Module II: The Poisson process

For Network Theory

Pablo Serrano and

José Alberto Hernández

{pablo, jahgutie}@it.uc3m.es

Bachelor’s Degree in
Telematics Engineering
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Contents
The exponential random variable
Definition
Simulation using Matlab
Memoryless
Other properties
The Poisson proccess
Counting processes
Definition of Poisson processes
Properties
Addition and sampling of Poisson processes
Non-homogeneous Poisson process
Theorem of Palm-Kintchine
Summary

2 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Exppdf: Definition and Moments

I A random variable X is said to be exponentially distributed with
rate λ if its PDF follows:

fX (t) = λe −λt , t≥0 (1)

I Its moments are:

E (X ) = 1/λ (2)
2
Var (X ) = 1/λ (3)

I Its CDF is:

Z t
FX (t) = P(X ≤ t) = fX (τ )dτ = 1 − e −λt , t≥0 (4)
0

3 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Matlab’s exppdf
I ... and its CCDF or Survival function:

SX (t) = P(X > t) = 1 − FX (t) = e −λt , t≥0 (5)

I Hence, its CCDF when plotted in a semilogy axis appears as a

straight line with slope −λ (Remark that log SX (t) = −λt.
I Function exprnd from Matlab can be used to generate
exponentially distributed random variables.
I Example: exprnd(10,1,1e3) gives us an 1 × 103 array of
exponential random variables with mean EX = 1/λ = 10.
I The easiest way to estimate λ from a measurement set
{e1 , . . . , eN } is by computing the inverse of the average value:

N
!−1
1 X
λ̂ = ei (6)
N
k=1
4 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Memoryless of exponential random variables

I Let X model the duration of a lightbulb, and let us compute

Pr(X > s + t|X > t), that is, the probability that such a lightbulb
lasts for t + s seconds assuming that it already survived t seconds:

P(X > s + t) e −λ(s+t)

Pr(X > s + t|X > t) = =
P(X > t) e −λt
= e −λs = P(X > s) (7)

hence, the duration of the lightbulb does not depend on how long
t it has survived so far (Memorylessness).
I Furthermore, the exponential distribution is the only memoryless
probability distribution.

5 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Minimum and comparison of exponential variables

I Let X1 , . . . , Xn be a set of n independent and exponentially

distributed random variables with rates λ1 , . . . , λn respectively.

Pr(min{X1 , X2 , . . . , Xn } > t) = e −( λi )t
P
i (8)

I Let X1 ∼ exp(λ1 ) and X2 ∼ exp(λ2 )

λ1
Pr(X1 < X2 ) = (9)
λ1 + λ2

6 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Example: Reliability of cars

I Let us consider a car with three components subject of failure: The

engine with rate 1 failure every 10 years, the electricity system with
rate 2 failures every 5 years and the air conditioning with rate 1
failure every 2 years.
1. Assume we start at time t = 0. Find the probability of at least one
failure in the first year.
2. Find the probability that the engine fails before any other component.
3. Find the probability that the engine fails twice in a row before any
other failure.

7 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Example: Reliability of a car (solution)

1. We use random variable Xfailure to model the time until the first
failure. This is exponentially distributed with rate:
1
λt = 10 + 25 + 12 = 1 failure per year
Therefore: P(Xfailure ≤ 1year ) = 1 − e −1
2. Comparison between exponentially-distributed random variables:
1/10
P(X1 ≤ min{X2 , X3 }) = λ1 +(λλ12 +λ3 ) = 1/10+2/5+1/2 1
= 10
3. Previous case twice:
1 1
P(X1 ≤ min{X2 , X3 })P(X1 ≤ min{X2 , X3 }) = 10 10

8 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Introduction to counting processes

Counting process A counting process N(t) is a random process that
accounts for the number of events occurred by time t.
The counting process N(t) must satisfy:
(i) N(t) ≥ 0
(ii) N(t) is integer-valued
(iii) If s < t, then N(s) ≤ N(t)
(iv) The value N(t) − N(s) (when s < t) refers to the
number of occurred within the interval (s, t].

e3 e4 e6 Xi = Inter-arrival times
e1 e2 e5 ei = Arrival times of events
X1 X2 X3 Si = Waiting times

t=0 S1=X1 S2=X1+X2 S3=X1+X2+X3 time t

9 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Independent and stationary increments

I Important properties of a counting process N(t) is to see whether

or not it has got stationary and independent increments.
Independent increments N(t) is said to have independent increments
if the number of events occurred in disjoint time
intervals are independent.
Stationary increments N(t) is said to have stationary increments if
the probability distribution of the number of events
occurred on a time interval only depends on its length.

