Statistics 351 (Fall 2015) September 28, 2015

Prof. Michael Kozdron

Lecture #9: The Gamma Function

Suppose that p > 0, and define

Z 1
(p) := up 1
e u
du.
0

We call (p) the Gamma function and it appears in many of the formulæ of density functions
for continuous random variables such as the Gamma distribution, Beta distribution, Chi-
squared distribution, t distribution, and F distribution.
The first thing that should be checked is that the integral defining (p) is convergent for
p > 0. For now, we will assume that it is true that the Gamma function is well-defined. This
will allow us to derive some of its important properties and show its utility for statistics.
The Gamma function may be viewed as a generalization of the factorial function as this first
result shows.

Proposition 1. If p > 0, then (p + 1) = p (p).

Proof. This is proved using integration by parts from first-year calculus. Indeed,
Z 1 Z 1 1 Z 1
p+1 1 u p u p u
(p + 1) = u e du = u e du = u e + pup 1 e u du = 0 + p (p).
0 0 0 0

To
R do the integration
R by parts, let w = up , dw = pup 1 , dv = e u
,v= e u
and recall that
w dv = wv v dw.

If p is an integer, then we have the following corollary.

Corollary 2. If n is a positive integer, then (n) = (n 1)!.

Proof. Using the previous proposition, we see that

(n) = (n 1) (n 1) = (n 1)(n 2) (n 2) = · · · = (n 1)(n 2) · · · 2 · (1).

However, Z Z
1 1 1
0 u u u
(1) = ue du = e du = e =1 (1)
0 0 0
and so
(n) = (n 1)(n 2) · · · 2 · 1 = (n 1)!
as required.

The next proposition shows us how to calculate (p) for certain fractional values of p.

9–1
 p
Proposition 3. (1/2) = ⇡.

Proof. By definition, Z 1
1/2 u
(1/2) = u e du.
0

Making the substitution u = v 2 so that du = 2v dv gives

Z 1 Z 1 Z 1 Z 1
1/2 u 1 v2 v2 v2
u e du = v e 2v dv = 2 e dv = e dv
0 0 0 1

2
where the last equality follows since e v is an even function. We now recognize this as the
density function of a N (0, 1/2) random variable. That is,
Z 1
1 v2
p e 2 2 dv = 1
2⇡ 1
and so Z 1
v2 p
e 2 2 dv = 2⇡.
1
2
Choosing = 1/2 gives Z 1 p
v2
e dv = ⇡
1
p
and so we conclude that (1/2) = ⇡ as claimed.

This proposition can be combined with Proposition 1 to show, for example, that
p
⇡
(3/2) = (1/2 + 1) = 1/2 · (1/2) =
2
and p
3 ⇡
(5/2) = (3/2 + 1) = 3/2 · (3/2) = .
4
For students, though, perhaps the most powerful use of the Gamma function is to compute
integrals such as the following.
Example 4. Suppose that Y ⇠ Exp(✓). Use Gamma functions to quickly compute E(Y 2 ).
Solution. By definition, we have
Z 1 Z 1
1
E(Y ) =
2 2
y fY (y) dy = y2e y/✓
dy.
1 ✓ 0

Make the substitution u = y/✓ so that dy = ✓du. This gives

Z Z Z 1
1 1 2 y/✓ 1 1 2 2 u 2
y e dy = ✓ u e ✓ du = ✓ u2 e u
du = ✓2 (3).
✓ 0 ✓ 0 0

By Corollary 2, (3) = (3 1)! = 2 and so E(Y 2 ) = 2 ✓2 .

9–2
Example 5. If Y ⇠ Exp(✓), then this method can be applied to compute E(Y k ) for any
positive integer k. Indeed,
Z Z
1 1 k y/✓ 1 1 k k u
E(Y ) =
k
y e dy = ✓ u e ✓ du = ✓k (k + 1) = k! ✓k .
✓ 0 ✓ 0

Example 6. If X 2 N (0, 1), determine the distribution of X 2 .

