You are on page 1of 16

EE448/528

Version 1.0

John Stensby

Chapter 4: Matrix Norms

The analysis of matrix-based algorithms often requires use of matrix norms.

These

algorithms need a way to quantify the "size" of a matrix or the "distance" between two matrices.
For example, suppose an algorithm only works well with full-rank, nn matrices, and it produces
inaccurate results when supplied with a nearly rank deficit matrix. Obviously, the concept of rank (also known as numerical rank), defined by
rank ( A , ) = min rank ( B)
AB

(4-1)

is of interest here. All matrices B that are "within" of A are examined when computing the rank of A.
We define a matrix norm in terms of a given vector norm; in our work, we use only the pr
vector norm, denoted as X p . Let A be an mn matrix, and define
r
AX p
,
A p = sup
r
r
Xp
X 0

(4-2)

where "sup" stands for supremum, also known as least upper bound. Note that we use the same
p notation for both vector and matrix norms. However, the meaning should be clear from
context.
Since the matrix norm is defined in terms of the vector norm, we say that the matrix norm
is subordinate to the vector norm. Also, we say that the matrix norm is induced by the vector
norm.

r
Now, since X p is a scalar, we have

r
r r
AX p
/ Xp .
A
X
A p = sup
=
sup
r
r
r
p
X
X 0
X 0
p

CH4.DOC

(4-3)

Page 4-1

EE448/528

Version 1.0

John Stensby

r r
In (4-3), note that X / X p has unit length; for this reason, we can express the norm of A in terms
of a supremum over all vectors of unit length,
A p = rsup

X p =1

r
AX p .

(4-4)

r
r
That is, A p is the supremum of AX p on the unit ball X p = 1.
Careful consideration of (4-2) reveals that
r
r
AX p A p X p

(4-5)

r
r
r
r
for all X. However, AX p is a continuous function of X, and the unit ball X p = 1 is closed and
bounded (real analysis books, it is said to be compact). Now, on a closed and bounded set, a
continuous function always achieves its maximum and minimum values. Hence, in the definition
of the matrix norm, we can replace the "sup" with "max" and write
r
AX p
r
= max
A p = max
r
r
r
AX p .
X 0
Xp
X p =1

(4-6)

When computing the norm of A, the definition is used as a starting point. The process has
two steps.
r
r
r
1) Find a "candidate" for the norm, call it K for now, that satisfies AX p K X p for all X.
r
r
r
2) Find at least one nonzero X0 for which AX 0 p = K X 0 p
Then, you have your norm: set A p = K.

MatLab's Matrix Norm Functions

From an application standpoint, the 1-norm, 2-norm and the -norm are amoung the most

CH4.DOC

Page 4-2

EE448/528

Version 1.0

John Stensby

important; MatLab computes these matrix norms. In MatLab, the 1-norm, 2-norm and -norm
are invoked by the statements norm(A,1), norm(A,2), and norm(A,inf), respectively.
The 2-norm is the default in MatLab. The statement norm(A) is interpreted as norm(A,2) by
MatLab. Since the 2-norm used in the majority of applications, we will adopt it as our default. In
what follows, an "un-designated" norm A is to be intrepreted as the 2-norm A 2 .

The Matrix 1-Norm

Recall that the vector 1-norm is given by
n
r
X 1 = xi .

(4-7)

i =1

Subordinate to the vector 1-norm is the matrix 1-norm

A 1 = max
j

F a I.
GH i ij JK

(4-8)

That is, the matrix 1-norm is the maximum of the column sums. To see this, let mn matrix A
be represented in the column format
r
r
r
A = A1 A 2 L A n .

