Chapter 4 Solving Eigenvalues and

Eigenvectors of Matrix
Definition 4.1 For an n-order matrix A ， if there
exists  and vector x (nonzero) such that
Ax=x
We say  is an eigenvalue of A, and x is an eigenvector
A corresponding to .
Note. If  is an eigenvalue of A, then
det( I  A)  0

24/6/19 1
 Theorem 4.1 1 , 2 , , n are eigenvalues of A, and p(x)
is a polynomial, then p(A) has eigenvalues p (1 ), p(2 ), , p(n ) .

(1) Ak has ei genval ues 1k , 2k , , nk ; (2)I f A i s i nver t i bl e, t hen
1 1 1
， ，， i s ei genval ues of A1 wi t h t he s ame ei genvect or s .
1 2 n

Theorem 4.2 If A is real symmetric, then all

eigenvalues of A are real numbers and all eigenvectors
are orthogonal each other. Furthermore, there exists
orthogonal matrix Q, such that
QT AQ  diag (1 , 2 , , n )
Theorem 4.3 If|P|0 ， B=P-1AP ， we say A is similar to B.

24/6/19 2
 §1 The Power Method

1.1 The Power Method

Using power method, we can obtain the eigenvalue
with the greatest modulus (dominant eigenvalue)and
eigenvector with respect to it.
Theorem 4.4 A has n linear independent eigenvectors
x1,x2,…,xn ， and  1,  2,… ,  n are corresponding
eigenvalues, such that
| 1 || 2 || 3 |  | n |
then for any an initial vector v0,
vk  Avk 1 =  =A v0
k
k  1, 2,  ,

the series {vk } satisfy

24/6/19 3
 vk
(1) lim  a1 x1
k   1
k

(vk 1 ) m
(2) lim  1
k  (v )
k m

wher e (vk ) m i s m- t h ent r y of vk .

Proof.
Because x1,x2,…,xn are linear independent ， for any
vector v0, we have
v0  a1 x1  a2 x2    an xn (a1  0)
vk  Ak v0  a1 Ak x1  a2 Ak x2    an A k xn
 a11k x1  a2 2k x2    an nk xn

 1k [ a1 x1  a2 ( 12 ) k x2    an ( 1n ) k xn ]
24/6/19 4
 2 n
Supposeα1≠0 ， and | 1 | 1, ,| 1 | 1
i k
n
 [ a1 x1   ai ( ) xi ]
k

vk 1
i 2 1
lim k  lim
k   k   k
1 1
n

 lim[ a1 x1   ai ( i ) k xi ]  a1 x1
k 
i 2 1

n
i k 1
{ [ a1 x1   ai ( )
k 1
xi ]}m
(vk 1 ) m 1
i 2 1
lim  lim
k  (v ) k  n
i k
k m
{1 [ a1 x1   ai ( ) xi ]}m
k

i 2 1
( a1 x1 ) m ( x1 ) m
 1  1  1
( a1 x1 ) m ( x1 ) m
24/6/19 5
Note 1. The speed of convergence is determined by the
2
value | | . When the value is small, the speed is fast.
1
Note 2.Even the dominant eigenvalue is multiple root, the
result still holds.
Let 1  2    r，and |r || r 1 |  | n |
r n
vk  A v0   [ ai xi  a(
k k j k
1 j 1) xj ]
i 1 j  r 1
r n j k
 [ ai xi   a j ( ) x j ] r
k

vk
1
1
  ai xi
i 1 j  r 1
lim k  lim
k   k  1 k
1 i 1
r n  j k 1 r
{1 [ ai xi   a j ( ) x j ]}m
k 1
[ ai xi ]m
(v ) i 1 j  r 1 1
lim k 1 m  lim   i 1
 1
k  (v ) k  r n j k 1 r
k m
{1 [ ai xi   a j ( ) x j ]}m
k
[ ai xi ]m
i 1 j  r 1 1 i 1
24/6/19 6
Note 3 When |λ1|>1,elements of {vk} increase with |
λ1|k tend to infinite. On the contrary, if|λ1|<1,all
elements tend to 0.
For dealing with the problem, we only need normalize
all vector series.

