Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Look up keyword
Like this
1Activity
0 of .
Results for:
No results containing your search query
P. 1
Matrix Perturbation

Matrix Perturbation

Ratings: (0)|Views: 3 |Likes:

More info:

Published by: Julio Cesar Barraza Bernaola on Nov 23, 2012
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

09/08/2014

pdf

text

original

 
15
Matrix PerturbationTheory
Ren-Cang Li
University of Texas at Arlington
15.1 Eigenvalue Problems
...............................
15
-
1
15.2 Singular Value Problems
............................
15
-
6
15.3 Polar Decomposition
...............................
15
-
7
15.4 Generalized Eigenvalue Problems
...................
15
-
9
15.5 Generalized Singular Value Problems
...............
15
-
12
15.6 Relative Perturbation Theory for EigenvalueProblems
..........................................
15
-
13
15.7 Relative Perturbation Theory for Singular ValueProblems
..........................................
15
-
15
References
................................................
15
-
16
There is a vast amount of material in matrix (operator) perturbation theory. Related books that are worthmentioning are [SS90], [Par98], [Bha96], [Bau85], and [Kat70]. In this chapter, we attempt to include themost fundamental results up to date, except those for linear systems and least squares problems for whichthe reader is referred to Section 38.1 and Section 39.6.Throughout this chapter,
·
UI
denotes a general unitarily invariant norm. Two commonly used onesare the spectral norm
·
2
and the Frobenius norm
·
F
.
15.1 Eigenvalue Problems
The reader is referred to Sections 4.3, 14.1, and 14.2 for more information on eigenvalues and theirlocations.
Definitions:
Let
A
C
n
×
n
. A scalar–vector pair (
λ
,
)
C
×
C
n
is an
eigenpair
of 
A
if 
=
0 and
A
=
λ
.
Avector–scalar–vector triplet (
 y 
,
λ
,
)
C
n
×
C
×
C
n
is an
eigentriplet
if 
=
0,
=
0, and
A
=
λ
,
 y 
 A
=
λ
 y 
. The quantity cond(
λ
)
=
2
 y 
2
|
 y 
|
is the
individualconditionnumber
for
λ
, where (
 y 
,
λ
,
)
C
n
×
C
×
C
n
is an eigentriplet.Let
σ 
(
 A
)
= {
λ
1
,
λ
2
,
...
,
λ
n
}
, the multiset of 
A
’s eigenvalues, and set
=
diag(
λ
1
,
λ
2
,
...
,
λ
n
),
τ 
=
diag(
λ
τ 
(1)
,
λ
τ 
(2)
,
...
,
λ
τ 
(
n
)
),
15
-
1
 
15
-
2
Handbook of Linear Algebra
where
τ 
is a
permutation
of 
{
1,2,
...
,
n
}
. For real
, i.e., all
λ
 j 
’s are real,
=
diag(
λ
1
,
λ
2
,
...
,
λ
n
)
.
is in fact a
τ 
for which the permutation
τ 
makes
λ
τ 
(
 j 
)
=
λ
 j 
for all
.Given two square matrices
A
1
and
A
2
, the
separation
sep(
 A
1
,
A
2
) between
A
1
and
A
2
is defined as [SS90,p. 231]sep(
 A
1
,
A
2
)
=
inf 
 X 
2
=
1
 XA
1
A
2
 X 
2
.
 A
is perturbed to
 A
=
A
+
 A
.
The same notation is adopted for 
 A, except all symbols with tildes
.Let
,
C
n
×
k
with rank(
 X 
)
=
rank(
)
=
k
. The
canonical angles
between their column spaces are
θ 
i
=
arc cos
σ 
i
, where
{
σ 
i
}
ki
=
1
are the singular values of (
)
1
/
2
 X 
(
 X 
 X 
)
1
/
2
. Define the
canonicalanglematrix 
between
and
as
(
 X 
,
)
=
diag(
θ 
1
,
θ 
2
,
...
,
θ 
k
)
.
For
k
=
1, i.e.,
,
C
n
(both nonzero), we use
(
,
 y 
), instead, to denote the canonical angle betweenthe two vectors.
Facts:
1. [SS90, p. 168] (
Elsner
) max 
i
min
 j 
|
λ
i
λ
 j 
| ≤
(
 A
2
+
 A
2
)
1
1
/
n
 A
1
/
n
2
.
2. [SS90, p. 170] (
Elsner
) There exists a permutation
τ 
of 
{
1,2,
...
,
n
}
such that
τ 
2
2
n
2
(
 A
2
+
 A
2
)
1
1
/
n
 A
1
/
n
2
.
3. [SS90, p. 183] Let (
 y 
,
µ
,
) be an eigentriplet of 
A
.
 A
changes
µ
to
µ
+
µ
with
µ
=
 y 
(
 A
)
 y 
+
O
(
 A
22
),and
|
µ
| ≤
cond(
µ
)
 A
2
+
O
(
 A
22
).4. [SS90, p. 205] I
A
and
A
+
 A
are Hermitian, then
UI
≤ 
 A
UI
.
5. [Bha96,p.165](
Hoffman–Wielandt
)If 
A
and
A
+
 A
arenormal,thenthereexistsapermutation
τ 
of 
{
1,2,
...
,
n
}
such that
τ 
F
≤ 
 A
F
.6. [Sun96] If 
A
is normal, then there exists a permutation
τ 
of 
{
1,2,
...
,
n
}
such that
τ 
F
√ 
n
 A
F
.7. [SS90, p. 192] (
Bauer–Fike
) If 
A
is diagonalizable and
A
=
 X 
1
is its eigendecomposition,thenmax 
i
min
 j 
|
λ
i
λ
 j 
| ≤
 X 
1
(
 A
)
 X 
 p
κ
 p
(
 X 
)
 A
 p
.
8. [BKL97] Suppose both
A
and
 A
are diagonalizable and have eigendecompositions
A
=
 X 
1
and
 A
=
 X 
 X 
1
.(a) There exists a permutation
τ 
of 
{
1,2,
...
,
n
}
such that
τ 
F
 
