Transcript Numerical Methods - Hanyang University

Numerical Analysis –
Eigenvalue and Eigenvector
Hanyang University
Jong-Il Park
Eigenvalue problem
Ax  x
 : eigenvalue
x : eigenvector
※ spectrum: a set of all eigenvalue
Department of Computer Science and Engineering, Hanyang University
Eigenvalue
 Eigenvalue
( A  I ) x  0
if det ( A  I )  0
 x=0(trivial solution)
To obtain a non-trivial solution,
det ( A  I )  0
 a 
a

a
11
a21

an1
12
1n
a22   


an 2
a2 n
a11
0
 ann  
;Characteristic equation
Department of Computer Science and Engineering, Hanyang University
Properties of Eigenvalue
n
n
i 1
i 1
1) Trace A   aii   i
n
2) det A   i
i 1
3) If A is symmetric, then the eigenvectors are
i j
 0,
orthogonal: T
xi x j  
i j
Gii
4) Let the eigenvalues of A  1 , 2 ,n
then, the eigenvalues of (A - aI)
 1  a, 2  a,, n  a,
Department of Computer Science and Engineering, Hanyang University
Geometrical Interpretation of
Eigenvectors
 Transformation
Ax
Ax  x : The transformation of an eigenvector
is mapped onto the same line of.
 Symmetric matrix  orthogonal eigenvectors
 Relation to Singular Value
if A is singular  0
{eigenvalues}
Department of Computer Science and Engineering, Hanyang University
Eg. Calculating Eigenvectors(I)
 Exercise
1)  5 2  ; symmetric, non-singular matrix
 2  2
(  = -1, -6)


2)   5 1  ; non-symmetric, non-singular matrix
  2  2
(  = -3, -4)


Department of Computer Science and Engineering, Hanyang University
Eg. Calculating Eigenvectors(II)
3) 1 2
2 4


; symmetric, singular matrix
(  = 5, 0)
4) 1 2
3 6


; non-symmetric, singular matrix
(  = 7, 0)
Department of Computer Science and Engineering, Hanyang University
Discussion
 symmetric matrix
=> orthogonal eigenvectors
 singular matrix
=> 0
{eigenvalue}
 Investigation into SVD
Department of Computer Science and Engineering, Hanyang University
Similar Matrices
 Eigenvalues and eigenvectors of similar matrices
Eg. Rotation matrix
Department of Computer Science and Engineering, Hanyang University
Similarity Transformation
 Coordinate transformation
 x’=Rx, y’=Ry
 Similarity transformation
 y=Ax
 y’=Ry=RAx=RA(R-1 x’)= RAR-1 x’=Bx’
B = RAR-1
Department of Computer Science and Engineering, Hanyang University
Numerical Methods(I)
 Power method

Iteration formula
Ax( k )  y ( k 1)  ( k 1) x( k 1)

for obtaining large 
Department of Computer Science and Engineering, Hanyang University
Eg. Power method
Department of Computer Science and Engineering, Hanyang University
Numerical Methods (II)
 Inverse power method

Iteration formula
1 ( k )
A x
y
( k 1)
 Ay
( k 1)
x
(k )
 Lc  x ( k ) ,Uy( k 1)  c

for obtaining small
Department of Computer Science and Engineering, Hanyang University
Exploiting shifting property
Let the eigenvalues of A  1 , 2 ,n
then, the eigenvalues of (A - aI)
 1  a, 2  a,, n  a,
• Finding the maximum eigenvalue with opposite sign after obtaining

• Accelerating the convergence when an approximate eigenvalue is available
Department of Computer Science and Engineering, Hanyang University
Deflated matrices
 It is possible to obtain eigenvectors one after another
 Properly assigning the vector x is important
 Eg. Wielandt’s deflation
Department of Computer Science and Engineering, Hanyang University
Eg. Using Deflation(I)
Department of Computer Science and Engineering, Hanyang University
Eg. Using Deflation(II)
Department of Computer Science and Engineering, Hanyang University
Numerical Methods (III)
 Hotelling's deflation method

Iteration formula:
Ai 1  Ai  i xi xiT


given i , xi
for symmetric matrices
deflation from large to small

Department of Computer Science and Engineering, Hanyang University
Numerical Methods (IV)
 Jacobi transformation

Successive diagonalization without changing  .

for symmetric matrices
Department of Computer Science and Engineering, Hanyang University
Homework #7
[Due: Nov. 19]
 Generate a 9x9 symmetric matrix A by using random
number generator(Gaussian distribution with mean=0
and standard deviation=1.1]). Then, compute all
eigenvalues and eigenvectors of A using the routines in
the book, NR in C. Print the eigenvalues and their
corresponding eigenvectors in the descending order.

You may use
 jacobi(): Obtaining eigenvalues using the Jacobi
transformation
 eigsrt(): Sorting the results of jacobi()
Department of Computer Science and Engineering, Hanyang University