tenth immigrant arrives?
(ii) What is the probability that the elapsed time between the tenth and the eleventh
arrival exceeds two days?
(iii) What is the probability that the number of arrivals is 2 between day 1 and day 4
and the number of arrivals is 3 between day 3 and day 5?
Solution. (i) Since S10 ∼ Gamma(10, 1), E (S10 ) = 10/1 = 10 days.
(ii) Since T11 ∼ Exponential(1), P (T11 > 2) = e−1×2 = e−2 = 0.1353.
(iii) The desired probability is

P(N(4) = 2, N(5) − N(2) = 3)

2
= ∑ P(N(4) = 2, N(5) − N(2) = 3 | N(4) − N(2) = k)P(N(4) − N(2) = k)
k=0
2
= ∑ P(N(2) = 2 − k, N(5) − N(4) = 3 − k | N(4) − N(2) = k)P(N(4) − N(2) = k)
k=0
2
= ∑ P(N(2) = 2 − k, N(5) − N(4) = 3 − k)P(N(4) − N(2) = k)
k=0
2
= ∑ P(N(2) = 2 − k)P(N(5) − N(4) = 3 − k)P(N(4) − N(2) = k)
k=0
2
(2λ )2−k λ 3−k (2λ )k
= ∑ e−2λ (2 − k)! e−λ (3 − k)! e−2λ k!
k=0
λ5
 
−5λ
=e + 2λ 4 + 2λ 3 = 0.0292.
3
t
u
Example 3.8. (The Coupon Collecting Problem) There are m different types of
coupons. Each time a person collects a type j coupon independently with probabil-
ity p j . Let K denote the number of coupons one needs to have a complete collection
of at least one of each type. Find E [K].
Solution. We model the number of collected coupons by a Poisson process N(t) with
rate λ = 1. Let X be the time at which a complete collection is obtained. Then, it is
easy to see that
K
X = ∑ Ti ,
i=1

where Ti , i = 1, 2, · · · , K are interarrival times of N(t). By using the argument of

conditional expectation, we can show that
56 3 Poisson process

E [X] = E [K] E [Ti ] = E [K] .

Here, we have used the fact that {Ti } is a sequence of i.i.d. Exponential(1) random
variables.
Now, our objective is to calculate E [X]. We say that an event of the Poisson
process N(t) is of type j, 1 ≤ j ≤ m, if the coupon obtained at that time is a type
j coupon. Let N j (t) denote the number of type j coupons collected by time t. It
follows from Proposition 3.3 that {N j (t) ,t ≥ 0} for j = 1, . . . , m are independent
Poisson processes with respective rates λ p j = p j .
Furthermore, let X j1 denote the time of the first event of the jth process. Then,
we know that
X = max X j1 .
1≤ j≤m

Since the X j1 are independent Exponential(p j ) random variables, it follows that

  m
P (X < t) = P max X j1 < t = P X j1 < t for j = 1, . . . , m = ∏ 1 − e−p j t .



1≤ j≤m j=1

Therefore,
Z ∞ Z ∞ m
{1 − ∏ 1 − e−p j t }dt.


E [X] = P (X > t) dt =
0 0 j=1

Here, we have used a fact that for any non-negative random variable X with a c.d.f.
F(·), Z ∞
E[X] = (1 − F(x))dx.
0
t
u

3.5 Conditional distribution of waiting time

In this section, we consider the (joint) conditional distribution of {Sn ; n = 1, 2, . . .}.

Suppose that exactly one event of a Poisson process has taken place by time t. Since
a Poisson process possesses stationary and independent increments it seems rea-
sonable 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) , no events in [s,t])
=
P (N (t) = 1)
3.5 Conditional distribution of waiting time 57

P (1 event in [0, s)) P (no events in [s,t])

=
P (N (t) = 1)
λ se−λ s e−λ (t−s)
=
λte−λt
s
= .
t
For the general case, we want to derive the joint distribution of the n arrival (waiting)
times S1 , S2 , . . . , Sn given N (t) = n, i.e.,

P (S1 < s1 , S2 < s2 , . . . , Sn < sn | N (t) = n) ,

where S1 ≤ S2 ≤ · · · ≤ Sn and S j = T1 + · · · + T j .
To fulfill it, we need to introduce the concept of order statistics. Let Y1 , Y2 , . . ., Yn
be n random variables. Then, Y(1) , Y(2) , . . ., Y(n) are the order statistics corresponding
to Y1 , Y2 , . . ., Yn if Y(k) is the kth smallest value among Y1 , Y2 , . . ., Yn , k = 1, 2, . . . , n.
If Yi , i = 1, . . . , n, are i.i.d. continuous random variables with probability density f ,
then the joint density of the order statistics Y(1) , Y(2) , . . ., Y(n) is given by
n
f (y1 , y2 , . . . , yn ) = n! ∏ f (yi ) , y1 < y2 < · · · < yn .
i=1

