MA2216/ST2131 Probability

Notes 5
Distribution of a Function of a Random Variable
and
Miscellaneous Remarks
1. Change-of-Variable.
To begin with, let us recall an ane transformation of a normal r.v.
mentioned in 8.8 of Lecture Notes 4.
To be precise, let X N (, 2 ). Put
Y = aX + b.
Then, it has been pointed out in 8.8 of Notes 4 that, for a = 0,
Y N (a + b, a2 2 ).
Equivalently, for y IR, the p.d.f. of Y is given by
(yab)2
1

fY (y) =
e 2(a)2 .
2
2(a)

Derivation of (1.1).

(1.1)

Recall that
(x)2
1
e 22 .
fX (x) =
2 2

It takes two steps for us to determine the distribution of Y , i.e., to nd

FY , the d.f. of Y , rst, and then to dierentiate FY to get fY , the p.d.f.
of Y .
Consider the case a > 0 rst.
FY (y) = IP (Y y)
= IP (aX + b y)
yb
= IP (X
)
a
)
(
yb
.
= FX
a
1

To identify the distribution of Y , we need to nd its p.d.f., fY , which is

the derivative of FY : for y IR,
d
d
fY (y) =
FY (y) =
FX
dy
dy
(
)
yb
1
= fX

a
a

yb
a

yb

)2
(
a
1
1
e 22
=
a
2 2
(yab)2
1

=
e 2(a)2 .
2(a)2

We recognize that it is the p.d.f. of N (a + b, a2 2 ), which is (1.1).

If a < 0, then
)
(
)
(
yb
yb
FY (y) = IP X
= 1 FX
.
a
a
Hence,
yb

(
)2
a
1
1
fY (y) =
e 22
a
2 2
(yab)2
1

=
e 2(a)2 ,
2(a)2

which is obviously the p.d.f. of the normal distribution N (a+b, a2 2 ),

and therefore, we are done.
In particular, if we set
Z=

X
X

=
,

then Z is a standard normal r.v., i.e., Z N (0, 1). This is perhaps the
most frequently used transformation.

In the following, we will see a few additional examples on the derivation

of the distribution of a transformed r.v.
1. Example.
Y = Z 2?

Let Z N (0, 1). What are the d.f. and p.d.f. of Y , where

Note that, Y is known in the literature as chi-square (2 ) r.v. of degree

1.
Sol. First note that Y takes nonnegative values. Obviously, Y is continuous, and hence we determine its c.d.f. rst. For y > 0, we have
FY (y) = IP (Y y) = IP (Z 2 y)
1

= IP ( y Z y) =
2
y
2
2
=
ez /2 dz,
2 0
hence
fY (y) =

ez

/2

dz,

dFY (y)
2
1 y/2
y 1/2 ey/2

=
e
=

dy
2 2 y
2
1

( 12 ) 2 1 1 1 y
=
y2 e 2 .
( 12 )
We conclude that

0,
FY (y) =

for y 0,
y

eu

/2

du, for y > 0.

0,

and
fY (y) =

for y 0,
1

( 2 )12 y 12 1 e 12 y , for y > 0.

( )
2

Remark. Do you recognize the distribution of the above Y ?

In fact, the distribution of Z 2 with Z N (0, 1) is ( 12 , 12 ). Thus,
(
21

1 1
,
2 2

)
.

(1.2)

This is a very useful result, which will be quoted often throughout.

3

Addendum. Refer to 1.4 of Notes 7, in which the following property

will be established:
If Xi , i = 1, 2, . . . , n are independent gamma r.v.s with respective parameters (i , ), i = 1, 2, . . . , n, then
( n
)
n

i , .
(1.3)
Xi
i=1

i=1

Now, apply this proposition to the sum,

Zi2 , where Z1 , Z2 , . . . , Zn is

i=0

a sequence of i.i.d. standard normal random variables. Put

X=

Zi2 .

i=0
2
The distribution of X is called the chi-square
( n 1 )(n ) distribution of n
degrees of freedom, which is the same as 2 , 2 . That is,
(
)
n 1
2
,
n
.
(1.4)
2 2

2. Example. Let X N (0, 1). Dene Y = eX , commonly known as the

lognormal random variable. Find the p.d.f. fY .
Sol. First note that Y takes nonnegative values. Therefore, for y > 0,
we have
FY (y) = IP (Y y)
= IP (eX y)
= IP (X ln y)
ln y
2
1
=
ex /2 dx,
2
hence
fY (y) =

2
dFY (y)
1
1
= e(ln y) /2 .
dy
2 y

To be precise,

0,
fY (y) =

1
y 2

for y 0,
e(ln y)

/2

, for y > 0.

