Module 30

DISTRIBUTIONS BASED ON SAMPLING FROM A

NORMAL DISTRIBUTION

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

1 / 30
 We will introduce two new probability distributions, called the
Student t-distribution and the Snedecor F -distribution, which arise as
probability distributions of various statistics based on a random
sample from normal distribution.

Definition 1
(i) For a given positive integer m, a random variable X is said to have
the Student t-distribution with m degrees of freedom (written as
X ∼ tm ) if the p.d.f. of X is given by
 
m+1
Γ 2 1
fX (x) = √  m+1 , −∞ < x < ∞.
mπΓ( m2 )  2 2
1 + xm

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

1 / 30
 (ii) The Student t-distribution with 1 degree of freedom is also called
the standard Cauchy distribution.

(iiii) For positive integers n1 and n2 , a random variable X is said to

have the Snedecor F -distribution with (n1 , n2 ) degrees of freedom
(written as X ∼ F (n1 , n2 )) if the p.d.f. of X is given by
n1
( nn12 x) 2 −1
( nn12 )
fX (x) =    n1 +n2 I(0,∞) (x).
n1 n2 2
β 2, 2 1 + nn12 x

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

2 / 30
Remark 1

The following observations are obvious:

d
(i) X ∼ tm ⇒ X = −X (since fX (x) = fX (−x), ∀ x ∈ R),
i.e., the distribution of X ∼ tm is symmetric about 0. Moreover the
distribution of tm is unimodal with mode at 0.;
(ii) The p.d.f. of Cauchy distribution is given by

1 1
f (y ) = · , −∞ < y < ∞.
π 1 + y2

If a random variable X has the Cauchy distribution (i.e., if X ∼ t1 )

then E (X ) does not exist;
n1
X
(iii) X ∼ Fn1 ,n2 , ⇒ Y = n2
n1
1+ n X
∼ Beta( n21 , n22 ), the beta distribution with
2
n1 n2
shape parameter ( 2 , 2 ).

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

3 / 30
Result 1:
(i) Let Z ∼ N(0, 1) and Y ∼ χ2m (where m ∈ {1, 2, . . .}) be independent
random variables. Then
Z
T = q ∼ tm .
Y
m

(ii) For positive integers n1 and n2 , let X1 ∼ χ2n1 and X2 ∼ χ2n2 be

independent random variables. Then
X1 /n1
U= ∼ Fn1 ,n2 .
X2 /n2
(iii) Let m and r be positive integers and let X ∼ tm . Then E (X r ) is not
finite if r ∈ {m, m + 1, . . .}. For r ∈ {1, 2, . . . , m − 1} and m ≥ r + 1

0, if r is odd
r
E (X ) = r
m 2 r !Γ( m−r ) .
2 ( )!Γ( m ) , if r is even
 r r 2
2 2

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

4 / 30
(iv) If X ∼ tm , then
0
µ1 = E (X ) = 0, for m ∈ {2, 3, . . .}
m
µ2 = Var(X ) = , for m ∈ {3, 4, . . .}
m−2
β1 = coefficient of skewness = 0, for m ∈ {4, 5, . . .}
3(m − 2)
and γ1 = kurtosis = , for m ∈ {5, 6, . . .}.
m−4
(v) Let n1 , n2 and r be positive integers and let X ∼ Fn1 ,n2 . Then, for
n2 ∈ {1, 2, . . . , 2r } and r ≥ n22 , E (X r ) is not finite. For
n2 ∈ {2r + 1, 2r + 2, . . .}
 r 
r Y 
n2 n1 + 2(i − 1)
E (X r ) = .
n1 n2 − 2i
i=1

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

5 / 30
(vi) If X ∼ Fn1 ,n2 then
0 n2
µ1 = E (X ) = , if n2 ∈ {3, 4, . . .}
n2 − 2
2n22 (n1 + n2 − 2)
µ2 = Var(X ) = , if n2 ∈ {5, 6, . . .}
n1 (n2 − 2)2 (n2 − 4)
β1 = coefficient of skewness
s
2(2n1 + n2 − 2) 2(n2 − 4)
= , if n2 ∈ {7, 8, . . .}
n2 − 6 n1 (n1 + n2 − 2)
and γ1 = kurtosis
12[(n2 − 2)2 (n2 − 4) + n1 (n1 + n2 − 2)n1 (5n2 − 22)]
= + 3,
n1 (n2 − 6)(n2 − 8)(n1 + n2 − 2)
if n2 ∈ {9, 10, . . .}.