1.2 Modified Power Method

Let v is a nonozero vector, normalize it

v
u
max(v)
where max(v) is the entry with greatest absolute value
of v. v0
u0 
max(v0 )
24/6/19 7
 Av0 v1 Av0
v1  Au0  , u1  
max(v0 ) max(v1 ) max( Av0 )
A2 v0 v2 A 2
v0
v2  Au1  , u2  
max( Av0 ) max(v2 ) max( A2v0 )
  
Ak v0 vk Ak v0
vk  Auk 1  k 1
, uk  
max( A v0 ) max(vk ) max( Ak v0 )
Finally,
n
i k
 [a1 x1   ai ( ) xi ]
k
1
i2 1 x1
lim uk  lim 
k  k  n
i k max( x1 )
max{1 [a1 x1   ai ( ) xi ]}
k

i 2 1
24/6/19 8
 n
i k
 [a1 x1   ai ( ) xi ]
k
1
i 2 1 1 x1
lim vk  lim 
k  k  n
i k 1 max( x1 )
max{1 [a1 x1   ai ( ) xi ]}
k 1

i2 1
lim max(vk )  1
k 

24/6/19 9
Eg. 4.1 Show the dominant eigenvalue and eigenvector
of matrix 2 4 6
 
A   3 9 15 
 4 16 36 
 
Solution. As showed in the table
k vk uk
0 1 1 1 1 1 1
1 12.00 27.00 56.00 0.2143 0.4821 1
2 8.357 19.98 44.57 0.1875 0.4483 1
3 8.168 19.60 43.92 0.1860 0.4463 1
4 8.157 19.57 43.88 0.1859 0.4460 1
5 8.157 19.57 43.88 0.1859 0.4460 1
24/6/19 10
 1.3 Inverse Iteration
We can use the inverse iteration to solve the eigenvalue
with smallest absolute value and its eigenvector. Let
| 1 || 2 |  | n 1 || n |
Corresponding eigenvectors x1 , x2 , , xn
Then the eigenvalues of A-1 satisfy
1 1 1
 
| n | | n 1 | | 1 |
When we solve the dominant eigenvalue of A-1, we obtain
the eigenvalue with smallest absolute value of A.
For any initial nonzero vector v0 ， formulate the series
v0
u0 
max(v0 )
24/6/19 11
 1 vk
vk  A uk 1 , uk  , k  0,1, 2, 
max(vk )
xn 1
lim uk  , lim max(vk ) 
k  max( xn ) k  n
Note.
We can get vk through solving the system Avk  uk 1

24/6/19 12
 §2 Jacobi’s Method
Jacobi’s method is a global method and can
obtain all eigenvalues and eigenvectors of a
matrix. When the matrix A is symmetric, its all
eigenvalues are real numbers and all eigenvectors
are orthogonal each other. Furthermore, there
exists a orthogonal matrix Q, such that

QT AQ  diag (1 , 2 , , n )

24/6/19 13
 2.1 Diagonalization of a real symmetric matrix

The idea of Jacobi’s method is finding s series of

orthogonal matrices {S k } , such that

lim S1S 2  S k  Q
k 

We have
Tk  SkT SkT1  S 2T S1T AS1S2  Sk 1Sk  SkT Tk 1Sk
 diag (1 , 2 , , n )

The series {Tk} are similar each other.

tij( k ) , sij( k ) ar e ent r i es of Tk and S k .
24/6/19 14
Definiton 4.2
n n n n
vk   (tij( k ) ) 2 , wk   (tij( k ) ) 2 k  1, 2,
i 1 j 1 i 1 j 1
j i
We restrict {Tk} satisfying
(1) wk 1  wk , vk 1  vk k
(2) lim vk  0
k 

we have
lim Tk  diag (1 , 2 , , n )
k 

24/6/19 15
 p q

1 
  
 p
 cos  k  sin  k 
 
Sk  S ( p, q ) 
    
q
  sin  k  cos  k 
 
  
 1 


S(p,q) is orthogonal, S(p,q)TA S(p,q) only change

p-th and q-th row, p-th and q-th column of A.
S(p,q) is called Givens matrix.
24/6/19 16
 Tk  S kT Tk 1S k , so we have
t (pjk )  t (pjk 1) cos  k  tqj( k 1) sin  k
 (k )
tqj  t (pjk 1) sin  k  tqj( k 1) cos  k j  p, q
 (k )
tip  tip( k 1) cos  k  tiq( k 1) sin  k
 (k )
tiq  tip( k 1) sin  k  tiq( k 1) cos  k i  p, q
t ( k )  t (ppk 1) cos 2  k  tq( qk 1) sin 2  k  2t (pqk 1) sin  k cos  k
 pp
tqq
(k )
 t (ppk 1) sin 2  k  tqq
( k 1)
cos 2  k  2t (pqk 1) sin  k cos  k