(4-9)

Then we can write

n r
r
r
r
r r
AX = A1 A 2 L A n X = A k x k ,

(4-10)

k =1

CH4.DOC

Page 4-3

EE448/528

Version 1.0

John Stensby

where xk, 1 k n, are the components of arbitrary vector X. The triangle inequality and
standard analysis applied to the norm of (4-10) yields
n r
x
r
r
AX 1 = A k x k x k A k 1
k =1

r
max A j
j

k =1

Fnx I
1GH k JK
k =1

(4-11)

r
r
= max A j
X
1.
1
j

With the development of (4-11), we have completed step #1 for computing the matrix 1-norm.
That is, we have found a constant
r
K = max A j
1
j
r
r
r
such that AX 1 K X 1 for all X. Step 2 requires us to find at least one vector for which we
r
r
have equality in (4-11). But this is easy; to maximize AX 1 , it is natural to select the X0 that puts
"all of the allowable weight" on the component that will "pull-out" the maximum column sum.
That is, the "optimum" vector is
r
X0 = 0 0 L 0 1 0 L 0
jth-position
where j is the index that satisfies
r
r
A j = max A k
1
1
k

CH4.DOC

Page 4-4

EE448/528

Version 1.0

John Stensby

(the sum of the magnitudes in the jth column is equal to, or larger than, the sum of the magnitudes

in any column). When X0 is used, we have equality in (4-11), and we have completed step #2, so
(4-8) is the matrix 1-norm.

The Matrix -Norm

Recall that the vector -norm is given by
r
X = max x k ,

(4-12)

the vector's largest component. Subordinate to the vector -norm is the matrix -norm

A = max
i

F aI
GG ij JJ .
Hj K

(4-13)

That is, the matrix -norm is the maximum of the row sums. To see this, let A be an arbitrary
mn matrix and compute

a1k xk
k

a2k x k

r
k
AX =

a mk xk
k

max
i

CH4.DOC

= max
i

aik x k max
k

F axI
GH k ik k JK
(4-14)

F a I max x
GH k ik JK k k

Page 4-5

EE448/528

Version 1.0

= max
i

F a I
GH k ik JK

John Stensby

r
X

We have completed the first step. We have found a constant K = max

i

r
r
AX K X

F a I for which
GH k ik JK
(4-15)

for all X.

Step #2 requires that we find a non-zero X0 for which equality holds in (4-14) and (4-15).

A close examination of these formulas leads to the conclusion that equality prevails if X0 is
defined to have the components

xk

=
=

aik
, a ik 0, 1 k n,
a ik
1,

a ik = 0

(4-16)

(the overbar denotes complex conjugate) where i is the index for the maximum row sum. That is,
in (4-16), use index i for which

aik a jk
k

, i j.

(4-17)

The Matrix 2-Norm

Recall that the vector 2-norm is given by

CH4.DOC

Page 4-6

EE448/528
n

r
X2=

Version 1.0

xi

John Stensby

r r
X, X .

(4-18)

i =1

Subordinate to the vector 2-norm is the matrix 2-norm

A 2 = largest eigenvalue of A A .

(4-19)

Due to this connection with eigenvalues, the matrix 2-norm is called the spectral norm.
To see (4-19) for an arbitrary mn matrix A, note that A*A is nn and Hermitian. By
Theorem 4.2.1 (see Appendix 4.1), the eigenvalues of A*A are real-valued. Also, A*A is at least

positive semi-definite since X*(A*A)X = (AX)*(AX) 0 for all X. Hence, the eigenvalues of
A*A are both real-valued and non-negative; denote them as

12 22 32 L 2n 0 .

(4-20)

Note that these eigenvalues are arranged according to size with 12 being the largest. These
eigenvalues are known as the singular values of matrix A.
Corresponding to these eigenvalues are n orthonormal (hence, they are independent)

r r

eigenvectors U1, U2, ... , Un with

(A*A)Uk = (k2)Uk , 1 k n.