10 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

The Poisson process. Definition

I The Poisson process is a counting process with independent and
stationary increments.
Poisson process A Poisson process N(t) with rate λ is a counting
process that follows:
(i) N(0) = 0
(ii) N(t) has independent increments.
(iii) The number of events occurred within a time interval
of length t follows the Poisson pdf with rate λt:

(λt)n −λt
Pr({N(t+s)−N(s)} = n) = e , n = 0, 1, . . .
n!
(10)

I Additionally, E (N(t)) = λt = Var (N(t)).

11 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

A more formal definition

Poisson process (formal) A Poisson process N(t) with rate λ is a

counting process that follows:
(i) N(0) = 0
(ii) N(t) has independent and stationary increments.
(iii) Pr(N(h) = 1) = λh + o(h)
(iv) Pr(N(h) ≥ 2) = o(h)
where o(h) is something very small (increases smaller than h)

12 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

A more formal definition

I Now, let Pn (t) = Pr(N(t) = n)
I Thus:

P0 (t + h) = Pr(N(t + h) = 0)
= Pr(N(t) = 0) Pr({N(t + h) − N(t)} = 0)
= Pr(N(t) = 0) Pr(N(h) = 0) = P0 (t)(1 − λh + o(h))
(11)

I This yields

P0 (t + h) − P0 (t) o(h)
= −λP0 (t) +
h h
(12)

13 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

A more formal definition

I Taking the limit when h goes to zero on the above:

h → 0 ⇒ P00 (t) = −λP0 (t) (13)

I The solution to this differential equation is: P0 (t) = Ke −λt for

some K
I K = 1 from P(N(0) = 0) = 1, hence: P0 (t) = e −λt
I Similarly, it can be demonstrated that:

(λt)n −λt
Pn (t) = e , n = 0, 1, . . . (14)
n!
I Thus, the number of events occurred by time t follows the Poisson
(λt)n −λt
distribution: Pr(N(t) = n) = n! e .

14 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Inter-arrival times

I Let X1 denote the time elapsed until the first event occurs.
I Clearly, the first event occurs after time t (with probability
P((X1 ) > t)) only if there are no events before time t (with
probability P(N(t) = 0)), which brings:

(λt)0 −λt
Pr(X1 > t) = Pr(N(t) = 0) = e = e −λt (15)
0!
I which is the survival function of an exponential random variable
with rate λ.
I Thus, the time until the first event occurs is exponentially
distributed with mean EX1 = 1/λ.

15 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Inter-arrival times
I Similarly, the inter-arrival time between the first event and the
second event, X2 , arises from:

Pr(X2 > s + t|X1 = s) = Pr(0 events in (s, s + t)|X1 = s)

= Pr(0 events in an interval of length t)
(λt)0 −λt
= e = e −λt (16)
0!
I The inter-arrival time between the first and the second events is
also exponentially distributed.
I Similarly, it can be shown that the pdf of the inter-arrival times
between all events is exponentially distribtued with mean
E (Xi ) = 1/λ.
I In conclusion, the Poisson process has no memory and it renews
itself at all times.
16 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

WaitingPtimes
n
I Sn = i=1 Xi : waiting time until the n-th event occurs.
∞
X (λt)j
Pr(Sn ≤ t) = Pr(N(t) > n) = e −λt (17)
j!
j=n

I Thus:
d Pr(Sn ≤ t)
fSn (t) = =
dt
∞ ∞
X (λt)j −λt X (λt)j−1 −λt
= − λ e + jλ e
j! j!
j=n j=n
λ t n−1
n
= e −λt = Γt (λ, n) (18)
(n − 1)!
I This is a Gamma distribution with rate λ and n degrees of freedom.
17 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Conditional distribution of arrivals

I Assume we are told that one event occurred at some point within
(0, t). The question is: when is it more likely for that event to
have occurred?
Pr(Y1 < s, N(t) = 1)
Pr(Y1 < s|N(t) = 1) =
Pr(N(t) = 1)
Pr(1 event in (0, s)) Pr(0 events in (s, t))
=
Pr(1 event in (0, t)
(λs)1 −λs (λ(t−s))0 −λ(t−s) =
1! e × 0! e s
= 1
(λt) −λt
(19)
t
1! e

