28

# normalisation
u:=w/LinearAlgebra[VectorNorm](w,Euclidean);
# iteration
w:=A.u;
# eigenvalue estimate
lambda:=w.u;
end do;
# output
return lambda,u;
end proc:

c) Deflation

So far: given A with |(1) | > |(2) | > . . . > |(n) | we can compute (1) and u(1) . How to
obtain the other eigenvalues (2) , (3) ,. . . , (n) ?

Idea: Construct a matrix B with eigenvalues 0, (2) , (3) , . . . , (n) (deflate the matrix A,
remove (1) ). Then (2) can be obtained by the power method.

3 0
Example 4.3: A = , eigenvalues and eigenvectors
0 2

1 0
(1) = 3, u(1) = (leading), (1) = 2, u(2) =
0 1

3 0 3 0 0 0
B= =
0 2 0 0 0 2
| {z }

1
(3, 0)
0
has eigenvalues 0 and 2. If aT = (3, 0) denotes the 1st row of A (with the 1st component
of u(1) being 1) we have B = A u(1) aT . How to generalise this example?

Construction of B:

A given, (1) and u(1) known (e.g., by power method).

 29

(1)
Choose a nonzero component of u(1) , say up , (fixes p !).

Denote by aT the pth row of A, i.e., aT = (Ap1 , Ap2 , . . . , Apn ). Because of the
eigenvalue equation Au(k) = (k) u(k) we have


(k) u(k)
p = Au
(k)
p
= aT u(k) .

Consider

1 1
B =A (1)
u(1) aT , i.e., Bn = An (1)
u(1) u(1)
n Ap .
up up
B has the required properties. Namely:

1

Bu(1) = A u(1) (1) u(1) aT u(1) = 0
| {z } up | {z }
(1) (1)
u (1) u(1)
p

i.e., 0 is eigenvalue of B (instead of (1) ), and for k = 2, 3, . . . , n

!
(k)
1
up
Bu(k) = A u(k) (1) u(1) aT u = (k) u(k) (1) u(1)
(k)
| {z } up | {z } up
(k) u(k) (k) u(k)
p

that means ! !
(k) (k)
up up
B u(k) (1)
u(1) = (k) u(k) (1)
u(1) .
up up
Thus (k) , k = 2, 3, . . . , n are still eigenvalues of B (but the eigenvectors have
changed).

Example 4.4:
4 14 0


A = 5 13 0

1 0 2
Eigenvalues and eigenvectors

4 1 0

(1) (1)
(2) = 3, u(2) (3) = 2, u(3)

= 6, u = 20/7 , = 1/2 , = 0

1 1 1

Deflate A: choose, e.g., p = 1

(1)
up=1 = 4, aT = (4, 14, 0)
 30

4 14 0 4
1

B = 5 13 0 20/7 (4, 14, 0)
(4)
1 0 2 1

4 14 0 16 56 0
1

= 5 13 0 + 80/7 40 0
4
1 0 2 4 14 0

0 0 0


= 15/7 3 0

2 7/2 2

Eigenvalues and eigenvectors of B

0 0 4


3, 2/7 , 2, 0 , 0, 20/7 (non leading)

1 1 1

0


Deflate again to make the next eigenvalue the leading one: now (1) = 3, u(1) = 2/7 .

1
(1)
Choose e.g., p = 3, then up = 1, aT = (2, 2/7, 2) and

0 0 0 0 0 0 0
1

B = 15/7 3 0 2/7 (2, 7/2, 2) = 11/2 2 4/7 .
1
2 7/2 2 1 0 0 0

Flow chart (for a single deflation step):

choose p OUTPUT
INPUT
A,u (1) such that aT=(A ,...,A ) (1) T (1)
B=Au a /u
(1) p1 pn
u =0 /
p

(another procedure)

(1)
choose p such that up 6= 0 is another procedure:
 31

yes
INPUT (i) OUTPUT
i:=1 v =0
/
v i
no
next i

(1)
Maple code: i) procedure to compute p with up 6= 0:

pindex:=proc(v)
local i,ndim;
# number of components
ndim:=LinearAlgebra[Dimension](v);
# loop over all components
for i from 1 to ndim do
if abs(v[i])>0 then
# nonzero component
return i;
end if;
end do;
end proc:

ii) deflation step

simple_deflat:=proc(A,u)
local p,ap;
# compute p value
p:=pindex(u);
# pth row of A
ap:=A[p];
# output
return A-(u.ap)/u[p];
end proc:

Remark: The procedure pindex is ill-conditioned (see 2). Thus, it has to be replaced