Finding Eigenvalues and Eigenvectors

Transcription

Finding Eigenvalues and
Eigenvectors
What is really
important?
Approaches

Find the characteristic polynomial
–

Find the largest or smallest eigenvalue
–
–

Power Method
Inverse Power Method
Find all the eigenvalues
–
–
–
–
2
Leverrier’s Method
Jacobi’s Method
Householder’s Method
QR Method
Danislevsky’s Method
1/12/2017
DRAFT Copyright, Gene A
Tagliarini, PhD
Finding the Characteristic Polynomial



Reduces to finding the coefficients of the
polynomial for the matrix A
Recall |lI-A| = ln+anln-1+an-1ln-2+…+a2l1+a1
Leverrier’s Method
–
–
Set Bn = A and an = -trace(Bn)
For k = (n-1) down to 1 compute


3
Bk = A (Bk+1 + ak+1I)
ak = - trace(Bk)/(n – k + 1)
1/12/2017
Tagliarini, PhD
Vectors that Span a Space and Linear
Combinations of Vectors


4
Given a set of vectors v1, v2,…, vn
The vectors are said span a space V, if given
any vector x ε V, there exist constants
c1, c2,…, cn so that
c1v1 + c2v2 +…+ cnvn = x and x is called a
linear combination of the vi
1/12/2017
Tagliarini, PhD
Linear Independence and a Basis



5
Given a set of vectors v1, v2,…, vn and
constants c1, c2,…, cn
The vectors are linearly independent if the
only solution to c1v1 + c2v2 +…+ cnvn = 0 (the
zero vector) is c1= c2=…=cn = 0
A linearly independent, spanning set is called
a basis
1/12/2017
Tagliarini, PhD
Example 1: The Standard Basis




6
Consider the vectors v1 = <1, 0, 0>, v2 = <0, 1, 0>,
and v3 = <0, 0, 1>
Clearly, c1v1 + c2v2 + c3v3 = 0  c1= c2= c3 = 0
Any vector <x, y, z> can be written as a linear
combination of v1, v2, and v3 as
<x, y, z> = x v1 + y v2 + z v3
The collection {v1, v2, v3} is a basis for R3; indeed, it
is the standard basis and is usually denoted with
vector names i, j, and k, respectively.
1/12/2017
Tagliarini, PhD
Another Definition and Some Notation



7
Assume that the eigenvalues for an n x n
matrix A can be ordered such that
|l1| > |l2| ≥ |l3| ≥ … ≥ |ln-2| ≥ |ln-1| > |ln|
Then l1 is the dominant eigenvalue and |l1|
is the spectral radius of A, denoted r(A)
The ith eigenvector will be denoted using
superscripts as xi, subscripts being reserved
for the components of x
1/12/2017
Tagliarini, PhD
Power Methods: The Direct Method


8
Assume an n x n matrix A has n linearly
independent eigenvectors e1, e2,…, en
ordered by decreasing eigenvalues
|l1| > |l2| ≥ |l3| ≥ … ≥ |ln-2| ≥ |ln-1| > |ln|
Given any vector y0 ≠ 0, there exist constants
ci, i = 1,…,n, such that
y0 = c1e1 + c2e2 +…+ cnen
1/12/2017
Tagliarini, PhD
The Direct Method (continued)





9
If y0 is not orthogonal to e1, i.e., (y0)Te1≠ 0,
y1 = Ay0 = A(c1e1 + c2e2 +…+ cnen)
= Ac1e1 + Ac2e2 +…+ Acnen
= c1Ae1 + c2Ae2 +…+ cnAen
Can you simplify the previous line?
1/12/2017
Tagliarini, PhD