(4-21)

The n eigenvectors form the columns of a unitary nn matrix U that diagonalizes matrix A*A
under similarity (matrix U*(A*A)U is diagonal with eigenvalues (4-20) on the diagonal).

r r

Since the n eigenvectors U1, U2, ... , Un are independent, they can be used as a basis, and

vector X can be expresssed as

CH4.DOC

Page 4-7

EE448/528

Version 1.0

John Stensby

n
r
r
X = ck U k ,

(4-22)

k =1

where the ck are X-dependent constants. Multiply X, in the form of (4-22), by A*A by to obtain
n
n
r
r
r
A AX = A A ck U k = c k 2k U k ,
k =1

(4-23)

k =1

r 2
r
r r
r
AX = ( AX ) AX = X ( A AX ) =
2

12

F n c Ur I F n c 2 Ur I = n c 2 2
GH k=1 k k JK GH j=1 j j jJK k=1 k k

F n c 2I
GH k=1 k JK

(4-24)

r 2
= 12 X
2

for arbitrary X. Hence, we have completed step#1: we found a constant K = 12 such that
r
r
r
r
AX 2 K X 2 for all X. Step#2 requires us to find at least one vector X0 for which equality
r
r
r
holds; that is, we must find an X0 with the property that AX 0 2 = K X 0 2 . But, it is obvious

that X0 = U1, the unit-length eigenvector associated with eigenvalue 12 , will work. Hence, the
matrix 2-norm is given by A 2 =

12 , the square root of the largest eigenvalue of A*A.

The 2-norm is the default in MatLab. Also, it is the default here. From now on, unless
specified otherwise, the 2-norm is assumed: A means A 2 .
Example

A=

LM1 1OP
N2 1Q

CH4.DOC

Page 4-8

EE448/528

Version 1.0

John Stensby

r
r
Find A 2 , and find the unit-length vector X0 that maximizes AX . First, compute the product
2
5 3
.
A A = A TA =
3 2

LM OP
N Q

The eigenvalues of this matrix are 12 = 6.8541 and 22 = .1459; note that ATA is positive definite
symmetric. The 2-norm of A is A 2 = 12 = 2.618 , the square root of the largest eigenvalue of
ATA.

The corresponding eigenvectors are U1 = [-.8507 -.5257] and U2 = [.5257 -.8507]. They

are columns in an orthogonal matrix U = [U1 U2]; note that UTU = I or UT = U-1. Furthermore,
matrix U diagonalizes ATA

L
12
( A A )U = M
MN 0
T

OP L
PQ MN

6.8541
0
0
=
.1459
0
22

r
The unit-length vector that maximizes AX
r
max
AX
r
X =1

OP
Q
r

is X0 = U1, and

= 12 = 6.8541

p-Norm of Matrix Product

For mn matrix A we know that
r
r
AX p A p X p

(4-25)

for all X. Now, consider mn matrix A and nq matrix B. The product AB is mq. For q-

dimensional X, we have

CH4.DOC

Page 4-9

EE448/528

Version 1.0

r
r
r
r
ABX p = A ( BX ) p A p BX p A p B p X p .

John Stensby
(4-26)

r
ABX p
Ap Bp
r
Xp

(4-27)

AB p A p B p ,

(4-28)

a useful, simple result.

2-Norm Bound
Let A be an mn matrix with elements aij, 1 i m, 1 j n. Let
max a ij
i, j

denote the magnitude of the largest element in A. Let (i0, j0) be the indices of the largest element
so that
a i j a ij
0 0

(4-29)

for all i, j.

CH4.DOC

Page 4-10

EE448/528

Version 1.0

John Stensby

max a ij A 2

mn max a ij

i, j

(4-30)

i, j

Proof: Note that

r
AX

2
2

a1k x k

A 2 = max
r

X =1

a2k x k
k

+L+

a mk x k

(4-31)

r
AX .
2

Define X0 as the vector with 1 in the j0 position and zero elsewhere (see (4-29) for definition of i0
and j0). Since
r
r
AX = A 2
AX 0 2 max
r
2
X 2 =1

(4-32)

we have

ai j
0 0

a k j0 2
k

r
max
AX 2 = A 2
r
X =
2

(4-33)

Hence, we have
max a ij A 2
i, j

(4-34)

as claimed. Now, we show the remainder of the 2-norm bound. Observe that

CH4.DOC

Page 4-11

EE448/528

Version 1.0
2

A 2 = max
r

X 2 =1

a1k x k

max
r

X 2 =1

a2k x k

+L+

a1k xk

John Stensby

a mk x k
k

+ a 2 k x k + L + a mk x k
2

(4-35)

a1k 2 + a2k 2 + L + a mk 2
k

mn max a ij .
i, j

i, j

mn max a ij .
i, j

Triangle Inequality for the p-Norm

Recall the Triangle inequality for real numbers: = + . A similar result is
valid for the matrix p-norm.

