STAT2001_CH03C Page 1 of 9

The Poisson distribution

We have delayed talking about this distribution because its definition involves
mention of expectation, which was therefore covered first.
Example 19 At a certain traffic intersection, two accidents occur each year,
on average.
Find the probability that 4 accidents will occur at this intersection next
year, supposing that either 0 or 1 accidents can occur in each period of:
(a)
a month
(b)
a week
(c)

a day.

Assume that accidents occur independently from period to period.

(a)

Let Y = number of accidents per year.

Then Y ~ Bin(12, 2/12).
12 2 4 10 8
So P(Y = 4) = = 0.0888 .
4 12 12

(Then, EY = 12(2/12) = 2.)

(b)

In this case, Y ~ Bin(52, 2/52).

(Then, EY = 52(2/52) = 2.)
4
48
52 2 50
So P(Y = 4) = = 0.0902.
4 52 52

(c)

Now, Y ~ Bin(365, 2/365).

(Then, EY = 365(2/365) = 2.)
4
361
365 2 363

= 0.0902 (same, to 4 decimals).
So P(Y = 4) =
4 365 365

A limit appears to have been reached.

For example, we would get the same answer (to 4 decimals) if we divided the year up
into 365 24 = 8760 hours and worked on the assumption that
8760 2 4 8758 8756

Y ~ Bin(8760, 2/8760). In that case, P(Y = 4) =

= 0.0902.
4 8760 8760

STAT2001_CH03C Page 2 of 9

More generally, suppose that 0 or 1 accidents can occur in each of n time periods,
and that the total expected number of accidents is fixed at 2. Then in the limit as n
tends to infinity, we find that the required probability can be written
e2 2 4
P(Y = 4) =
= 0.0902.
4!

Proof: Let Y ~ Bin(n,2/n).

n 2 4 2 n4

Then P(Y = 4) = 1
4 n n
4

n!
24 2 2
=
1 1
4!(n 4)! n 4 n n
n
4
n n 1 n 2 n 3 (n 4)! 24 2 2
=

1
1

n n n n (n 4)! 4! n n
m

2 4 2
x
x

1 e 1 as n .
(Note that e = lim 1 + .)
m
m
4!

Similarly, P (Y = 3) =

e2 23
, etc.
3!

We say that Y has a Poisson distribution (with parameter 2).

When it makes sense to take a limit as here, we say that a Poisson process is involved.
A random variable Y has the Poisson distribution with parameter if its pdf is of the
form

p( y) =

e y
,
y!

y = 0,1, 2,

( 0 ).

We write Y ~ Poi( ) and p ( y ) = pPoi( ) ( y ) .

Summarising, pPoi( ) ( y ) = lim pBin ( n , p ) ( y ) .
n
= np

Lets verify that the mean of the Poisson distribution with parameter is :

EY = y
y =0

e y
y1
x
= e
= e
= e e = .
y!
x =0 x !
y =1 ( y 1)!

STAT2001_CH03C Page 3 of 9

Exercise:

Show that the variance of the Poisson distribution with parameter is

also . (Hint: Either use the mgf technique, or first find E{Y(Y 1)}.)

The Poisson distribution is useful for modelling all sorts of count data, not just traffic
accident frequencies.

Example 20 A total of 27 defects were found in a 1 km coil of thread.

Find the probability (approximately) that a 10 m coil of the same type of thread
(taken randomly from another coil) will have at least one defect.
Let Y be the number of defects in the 10 m coil.
Then (approximately), Y ~ Poi( ) , where = 0.27.
(Each 1 metre length of the 1 km coil has 27/1000 = 0.027 defects on average.
So the expected number of defects in the 10 m coil is (about) 0.027 10 = 0.27.)

Thus P (Y 1) = 1 P (Y = 0) = 1

e0.27 0.27 0
= 1 e0.27 = 1 0.763 = 0.237 .
0!

Note: We have assumed that the number of defects in the thread follows a
Poisson process. This means that it is reasonable to consider the rope as being
comprised of a very large number of tiny lengths, such that:
(a)

each length can have either 0 or 1 defects (but not 2 or 3, say)

(b)

each length has the same chance of containing a defect

(c)

whether a certain length has a defect is independent of

whether any other lengths have a defect.

