numerical analysis

60
Numerical Analysis 1 EE, NCKU Tien-Hao Chang (Darby Chang)

Upload: natala

Post on 11-Jan-2016

36 views

Category:

Documents


2 download

DESCRIPTION

Numerical Analysis. EE, NCKU Tien-Hao Chang (Darby Chang). In the previous slide. Accelerating convergence linearly convergent Newton’s method on a root of multiplicity >1 (exercises) Proceed to systems of equations linear algebra review pivoting strategies. In this slide. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Numerical Analysis

Numerical Analysis

1

EE, NCKUTien-Hao Chang (Darby Chang)

Page 2: Numerical Analysis

In the previous slide Accelerating convergence

– linearly convergent

– Newton’s method on a root of multiplicity >1

– (exercises)

Proceed to systems of equations– linear algebra review

– pivoting strategies

2

Page 3: Numerical Analysis

In this slide Error estimation in system of equations

– vector/matrix norms

LU decomposition– split a matrix into the product of a lower and a upper

triangular matrices

– efficient in dealing with a lots of right-hand-side vectors

Direct factorization– as an systems of n2+n equations

– Crout decomposition

– Dollittle decomposition

3

Page 4: Numerical Analysis

3.3

4

Vector and Matrix Norms

Page 5: Numerical Analysis

Vector and matrix norms Pivoting strategies are designed to

reduce the impact roundoff error The size of a vector/matrix is

necessary to measure the error

5

Page 6: Numerical Analysis

Vector norm

6

Page 7: Numerical Analysis

7

The two most commonly used norms in practice

Page 8: Numerical Analysis

8

Page 9: Numerical Analysis

9

Page 10: Numerical Analysis

Vector norm

Equivalent One of the other uses of norms is to establish

the convergence

Two trivial questions:– converge or diverge in different norms?

– converge to different limit values in different norms?

The answer to both is no– all vector norms are equivalent

10

0lim )(

xx k

k

Page 11: Numerical Analysis

11

Page 12: Numerical Analysis

12

The Euclidean norm and the maximum norm are equivalent

Page 13: Numerical Analysis

Matrix norms

Similarly, there are various matrix norms, here we focus on those norms related to vector norms– natural matrix norms

13

Page 14: Numerical Analysis

Matrix norms

Natural matrix norms

14

Page 15: Numerical Analysis

15

Page 16: Numerical Analysis

Natural matrix norms

Computing maximum norm

16

Page 17: Numerical Analysis

17

Page 18: Numerical Analysis

18

Page 19: Numerical Analysis

Natural matrix norms

Computing Euclidean norm Is, unfortunately, not as

straightforward as computing maximum matrix norms

Requires knowledge of the eigenvalues of the matrix

19

Page 20: Numerical Analysis

Eigenvalue review

20

later

Page 21: Numerical Analysis

21

Page 22: Numerical Analysis

Eigenvalue review

22

Page 23: Numerical Analysis

23

Page 24: Numerical Analysis

24

In action

http://thomashawk.com/hello/209/1017/1024/Jackson%20Running.jpg

Page 25: Numerical Analysis

25

Page 26: Numerical Analysis

26

Page 27: Numerical Analysis

Any Questions?

27

3.3 Vector and Matrix Norms

Page 28: Numerical Analysis

3.4

28

Error Estimates and Condition Number

Page 29: Numerical Analysis

Error estimation A linear system Ax=b, and x’ is an

approximate solution The error, e=x’-x, cannot be directly

computed (x is never known) The residue vector, r=Ax’-b, can be

easily computed– r=0 x’=x e=0

29

Page 30: Numerical Analysis

Any Questions?

30

Page 31: Numerical Analysis

Is ||r|| a good estimation of ||e||? Construct the relationship between r

and e From the definition r=Ax’-b=Ax’-Ax=A(x’-x)=Aeanswer

31

hint#1

hint#2

hint#4hint#3

Page 32: Numerical Analysis

32

Page 33: Numerical Analysis

33

Page 34: Numerical Analysis

34

Page 35: Numerical Analysis

35

Page 36: Numerical Analysis

Condition number

36

Page 37: Numerical Analysis

37

Page 38: Numerical Analysis

Perturbations (skipped)

38

.

.

.

Page 39: Numerical Analysis

Any Questions?

39

3.4 Error Estimates and Condition Number

Page 40: Numerical Analysis

3.5

40

LU Decomposition

Page 41: Numerical Analysis

LU decomposition

Motivation Gaussian elimination solve a linear system, Ax=b, with n unknowns– (2/3)n3 + (3/2)n2 – (7/6)n

– with back substitution

– the minimum number of operations

If there are a lots of right-hand-side vectors– how many operations for a new RHS?

– with Gaussian elimination, all operations are also carried out on the RHS

41

Page 42: Numerical Analysis

LU decomposition

Given a matrix A, a lower triangular matrix L and an upper triangular matrix U for which LU=A are said to form an LU decomposition of A

Here we replace mathematical descriptions with an example to show how use Gaussian elimination to obtain an LU decomposition

42

Page 43: Numerical Analysis

43

Page 44: Numerical Analysis

44

Page 45: Numerical Analysis

45

Page 46: Numerical Analysis

46

Page 47: Numerical Analysis

Any Questions?

47

Page 48: Numerical Analysis

Is there any other LU decompositions in addition to using modified Gaussian elimination?– degree of freedoms (number of

unknowns)

– An2, LUn2+n

Direct factorization (3.6)– as an systems of n2+n equations

48

hint

answer

Page 49: Numerical Analysis

Solving a linear system A LU When a new RHS comes

– Ax=b PAx=Pb LUx=Pb

– with z=Ux, actually to solve Lz=Pb and Ux=z • both steps are easy

• notice that Pb does not require real matrix-vector multiplication

49

Page 50: Numerical Analysis

50

Page 51: Numerical Analysis

Solving a linear system

In summary Anyway, the two-step algorithm (LU

decomposition) is superior to Gaussian elimination with back substitution

51

Page 52: Numerical Analysis

Any Questions?

52

3.5 LU Decomposition

Page 53: Numerical Analysis

3.6Direct Factorization

53

Page 54: Numerical Analysis

Is there any other LU decompositions in addition to using modified Gaussian elimination?– degree of freedoms (number of

unknowns)

– An2, LUn2+n

Direct factorization (3.6)– as an systems of n2+n equations

54http://www.dianadepasquale.com/ThinkingMonkey.jpg

Recall that

Page 55: Numerical Analysis

Direct factorization Just add more n equations

– ex: diagonal must be 1

Crout decomposition– lii=1 for each i=1, 2, …, n

Dollittle decomposition– uii=1 for each i=1, 2, …, n

55

Page 56: Numerical Analysis

56

Page 57: Numerical Analysis

57

Page 58: Numerical Analysis

58

Page 59: Numerical Analysis

59

Page 60: Numerical Analysis

Any Questions?

60

3.6 Direct Factorization