10
If y0 is not orthogonal to e1, i.e., (y0)Te1≠ 0,
y1 = Ay0 = A(c1e1 + c2e2 +…+ cnen)
= Ac1e1 + Ac2e2 +…+ Acnen
= c1Ae1 + c2Ae2 +…+ cnAen
y1 = c1l1e1 + c2l2e2 +…+ cnlnen
What is y2 = Ay1?
1/12/2017
Tagliarini, PhD
n
y  c1l e  c2 l e  ...  cn l e   ck l2k e k
2
2 1
1
2
2
2
2
n
n
k 1
in general,
n
y i  c1l1i e1  c2 li2e 2  ...  cn lin e n   ck lik e k
k 1
or
n

 lk
i
i
1
y  l1 c1e   ck 

k 2
 l1

11
1/12/2017
 k 
 e



i
Tagliarini, PhD
i
n




l
y i  l1i  c1e1   ck  k  ek 

l1  
k 2



So what?
i
 lk 
lk
Recall that
 1, so    0 as i  
l1
 l1 
12
1/12/2017
Tagliarini, PhD
Since
i
n




l
y i  l1i  c1e1   ck  k  e k 

l1  
k 2



then
y i  l1i c1e1 as i  
13
1/12/2017
Tagliarini, PhD




14
Note: any nonzero multiple of an eigenvector
is also an eigenvector
Why?
Suppose e is an eigenvector of A, i.e., Ae=le
and c0 is a scalar such that x = ce
Ax = A(ce) = c (Ae) = c (le) = l (ce) = lx
1/12/2017
Tagliarini, PhD
Since y i  l1i c1e1 and e1 is an eigenvecto r
i
y will become arbitraril y close to an eigenvecto r
and we can prevent y i from growing inordinate ly
by normalizin g
i 1
Ay
yi 
i 1
Ay
15
1/12/2017
Tagliarini, PhD
Direct Method (continued)




16
Given an eigenvector e for the matrix A
We have Ae = le and e0, so eTe  0 (a
scalar)
Thus, eTAe = eTle = leTe  0
So l = (eTAe) / (eTe)
1/12/2017
Tagliarini, PhD
Direct Method (completed)
i

y  Ay

y  y
i T
i T
i
i
approximat es the dominant eigenvalue
r i  i I  A y i is the residual error vect or
and
r i  0 when i is an eigenvalue with eigenvecto r y i
17
1/12/2017
Tagliarini, PhD
Direct Method Algorithm
1. Choose   0, m  0, and y  0.
Set i  0 and compute x  Ay and .
2. Do
y
x
x
x  Ay
yT x
 T
y y
r  y - x
i  i 1
3. While ( r   ) and (i  m)
18
1/12/2017
Tagliarini, PhD
Jacobi’s Method





19
Requires a symmetric matrix
May take numerous iterations to converge
Also requires repeated evaluation of the
arctan function
Isn’t there a better way?
Yes, but we need to build some tools.
1/12/2017
Tagliarini, PhD
What Householder’s Method Does



20
Preprocesses a matrix A to produce an
upper-Hessenberg form B
The eigenvalues of B are related to the
eigenvalues of A by a linear transformation
Typically, the eigenvalues of B are easier to
obtain because the transformation simplifies
computation
1/12/2017
Tagliarini, PhD
Definition: Upper-Hessenberg Form

A matrix B is said to be in upper-Hessenberg
form if it has the following structure:
 b1,1 b1,2
b b
 2,1 2,2
 0 b3,2
B
0 0
 


 0 0
21
b1,3  b1,n -1
b 2,3  b 2,n -1
b3,3  b3,n -1
b 4,3 

  b n -1,n -1

0
b n,n -1
1/12/2017
b1,n 
b 2,n 
b3,n 

 
b n -1,n 

b n,n 
Tagliarini, PhD
A Useful Matrix Construction



Assume an n x 1 vector u  0
Consider the matrix P(u) defined by
P(u) = I – 2(uuT)/(uTu)
Where
–
–
–
22
I is the n x n identity matrix
(uuT) is an n x n matrix, the outer product of u
with its transpose
(uTu) here denotes the trace of a 1 x 1 matrix and
is the inner or dot product
1/12/2017
Tagliarini, PhD
Properties of P(u)