We could test the Poisson assumption by counting the number of defects in each of
the one hundred 10 m lengths that make up the 1 km coil. This would result in a
sequence of 100 numbers like 0,0,1,0,2,0,0,0,0,1,....,0,1,0,0.
If the Poisson assumption is justified, about 24 (23.7%) of these numbers will be
e0.27 0.27 2
nonzero. Also, the number of 2s will be about
100 = 2.78 (ie, 3 or so),
2!
etc. If we find large discrepancies between the expected and observed numbers, then
there will be good reason to doubt the Poisson assumption.

STAT2001_CH03C Page 4 of 9

The Poisson approximation to the binomial

As a consequence of the definition of the Poisson dsn in terms of the binomial dsn,
the Bin(n,p) can be approximated by the Poi ( = np ) when n is 'large' and p is 'small'.
Example: We roll 2 dice together 30 times. Find the pr. that one double six comes up.
Solution: Number of double sixes = Y ~ Bin(30,1/36) Poi(5/6)
5/ 6

So P(Y = 1)

(since 30(1/36) = 5/6).

30 1 35 29
(5 / 6)
= 0.3622. (The exact prob. is = 0.3681.)
1!
1 36 36
1

Example 2: We roll 3 dice together 30 times. Find the pr. that one triple six comes up.
Solution: Y = number of triple sixes ~ Bin(30,1/216) Poi(5/36)
(since 30(1/216) = 5/36).
30 1 215 29
e5/36 (5 / 36)1
= 0.1214.)
= 0.1209. (The exact pr is
So P(Y = 1)

1!
1 216 216
We see that the approximation here improves as p gets smaller. Generally, the Poisson
approximation should be considered only when the exact binomial probability is hard
or impossible to calculate. As a rule of thumb, the Poisson approximation is 'good' if
n is at least 20 and p is at most 0.05, or if n is at least 100 and np is at most 10.

Tchebysheffs theorem (or Chebyshevs theorem)

Let Y be a rv with mean and variance 2 (assumed to be finite).
Also let k be a positive constant. Then
1
P (| Y |< k ) 1 2 .
k
1
(Equivalently, P (| Y | k ) 2 .)
k

Example 21 The number of customers, Y, who visit a certain store per day has been
observed over a long period of time and found to have
mean and variance both equal to about 50.
Find a lower bound for the probability that from 40 to 60 customers
will visit the store tomorrow (inclusive).

STAT2001_CH03C Page 5 of 9
P (Y {40,41,...,60}) = P (39 < Y < 61)

(*)

= P (| Y 50 |< 11)
= P (| Y |< k ) where:

= 50
=

50 so that

k = 11/ 50
1

= 1

1
k2

1
(11/ 50) 2

= 1 50 /121
= 0.5868.
So the prob. that 40 to 60 people will visit the store tomorrow is at least 58.68%.
Note 1: At (*) above we could also have written
P (Y {40, 41,..., 60}) = P (39.9 < Y < 60.1)

= P (| Y 50 | < 10.1)
= P (| Y |< k ) where:

= 50
=

50 so that

k = 10.1/ 50

1
k2
= 0.5099.
1

So the probability in question is at least 50.99%. This is equally true, but not so
useful, as saying that the probability is at least 58.68%. It is for this reason that at (*)
we used the widest possible interval we could have, i.e. (39,61), before proceeding.
Note 2: Recall that count data often has a Poisson distribution. So it is of interest to

work out the probability exactly, under the assumption that Y has this distribution.
Suppose that Y ~ Poi(50). Then
e50 50 40 e50 5041
e50 5060
P (40 Y 60)
=
+
+ ... +
= 0.8633
40!
41!
60!
(which is indeed at least as big as 0.5868, the lower bound worked out earlier).
Note that if Ys mean and variance were different, it would be inappropriate to model
Ys dsn as Poisson. But we could still apply Chebyshev's thm to get a lower bound.

STAT2001_CH03C Page 6 of 9

Proof of Chebyshevs theorem

2 = E (Y ) 2 = ( y ) 2 p( y )
y

( y ) 2 p ( y ) +

0 p( y ) +

y:| y|<k

( y ) 2 p( y )

y:| y|k

y:| y|<k

(k )2 p( y )

y:| y|k

since ( y ) 2 0 for all y ,

and ( y )2 (k )2 for all y such that | y | k
= 0 + k 2 2

p( y )

y:| y|k

= k 2 2 P (| Y | k ) .