Solution. Let Y = g(X) = X 2 so that

FY (y) = P {Y  y} = P {X 2  y}.

Notice that g(x) = x2 is NOT monotone on 1 < x < 1. Therefore,

Z p
y
2 p p
P {X  y} = P { yX y} = p
fX (x) dx
y
Z 0 Z p
y
= p
fX (x) dx + fX (x) dx
y 0
Z p
y Z p
y
= fX (x) dx fX (x) dx,
0 0

and so
p 1 p 1 1 p p
fY (y) = FY0 (y) = fX ( y) · p fX ( y) · p = p [fX ( y) fX ( y)] .
2 y 2 y 2 y

Since g(x) = x2 is two-to-one we have two terms. In particular,

1 y/2 1 1 y/2 1 1 1 y/2 1
fY (y) = p e · p +p e · p =p p e = y 1/2 1 e y/2
, y > 0,
2⇡ 2 y 2⇡ 2 y 2⇡ y 21/2 (1/2)
2
i.e., Y 2 (1/2, 2) or Y 2 (1).

Remark. We will not cover many-to-one functions beyond this example. Therefore, omit
everything following Example 2.6 on page 23 (including Example 2.7 on page 24).

Theorem 7. For p > 0, the integral

Z 1
up 1
e u
du
0

is absolutely convergent.

Proof. Since we are considering the value of the improper integral

Z 1
up 1 e u du
0

for all p > 0, there is need to be careful at both endpoints 0 and 1.

9–3
We begin with the easiest case. If p = 1, then
Z 1 Z 1 Z N
0 u u u N
u e du = e du = lim e du = lim (1 e ) = 1.
0 0 N !1 0 N !1

For the remaining cases 0 < p < 1 and p > 1 we will consider the integral from 0 to 1 and
the integral from 1 to 1 separately.
If 0 < p < 1, then the integral Z 1
up 1
e u
du
0
is improper. Thus,
Z 1 Z 1 Z 1
p 1 u p 1 u 1 ap 1
u e du = lim u e du  lim up 1
du = lim =
0 a!0+ a a!0+ a a!0+ p p
u
since e  1 for 0  u  1.
Furthermore, if 0 < p < 1, then 0 < up 1  1 for u 1 and so
Z 1 Z N Z N
p 1 u p 1 u
u e du = lim u e du  lim e u du = lim (1 e N
) = 1.
1 N !1 1 N !1 1 N !1

Thus, we can conclude that for 0 < p < 1,

Z 1 Z 1 Z 1
p 1 u p 1 u 1
u e du = u e du + up 1
e u
du  + 1 < 1.
0 0 1 p

If p > 1, then up 1
2 [0, 1] and e u  1 for 0  u  1. Thus,
Z 1 Z 1 1
p 1 u up 1
u e du  up 1 du = = .
0 0 p 0 p

bpc 1
On the other hand, if p > 1, then notice that p bpc 2 [0, 1) so that 0 < up  1 for
u 1. We then have
Z N Z N Z N
p bpc 1 bpc
p 1 u
u e du = u u
u e du  ubpc e u du.
1 1 1

Thus, integration by parts bpc times (the so-called reduction formula) gives
Z N
ubpc e u du
1
N
= e u
ubpc + bpcubpc 1
+ bpc · (bpc 1)ubpc 2
+ · · · + bpc · (bpc 1) · · · 2 · u
1
Z N
u
+ bpc · (bpc 1) · · · 2 · 1 · e du
1

9–4
and so Z N
lim ubpc e u
du = bpc !.
N !1 1
Thus, we can conclude that for p > 1,
Z 1 Z 1 Z 1
p 1 u 1
u e du = up 1 e u
du + up 1
e u
du  + bpc ! < 1.
0 0 1 p

In every case we have up 1

e u
0 and so
Z 1 Z 1
p 1 u
u e du = up 1
e u
du < 1.
0 0

That is, this integral is absolutely convergent, and so (p) is well-defined for p > 0.

9–5