Chapter 6, Section 5

Transformations of Variables

Method of Moment-Generating
Functions

 John J Currano, 04/16/2010

1
 Method of Moment-Generating Functions

The crucial theorem is:

Theorem 6.1 (p. 318): If X and Y are random variables which

both have moment-generating functions, and if

mX(t) = mY(t) for all t in some interval around t = 0,

then X and Y have the same probability distribution.

2
 Method of Moment-Generating Functions

Some other useful facts:

1. If U = aY + b, then

   
m U (t )  E (e tU )  E e aYt  bt  e bt E eY (at )  e bt m Y (at ).

2. If Y1, Y2, … , Yn are independent and U = Y1 + Y2 +  + Yn, then

(
m U (t )  E (e tU )  E e tY1 )(
 tY2  tYn
)
 E e tY1 e tY2 e tYn

( )( ) ( )
 E etY1 E e tY2  E e tYn by independence

 mY1 (t ) mY2 (t )  mYn (t )

3
 Method of Moment-Generating Functions

Some other useful facts:

n
3. If Y1, Y2, … , Yn are independent and U   aiYi ,
i 1
n
m U (t )   m Yi (ai t ).
then i 1

4. If Y1, Y2, … , Yn are independent and identically distributed (iid)

with common distribution Y, i.e., a Random Sample from Y, and

U = Y1 + Y2 +  + Yn, then [ ].

m U (t )  m Y (t )
n

4
 Method of Moment-Generating Functions

Example 1. Suppose that Y1, Y2, … , Ym are independent binomial RVs

with Yi ~ bin(ni, p) [same p]. Then for i = 1, 2, … , m,

m Yi (t )  q  pe t  ni

m
Let Y  Yi . Then by Property 2 on slide 3,
i 1

m n1  n2  nm

i 1
( )
mY (t )   m Yi (t )  q  pe t ,

m
so Y has a binomial distribution with n  ni trials and probability of
i 1
success, p. Thus, the sum of independent binomials with the same
probability of success, p, is also binomial.
5
Example 2. Let Y ~ NegBin(r, p),
X0 = 0, and
Xi = # of trial on which i th success occurs.
Then
Yi = Xi  Xi 1 ~ Geom(p) for i = 1, 2, … , r, and
Y1, Y2, … , Yr are independent.
Also, r r
Yi   X i  X i 1  X r  X 0  X r  0  Y .
i 1 i 1
telescoping sum

r
r  pet 
Thus, mY (t )  mYi (t )   t

 by Property 2 on slide 3.
i 1 1  qe 
We have found the moment-generating functions of the negative
binomial distributions using those of the geometric distributions.
6
 r
 pe  t
m ( t )   
Example 2. Y ~ NegBin(r, p), Y  1  qe t  .
 
We can now use the mgf of Y to find its mean and variance.
For example,
 pe 0 
E (Y )  mY (0)  r  
r 1
 pe  1  qe    pe  qe 
0 0 0 0

0
 1  qe  1  qe 
0 2

r 1
( )
 p  p (1  q) pq
r  
1 q  ( )
1  q
2

r 1
 p p(p  q)
r   
p p2

 r
.
p

7
 Y  1 
Example 3. Let Y ~ N( ,  ) and Z 
2  Y  .
  
 2 2
Then m Y (t )  exp  t  21  t ,
   1 
so m Z (t )  exp   t  m Y  t  by Property 1 on slide 3
    

    1  1 
2
 exp   t  exp    t   21  2  t 

        
  t  2 t 2  
 exp     exp
t
 



 1
2

 2






( )
1
2
t2

So Z ~ N(0, 1 ) : Z has a standard normal distribution. We proved this

fact in Chapter 4 using the Distribution Function method – this is simpler.

Transforming a random variable Y by subtracting its mean and dividing

by its standard deviation is called standardizing Y.
8
Example 4. Let Z ~ N(0,1) and Y =  +  Z .

Then m Z (t )  exp  1
2

t2 ,

so mY (t )  exp  t  m Z  t  by Property 1 on slide 3

 exp  t  exp
( 1
2 )
 t 2

(
 exp  t  1
2
2 t2
)
So Y ~ N( ,  2) : Y has a normal distribution with mean  and
variance  2. We also proved this in Chapter 4 using the
Distribution Function method – again this method is simpler.

Transforming a standard normal random variable Z in this

fashion is a way of simulating other normal random variables.

9
Example 5. If Yi ~ N( i ,  i 2) are independent for i = 1, 2, . . . , n,


then m Yi (t )  exp  i t 
1
2
 i2 t 2  m
for 1 i  n. Let Y   ai Yi .
i 1

Then by the Property 3 on slide 4,

m m
( (a t)).
2
m Y (t )   m Yi ( ai t )   exp  i a i t  1
2
 i2 i
i 1 i 1

 m m 
 exp  t a i  i  2 t  ai2 i2  ,
1 2
 i 1 i 1 

m m 
so Y ~ N   a i  i ,  ai2  i2  . Thus, a linear combination of
 i 1 i 1 

independent normal random variables is also normal.

10
Special Case. Suppose Y1, Y2, . . . , Yn are a random sample from
1 n
a N(  ,  2) - distribution, and
Y   Yi , the sample mean.
n i 1


Then by Example 5, Y ~ N  ,  2 n . 

Thus the sample mean of a random sample of size n from a normal

distribution also has a normal distribution.

Its mean is the same as that of the original distribution and its
variance is smaller   2/n instead of  2.

This will be used often next semester.

11
Example 6. (p. 319, Example 6.11)
1 z2 2
Let Z ~ N(0,1), so f ( z )  e , and let Y  Z 2 . Then
2

mY (t )  E e 
tZ2

 e
t z2 1
2
e z2 2
dz



1 (1 2t ) z 2 2
  e dz
 2


1  z 2 (2(1 2t ) 1)
  e dz
 2

12
 2
Example 6. (p. 319, Example 6.11) Let Z ~ N(0,1) and Y  Z .

Then m Y (t )  
1
e  z2  2(12t )1  dz .
 2

The integrand is proportional to the density function of the normal

distribution with  = 0 and  2 = (1 – 2t)1. Thus, if we multiply by

1
 1  2t   
12 1 2
inside the integral and by   1  2t outside,

we obtain
m Y (t )  1  2t   1  1  2t 
1 2 1 2
.

Therefore, Y  Z 2 has a  2 (1) - distribution.

We proved this in Chapter 4 using the Distribution Function method.

Once again, this method is simpler.
13