3. Example. Let X be a continuous random variable with p.d.f. fX . Let

Y = X n where n is odd. Find p.d.f. fY .
Sol. Let y IR. We have
FY (y) = IP (Y y) = IP (X n y)
= IP (X y 1/n )
= FX (y 1/n ),
hence
fY (y) =

1 1
dFY (y)
= y n 1 fX (y 1/n ).
dy
n

4. Example. Let X be an exponential r.v. with parameter . That is,

X Exp(). Set
Y = [[X]] + 1.
(Note that [[x]] = k i k x < k +1 for integers k.) Find the distribution
of Y .
Sol. Obviously, Y is a discrete r.v., taking values in the set of positive
integers. For y = 1, 2, . . .,
fY (y) = IP {Y = y} = IP {[[X]] + 1 = y} = IP {[[X]] = y 1}
= IP {y 1 X < y} = IP {X y 1} IP {X y}
= e(y1) ey
(
)
= e(y1) 1 e .
Let p = 1 e and q = e . The above p.d.f. can now be written as
fY (y) = q y1 p
for y = 1, 2, . . ., and we conclude that Y Geom(p).

5. Let X Exp(1). Set Y = X/, where > 0. Determine the distribution

of Y .
Sol. Obviously, Y is continuous, taking non-negative values. For y > 0,
FY (y) = IP {Y y} = IP {X/ y}
= IP {X y} = 1 ey .
The p.d.f. fY () is thus obtained: for y > 0,
fY (y) =

dFY (y)
= ey .
dy

We conclude that Y Exp().

Alternatively, if Y Exp() is given, and we let X = Y , then
X Exp(1).

6. Example.

Let X be a r.v. of which its c.d.f. is given by

0,

for x ,
{ (
) }
FX (x) = 1 exp x
, for x > .

(Note that this distribution is called a Weibull distribution with parameters , > 0, and > 0. Such a distribution is widely used in
engineering practice due to its versatility. Dierentiating FX yields
0,
for x ,
{ (
(
)1
) }
x
fX (x) = x
exp
, for x > ,

which is the p.d.f. of X.)

(
Let Y =

)
. Determine the distribution of Y .

Sol. Y takes non-negative values. For y > 0,

{(
}
)
X
FY (y) = IP {Y y} = IP
y

{
}
= IP X y 1/ +
= 1 ey ,
which is the c.d.f. of an exponential distribution with parameter = 1.
Thus,
Y Exp(1).

7. Theorem. Let X be a continuous r.v. having p.d.f. fX . Suppose

that g(x) is a strictly monotone (increasing or decreasing) dierentiable
function of x. Then the r.v. Y dened by Y = g(X) has a p.d.f. given
by
{


 d 1 
1
fX [g (y)]  dy g (y) , if y = g(x) for some x,
fY (y) =
0,
if y = g(x) for all x,
where g 1 (y) is dened to equal that value of x such that g(x) = y, i.e.,
g 1 (y) is the inverse function of g.
Sketch of a Proof.
of Y rst.

Let us try to determine the ditribution function

FY (y) = IP (Y y) = IP (g(X) y)
= IP (X g 1 (y)) (assuming g is increasing)
= FX (g 1 (y)).
To nd fY (y), we dierentiate FY (y) w.r.t. y, and get
d
FX (g 1 (y))
dy
d
d 1
=
FX (x)|x=g1 (y)
g (y)
dx
dy
d 1
= fX [g 1 (y)]
g (y).
dy

fY (y) =

If g is decreasing, we have to modify the above derivation a little bit:

FY (y) = IP (Y y) = IP (g(X) y)
= IP (X g 1 (y)) (assuming g is decreasing)
= 1 FX (g 1 (y)).
7

Thus,

]
d [
1 FX (g 1 (y))
dy
d
d 1
= FX (x)|x=g1 (y)
g (y)
dx
dy
[
]
d 1
1
= fX [g (y)] g (y)
dy



d 1 
1

= fX [g (y)]  g (y) .
dy

fY (y) =

8. Remark. Let X N (, 2 ). Put Y = aX + b = g(X), where

g(x) = ax + b.
Assume a = 0. Thus, the function g is a strictly monotone dierentiable
function of x, i.e., g() satises the assumption in the above theorem.
Obviously,
yb
g 1 (y) =
,
a
d 1
1
with g (x) = a and
g (y) = .
dy
a
As a result, we arrive at the p.d.f. of Y :
fY (y) = fX [g



 d 1 
(y)]  g (y)
dy
yb

(
)2
a
1
1
e 22 | |
=
a
2 2
[y(a+b)]2
1
e 2a2 2
=
2a2 2
(yab)2
1

=
e 2(a)2 .
2(a)2

9. Remark. Can the above theorem be applied to the example in item

1?
10. Theorem 7 is concerning the change of variable for one-dimensional setting. Later in the course, we will learn how to make change of variables
for the multi-dimensional setting. (See Lecture Notes 8.)
8

2. Remarks & Examples

Several examples/remarks are presented below for reference only. (We
wont go through them in class.)
1. Stirlings Formula & De Moivre-Laplaces Thm.
(Refer to Kai Lai Chungs book on Elementary Probability Theory with
Stochastic Processes, Appendix 2 of Chapter 7.)
n!
as n , i.e.,
lim

2 nn+(1/2) en
n! en

n nn+(1/2)

2.

