Scribd
Upload
68 views13 pages

Reduction To Hessenberg or Tridiagonal Form: Naeemullah Khan

This document discusses two algorithms for reducing a matrix to Hessenberg or tridiagonal form: Householder reduction and Gram-Schmidt orthogonalization. Householder reduction uses Householder reflections to sequentially zero out subdiagonal entries until the matrix is upper Hessenberg. Gram-Schmidt orthogonalization uses orthogonal transformations from the Gram-Schmidt process to zero out entries. Both algorithms are demonstrated with MATLAB code. The reduced forms are useful for efficiently computing eigenvalues and have computational costs of O(n^3) flops. The process is numerically stable up to machine precision.

Uploaded by

Naeemullah Khan
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
68 views13 pages

Reduction To Hessenberg or Tridiagonal Form: Naeemullah Khan

This document discusses two algorithms for reducing a matrix to Hessenberg or tridiagonal form: Householder reduction and Gram-Schmidt orthogonalization. Householder reduction uses Householder reflections to sequentially zero out subdiagonal entries until the matrix is upper Hessenberg. Gram-Schmidt orthogonalization uses orthogonal transformations from the Gram-Schmidt process to zero out entries. Both algorithms are demonstrated with MATLAB code. The reduced forms are useful for efficiently computing eigenvalues and have computational costs of O(n^3) flops. The process is numerically stable up to machine precision.

Uploaded by

Naeemullah Khan
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 13

Reduction to Hessenberg or

Tridiagonal Form
Naeemullah Khan

Outline

Eigenvalue Computation Problems


Hessenberg or Tridiagonal Form
Algorithm 1 (Householder Reduction)
Matlab Code
Results
Algorithm 2 (Gram-Schmidt Orthogonalization)
Matlab Code
Results
Uses
Stability analysis

Eigenvalue Computation
Most of the general purpose eigenvalue problem
proceed by computing Schur Factorization.
A
Q j* A Q j
Where A converges to an upper-triangular matrix as
j
The above transformation is split into 2 phases
xxxx
xxxx
xxxx
xxxx

AA*

Phase 1

xxxx
xxxx
xxx
xx

Phase 2

xxxx
xxx
xx
x

Hessenberg or Tridiagonal Form


To compute Schur Factorization A=QTQ*, we apply unitary similarity
transformation to A.
x
x
x
x

x
x
x
x

x
x
x
x

x
x
x
x

Q1*.

x x
x
x
x

Q1*A

x
x
x
x

Q1*A

A
x x x x
x x x
x x
x

x
x
x
x

.Q1

x
x
x
x

x
x
x
x

x
x
x
x

x
x
x
x

Q1*A Q1

What should we do then? Try next best thing.


x x x x
x x x x
x x x x
x x x x

x x x x
Q1*.

A
x x x
x x x
x x
x x

x
x
x
x

x x x x
x x x
x x x

Q1*A

.Q

x x x
x x x
x x
x x

x
x
x
x

Q1*A
Q1*A Q1
After repeating the process m-2 times, we have the
product in Hessenberg form.

Algorithm: Householder Reduction to


Hessenberg form
for k=1 to m-2 do
x=A(k+1:m,k)
vk=sign(x1) llxkll2e1+x
vk=vk/llvkll2
A(k+1:m,k:m)=A(k+1:m,k:m)-2vkvk*A(k+1:m,k:m)
A(1:m,k+1:m)=A(1:m,k+1:m)-2(A(1:m,k+1:m)vk)vk*
end for

Matlab Code (Householder Reduction)


clc
clear
A=5*rand(4)
B=A
A1=A
m=length(A)
for k=1:m-2
x=A(k+1:m,k);
v=x;
v(1)=sign(x(1))*norm(x)+x(1);
v=v/norm(v);
F=v*v'/norm(v);
F=eye(size(F))-2*F;
q=eye(4);
q(k+1:m,k+1:m)=F;
B=q*B*q';
A(k+1:m,k:m)=A(k+1:m,k:m)-2*v*(v'*A(k+1:m,k:m));
A(1:m,k+1:m)=A(1:m,k+1:m)-2*(A(1:m,k+1:m)*v)*v';
end
A;
[Q,AA]=hess(A1);