P2(u) = I
–
–

P-1(u) = P(u)
–

–
23
P(u) is its own inverse
PT(u) = P(u)
–

The notation here P2(u) = P(u) * P(u)
Can you show that P2(u) = I?
P(u) is its own transpose
Why?
P(u) is an orthogonal matrix
1/12/2017
Tagliarini, PhD
Householder’s Algorithm


Set Q = I, where I is an n x n identity matrix
For k = 1 to n-2
–
–
–
–
–

24
a = sgn(Ak+1,k)sqrt((Ak+1,k)2+ (Ak+2,k)2+…+ (An,k)2)
uT = [0, 0, …, Ak+1,k+ a, Ak+2,k,…, An,k]
P = I – 2(uuT)/(uTu)
Q = QP
A = PAP
Set B = A
1/12/2017
Tagliarini, PhD
Example
1 2 3
Let A  3 5 6 .
4 8 9 3 x 3
Clearly, n  3 and since n - 2  1, k takes only the value k  1.
2
2
Then a  sgn(a 21 ) a21
 a31
 sgn(3 ) 32  4 2  1 5  5
25
1/12/2017
Tagliarini, PhD
Example
uT  [0,...,0, a21  a , a31,..., an1 ]  [0, a21  a , a31 ]
 [0,3  5,4]  [0,8,4]
0 
80 8 4
 
1
0
0


4
uu T 

P  I  2 T  0 1 0   2
?
u u
0 
0 0 1
0 8 48
4
26
1/12/2017
Tagliarini, PhD
Example
0 
8 0 8 4
 
1 0 0
T
 4
uu
P  I  2 T  0 1 0  2  
u u
0 
0 0 1
0 8 48
4


0 
1 0 0
0 0 0  1 0

2 
 3  4



2
 0 1 0  0 64 32  0
. Find P  ?
80
5
5 
0 0 1
0 32 16  
0  4 3 

5
5 
27
1/12/2017
Tagliarini, PhD
Example
Initially, Q  I, so

1 0 0 1



Q  QP  0 1 0 0
0 0 1 
0

28
1/12/2017
0
3
5
4
5
 
0  1
 4 
  0
5  
3  
0


5  
0
3
5
4
5

0 
 4

5 
3 
5 
Tagliarini, PhD
Example
Next, A  PAP , so

1

A  0

0

29
0
3
5
4
5


0  1 2 3 1

 4 

  3 5 6  0
5 



4
8
9
3 

0


5 

1/12/2017
0
3
5
4
5

0 
 4
?
5 
3 
5 
Tagliarini, PhD
Example
Hence,

1

A  0

0

0
3
5
4
5
- 18

1

5

357
 - 5
25

 0 - 24

25
30


0  1 2 3 1

 4 

 3 5 6  0
5 

3  4 8 9 
0


5 

1
5
26 

25 
-7
25 
0
3
5
4
5
1/12/2017
 
0  1
 4 
   5
5  
3  
0


5  
2
 47
5
4
5

3  1
 54  
 0
5 
3 
0


5 
0
3
5
4
5
Tagliarini, PhD

0 
 4

5 
3 
5 
Example
Finally, since the loop only executes once
- 18

1
5

357
B  A  - 5
25

 0 - 24

25
So what?
31
1/12/2017
1
5
26 
.
25 
-7
25 
Tagliarini, PhD
How Does It Work?




Householder’s algorithm uses a sequence of
similarity transformations
B = P(uk) A P(uk)
to create zeros below the first sub-diagonal
uk=[0, 0, …, Ak+1,k+ a, Ak+2,k,…, An,k]T
a = sgn(Ak+1,k)sqrt((Ak+1,k)2+ (Ak+2,k)2+…+ (An,k)2)
By definition,
–
–
32
sgn(x) = 1, if x≥0 and
sgn(x) = -1, if x<0
1/12/2017
Tagliarini, PhD
How Does It Work? (continued)