Particularly, if {Y1 ,Y2 , · · · ,Yn } is a sequence of i.i.d. U(0,t) random variables,

n!
f (y1 , y2 , . . . , yn ) = , y1 < y2 < · · · < yn .
tn
Theorem 3.3. Given that N (t) = n, the conditional distribution of the arrival times
(S1 , S2 , . . . , Sn ) is the same as that of (U(1) ,U(2) , · · · ,U(n) ), where {U(1) ,U(2) , · · · ,U(n) }
are the ordered statistics of {U1 ,U2 , · · · ,Un }, and {U1 ,U2 , · · · ,Un } is a sequence of
i.i.d. U(0,t) random variables.

Proof. For 0 < s1 < · · · < sn < t, take hi such that si + hi < si+1 for i = 1, 2, · · · , n.
Then,

P (si < Si ≤ si + hi for i = 1, 2, · · · , n | N(t) = n)

P (si < Si ≤ si + hi for i = 1, 2, · · · , n, N(t) = n)
=
P(N(t) = n)
P (one event occurs in (si , si + hi ] for i = 1, 2, · · · , n, no events occur in other places)
=
P(N(t) = n)
[λ h1 e−λ h1 ][λ h2 e−λ h2 ] · · · [λ hn e−λ hn ]e−λ (t−h1 −h2 −···−hn )
=
e−λt (λt)n /n!
n!
= n h1 h2 · · · hn .
t
Hence,
58 3 Poisson process

P (si < Si ≤ si + hi for i = 1, 2, · · · , n | N(t) = n) n!

= n.
h1 h2 · · · hn t
Letting hi → 0, it follows that
n!
f (s1 , s2 , . . . , sn | N(t) = n) = , 0 < s1 < · · · < sn < t.
tn
This completes the proof. t
u
Remark 3.5. Using the similar arguments as before, we can show that the marginal
conditional density of Si given N(t) = n is

si−1
i (t − si )
n−i n!
f (si | N(t) = n) = , 0 < si < t.
t n (n − i)!(i − 1)!

Particularly, it follows that

it
E(Si | N(t) = n) = for i = 1, 2, · · · , n.
n+1
Example 3.9. (Discounted cost of claims) Suppose that the number of insurance
claims made is a Poisson process with rate λ . Usually, claim amounts are i.i.d. ran-
dom variables with mean µ, and are independent of (i) the times at which the claims
are made and (ii) the number of insurance claims made. If the cost of a claim made
at time s is discounted by a factor of e−αs , what is the expected total discounted cost
of all claims made up to time t?
Solution. Let N(t) =# of insurance claims made by time t. Then, {N(t);t ≥ 0} is a
Poisson process with rate λ . Let Ci and Si be the amount and arrival time of the ith
claim. Then, the total discounted cost by time t, denoted by D(t), can be expressed
as
N(t)
D(t) = ∑ Ci e−αSi .
i=1

Since Ci are independent of Si and N(t),

n
E[D(t) | N(t) = n] = ∑ E[Ci ]E[e−αSi | N(t) = n]
i=1
n
= µ ∑ E[e−αSi | N(t) = n]
i=1
n
= µ ∑ E[e−αU(i) ]
i=1
n
= µ ∑ E[e−αUi ]
i=1
n Z t
e−αx 1 − e−αt
=µ∑ dx = nµ .
i=1 0 t αt
3.5 Conditional distribution of waiting time 59

Therefore,

1 − e−αt λ µ(1 − e−αt )

 
E(D(t)) = E(E(D(t) | N(t))) = E N(t)µ = .
αt α
t
u
Example 3.10. (Total Passenger Waiting Time) Suppose that passengers arrive at a
station according to a Poisson process with rate λ . A bus departs at a fixed deter-
ministic time T .
(i) Find the expected total waiting time of all the passengers who get on the bus.
(ii) Suppose that an additional bus is scheduled to depart at time t ∈ (0, T ). When
should this bus depart so that the total passenger waiting time can be minimized?
Solution. (i) Let Si be the arrival time of the ith passenger. Then, the total waiting
time is
N(T )
W= ∑ (T − Si ).
i=1

By using the similar argument as for the previous example, we have

n n
nT
E(W | N(T ) = n) = ∑ [T − E(U(i) )] = ∑ [T − E(Ui )] = .
t=1 t=1 2

Therefore, it follows that

λT2
 
N(T )T
E(W ) = E(E(W | N(T ))) = E = .
2 2

(ii) From (i), the expected total waiting time of all the passengers who get on the
buses can be determined as
!
λt 2 λ (T − t)2 T 2 T2
 