(2.1)

2. It has been shown that an exponential distribution possesses the socalled memorylessness property. Indeed, it is the unique distribution
(on IR+ ) possessing this property.
Proof. To see this, suppose X is a non-negative real-valued r.v. which
admits the memorylessness property and let
F (x) = IP {X > x}.
Then, the memorylessness yields
F (s + t) = F (s) F (t).
In other words, F () satises the functional equation
g(s + t) = g(s) g(t).
Obviously,

( )
(
)
( )2
2
1
1
1
g
=g
+
=g
,
n
n n
n
( )m
(m)
1
and repeating this yields g
=g
.
n
n
Similarly,

(
g(1) = g

1
1
+ +
n
n
9

( )n
1
=g
,
n

( )
1
g
= g(1)1/n .
n

or
(m)

= g(1)m/n . Recall that the distribution is dened to be

n
right-continuous. Clearly, F () = 1 FX () is right-continuous as well.
By making use of the right-continuity of F , we claim that
Hence, g

g(x) = g(1)x .
In turn, we conclude that for x > 0,
g(x) = ex ,
where = ln g(1).

3. A variation of the exponential distribution is the distribution of a r.v.

that is equally likely to be either positive or negative and whose absolute
variable is exponentially distributed with parameter > 0. Determine
the d.f. and p.d.f. of this r.v.
Sol. Let X denote such a r.v. By assumption, X is equally likely to
be either positive or negative, and hence X and X have the identical
distribution. Thus, for x > 0,
IP {X < x} = IP {X > x}.
On the one hand, for x > 0,
IP {|X| x} = IP {x X X}
= 1 2 IP {X > x},
and on the other hand, IP {|X| x} = 1 ex . We obtain for x > 0,
IP {X > x} =

1 x
e
.
2

The c.d.f. of X is hence given by

{ 1 |x|
e
,
x < 0,
IP {X x} = 2 1 x
1 2e
, x > 0.
Dierentiating w.r.t. x yields the p.d.f. which is given by
fX (x) =
10

1
e|x| ,
2

for x IR.
Remark. The above distribution is said to have a Laplacian distribution with parameter . It is sometimes called the double exponential
distribution with parameter .
4. Suppose that if you are s minutes early for an appointment, then you
incur a cost cs, and if you are s minutes late, then you incur a cost
ks. Suppose the travelling time from where you are to the location
of your appointment is a continuous random variable having p.d.f. f .
Determine the time at which you should depart if you want to minimize
your expected cost.
Sol. Let X be your travel time. If you leave t minutes before your
appointment, then the cost, call it Ct (X), is given by
{
c(t X), if X t,
Ct (X) =
k(X t), if X > t.
Therefore,

E [Ct (X)] =

Ct (x) f (x) dx

c(t x) f (x) dx +
k(x t) f (x) dx
0
t
t
t

= ct
f (x) dx c
x f (x) dx + k
x f (x) dx
0
0
t

kt
f (x) dx
t
t

= ct F (t) c
x f (x) dx + k
x f (x) dx
=

kt [1 F (t)].
The value of t which minimizes E [Ct (X)] is obtained by calculus. Differentiation yields
dE
E [Ct (X)]
= c F (t) + ct f (t) c tf (t) kt f (t) k [1 F (t)] + kt f (t)
dt
= (k + c) F (t) k.
Equating to 0, the minimal expected cost is obtained when you leave t
minutes before your appointment, where t satises
F (t ) =
11

k
,
k+c

k
that is, t = F 1 ( k+c
) if F 1 exists.

Question. How do we know that this t gives us a minimum and not

a maximum?
5. A stick of length 1 is broken at random. Determine the expected length
of the piece that contains the point P of length p from one end, call A,
where 0 p 1.
Sol. Let U be the point (measuring from A) where the stick is broken
into two pieces. Then, U is uniformly distributed, i.e., the p.d.f. of U is
given by
{
1, for 0 < u < 1,
fU (u) =
0, otherwise.
Let Lp (U ) denote the length of the piece which contains the point P ,
and note that
{
1 U, if U < p,
Lp (U ) =
U,
if U > p.
Hence,

E [Lp (U )] =

Lp (u) fU (u) du

Lp (u) 1 du

=
0

(1 u) du +

u du
p

1
+ p(1 p).
2

6. Buses arrive at a specied stop at 15-minute intervals starting at 7 am.

That is, they arrive at 7, 7:15, 7:30, 7:45, and so on. If a passenger
arrives at the stop at a time that is uniformly distributed between 7 and
7:30, nd the probability that he waits
(i) less than 5 minutes for a bus;
(ii) more than 10 minutes for a bus.
Sol. Let X denote the arrival time of the passenger (after 7 in minutes).
Then,
X U (0, 30).
(i) Probability that the passenger waits less than 5 minutes for a bus
when and only when he arrives (a) between 7:107:15 or (b) 7:25
12

7:30. Therefore the desired probability is given by

15 10 30 25
+
30
30
1
= .
3

IP (10 < X < 15) + IP (25 < X < 30) =

(ii) Probability that the passenger waits more than 10 minutes for a bus
when and only when he arrives (a) between 7:007:05 or (b) 7:15
7:20. And so the desired probability is
IP (0 < X < 5) + IP (15 < X < 20) =

13

20 15
1
5
+
= .
30
30
3