The matrix Q is orthogonal
–
–
–

B = QT A Q hence Q B = Q QT A Q = A Q
–
33
the matrices P are orthogonal
Q is a product of the matrices P
The product of orthogonal matrices is an
orthogonal matrix
Q QT = I (by the orthogonality of Q)
1/12/2017
Tagliarini, PhD
How Does It Work? (continued)



If ek is an eigenvector of B with eigenvalue
lk, then B ek = lk ek
Since Q B = A Q,
A (Q ek) = Q (B ek) = Q (lk ek) = lk (Q ek)
Note from this:
–
–
34
lk is an eigenvalue of A
Q ek is the corresponding eigenvector of A
1/12/2017
Tagliarini, PhD
The QR Method: Start-up

Given a matrix A
Apply Householder’s Algorithm to obtain a
matrix B in upper-Hessenberg form

Select >0 and m>0

–
–
35
 is a acceptable proximity to zero for subdiagonal elements
m is an iteration limit
1/12/2017
Tagliarini, PhD
The QR Method: Main Loop
Do {
Set Q T  I
For k  1 to n - 1{
Bk ,k
c
Bk2,k  Bk21,k
; s
Bk 1,k
2
k ,k
B
B
2
k 1, k
;
Set P  I; Pk,k  Pk 1,k 1  c; Pk 1,k   Pk,k 1  s;
B  PB ;
Q T  PQ T ;
}
B  BQ;
i  ;
} While (B is not upper block tria ngular) and (i  m)
36
1/12/2017
Tagliarini, PhD
The QR Method: Finding The l’s
Since B is upper block tria ngular, one may compute lk from
the diagonal blocks of B. Specifical ly, the eigvalues of B are
the eigenvalue s of its diagonal blocks B k .
If a diagonal block B k is 1x1, i.e., B k  a , then lk  a.
a b 
If a diagonal block B k is 2x2, i.e., B k  
,

c d 
trace(B k )  trace 2 (B k )  4 det( B k )
then lk ,k 1 
.
2
37
1/12/2017
Tagliarini, PhD
Details Of The Eigenvalue Formulae
a b 
Suppose B k  
.

c d 
l  a  b 
lI  B k  

 c l  d
lI  B k  ?
38
1/12/2017
Tagliarini, PhD
Details Of The Eigenvalue Formulae
l  a  b 
Given lI  B k  

 c l  d
lI  B k  (l  a)(l  d )  bc
 l  (a  d )l  ad  bc
2
 l  trace(B k )l  det( B k )
2
39
1/12/2017
Tagliarini, PhD
Finding Roots of Polynomials



40
Every n x n matrix has a characteristic
polynomial
Every polynomial has a corresponding n x n
matrix for which it is the characteristic
polynomial
Thus, polynomial root finding is equivalent to
finding eigenvalues
1/12/2017
Tagliarini, PhD
Example Please!?!?!?

Consider the monic polynomial of degree n
f(x) = a1 +a2x+a3x2+…+anxn-1 +xn
and the companion matrix
 an  an1
 1
0

A 0
1

0
 0
 0
0
41
  a2  a1 
 0
0 
 0
0 

 0
0 
0 1
0 
1/12/2017
Tagliarini, PhD
Find The Eigenvalues of the
Companion Matrix
 an  an1
 1
0

lI  A  lI   0
1

0
 0
 0
0
42
1/12/2017
  a2  a1 

 0
0 
 0
0  ?

 0
0 
0 1
0 
Tagliarini, PhD
Find The Eigenvalues of the
Companion Matrix
l  an an1  a2
1
l  0
lI  A  0
1  0
0
0  l
0
0 0 1
a1
0
0
0
l
an1  a2 a1

 (l  an )ln1  (1)
1  0
0  l
0
43
0
0
0 1 l
1/12/2017
Tagliarini, PhD

Finding Eigenvalues and Eigenvectors

Transcription

Similar documents