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

6 / 30
Proof :

(i) The joint p.d.f. of (Y , Z ) is given by

fY ,Z (y , z) = fY (y )fZ (z)
 2
 m+1 1 − y +z m
−1
m √ e 2 y 2 , if (y , z) ∈ (0, ∞) × R
= 2 2 Γ( 2 ) π .
0, otherwise


Clearly SY ,Z = (0, ∞] × R. Consider the transformation q

y
h = (h1 , h2 ) : SY ,Z → R2 defined by h1 (y , z) = √z y and h2 (y , z) = m .
m
q
Then T = h1 (Y , Z ) = qZY . Let U = h2 (Y , Z ) = Y m.
m

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

7 / 30
Clearly the transformation h = (h1 , h2 ) : SY ,Z → R2 is one-to-one with
inverse transformation h−1 = (h1−1 , h2−1 ), where for (t, u) ∈ h(SY ,Z ),

h1−1 (t, u) = mu 2 and h2−1 (t, u) = tu.

The Jacobian determinant is

 −1
∂h1−1 
 
 ∂h1 
 0 2mu 
= −2mu 2 .
 ∂t ∂u
J =  ∂h−1 ∂h−1  = 
 2 2  u t 
∂t ∂u

Also

h(SY ,Z ) = {(t, u) ∈ R2 : (h1−1 (t, u), h2−1 (t, u)) ∈ SY ,Z }

= {(t, u) ∈ R2 : mu 2 ∈ [0, ∞), u > 0, tu ∈ R}
= {(t, u) ∈ R2 : t ∈ R, u > 0}
= R × (0, ∞)
= A, say.

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

8 / 30
Therefore the joint p.d.f. of (T , U) is given by

fT ,U (t, u) = fY ,Z (h1−1 (t, u), h2−1 (t, u))|J|Ih(SY ,Z ) (t, u)

= fY ,Z (mu 2 , tu)| − 2mu 2 |IA (t, u)
m (m+t 2 )u 2

m2 m − , if (t, u) ∈ R × (0, ∞)
√ m−1 m u e 2

= π2 2 Γ( 2 ) .
0, otherwise


Consequently the p.d.f. of T is given by

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

9 / 30
 Z∞
fT (t) = fT ,U (t, u)du
−∞
m Z∞
m2 (m+t 2 )u 2
= √ m−1 um e − 2 du, t ∈ R
π2 2 Γ( m2 )
0
Z∞
1 m−1
= √
m+1 y 2 e −y dy
t2
mπΓ( m2 )
1+ m 0
2

 
Γ m+1
2 1
= √ m · , t ∈ R,
mπΓ( 2 ) 1 + t 2
m+1
2
m

which is the p.d.f. of Student’s t-distribution with m degrees of freedom.

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

10 / 30
 (ii) The joint p.d.f. of X = (X1 , X2 ) is given by

fX1 ,X2 (x1 , x2 ) = fX1 (x1 )fX2 (x2 )

n1 n2
1 x +x
− 1 2 2 2 −1 2 −1
= n1 +n2 e x1 x2 I(0,∞)2 (x1 , x2 ).
2 2 Γ( n21 )Γ( n22 )

