Scribd
Upload
2 views

lec12 (1)

This document discusses computational methods for eigenvalues, focusing on Householder's method and the QR algorithm for matrix decomposition. It explains the QR decomposition using Householder transforms, the transformation of matrices into upper Hessenberg or tridiagonal forms, and the application of the QR algorithm for finding eigenvalues. Additionally, it covers initial value problems for ordinary differential equations (ODEs) and various numerical methods for their solutions, including forward and backward Euler methods.

Uploaded by

kedir
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
2 views

lec12 (1)

This document discusses computational methods for eigenvalues, focusing on Householder's method and the QR algorithm for matrix decomposition. It explains the QR decomposition using Householder transforms, the transformation of matrices into upper Hessenberg or tridiagonal forms, and the application of the QR algorithm for finding eigenvalues. Additionally, it covers initial value problems for ordinary differential equations (ODEs) and various numerical methods for their solutions, including forward and backward Euler methods.

Uploaded by

kedir
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/ 5

Computational methods-Lecture 12

Householder’s method and QR method for eigenvalues

1 First approach for QR decomposition using House-


holder transform
Using the Householder transforms, one can obtain the following result:

Theorem 1 (QR decomposition). If A is nonsingular, then there exists an


orthogonal matrix Q such that

A = QR,

where R is upper triangular. If one requires the diagonal elements of R to


be positive, then the decomposition is unique.
 Moreover,
 if A is of size m × n (m > n) and with rank n, then QA =
R
.
0
The proof is straightforward: we apply Householder’s transform repeat-
edly.
We first pick Q1 = H1 that maps the the first column of A to −σ1 e1 .
Let
A1 = Q1 A.
Next, one is going to construct matrix of the form
 
1 0
Q2 =
0 H2

where H2 will map the first column of the (2 : n, 2 : n) submatrix of A1 into


−σ2 e1 . Then,
A2 = Q2 A1 .
Repeating this process, the theorem can be proved.
Another approach that is useful for the QR decomposition, and
in fact used in practice for tridiagonal matrices is based on Givens
transform. This is not required in this course.

1
2 Householder’s method
We have seen that using Householder transforms one can have the QR
decomposition
A = QR.
However. this is not a similar transform. A question is: can we use
Householder transform to obtain a similar transform to obtain a band matrix
with small bandwidth?
Below, we try to argue that using Householder transform, we can find
similar transformations to change

• a real square matrix into Upper Hessenberg matrix (i.e. hij = 0 for
i ≥ j + 2)

• a real symmetric matrix into a tridiagonal matrix.

Let us make a first attempt: take Q1 = H1 that maps the first column
of A into −σ1 e1 . Then, consider the similar transform:

H1 AH1T = [−σ1 e1 v2 · · · vn ]H1

Unfortunately, after we multiply H1 on the right, the matrix will generally


be full again. Hence, we cannot desire an orthogonal similar transform that
changes A into upper triangular or so.
An important observation is that though we cannot change the first
column into a vector with only one nonzero component, we can change it to
a vector with two nonzero components!
Let the first column of A be
 
a11
u1 =
c1

We take H1 such that H1 c1 = −σ1 e1 . Consider the transform


 
1 0
Q1 =
0 H1

Then, we have !
(1)
a11 A12
Q1 AQT1 = (1) .
H1 c1 H1 A22 H1
One can see that Hc1 has only one nonzero entry.
Note that Q1 is also a reflection matrix.

2
Next, one considers  
I2 0
Q2 =
0 H2
where H2 is of size (n − 2) × (n − 2). We use this matrix to change the vector
(1)
below the first entry in the first column of H1 A22 H1 into −σ2 e1 .
Repeat this process, we have
Theorem 2. Let A be a real square matrix. Then, there exists reflection
matrices Q1 , · · · , Qn−2 such that
Qn−2 · · · Q1 AQ1 · · · Qn−1
is upper Hessenberg. If A is symmetric, this matrix is tridiagonal.

3 The QR Algorithm for finding all eigenvalues


The QR algorithm can be used for find all eigenvalues of matrices pro-
vided some conditions. The QR algorithm is as following.
QR algorithm
1. Let A1 ← A
2. For k = 1, · · · ,
Ak = Qk Rk
Ak+1 = Rk Qk .
Below is a simple observation:
Lemma 1. Ak is similar to Ak+1 .
In fact, Ak+1 = Rk Qk = (Q−1 T
k Ak )Qk = Qk Ak Qk .
That means the sequence {Ak } will preserve eigenvalues!
The following fact then guarantees the convergence (for general case,
read the reference books):
Theorem 3. Suppose that A has eigenvalues |λ1 | > |λ2 | > · · · > |λn |, and
A = XDX −1 such that X −1 has LU decomposition. Then, the diagonal
elements of Ak converge to the eigenvalues.
Though we have a way for finding QR decomposition using Householder
transform, it is not simple. In the case that the matrix is upper Hes-
senberg or tridiagonal, one can use Givens transform to find the QR
decomposition efficiently.
Hence, the for general matrices, the approach is as following:

3
1. Use the Householder’s method to transform a matrix into upper Hes-
senberg or tridiagonal.

2. Apply the QR algorithm (using Givens transform) to find all eigenval-


ues.

4 Initial value problems for ODEs


The initial value problem (also called Cauchy problem) of an ODE is
Consider the ODE

u0 (t) = f (t, u(t)), u(t0 ) = u0 ,

where u could be a vector valued function. Any ODE can be reduced to a


first order system, so this is general enough.
For studying ODEs, we often assume

Assumption 1. The function f is Lipschitz continuous in u with a uniform


Lipschitz constant. In other words, there exists L > 0 such that

|f (t, x) − f (t, y)| ≤ L|x − y|

This condition guarantees the existence and uniqueness of the solution


to the ODE for a given initial value.
Suppose we care the solution at T . For discretization, we take the time
step:
T
h= .
N
Define
tn := nh.
The ODE solvers are all approximations to
Z tn+1
u(tn+1 ) = u(tn ) + f (s, u(s))ds.
tn
R tn+1
tn f (s, u(s)) will be approximated by data u0 , u1 , . . . , un+1 .
Below we consider some simple approximations.

• If we approximate f (s, u(s)) ≈ f (tn , un ), then we have the forward


Euler:
un+1 = un + hf (tn , un )

4
• f (s, u(s)) ≈ f (tn+1 , un+1 ), we have the backward Euler:

un+1 = un + hf (tn+1 , un+1 )

• f (s, u(s)) ≈ 12 (f (tn , un ) + f (tn+1 , un+1 )), then we have the trapezoidal
method:
h
un+1 = un + (f (tn , un ) + f (tn+1 , un+1 ))
2
• (*) The higher order Taylor methods (see below, not required)

The forward Euler and backward Euler methods are different in the sense
that the foward Euler is an explicit scheme while the backward Euler is an
implicit scheme. By “implicit”, we mean the right hand side contains the
desired solution un+1 , so un+1 cannot be evaluated directly. Instead, one
must solve the algebraic equation to obtain un+1 .
Example For the equation y 0 = λy. Apply forward Euler scheme and
backward Euler scheme. Take h → 0, does it give the true solution?
The forward Euler has

yn = y0 (1 + λh)n ,

while the backward Euler gives


y0
yn = .
(1 − λh)n

For any λ, we have


y0
y0 (1 + λh)n ≤ y0 eλnh ≤ .
(1 − λh)n

This means that the explicit method often gives a solution that is below the
true solution while an implicit method often gives a solution above the true
solution.
Taking h → 0, we have
yt/h → eλt .

You might also like