Statistics 351 (Fall 2015)

Prof. Michael Kozdron

September 16, 2015

Lecture #4: Functions of Random Variables

Reference. 1.2 pages 1924

Example. Let X 2 Exp(1/ 1 ) and Y 2 Exp(1/ 2 ) be independent. Compute P {X < Y }.

Solution. Since X and Y are independent, we know that the joint density fX,Y (x, y) is
simply the product of the marginal densities fX (x) and fY (y). That is,
fX,Y (x, y) = fX (x) fY (y) =

1x

1e

2y

2e

1x

1 2e

2y

provided that x > 0 and y > 0. Recall that in the one-dimensional case, if we want to
compute the probability of an event such as P {a < X < b} we simply need to integrate the
density function for X over the appropriate region. That is,
Z b
Z b
1x
P {a < X < b} =
fX (x) dx =
dx = e a 1 e b 1 .
1e
a

In the multi-dimensional case, we do the same thing. That is, we integrate the joint density
over the appropriate region:
ZZ
ZZ
P {X < Y } =
fX,Y (x, y) dx dy = 1 2
e 1 x 2 y dx dy.
{x<y}

{x<y}

There are two equivalent ways to compute this iterated integral. The first is
ZZ
Z 1Z y
1x
2y
e
dx dy = 1 2
e 1 x 2 y dx dy
1 2
{x<y}

=
=
=

Z
Z
Z

1
0

1
1

2e

2y

2e

2y

2e

2y

2
1

{x<y}

2y

dx dy =

1 2

1
0

1x

+
2y

1y

dy

2
1
1

dy dx =

=
=
=
41

dx dy

=
2

1x

1e

=1
The second is
ZZ
e 1x
1 2

Z
Z

dy

y(

1+ 2)

dy

.
2
1

1x

1e

0
1
0

1x

1e

x(

1e

0
1
1+

.
2

Z
e

2e

x
2x

1+ 2)

dx

dx

2y

dy dx

In either case, we find

1

P {X < Y } =
Notice that if

2,

1+

then P {X < Y } = 1/2 as expected. (Why?)

Motivation. Sometimes we are interested in the distribution of a function of a random

variable.
Stat 252 Motivation. The pivot method for constructing confidence intervals required one
to do exactly that. We did, however, restrict ourselves to strictly increasing functions of one
dimensional random variables.
Example. Suppose that X 2 Exp( ). Determine the distribution/density of Y = eX .
Solution. If X 2 Exp( ), then
FX (x) = 1

x/

x > 0,

and so fX (x) =

x/

x > 0.

Therefore,
FY (y) = P {Y y} = P {eX y} = P {X log y}
Z log y
1 x/
=
e
dx
0

log y

x/

=1

=1

0
log y/
1/

y > 1.

We now find fY (y).

Method #1:
fY (y) =

d
d
FY (y) =
(1
dy
dy
1
= y 1

Method #2:
d
d
fY (y) =
FY (y) =
dy
dy
1
= e
=

log y

x/

y
1/

1/

dx

log y/

1/

42

d
(log y) (chain rule)
dy

1
as above.
y

Remark. We observe that Method #2 can be generalized to any strictly increasing function
g provided that its derivative g 0 exists.
Theorem and Proof. If X is a continuous random variables and g is a strictly increasing,
dierentiable function, then if Y = g(X),
FY (y) = P {Y y} = P {g(X) y} = P {X g 1 (y)} = FX (g 1 (y)), and
Z g 1 (y)
d
d
d 1
fY (y) =
FY (y) =
fX (x) dx = fX (g 1 (y))
g (y).
dy
dy 1
dy
On the other hand, if g is strictly decreasing, then
fY (y) =
(The extra minus sign is needed since

fX (g 1 (y))
d
g 1 (y)
dy

d 1
g (y).
dy

< 0.)

Summary. If g is strictly monotone, then

fY (y) = fX (g 1 (y))

d 1
g (y) .
dy

Exercise. Read Examples 2.1, 2.2 on pages 1920. Although Gut [1] skips some steps, you
can now fill them in.
Remark. When you need to change variables, dont try to just plug into a memorized
formula. Instead, follow Method #2 directly as in the example.
Problem. You can now solve Problems #1, #2, #3, #4, #6, #7 on page 24 of [1]. (Note
that Problem #6 is very easy and does not require this result.)
Remark. We will examine the multi-dimensional case next class.

43