We have SX1 ,X2 = [0, ∞)2 . Consider the one-to-one transformation

h = (h1 , h2 ) : SX1 ,X2 → R2 given by
n2 x1 x2
h1 (x1 , x2 ) = and h2 (x1 , x2 ) = .
n1 x2 n2
X1 /n1 X2
Define U = h1 (X1 , X2 ) = X2 /n2 and V = h2 (X1 , X2 ) = n2 . Then the
inverse of transformation h = (h1 , h2 ) : SX1 ,X2 → R is h −1
= (h1−1 , h2−1 ),
where for (u, v ) ∈ h(SX1 ,X2 ),

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

11 / 30
 h1−1 (u, v ) = n1 uv and h2−1 (u, v ) = n2 v .
The Jacobian determinant is
 −1
∂h1−1 
 
 ∂h1 
 n v n1 u 
J =  ∂h∂u−1 ∂h∂v−1  =  1 = n1 n2 v .

 2 2  0 n2 
∂u ∂v

Also

h(SX1 ,X2 ) = {(u, v ) ∈ R2 : (h1−1 (u, v ), h2−1 (u, v )) ∈ SX1 ,X2 }

= {(u, v ) ∈ R2 : n1 uv > 0, n2 v > 0}
= {(t, u) ∈ R2 : t > 0, u > 0}
= (0, ∞)2 ,

and therefore, the joint p.d.f. of (U, V ) is given by

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

12 / 30
fU,V (u, v ) = fX1 ,X2 (h1−1 (u, v ), h2−1 (u, v ))|J|Ih(SX ) (u, v )
1 ,X2

= fX1 ,X2 (n1 uv , n2 v )|n1 n2 v |I(0,∞]2 (u, v )

n1 n2
n12 n22 n1 n +n (n +n u)v
−1 1 2 2 −1 − 2 2 1
= ( n1 +n2 n1 n2 u
2 v e I(0,∞)2 (u, v ).
2 2 )Γ( 2 )Γ( 2 )

Consequently the p.d.f. of U is given by

Z∞
fU (u) = fU,V (u, v )dv .
−∞

Clearly fU (u) = 0, if u ≤ 0. For u > 0

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

13 / 30
 n1 n2
Z∞
n12 n22 n1
−1
n1 +n2
−1
(n2 +n1 u)v
fU (u) = n1 +n2 n n u 2 v 2 e− 2 dv
Γ( 21 )Γ( 22 )
2 2
0
n1
n1 +n2 −1

n
Γ 1
( nn12 u) 2
= n1

2 n2
n2

n1 +n2 , 0 < u < ∞.
Γ 2 Γ (1 + n1
n2 u)
2 2

Therefore
X1 /n1
U= ∼ Fn1 ,n2 .
X2 /n2

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

14 / 30
 (iii) For m ∈ {1, 2, . . .}, by (i),
d Z
X =q ,
Y
m

where Z ∼ N(0, 1) and Y ∼ χ2m are independent random variables. Thus,

for m ∈ {1, 2, . . .} and r > 0,
r r r r
E (X r ) = m 2 E (Z r Y − 2 ) = m 2 E (Z r )E (Y − 2 ),

(since Y and Z are independent)

provided the expectations are finite. We have,

0, if r is odd
E (Z r ) = r!
.
 2r r , if r is even
2 2
!

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

15 / 30
Moreover, for r ∈ {1, 2, . . .},
Z∞
− 2r 1 m−r
−1 − y2
E (Y )= m
m

y 2 e dy ,
2 Γ
2
2 0

which is finite if, and only if, m > r . Also, for m > r
m−r
m−r m−r



− 2r 2 2 Γ 2 Γ 2
E (Y )= m
m
= r
.
2 2 Γ m2


22Γ 2
Thus E (X r ) is finite if r ∈ {1, 2, . . . , m − 1}. For r ∈ {1, 2, . . . , m − 1}
and m ≥ r + 1

0,


if r is odd
r r
E (X ) = m 2 r !Γ m−r2
.
 2r r
!Γ m
, if r is even

2 2

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

16 / 30
 (iv) Using (iii), we have

0
µ1 = E (X ) = 0, if m ∈ {2, 3, . . .}
0 m
µ2 = µ2 = E (X 2 ) = , if m ∈ {3, 4, . . .}
m−2
0
µ3 = µ3 = E (X 3 ) = 0 if m ∈ {4, 5, . . .}
0 3m2
and µ4 = µ4 = E (X 4 ) = , if m ∈ {5, 6, . . .}.
(m − 2)(m − 4)