E(W ) = + =λ t− +
2 2 2 4

T
which is minimized at t = . t
u
2
To end this section, we apply Theorem 3.3 to solve some problems in practice.
Suppose that there are k possible types of events and that the probability that an
event is classified as a type i event for i = 1, . . . , k, depends on the time the event
occurs. Suppose that if an event occurs at time s then it will be classified as a type
i event, independently of anything that has previously occurred, with probability
Pi (s) for i = 1, . . . , k, where ∑ki=1 Pi (s) = 1. Using Theorem 3.3, we can prove the
following proposition.
Proposition 3.4. If Ni (t) for i = 1, . . . , k represents the number of type i events oc-
curring by time t, then Ni (t) for i = 1, . . . , k, are independent Poisson random vari-
ables having means
60 3 Poisson process
Z t
E [Ni (t)] = λ Pi (s) ds.
0

Proof. Conditioning on N (t), we have

!
k


P (N1 (t) = n1 , . . . , Nk (t) = nk ) = P N1 (t) = n1 , . . . , Nk (t) = nk N (t) = ∑ ni

i=1
!
k
× P N (t) = ∑ ni .
i=1

Now consider an arbitrary event that occurred in the interval [0,t]. Suppose that it
had occurred at time S, then the probability that it would be a type i event would be
Pi (S). By Theorem 3.3, S is uniformly distributed on (0,t), Hence, this event has
the following probability to be a type i event:
Z t
1
Pi = E(Pi (S) | N(t) = 1) = E[Pi (U(1) )] = Pi (s) ds,
t 0

and it is independent of other type events. It follows that

!
k


P Ni (t) = ni , i = 1, . . . , kN (t) = ∑ ni

i=1

will just equal the multinomial probability of ni type i outcomes for i = 1, . . . , k

when each of ∑ki=1 ni independent trials results in outcome i with probability Pi for
i = 1, . . . , k. That is,
!
∑ki=1 ni ! n1
 k


 n
P N1 (t) = n1 , . . . , Nk (t) = nk N (t) = ∑ ni =
 P1 · · · Pk k .
i=1 n 1 ! · · · nk !

Consequently, since ∑i Pi = 1,

(∑i ni )! n1 n (λt)∑i ni
P (N1 (t) = n1 , . . . , Nk (t) = nk ) = P1 · · · Pk k e−λt
n1 ! · · · nk ! (∑i ni )!
k
= ∏ e−λtPi (λtPi )ni /ni !
i=1

and this completes the proof. t

u

Example 3.11. (Tracking the Number of HIV Infections) There is a relatively long
incubation period from the time when an individual becomes infected with the HIV
virus, which causes AIDS, until the symptoms of the disease appear. As a result, it
is difficult for public health officials to be certain of the number of members of the
population that are infected at any given time. We will now present a first approx-
imation model for this phenomenon, which can be used to obtain a rough estimate
3.5 Conditional distribution of waiting time 61

of the number of infected individuals.

Let us suppose that individuals contract the HIV virus in accordance with a Poisson
process whose rate λ is unknown. Suppose that an individual who becomes infected
at time s will show AIDS symptoms by time t with a probability G(t − s), where
G(·) is known.

Let N1 (t) denote the number of individuals who have shown symptoms of the dis-
ease by time t. Also, let N2 (t) denote the number who are HIV positive but have
not yet shown any symptoms by time t. By Proposition 3.4, N1 (t) and N2 (t) are
independent Poisson random variables with respective means
Z t Z t
E [N1 (t)] = λ G (t − s) ds = λ G (s) ds
0 0

and Z t Z t
E [N2 (t)] = λ [1 − G (t − s)]ds = λ [1 − G(s)]ds.
0 0
Now, if we knew λ , then we could use it to estimate N2 (t), the number of individuals
infected but without any outward symptoms at time t, by its mean value E [N2 (t)].
However, since λ is unknown, we must first estimate it. Now, we will presumably
know the value of E [N1 (t)], and so we can use its known value as an estimate of its
mean E [N1 (t)]. That is, if the number of individuals who have exhibited symptoms
by time t is n1 , then we can estimate that
Z t
n1 ≈ E [N1 (t)] = λ G (s) ds.
0

Therefore, we can estimate λ by the quantity λ̂ given by

.Z t
λ̂ = n1 G (s) ds
0

Using this estimate of λ , we can estimate the number of infected but symptomless
individuals at time t by
Z t Rt
n1 0 [1 − G (s)]ds
estimate of N2 (t) = λ̂ [1 − G (s)]ds = Rt .
0 0 G (s) ds

