Power Method for Approximating Eigenvalues - (4.

1)

1. Symmetric matrices:
If A is a symmetric n  n matrix, then eigenvalues of A are real numbers and there exist n eigenvectors
that are orthogonal.
2 −1 1 −1
Example A  . , v : 1, 3, , v T1 v 2  0.
−1 2 1 1

2. Gersgorin Circle:
Let A be an n  n matrix and R i be the circle in the complex plane C with center a ii and radius ∑ j1,j≠i |a ij |
n

such that
n
Ri  z ∈ C; |z − a ii | ≤ ∑ |a ij | .
j1, j≠i

The eigenvalues of A are contained within R   ni1

R i . Moreover, the union of any k of these circles that
do not intersect the remining n − k contains precisely k (counting multiplicities) of the eigenvlaues.
4 1 1
Example Let A  0 2 1 . Use the Gersgorin Circle Theorem to determine bounds for the
−2 0 9
eigenvalues of A.

2
y

R1  z; |z − 4| ≤ 2 ,
R2  z; |z − 2| ≤ 1 , 0
1 2 3 4 5 6 7 8 9 10 11
R3  z; |z − 9| ≤ 2 . x

-1

-2

Since R 1 and R 2 are disjoint from R 3 , there are two eigenvalues within R 1  R 2 and one within R 3 .
Hence,
1 ≤ | 1 |, | 2 | ≤ 6 and 7 ≤ | 3 | ≤ 11
Moreover, 7 ≤ A ≤ 11 and since 1 ≤ | 1 |, A is nonsingular.
Eigenvalues of A are: 8.48534949329733, 4.63182668375403, 1.88282382294864

1
 4 −1 −1
Example Let A  −1 4 −1 . Use the Gersgorin Circle Theorem to determine bounds for the
−1 −1 4
eigenvalues of A.
A is symmetric so all its eigenvalues are real. R 1  z ∈ R; |z − 4| ≤ 2 , 2 ≤ z ≤ 6. A is positive
definite (then it is also nonsingular).

3. Power Method:
Let A be an n  n real matrix, and  i , v i  for i  1, . . . , n be eigenpairs of A where
| 1 |  | 2 | ≥  ≥ | n | ≥ 0.
The Power Method is an iterative method that finds  1 asymptotically.
Idea: Assume that v 1 , . . . , v n are linearly independent. Let x be in R n and x  v 1    c n v n . Then
Ax  Av 1    c n Av n   1 v 1    c n  n v n
and
A k x   k1 v 1    c n  kn v n , for k  1, 2, . . . .
Since
k k
A k x   k1 v 1  2k   c n nk v n ,
1 1
 k2  kn
lim A k x  lim  k1 v 1    c n vn   k1 v 1 .
k→ k→  k1  k1
When k is very large, A k x ≈  k1 v 1 .
Algorithm: Given A, x and a stopping criterion , let x 0  1 x. Because ||x 0 ||  1, let
||x||  

|x p 0 |  ‖x 0 ‖   1.
Approximate  1 , v 1  as follows: For k  1, 2, . . . ,
(1) Compute y k  Ax k−1 .
k
(2) Compute p k such that y p k  ‖y k ‖  .
k
(3) Let r k  y p k and x k  r1k y k ‖x k ‖   1
(4) If ‖x k − x k−1 ‖   , then  1 ≈ r k and v 1  x k . Otherwise, k  k  1, repeat steps (1)-(4).

−4 14 0 1
Example A  −5 13 0 , x 0  1 .
−1 0 2 1

Symmetric Power Method: x 0  1 x.

‖x‖ 2
(1) Compute y k  Ax k−1 .
T
(2) r k  y k  x k−1 .
(3) Compute ‖y k ‖ 2 . If ‖y k ‖ 2  0, then A has a zero eigenvalue.

2
 (4) Compute x k  1
y k .
y k 2

(5) If ‖x k − x k−1 ‖ 2  , then  1 ≈ r k and v 1  x k . Otherwise, k  k  1, repeat steps (1)-(5).

4 −1 1 1
Example A  −1 3 −2 , x 0  0 .
1 −2 3 0

4. Aitken’s Δ 2 Method:
Let p n  be a sequence which converges to its limit p linearly. That is, there exists a positive number 
such that
p n1 − p
lim  .
n→
pn − p
Can the order of convergence of this sequence be improved? Observe the following. For sufficiently
large n,
p n1 − p
≈ .
pn − p
Assume the signs of p n1 − p and p n − p agree (either both are positive or both are negative) for all n.
Then
p n2 − p p n1 − p 2
p n1 − p ≈ p n − p  p n2 − p p n − p ≈ p n1 − p

p n2 p n − p n2 p − p p n  p 2 ≈ p 2n1 − 2p n1 p  p 2

p n2 p n − p n2 p − p p n ≈ p 2n1 − 2p n1 p
Solve for p :
2
p n p n2 − p 2n1 p n1 − p n
p n p n2 − p 2n1 ≈ p p n2 − 2p n1  p n   p≈  pn − .
p n2 − 2 p n1  p n p n2 − 2 p n1  p n
Define a new sequence p̂ n as
2
p n1 − p n
p n p n2 − p 2n1
p̂ n  p n − or p̂ n 
.
p n2 − 2p n1  p n p n2 − 2 p n1  p n
Algebraically these two formulas are equivalent, but numerically the first one is more stable than the
second one. The sequence p̂ n converges to p more rapidly. This method is called Aitken’s
Δ 2 Method. Observe that estimate p̂ n depends on estimates p n , p n1 , and p n2 . So, p̂ 0 can be
computed after p 2 is computed.
Steps for Aitken’s Δ 2 Method: let p n be generated by a method which has a linear convergence.
2
p1 − p0
Having p 0 , p 1 , and p 2 , compute p̂ 0  p 0 − , and for n  1, 2, . . .
p 2 − 2p 1  p 0
a. compute p n2 ,
2
p n1 − p n
b. compute p̂ n  p n − , and
p n2 − 2p n1  p n
3