Consequently
µ3
β1 = 3/2
= 0, if m ∈ {4, 5, . . .}
µ2
and
µ4 3(m − 2)
γ1 = 2
= , if m ∈ {5, 6, . . .}.
µ2 m−4

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

17 / 30
(v) Using (ii), we have

d n2 X1
X = ,
n1 X2
where X1 ∼ χ2n1 and X2 ∼ χ2n2 are independent random variables. Fix
r ∈ {1, 2, . . .}. Then
 r  r
n2 r −r n2
r
E (X ) = E (X1 X2 ) = E (X1r )E (X2−r ),
n1 n1
(since X1 and X2 are independent)
provided the expectations are finite. Since X1 ∼ χ2n1 , E (X1r ) is finite for
any r > 0 and
Z∞
1 n1 x
r
E (X1 ) = n1 x 2 +r −1 e − 2 dx
2 2 Γ n21


0
n1
+r n


2 2 Γ 21 + r
= n1
2 2 Γ n21

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

18 / 30
 n1
n1 n1
= 2r


+r −1 + r − 2 ...
2 2 2
= (n1 + 2(r − 1))(n1 + 2(r − 2)) . . . n1
Yr
= (n1 + 2(i − 1)), r ∈ {1, 2, . . .}.
i=1

Since X2 ∼ χ2n2 , E (X2−r )

is finite if, and only if, n2 > 2r . For n2 > 2r
n2
2 2 −r Γ n22 − r


−r 1
E (X2 ) = n2 = r .
n


22Γ 2 2 Q
(n2 − 2i)
i=1
It follows that, for n2 ∈ {1, 2, . . . , 2r } and r ≥ n22 , E (X r ) is not finite. For
n2 ∈ {2r + 1, 2r + 2, . . .}
 r Y r  
n2 n1 + 2(i − 1)
E (X r ) = .
n1 n2 − 2i
i=1

(vi) Follows on using (v) after some tedious calculations.

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO
19 / 30
Corollary 1:

Let X1 , . . . , Xn (n ≥ 2) be a random sample from N(µ, σ 2 ) distribution,

n
where µ ∈ (−∞, ∞) and σ > 0. Let X̄ = n1
P
Xi and
i=1
n
1
S2 = (Xi − X̄ )2 denote the sample mean and the sample variance
P
n−1
i=1
respectively. Then
√
n(X̄ − µ)
∼ N(0, 1)
σ
and √
n(X̄ − µ)
∼ tn−1 .
S

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

20 / 30
Proof.

2 2
X̄ ∼ N(µ, σn ) and (n−1)S
σ2
∼ χ2n−1 are independent random variables.
√
n(X̄ − µ) (n − 1)S 2
⇒ ∼ N(0, 1) and ∼ χ2n−1
σ σ2
are independent random variables.
√
n(X̄ −µ)
σ
⇒q ∼ tn−1 ,
(n−1)S 2 /σ 2
n−1

i.e., √
n(X̄ − µ)
∼ tn−1 .
S

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

21 / 30
Corollary 2:

Let X1 , . . . , Xm (m ≥ 2) be a random sample from N(µ1 , σ12 ) distribution

and let Y1 , . . . , Yn (n ≥ 2) be a random sample from N(µ2 , σ22 )
distribution, where µi ∈ (−∞, ∞) and σi > 0, i = 1, 2. Further suppose
that X = (X1 , . . . , Xm ) and Y = (Y1 , . . . , Yn ) are independent. Let
m n
1 P 1 P
S12 = m−1 (Xi − X̄ )2 and S22 = n−1 (Yi − Ȳ )2 be the sample
i=1 i=1
variances based on random samples X = (X1 , . . . , Xm ) and
m n
Y = (Y1 , . . . , Yn ), respectively; here X̄ = m1 1
P P
Xi and Ȳ = n Yi are
i=1 i=1
the sample means based on two random samples. Then