For example, suppose that G is exponential with mean µ. Then 1 − G (s) = e−s/µ ,
and a simple integration gives that

n1 µ 1 − e−t/µ


estimate of N2 (t) =
.
t − µ 1 − e−t/µ

If we suppose that t = 16 years, µ = 10 years, and n1 = 2200 thousand, then the

estimate of the number of infected but symptomless individuals at time 16 is
62 3 Poisson process

2200 1 − e−1.6


= 218.96.
16 − 10 (1 − e−1.6 )

That is, if the incubation period is exponential with mean 10 years and if the to-
tal number of individuals who have exhibited AIDS symptoms during the first 16
years of the epidemic is 2200 thousand, then we can expect that approximately 219
thousand individuals are HIV positive though symptomless at time 16. t u

3.6 Compound Poisson process

Definition 3.10. (Compound Poisson Process) A stochastic process {X (t) ;t ≥ 0}

is said to be a compound Poisson process if it can be represented as
N(t)
X (t) = ∑ Yi , t ≥ 0,
i=1

where {N (t) ;t ≥ 0} is a Poisson process, and {Yi ; i ≥ 1} is a family of independent

and identically distributed random variables which is also independent of {N (t) ;t ≥
0}.

A compound Poisson process can be viewed as the accumulation of “rewards”

from events occurring according to a Poisson process.

Example 3.12. (Consumer Shopping Behavior)

N(t) = # of customers visiting a supermarket by time t;
Yi = amount spent by the ith customer in the supermarket;
X(t) = total amount spent by customers visiting the supermarket by time t.

Example 3.13. (Stock Price Model)

N(t) = # of transactions in a certain stock by time t;
Yi = change in market price between the (i − 1)th and ith transactions;
X(t) = total price change of the stock up to time t.

Example 3.14. (Collective Risk Model)

N(t) = # of insurance claims produced from a portfolio by time t;
Yi = amount of the ith claim;
X(t) = aggregate claim generated by the portfolio by time t.

Property 3.3. For a compound Poisson process X(t),

E [X (t)] = λtE [Yi ] ,

Var [X (t)] = λtE Yi2 .

 

Proof. This is left as an exercise. t

u
3.6 Compound Poisson process 63

Example 3.15. Suppose that families migrate to an area at a Poisson rate λ = 2 per
week. If the number of people in each family is independent and takes on the values
1, 2, 3, 4 with respective probabilities 16 , 13 , 13 , 16 , then what is the expected value and
variance of the number of individuals migrating to this area during a fixed five-week
period?

Solution. Letting Yi denote the number of people in the ith family, we have that
1 1 1 1 5
E [Yi ] = 1 ·
+2· +3· +4· = ,
6 3 3 6 2
2 1 2 1 2 1 1 43
E Yi = 1 · + 2 · + 3 · + 42 · = .
 2
6 3 3 6 6
Hence, letting X (5) denote the number of immigrants during a five-week period, we
get that
5
E [X (5)] = 2 · 5 · = 25,
2
43 215
Var [X (5)] = 2 · 5 · = .
6 3
t
u

For the compound Poisson process, we have the following Normal approxima-
tion:
• Let {X(t);t ≥ 0} be a compound Poisson process built on discrete random vari-
able Yi , with pmf
P(Y = ck ) = pk , k = 1, . . . , m
• Let Nk (t) = # of event by time t such that the corresponding value of Yi is ck . Then
{Nk (t);t ≥ 0}, k = 1, . . . , m are independent Poisson processes with respective
rates λk = λ pk .
• For large t, Nk (t) is approximately distributed as normal.
• Since the compound Poisson process is a linear combination of the decomposed
Poisson processes:

X(t) = c1 N1 (t) + c2 N2 (t) + · · · + cm Nm (t),

we can approximate the distribution of X(t) by normal if t is large.

Example 3.16. Find the approximate probability that at least 240 people migrate to
the area within the next 50 weeks.

Solution. Since λ = 2, E [Yi ] = 5/2, E Yi2 = 43/6, we can see that

 

E [X (50)] = 250 and Var [X (50)] = 4300/6.

Now, the desired probability is

64 3 Poisson process

P (X (50) ≥ 240) = P (X (50) ≥ 239.5)

!
X (50) − 250 239.5 − 250
=P p ≥ p
4300/6 4300/6
≈ 1 − Φ (−0.3922)
= Φ (0.3922)
= 0.6525,

where Φ (−0.3922) is the probability that a standard normal is less than 0.3922. t
u