So P ( Y k )

1
1
, and hence also P (| Y |< k ) 1 2 .
2
k
k

Two other measures of central tendency (other than the mean)

The mode of a rv Y is any value y at which Ys pdf, p(y) (or f(y)) is a maximum.
(There may be more than one mode, and the mode may then also be defined as the set
of all such modes.)
The median of a rv Y is any value y such that
and P (Y y ) 1/ 2 .
P (Y y ) 1/ 2
(There may be more than one median, and the median may then also be defined as the
set of all such medians.)

STAT2001_CH03C Page 7 of 9

Example 22 Find the mode and median of the geometric distribution

with parameter 1/4.
y

P (Y y )

p(y)

P (Y y ) = 1 P (Y < y ) = 1 P (Y y 1)

-----------------------------------------------------------------------------------------------------( 1/2)

1/4

=16/64

16/64 (< 1/2)

(3/4)1/4 = 12/64

28/64 (< 1/2)

1 16/64 = 48/64 ( 1/2)

(9/16)1/4 = 9/64

37/64 ( 1/2)

1 28/64 = 36/64 ( 1/2)

(27/64)1/4 = 6.75/64 43.75/64 ( 1/2)

1 37/64 = 27/64 (< 1/2)

etc.
-----------------------------------------------------------------------------------------------------We see that Mode(Y) = 1 and Median(Y) = 3.
Note that P (Y 3) = 37 / 64 1/ 2 and P (Y 3) = 1 28 / 64 = 36 / 64 1/ 2 ,
and that no value other than 3 could be substituted here. So 3 is the only median.
Also, Ys mean is Mean(Y) = EY = 1/(1/4) = 4

(see Tutorial 5).

Some general interpretations

mean = average of a very large number of independent realisations of Y
mode = most likely value of Y
median = 'middle' value, such that Y is at least 50% likely to be above that
value and at least 50% likely to be below it.

An example of a rv with multiple medians and modes

Suppose that Y ~ Bern(1/2).
Then Median(Y) = any value in the interval [0,1].
Eg, 1 is a median because P (Y 1) = 1 1/ 2 and P (Y 1) = 1/ 2 1/ 2 .
Eg, 0.7 is a median because P (Y 0.7) = 1/ 2 1/ 2 and P (Y 0.7) = 1/ 2 1/ 2 .
Thus Y has an infinite number of medians. We may also write Median(Y) = [0,1].
Note that Y here also has multiple modes: 0 and 1.
We may say that Y is multimodal or bimodal.
We may also write Mode(Y) = {0,1}.

STAT2001_CH03C Page 8 of 9
Finally, note that Ys mean is 0(1/2) + 1(1/2) = 1/2.

Summary of discrete distributions

As an exercise, fill in the empty cells below and check against the back inside cover
of the text book. You may also wish to add two more columns, one for the mode and
one for the median, although these may not always have simple formulas.

distribution
Y~

pdf
p(y)

Binomial
Bin(n,p)

n y
p (1 p ) n y
y
y = 0,1,...,n

Bernoulli
Bern(p)
=?

Geometric

Hypergeometric

Poisson

mgf
m(t ) = EeYt

mean
= EY

np

variance
2 = VarY

STAT2001_CH03C Page 9 of 9

Completed summary of discrete distributions

distribution
Y~

pdf
p(y)

mgf
m(t ) = EeYt

mean
= EY

variance
2 = VarY

Binomial
Bin(n,p)

n y
p (1 p )n y
y

(1 p + pet ) n

np

np(1 p)

Bernoulli
Bern(p)

y = 0,1,...,n
1 p, y = 0

p, y = 1

1 p + pet

p(1 p)

pet
1 (1 p )et

1
p

1 p
p2

no simple

nr
N

nr ( N r )( N n)
N 2 ( N 1)

r
p

r (1 p )
p2

= Bin(1,p)

Geometric
Geo(p)
Hypergeometric
Hyp(N,r,n)

(1 p ) y1 p
y = 1,2,3,...
r

N r

n y
y
N

n

expression

y = 0,...,r
0 n y N r

Poisson
Poi( )
= lim Bin(n, p )
n

e y
y!
y = 0,1, 2,

e ( e 1)

y 1 r yr

p q
r 1

pet

t
1 (1 p )e

p( y) =

np =

Negative binomial
Neg(r, p)
Neg(1,p) = Geo(p)
(Section 3.6)

y = r, r + 1, ...