081002 多變量分析 Homework 01

Extra
Let X1 , X 2 ,..., X n be a random sample from a normal population
with mean  and covariance . Show their mle.
sol:
Assume that the p 1 vector X1 , X 2 ,..., X n are iid and each has N p (  , ).
n  
 Joint density   1  ( x j   )' 1 ( x j   ) / 2 
    1/ 2
e 
 of X1 , X 2 ,..., X n  j 1  (2 ) 
p/2
 
n

1   ( x j   )' 1 ( x j   ) / 2
 n/2
e j 1

(2 ) np / 2

 n 
1  tr  1

 ( x j   )( x j   )' 2
 n/2
e  j 1 

(2 ) np / 2 
 n 
1  tr  1

 ( x j  x )( x j  x )' n ( x   ) 1 ( x   )' 2
 n/2
e  j 1 

(2 ) np / 2 
And consider the Result 4.10.
Given a p  p symmetric positive definite matrix B and a scalar b  0, it follow that

1 1 1
b
e  tr(  B) / 2
 b
(2b)bp e bp
 B

for all postive define  p p , with equality hold only for  = (1/2b)B.

 n 
1  tr   1

 ( x j  x )( x j  x )' n ( x   ) 1 ( x   )' 2
L(  , )  n/2
e  j 1 

(2 ) np / 2 
Since  -1 is positive definite, so ( x   ) 1 (x   ) '  0 for all (x   )  0.
Thus take ˆ = x to maximize L(  , ).
 n 
1  tr  1

 ( x j  x )( x j  x )' 2
To maximeize L( ˆ , )  n/2
e  j 1 
over ,
(2 ) np / 2 
n
just by Result 4.10 with b  n / 2 and B   ( x j  x )( x j  x) ',
j 1
n
1
the maximun occurs at ˆ 
n
 (x
j 1
j  x)( x j  x ) '.
081002 多變量分析 Homework 01

4.9.
Let X 1 be N (0,1) and let

  X 1 if - c  X 1  c
X2  
 X 1 elsewhere

Show that c can be chosen so that Cov(X1 , X 2 )  0, but that the two random
variables are not independent.
Sol:
Clearly X2 dependence on X1.
Since X 1 be N (0,1) which is a synnetric distribution, then E ( X 1 )  0, E ( X 2 )  0.
Thus we have Cov( X 1 , X 2 )  E ( X 1 X 2 ).
 Cov( X 1 , X 2 )  E ( X 12 )  1 as c  0.
 Cov( X 1 , X 2 )  E (  X 12 )  1 as c  .

Let  (z) is the standard normal density, and for all c  0

H ( c )  Cov( X 1 , X 2 ) c c
c c 
  z 2 ( z )dz    z 2 ( z )dz   z 2 ( z )dz.
 c c
c
 1  4  z 2 ( z )dz
0

 H (c ) is continous for all c  0.

Thus we found that Cov( X 1 , X 2 ) bound on  -1,1 , continous for all c  0,

 Let N is large s.t. if c  N then H ( c )  -1.
 c  (0, N )  Cov( X 1 , X 2 )  0.

4.16
Let X1 , X 2 , X3 and X 4 be independent N p (  , ) random vector.
(a) Find the marginal distributions for each of the random vectors
1 1 1 1
V1  X1  X 2  X3  X 4
4 4 4 4
1 1 1 1
V2  X1  X 2  X3  X 4
4 4 4 4
(b) Find the joint density of the random vectors V1 and V2 .
081002 多變量分析 Homework 01

Sol:
By Result 4.8,
(a)
1 1 1 1 1
V1  X1  X 2  X3  X4 is distributed as N p (0, )
4 4 4 4 4
1 1 1 1 1
V2  X1  X 2  X3  X 4 is distributed as N p (0, )
4 4 4 4 4
(b)
 1 
 4 
 
 1 
 1 1 1 1  4 
 ,  , ,   1   0
 4 4 4 4   
 4
 1
  
 4
  1  
  0   4  0 
f ( V1 , V2 ) ~ N   ,  
  0   0 1  
  
  4  

4.24
Exercise 1.4 contains data for world’s 10 largest companies as of April 2005.

For the sales ( x1 ) and profits ( x2 ) data:

(a) Construct Q-Q plots. Do these data appear to be normally distributed?