Results
A=
1.1449
4.5667
0.7619
4.1291

Q1 =
2.6917
4.9807
0.3909
2.2134

0.5333
4.8095
0.0232
3.8746

4.0865
4.3435
0.4222
1.9989

H_A =
1.1449 -4.7669 -1.2014 0.2496
-6.2036 7.6190 0.7615 5.1709
0 -2.2754 0.0558 -0.1990
0
0 -0.0610 -0.6720
Hess_Q =
1.0000
0
0
0
0 -0.7361 0.6766 -0.0182
0 -0.1228 -0.1600 -0.9794
0 -0.6656 -0.7188 0.2009
Hess_A =
1.1449 -4.7669 -1.2014 0.2496
-6.2036 7.6190 0.7615 5.1709
0 -2.2754 0.0558 -0.1990
0
0
-0.0610 -0.6720

1.0000
0
0
0
0 -0.7361 -0.1228 -0.6656
0 -0.1228 0.9913 -0.0471
0 -0.6656 -0.0471 0.7448
A1 =
1.1449
-6.2036
0.0000
0.0000

-4.7669
7.6190
0.4730
2.2257

0.0056
-5.2162
-0.6934
-0.0398

1.2270
0.3301
0.0982
0.0772

Q2 =
1.0000
0
0
0
0 1.0000
0
0
0
0 -0.2079 -0.9782
0
0 -0.9782 0.2079
A2=
1.1449
-6.2036
-0.0000
0.0000

-4.7669
7.6190
-2.2754
0.0000

-1.2014
0.7615
0.0558
-0.0610

0.2496
5.1709
-0.1990
-0.6720

Matlab Code (Gram-Schmidt)


clc
clear
A=5*rand(4)
eig_A=eig(A)
hess(A)
m=length(A);
for i=2:m-1
Q=try1(A(i:m,i-1:m-1),m)
A=Q'*A*Q
end
eig_AH=eig(A)

function Q=try1(A,n)
m=length(A);
m=length(A)
for i=1:m
v(:,i)=A(:,i);
end
for i=1:m
r=norm(v(:,i));
q(:,i)=v(:,i)/r;
for j=i+1:m
r2=q(:,i)'*v(:,j);
v(:,j)=v(:,j)-r2*q(:,i);
end
end
Q=eye(n);
l=length(q);
Q(n-l+1:n,n-l+1:n)=q;
end

Results
A=
0.4930
0.7101
0.8413
0.9812

A1 =
1.5874
1.5821
1.0878
1.2552

4.4646
3.5161
2.7787
0.9222

1.0602
0.3867
4.5690
3.5336

eig_A =
7.5140
-0.1446
0.5090 + 1.6048i
0.5090 - 1.6048i
Hess_A
0.4930 -4.0166 -2.7283 0.0087
-1.4747 6.7104 -0.2253 -1.2007
0 1.6597 0.1011 -0.9926
0
0 2.8746 1.0830
Q1 =
1.0000
0
0
0
0 0.4815 0.8763 -0.0132
0 0.5704 -0.3019 0.7638
0 0.6654 -0.3754 -0.6453

0.4930
1.4747
-0.0000
0.0000

4.0166
6.7104
-0.2106
1.6463

-0.3548
-1.1624
0.8303
-0.8993

2.7051
-0.3758
2.9679
0.3538

Q2 =
1.0000
0
0
0
0 1.0000
0
0
0
0 -0.1269 0.9919
0
0 0.9919 0.1269
A 2=
0.4930
1.4747
0.0000
-0.0000

4.0166
6.7104
1.6597
-0.0000

eig_AH =
7.5140
-0.1446
0.5090 + 1.6048i
0.5090 - 1.6048i

2.7283
-0.2253
0.1011
2.8746

-0.0087
-1.2007
-0.9926
1.0830

Uses
Work for Hessenberg reduction: 10/3(m3)
For Hermitian case reduces to tridiagonal
form: 4/3(m3)
Phase 2 applied to a matrix A require O(m4)
flops
Phase 2 applied to Hessenberg matrix requires
O(m3) steps.
For Hermitian case for calculation of ONLY
eigenvalues phase 2 require O(m2) flops

Stability Analysis
If A Cmxm , Let A=QHQ* and the computed
factors are Q1H1Q1* then we have
Q1H1Q1*=A+A,

llAll/llAll=O(machine)

Questions?

You might also like