σ22 S12
∼ Fm−1,n−1 .
σ12 S22

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

22 / 30
Proof :

We have
(m − 1)S12 2 (n − 1)S22
∼ χ m−1 and ∼ χ2n−1 .
σ12 σ22
(m−1)S12
Also the independence of X and Y implies that σ12
(a function of X
(n−1)S22
alone) and σ22
(a function of Y alone) are independent. Thus

S12 /σ12
∼ Fm−1,n−1 ,
S22 /σ22

i.e.,
σ22 S12
∼ Fm−1,n−1 .
σ12 S22

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

23 / 30
Remark 2 :

(i) Suppose that X ∼ tm . Then

d Z
X =q ,
Y
m

where Z ∼ N(0, 1) and Y ∼ χ2m are independent random variables.

Therefore
d Z2 χ2 /1 E
X2 = = 21 independent
Y /m χm /m
Thus
X ∼ tm , ⇒ X 2 ∼ F1,m .

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

24 / 30
(ii) Suppose that X ∼ Fn1 ,n2 . Then,

d χ2n1 /n1 E
X = independent
χ2n2 /n2

1 d χ2n2 /n2 E
⇒ = 2 independent ∼ Fn2 ,n1
X χn1 /n1
Thus,
1
X ∼ Fn1 ,n2 ⇒ ∼ Fn2 ,n1 .
X

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

25 / 30
Note that if X ∼ tm then, the distribution of X is symmetric about 0
and its kurtosis is
3(m − 2)
γ1 = > 3, m > 4.
m−4
Thus a t-distribution with m (> 4) degrees of freedom is symmetric
and leptokurtic (i.e., it has shaper peak and longer fatter tails).
Note that the kurtosis ν1 decreases as m increases and γ1 → 3, as
m → ∞. This suggests that, for large degrees of freedom, Studentś
t-distribution behaves like N(0, 1). distribution. This is infact true.

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

26 / 30
Suppose that X ∼ tm and, for a fixed α ∈ (0, 1), let tm,α be the
(1 − α)-th quantile of X , i.e.,

FX (tm,α ) = P(X ≤ tm,α ) = 1 − α.

Then
d
FX (−tm,α ) = 1 − FX (tm,α ) = α (since X = −X ).
Now suppose that X ∼ Fn1 ,n2 and, for a fixed α ∈ (0, 1), let fn1 ,n2 ,α
be the (1 − α)-th quantile of X , i.e.,

FX (fn1 ,n2 ,α ) = P({X ≤ fn1 ,n2 ,α }) = 1 − α.

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

27 / 30
 1
Since X ∼ Fn2 ,n1 and P({X > 0}) = 1, follows that
 
1 1
P ≥ =1−α
X fn1 ,n2 ,α
 
1 1
⇒P ≤ = α = 1 − (1 − α)
X fn1 ,n2 ,α
1
⇒ fn2 ,n1 ,1−α = .
fn1 ,n2 ,α
i.e.,
fn1 ,n2 ,α × fn2 ,n1 ,1−α = 1.

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

28 / 30
Take Home Problems

(1) Let Z1 and Z2 be i.i.d. N(0, 1) r.v.s. Show that

Z1 d Z1
= .
Z2 |Z2 |
Z1
Hence show that Z = Z2 follows Cauchy distribution (i.e., Z ∼ t1 ).
(2) Let X1 and X2 be i.i.d. N(µ, σ 2 ) r.v.s. Show that X1 + X2 and
X1 − X2 are independent. Find the p.d.f. of
X1 + X2 − 2µ
Y = √ .
2|X1 − X2 |

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

29 / 30
 Thank you for your patience

() Module 30 DISTRIBUTIONS BASED ON SAMPLING FROM A NORMAL DISTRIBUTIO

30 / 30