Theorem 4.2 (Triangle Inequality for the matrix 2-norm)

Let A and B be mn matrices. Then
A+Bp A p+ Bp

(4-36)

Application of Matrix Norm: Inverse of Matrices Close to a Non-singular Matrix.

Let A be an nn non-singular matrix. Can we state conditions on the size(i.e., norm) of
nn matrix E which guarantees that A+E is nonsingular? Before developing these conditions, we
derive a closely related result, the geometric series for matrices. Recall the geometric series

1
= xk , x < 1
1 x k =1

(4-37)

for real variables. We are interested in the matrix version of this result.
CH4.DOC

Page 4-12

EE448/528

Version 1.0

John Stensby

Theorem 4.3 (Geometric Series for Matrices)

Let F be any nn matrix with F 2 < 1 . Let I denote the nn identity matrix. Then, the
difference I - F is nonsingular and

( I F)1 =

Fk

(4-38)

k=0

with
1
.
( I F)1 2
1 F 2

(4-39)

Proof: First, we show that I - F is nonsingular. Suppose that this is not true; suppose that I - F is
r
r
r
r
singular. Then there exists at least one non-zero X such that (I - F)X = 0 so that X 2 = FX 2 .
r
r
r
r
r
But since FX 2 F 2 X 2 , we must have X 2 = FX 2 F 2 X 2 which requires that F 2 1 ,
a contradiction. Hence, the nn matrix I - F must be nonsingular. To obtain a series for (I - F)-1,
consider the obvious identity

F N Fk I ( I F) = I F
GH k=0 JK

N +1

(4-40)

However, Fk 2 F 2 , and since F 2 < 1 , we have Fk 0 as k . As a result of this,

N
F
I
limit G Fk J ( I F) = I
N H k = 0 K

so that ( I F)

(4-41)

Fk

as claimed. Finally

k=0

CH4.DOC

Page 4-13

EE448/528
( I F)

1
2

Version 1.0

k =0

F2

John Stensby

Fk 2

k =0

(this is an "ordinary" scalar geometric series) .

(4-42)

k =0

1
1 F 2

as claimed.
In many applications, Theorem 4.3 is used in the form

( I F )

Fk k

k =0

for < 1 / F 2 ,

(4-43)

where is considered to be a "small" parameter. Next, we generalize Theorem 4.3 and obtain a

result that will be used in our study of how small errors in A and b influence the solution of the

linear algebraic problem AX = b.

Theorem 4-4
Assume that A is nn and nonsingular. Let E be an arbitrary nn matrix. If
= A 1E 2 < 1,

(4-44)

2

( A + E )1 A 1 2

E 2 A 1
2
1

(4-45)

CH4.DOC

Page 4-14

EE448/528

Version 1.0

John Stensby

(4-46)

Since = F 2 = A 1E 2 < 1, it follows from Theorem 4.3 that I - F is nonsingular and

( I F)1 2

1 F 2

1
.
1

(4-47)

From (4-46) we have (A + E)-1 = (I - F)-1A-1; with the aid of (4-47), this can be used to write

( A + E)

1
2

A 1

2 .

(4-48)

(A + E)-1 - A-1 = -A-1E(A+E)-1

(4-49)

by A + E to see this matrix identity. Finally, take the norm of (4-49) to obtain
( A + E )1 A 1

= A 1E( A + E )1

(4-50)
A

1
2

E 2 ( A + E)

1
2

Now use (4-48) in this last result to obtain

( A + E)

1
2

A 1

1
2

E2

A 1

2
2

E2

(4-51)

as claimed.

CH4.DOC

Page 4-15

EE448/528

Version 1.0

John Stensby

We will use Theorem 4.4 when studying the sensitivity of the linear equation AX = b.

That is, we want to relate changes in solution X to small changes (errors) in both A and b.

CH4.DOC

Page 4-16