κ
2
(
 X 
)
κ
2
(
 X 
)
 A
F
.(b)
UI
 
κ
2
(
 X 
)
κ
2
(
 X 
)
 A
UI
for real
and
.
 
Matrix Perturbation Theory 
15
-
3
9. [KPJ82] Let residuals
r
=
A
µ
and
s
=
 y 
 A
µ
 y 
, where
2
=
 y 
2
=
1, and let
ε
=
max 
{
r
2
,
s
2
}
. The smallest error matrix 
 A
in the 2-norm, for which (
 y 
,
µ
,
) is an exacteigentriplet of 
 A
=
A
+
 A
, satisfies
 A
2
=
ε
, and
|
µ
µ
| ≤
cond(
µ
)
ε
+
O
(
ε
2
) for some
µ
σ 
(
 A
).10. [KPJ82], [DK70],[Par98, pp. 73, 244] Suppose
A
is Hermitian, and let residual
r
=
A
µ
and
2
=
1.(a) ThesmallestHermitianerrormatrix 
 A
(inthe2-norm),forwhich(
µ
,
)isanexacteigenpairof 
 A
=
A
+
 A
, satisfies
 A
2
= 
r
2
.
(b)
|
µ
µ
| ≤
r
2
for some eigenvalue
µ
of 
A
.(c) Let
µ
be the closest eigenvalue in
σ 
(
 A
) to
µ
and
be its associated eigenvector with
2
=
1,and let
η
=
min
µ
=
λ
σ 
(
 A
)
|
µ
λ
|
. If 
η >
0, then
|
µ
µ
| ≤
r
22
η
, sin
(
,
)
r
2
η.
11. Let
A
be Hermitian,
C
n
×
k
have full column rank, and
C
k
×
k
be Hermitian havingeigenvalues
µ
1
µ
2
··· ≤
µ
k
. Set
R
=
AX 
X
.
There exist
k
eigenvalues
λ
i
1
λ
i
2
··· ≤
λ
i
k
of 
A
such that the following inequalities hold. Notethat subset
{
λ
i
 j 
}
k j 
=
1
may be different at different occurrences.(a) [Par98, pp. 253–260], [SS90, Remark 4.16, p. 207] (
Kahan–Cao–Xie–Li
)max 
1
 j 
k
|
µ
 j 
λ
i
 j 
|
R
2
σ 
min
(
 X 
),
 
k
 j 
=
1
(
µ
 j 
λ
i
 j 
)
2
R
F
σ 
min
(
 X 
)
.
(b) [SS90, pp. 254–257], [Sun91] If 
 X 
=
and
=
 AX 
, and if all but
k
of 
A
’s eigenvaluesdiffer from every one of 
’s by at least
η >
0 and
ε
F
= 
R
F
<
1, then
 
k
 j 
=
1
(
µ
k
λ
i
 j 
)
2
R
2
F
η
 
1
ε
2
F
.
(c) [SS90,pp. 254–257],[Sun91]If 
 X 
=
and
=
 AX 
,andthereisanumber
η >
0suchthat either all but
k
of 
A
’s eigenvalues lie outside the open interval (
µ
1
η
,
µ
k
+
η
) or all but
k
of 
A
’s eigenvalues lie inside the closed interval [
µ
+
η
,
µ
+
1
η
] for some 1
k
1,and
ε
= 
R
2
<
1, thenmax 
1
 j 
k
|
µ
 j 
λ
i
 j 
| ≤
R
22
η
√ 
1
ε
2
.
12. [DK70] Let
A
be Hermitian and have decomposition
 X 
1
 X 
2
A
[
 X 
1
2
]
=
 A
1
 A
2
,

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->