(b)Carry out a test of normality based on the correlation coefficient rQ. Set the
significance level at  = .10. Do the result of these tests corroborate the
result in Part (a) ?
Sol:

(a) From Q-Q plot profits ( x2 ) might form normal distribution but sales ( x1 ) may

not be.
081002 多變量分析 Homework 01

Normal Q-Q Plot for x1 Normal Q-Q Plot for x2

25
250

20
200
Sample Quantiles

Sample Quantiles
150

15
100

10
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

Theoretical Quantiles Theoretical Quantiles

(b)
For x1 , rQ  0.9372
By Table 4.2 Sample size = 10, significance level   .10
Q 0.9372  0.9351  Do not reject the hypothesis of normality at level   .10
 x1 might form normal distrbution.

For x2 , rQ  0.9692
By Table 4.2 Sample size = 10, significance level   .10
Q 0.9692  0.9351  Do not reject the hypothesis of normality at level   .10
 x2 might form normal distrbution.

4.30
Consider the used-car data in Exercise 4.26

(a) Determine the power transformation ̂1 that makes the x1 values

approximately normal. Construct a Q-Q plot for the transformed data.

(b)Determine the power transformation ̂2 that makes the x2 values

approximately normal. Construct a Q-Q plot for the transformed data.

(c) Determine the power transformation λ   1 , 2  that makes the x   x1 , x2 

ˆˆ

values jointly normal using (4-40). Compare the results with those obtained
in Parts a and b.
Sol:

(a) By Box-Cox transform, take ̂1 =0.37. Thus the transform data ( x1 ) would

approximately normal.
081002 多變量分析 Homework 01

Normal Q-Q Plot for x1 after transform

-10

4
-12

3
-14

Sample Quantiles
-16

2
L1

-18
-20

1
-22

lambda=0.37

0
-24

-2 -1 0 1 2 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

lambda Theoretical Quantiles

(b)By Box-Cox transform, ̂2 = 0.94 which is too close to 1. Thus we don’t do

any transform for data ( x2 ) , however the raw data would approximately

normal.
Normal Q-Q Plot for x2
-18

15
Sample Quantiles
-20
L2

10
-22
-24

lambda=0.94
5

-2 -1 0 1 2 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5

lambda Theoretical Quantiles

(c) Because marginal distributions is normal which don’t mean the joint
distribution still form normal. By Box-Cox transform, we found that
λˆ   1.27, 0.03 which is different with part (a) and (b). But the chi-square

plot for the transform by (a) & (b), would better than transform by (c). Thus
we consider the contour plot, it could be found that (0.37, 0.94) also fall in

the top region. It means that λ   0.37, 0.94  could be a nice choice too.
ˆ
081002 多變量分析 Homework 01

Chi-square plot for x1 & x2 transform by (a) & (b) Chi-square plot for x1 & x2 transform by (c)

6
5

5
4

4
Y.b

Y.a
3

3
2

2
1

1
0

0
0 1 2 3 4 5 6 0 1 2 3 4 5 6

q q
2

-30
1
lambda 2

-25

-15
-1

-20

-3 -25
5 -30
-2

-1 0 1 2 3

lambda 1

4.35 Examine the data on paper-quality measurements in Table 1.8 for marginal and
multitivariate normality.
Marginal normal Q-Q plot
Normal Q-Q Plot for Density Normal Q-Q Plot for Machine direction Normal Q-Q Plot for Cross direction
135

80
0.95

130

75
125
0.90

70
Sample Quantiles

Sample Quantiles
Sample Quantiles

120

65
0.85

115

60
110
0.80

55
105

50
0.75

-2 -1 0 1 2 -2 -1 0 1 2 -2 -1 0 1 2

Theoretical Quantiles Theoretical Quantiles Theoretical Quantiles

(a) Density (b) Machine direction (c) Cross direction

Chi-square plot
Chi-square plot for Density & Machine direction Chi-square plot for Machine direction & Cross direction Chi-square plot for Density & Cross direction
30
25

25
20

20
chi-square quantile

chi-square quantile

chi-square quantile
15

15
4
10

10
2
5

5
0

0 5 10 15 20 25 0 2 4 6 8 0 5 10 15 20 25 30

Mahalanobis distance Mahalanobis distance Mahalanobis distance

(d) Density & Machine direction (e) Machine & Cross direction (f) Density & Cross direction
Chi-square plot
081002 多變量分析 Homework 01

Chi-square plot for Density & Machine & Cross direction

30
25
chi-square quantile

20
15
10
5
0
0 5 10 15 20 25 30

Mahalanobis distance

(g) Density & Machine direction & Cross direction

By the above plots (b), (e) would be normal. However if we detect outlier, than
plot (a), (d), (f), (g) might be normal.