I Thus, Y is uniformly distributed on (0, t): Y ∼ U(0, t)

I This result was expected given the memoryless nature of the
Poisson process.
18 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Addition and sampling of Poisson processes

I Two more (very important) properties of the Poisson process:

Addition Let N1 (t) and N2 (t) be two Poisson processes with
rates λ1 and λ2 respectively. Then, the process
Ntot = N1 (t) + N2 (t) is also a Poisson process with rate
λtot = λ1 + λ2
Sampling Let N(t) be a Poisson process with rate λ, and let
0 ≤ p ≤ 1 be some sampling probability. Then, the
process obtained by sampling the events of N(t) with
probability p is also a Poisson process with rate λp. The
events left apart is also a Poisson process with rate
λ(1 − p).

19 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Addition and sampling of Poisson processes. Example

I Consider an Ethernet switch with 64 input ports and 2 output

ports. Each input port injects packets to the switch as a Poisson
process with rate λ = 10 packets/sec. The routing tables state
that 70% of the packets must go to output port number 1, while
the other 30% must go to the second output port.
I Characterise the packet process on the two output ports. Are they
Poisson? What rates?

20 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Addition and sampling of Poisson processes. Example

I Solution:
I The counting process at the switch is also Poisson because it is the
combination of 64 Poisson processes. The total rate is then
λtot = 64 × 10 = 640 packets/sec.
I The counting process of outgoing packets on the two ports are also
Poisson because the are random samples of a Poisson process. The
first output port has got a rate of λo1 = 0.7 × 640 = 448
packets/sec, and the second λo2 = 0.3 × 640 = 192 packets/sec.

21 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Non-homogeneous Poisson process

I Sometimes, a process behaves as a Poisson process (with all its
properties), but with variable rate λ(t).
I This is:
(i) N(0) = 0
(ii) N(t) has independent increments.
(iii) P(N(t + h) − N(t) = 1) = λ(t)h + o(h)
(iv) P(N(t + h) − N(t) ≥ 2) = o(h) where o(h) is something very small
(increases smaller than h)
I If we let: Z t
m(t) = λ(s)ds (20)
s=0
then it can be shown that (for n ≥ 0):

[m(t + s) − m(t)]n −(m(t+s)−m(t))

Pr(N(t + s) − N(t) = n) = e
n!
(21)
22 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Theorem of Palm-Kintchine
Theorem of Palm-Kintchine Let {A1 , . . . , An } be n independent
renewal processes with average inter-arrival times
1/λ1 , . . . , 1/λn respectively. Then, the process made by
the superposition of them:
n
∗
X  t
An (t) = Ai nt0 + , t0 > 0 (22)
n
i=1

Pn λi
I This has an average rate of λ∗n = i=1 n .
I If limn→∞ λ∗n = λ, then limn→∞ A∗n = A Poisson
I The conclusion is that the superposition of multiple counting
process (not necessarily Poisson processes) approximates the
Poisson process as long as these many processes are independent.

23 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Theorem of Palm-Khintchine. Example

I Consider an Ethernet switch with 64 input ports and one output
port. On each input port, packet inter-arrival times are measured
to follow a uniform distribution U(2, 12) ms. Describe the
aggregated process of the switch.
I Solution:
I Following the theorem of Palm-Khintchine, the aggregation of
many independent processes converges to a Poisson process with
rate:
64
X
λtot = λi
i=1
1
where λi = EX1 = 142.85 packets/sec
I Hence, λtot = 64 × 142.85 = 9142.4 packets/sec

24 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

Summary
I The exponential random variable is very important in modelling
many processes in real life, for instance:
I Packet arrivals at a network router.
I Failure events on a network.
I It has got the so-important property of memorylessness.
I The Poisson process is a counting process, where event interarrival
times follow the exponential distribution.
I It is very easy to simulate Poisson processes with Matlab.
I Adding and sampling Poisson processes are also Poisson processes.
I Non-homogeneous Poisson processes have a variable rate λ.
I The superposition of many independent processes converge to a
Poisson process.

25 of 26
 Module II: The Poisson process {pablo,
jahgutie}@it.uc3m.es

References

I Sheldom M. Ross: Stochastic processes, second edition. Wiley

Series in Probability and Mathematical Analysis, 1996.
I José Juan Pazos Arias et al: Teorı́a de colas y simulación de
eventos discretos, first edition. Prentice Hall, 2003.
I The Wikipedia for “Exponential distribution”, “Memorylessness”,
and “Poisson process”.

26 of 26