t ( k ) 1 ( k 1) ( k 1)
 pq  t  (t pp  tqq ) sin 2 k  t (pqk 1) cos 2 k
(k )
qp
2
 (k ) ( k 1)
tij  tij i  p , q; j  p , q
24/6/19 17
 Select k ,such that t (pqk )  0
( k 1)
tqq  t (ppk 1)  
a  cot 2 k  ( k 1)
,   2 k 
2t pq 2 2
Let t  tan  k，accor di ng t o tan  k  2 tan k cot 2k  1  0,
2

t 2  2at  1  0
 1
 a0
a  1  a2
t
 1
a0
 a  1 a 2

1

24/6/19 | a |  1  a2 18
 1
and cos  k  , sin  k  t  cos  k
1 t 2

Theorem 4.5 The linear transform constructed by the

above way satisfies
wk  wk 1 , vk  vk 1
and lim Tk  diag (1 , 2 , , n )
k 

2.2 The Classic Jacobi’s Method

The classic Jacobi’s method always change the
non-diagonal entries with the greatest absolute value
into 0.Normally,the process need infinite times
because the next transform could change 0 entry into
nonzero again.

24/6/19 19
The process of Algorithm
（ 1 ） Select the entry with the greatest value
| ti1 j1 | max | aij | 0
i j

Obtain the matrix S1  S (i1 , j1 ) such that T1  S1 AS1 with

T

ti(1)
1 j1
 t j1i1  0
(1)

（ 2 ） Similarly, select
| ti(1)
2 j2
| max | t ij | 0
(1)
i j
And matrix S 2  S (i2 , j2 ) such that T2  S 2 T1S 2 with
T

ti(2)
2 j2
 t j2i2  0
(2)

（ 3 ） Continue the above process until get a solution

satisfying precision.
24/6/19 20
Eg. 4.2 Show all eigenvalues and eigenvector of matrix
 1.0 1.0 0.50 
 
A   1.0 1.0 0.25  ,
 0.50 0.25 1.0 
 
using Jacobi’s method with precision 0.0005 。
Solution. （ 1 ） Change the entriesa12 , a21 of A into 0
a22  a11
a1  0  t1  1
2a12
 2
1   cos 1  sin 1 
4 2
 0.70711 0.70711 0 
 
S1   0.70711 0.70711 0 
 0 0 1 

24/6/19 21
  0 0 0.17678 
 
T1  S1T AS1   0 2 0.53038 
 0.17678 0.53038 2 
 
(1) (1)
（ 2 ） Change the entriest 23 ,t 32
of T2 into 0
(1)
t33  t22(1)
a2  (1)
0  t2  1
2t23
 2
2   cos  2  sin  2 
4 2
1 0  0
 
S2   0 0.70711 0.70711
0 0.70711 0.70711

 0 0.12500 0.12500 
 
T2  S 2T T1S 2   0.12500 1.4697 0 
 0.12500 0 2. 5304 
24/6/19   22
 (2) (2)
（ 3 ） Change the entriest12 and t21 of T2 into 0
(2)
t22  t11(2) 1
a3   5.8788  t3 
2t12(2) 5.8788  1  (5.8788) 2
1
cos  3   0.99645, sin  3  t3  cos 3  0.084145
1 t 2
3

 0.99645 0.84145 0 
 
S3   0.84145 0.99645 0 
 0 0 1 
 
 0.010556 0 0.12456 
 
T3  S3T T2 S3   0 1.4802 0.010518 
 0.12456 0.010518 2.5304 
 

Continue the process

24/6/19 23
  0.016647 0.000409 0.000005 
 
T5   0.000409 1.4801 0 
 0.000005 0 2.5336 
 
 0.72135 0.44404 0.53584 
 
Q  S1S 2 S3 S 4 S5   0.68616 0.56234 0.46147 
 0.093844 0.69757 0.71033 
 
Finally, we get three eigenvalues of A with precision 0.0005
1=-0.016647 ， 2=1.4801 and 3=2.5366, the three eigenvectors
are
 0.72135   0.44404   0.53584 
     
x1   0.68616  , x2   0.56234  , x3   0.46147 
 0.093844   0.69757   0.71033 
     

24/6/19 24