You are on page 1of 6

Jim Lambers

MAT 610
Summer Session 2009-10
Lecture 11 Notes
These notes correspond to Sections 5.4, 5.5 and 5.7 in the text.

The Rank-Decient Least Squares Problem


with Column Pivoting
When does not have full column rank, the property
span{a1 , . . . , a } = span{q1 , . . . , q }
can not be expected to hold, because the rst columns of could be linearly dependent, while
the rst columns of , being orthonormal, must be linearly independent.
Example The matrix

1 2 1
= 2 4 0
1 2 3

has rank 2, because the rst two columns are parallel, and therefore are linearly dependent, while
the third column is not parallel to either of the rst two. Columns 1 and 3, or columns 2 and 3,
form linearly independent sets.
Therefore, in the case where rank() = < , we seek a decomposition of the form = ,
where is a permutation matrix chosen so that the diagonal elements of are maximized at each
stage. Specically, suppose 1 is a Householder reection chosen so that

11
0

1 = .
, 11 = a1 2 .
.
.

0
To maximize 11 , we choose 1 so that in the column-permuted matrix = 1 , we have a1 2
a 2 for 2. For 2 , we examine the lengths of the columns of the submatrix of obtained by
removing the rst row and column. It is not necessary to recompute the lengths of the columns,
because we can update them by subtracting the square of the rst component from the square of
the total length.

This process is called with column pivoting. It yields the decomposition


[
]

=
0 0
where = 1 , = 1 , and is an upper triangular, matrix. The last
rows are necessarily zero, because every column of is a linear combination of the rst columns
of .
Example We perform with column pivoting on the matrix

1
3 5 1
2 1 2 1
.
=
1
4 6 1
4
5 10 1
Computing the 2-norms of the columns yields
a1 2 = 22,

a2 2 = 51,

We see that the third column has the largest


columns to obtain

0 0

0 1
(1) = 1 =
1 0
0 0

a3 2 = 165,

a4 2 = 4.

2-norm. We therefore interchange the rst and third



0
5
3 1 1
2 1 2 1
0
=
0 6
4 1 1
1
10
5 4 1

1
0
0
0

We then apply a Householder transformation 1 to (1) to make the rst column a multiple of e1 ,
which yields

12.8452 6.7729 4.2817 1.7905

0 2.0953
1.4080
0.6873
.
1 (1) =

0
0.7141 0.7759
0.0618
0 0.4765
1.0402 0.5637
Next, we consider the submatrix obtained by removing the rst row and column of 1 (1) :

2.0953
1.4080
0.6873
0.0618 .
(2) = 0.7141 0.7759
0.4765
1.0402 0.5637
We compute the lengths of the columns, as before, except that this time, we update the lengths of
the columns of , rather than recomputing them. This yields
(2)

(1)

(1)

(2)

(1)

(1)

(2)

(1)

(1)

a1 22 = a2 2 [12 ]2 = 51 (6.7729)2 = 5.1273,

a2 22 = a3 2 [13 ]2 = 22 (4.2817)2 = 3.6667,

a3 22 = a4 2 [14 ]2 = 4 (1.7905)2 = 0.7939.


2

The second column is the largest, so there is no need for a column interchange this time. We
2 to the rst column of (2) so that the updated column is a
apply a Householder transformation
multiple of e1 , which is equivalent to applying a 4 4 Householder transformation 2 = 2v2 v2 ,
where the rst component of v2 is zero, to the second column of (2) so that the updated column
is a linear combination of e1 and e2 . This yields

2.2643 1.7665 0.4978


2 (2) =
0 0.2559
0.2559 .

0
0.6933 0.6933
Then, we consider the submatrix obtained by removing the rst row and column of 2 (2) :
[
]
0.2559
0.2559
(3)

=
.
0.6933 0.6933
Both columns have the same lengths, so no column interchange is required. Applying a Householder
3 to the rst column to make it a multiple of e1 will have the same eect on the second
reection
column, because they are parallel. We have
]
[
0.7390 0.7390
(3)

.
3 =
0
0
It follows that the matrix (4) obtained by removing the rst row and column of 3 (3) will be
the zero matrix. We conclude that rank() = 3, and that has the factorization
]
[

,
=
0 0
where

0
0
=
1
0

0
1
0
0

1
0
0
0

0
0
,
0
1

12.8452 6.7729 4.2817


0
2.2643 1.7665 ,
=
0
0
0.7390

1.7905
= 0.4978 ,
0.7390

and = 1 2 3 is the product of the Householder reections used to reduce to uppertriangular form.
Using this decomposition, we can solve the linear least squares problem x = b by observing
that

2
[
]



2


b x2 = b
x
0 0
2
3


[
][
] 2

u

=
b 0 0
v 2
[
] [
] 2
c
u + v


=


d
0
2
= c u v22 + d22 ,
where

b=

c
d

x=

u
v

]
,

with c and u being -vectors. Thus min b x22 = d22 , provided that u + v = c. A basic
solution is obtained by choosing v = 0. A second solution is to choose u and v so that u22 + v22
is minimized. This criterion is related to the pseudo-inverse of .

Complete Orthogonal Decomposition


After performing the factorization with column pivoting, we have
]
[

=
0 0
where is upper triangular. Then

0
0

where is lower triangular. We apply Householder reections so that


]
] [
[
0

0
,
=
2 1
0 0
0
where is upper-triangular. Then

0
0 0

0
0

where = 1 . In other words,


=

where is a lower-triangular matrix of size , where is the rank of . This is the complete
orthogonal decomposition of .
Recall that is the pseudo-inverse of if
4

1. =
2. =
3. () =
4. () =
Given the above complete orthogonal decomposition of , the pseudo-inverse of , denoted + , is
given by
[ 1
]

0
+
=
.
0
0
Let = {xb x2 = min }. If x and we desire x2 = min , then x = + b. Note that in
this case,
r = b x = b + b = ( + )b
where the matrix ( + ) is a projection matrix . To see that is a projection, note that

+
=
[

=
0
[

=
0

0
0


]
0
.
0

1 0
0
0

It can then be veried directly that = and 2 = .

Applications of the SVD


Minimum-norm least squares solution
One of the most well-known applications of the SVD is that it can be used to obtain the solution
to the problem
b x2 = min , x2 = min .
The solution is
= + b = + b
x
where + is the pseudo-inverse of .

Closest Orthogonal Matrix


Let be the set of all orthogonal matrices. Given an matrix , we wish to nd the
matrix that satises
= min , , () = 1.
= , then
Given = , if we compute
2

= ( ) 2
= 2
= (1 1)2 + + ( 1)2

It can be shown that this is in fact the minimum.

Low-Rank Approximations
()

()

()

Let , be the set of all matrices of rank , and let , . We wish to nd , ,


where < , such that = min .
= where
To solve this problem, let = be the SVD of , and let

..

..

.
0
Then
2

= ( ) 2
= 2
2
+ + 2 .
= +1

such that
2 2 ,
We now consider a variation of this problem. We wish to nd

is minimized. We know that if = then


where the rank of
2
2 = +1
+ + 2 .

= is the solution if
It follows that
+1 + + 2 2 ,
Note that

(
+

+ 2

2 + + 2 > 2 .
1

2
+1

1
+ + 2

)
.