MNchap1

Embed Size (px)

Citation preview

  • 7/29/2019 MNchap1

    1/46

    Solving Nonlinear Equation

    0x1x2x

    )( 1xf

    )( 0xf

    Root r

  • 7/29/2019 MNchap1

    2/46

    Numerical Methods Wen-Chieh Lin 2

    Nonlinear Equations

    Given function f, we find value x for which

    Solution x is root of equation, or zero of

    function f So problem is known as root finding or zero

    finding

    0)( xf

  • 7/29/2019 MNchap1

    3/46

    Numerical Methods Wen-Chieh Lin 3

    Nonlinear Equations: two cases

    Single nonlinear equation in one unknown, where

    Solution is scalar x for which f(x) = 0

    System ofn coupled nonlinear equations in nunknowns, where

    Solution is vector x for which all components off

    are zero simultaneously, f(x) = 0

  • 7/29/2019 MNchap1

    4/46

    Numerical Methods Wen-Chieh Lin 4

    Example of 1-D nonlinear equation

    for which x = 0.3604 is one approximate solution

    Example of system of nonlinear equations in twodimensions

    for which x = [-1.8, 0.8] is one approximatesolution vector

    1

    4

    2

    2

    2

    2

    1

    1

    xe

    xx

    x

    Examples: Nonlinear Equations

    0)sin(3 xexx

  • 7/29/2019 MNchap1

    5/46

    Numerical Methods Wen-Chieh Lin 5

    Multiplicity

    Iff(R) = f(R) = f(R) = = f(m-1)(R) = 0 butf(m)(R) 0, then root R has multiplicity m

    Ifm = 1 (f(R) = 0, f(R) 0 ), then R is simpleroot

  • 7/29/2019 MNchap1

    6/46

    Numerical Methods Wen-Chieh Lin 6

    Bisection method begins with initial bracket andrepeatedly halves its length until solution hasbeen isolated as accurately as desired

    http://www.cse.uiuc.edu/iem/nonlinear_eqns/Bisection/

    Interval Halving (Bisection)

    while |b-a| > tol,

    m = a + (b-1)/2;

    If f(a)*f(m) < 0,

    b = m;

    else

    a = m;

    end;end;

    a b

    m

  • 7/29/2019 MNchap1

    7/46

    Numerical Methods Wen-Chieh Lin 7

    Bisection (cont.)

    Simple and guaranteed to work if fis continuous in [a, b]

    [a, b] brackets a root

    Needed iterations to achieve a specified accuracy is

    known in advance

    Error after n iterations < |b - a| /2n

    Slow to converge

    Good for initial guess for other root finding

    algorithms Finding the initial bracket may be a problem iffis not

    given explicitly

  • 7/29/2019 MNchap1

    8/46

    Numerical Methods Wen-Chieh Lin 8

    Use graphing to assist root finding

    Set the initial bracket

    Detect multiple roots

  • 7/29/2019 MNchap1

    9/46

    Numerical Methods Wen-Chieh Lin 9

    Can we find a root in a better way?

    Bisection only utilizes function values f(x)

    We can find a root with fewer iterations if

    other information is used

    linear approximation off(x) Secant line secant method

    Tangent line Newtons method

    Polynomial approximation off(x)

    Mullers method

  • 7/29/2019 MNchap1

    10/46

    Numerical Methods Wen-Chieh Lin 10

    Secant Method

    Approximate a function by a straight line

    Compute the intersection of the line and x-axis

    )()(

    )(

    )(

    )(

    10

    10

    1

    21

    xfxf

    xx

    xf

    xx

    )()(

    )()(

    10

    10112

    xfxf

    xxxfxx

    0x1x2x

    )( 1xf

    )( 0xf

    Root r

  • 7/29/2019 MNchap1

    11/46

    Numerical Methods Wen-Chieh Lin 11

    Secant Method (cont.)

    Update endpoints

    Repeat

    )()()()(

    1

    11

    nn

    nnnnn

    xfxfxxxfxx

    0

    x

    1x

    )( 1xf

    )( 0xfRoot r

    withswap,)()(if 1010 xxxfxf

  • 7/29/2019 MNchap1

    12/46

    Numerical Methods Wen-Chieh Lin 12

    Example

    f(x) = 3x + sin(x)

    exp(x) Find the root in [0, 1]

    http://www.cse.uiuc.edu/iem/nonlinear_eqns/Secant/

    0 0.5 1 1.5 2-1

    -0.5

    0

    0.5

    1

    1.5

  • 7/29/2019 MNchap1

    13/46

    Numerical Methods Wen-Chieh Lin 13

    Method of False Position

    Problem of secant method

    Remedy

    Always bracket a root in the interval [x0, x1]

    How to do this?

  • 7/29/2019 MNchap1

    14/46

    Numerical Methods Wen-Chieh Lin 14

    Newtons method

    Better approximation using the first derivative

    )( 0xf

    10 xx 0x1x

    10

    00

    )()(')tan(

    xx

    xfxf

    )('

    )(

    0

    001

    xf

    xfxx

    )('

    )(1

    n

    nnn

    xf

    xfxx

  • 7/29/2019 MNchap1

    15/46

    Numerical Methods Wen-Chieh Lin 15

    Interpretation of Newtons method

    Truncated Taylor series

    is a linear function ofh approximating fnear x

    Replace nonlinear function f by this function,whose zero is h = - f(x)/f(x)

    Zeros of original function and linear

    approximation are not identical, so repeat

    process

    )(')()( xhfxfhxf

    )('

    )(1

    n

    nnn

    xf

    xfxx

  • 7/29/2019 MNchap1

    16/46

    Numerical Methods Wen-Chieh Lin 16

    Example: Newtons method

    http://www.cse.uiuc.edu/iem/nonlinear_eqns/Newton/

  • 7/29/2019 MNchap1

    17/46

    Numerical Methods Wen-Chieh Lin 17

    Comparison of Secant and

    Newtons methods

    )()(

    )()(

    1

    11

    nn

    nnnnn

    xfxf

    xxxfxx

    )('

    )(1

    n

    nnn

    xf

    xfxx

    Secant method Newtons method

  • 7/29/2019 MNchap1

    18/46

    Numerical Methods Wen-Chieh Lin 18

    Pros and Cons of Newtons method

    Pros

    efficient

    Cons

    Need to know the derivative function

  • 7/29/2019 MNchap1

    19/46

    Numerical Methods Wen-Chieh Lin 19

    )('

    )(

    1n

    n

    nn xf

    xf

    xx

    When will Newtons method not converge?

    x1=x6, loop

    passing

    maximum/minimum

  • 7/29/2019 MNchap1

    20/46

    Numerical Methods Wen-Chieh Lin 20

    Mullers method Instead of linear approximation, Mullers

    method uses quadratic approximate

    Evaluation of derivatives are not required

    See the textbook for details

  • 7/29/2019 MNchap1

    21/46

    Numerical Methods Wen-Chieh Lin 21

    Fixed-point Iteration Method

    Rearrange f(x)=0 into an equivalent form x=g(x)f(x) = x g(x) = 0

    Ifris a root off, then f(r) = r-g(r) = 0r=g(r)

    In iterative form,

    Also called function iteration

    For given equation f(x)=0, there may be manyequivalent fixed-point problems x=g(x) with different

    choice ofg Will the method always converge?

    )(1 nn xgx

  • 7/29/2019 MNchap1

    22/46

    Numerical Methods Wen-Chieh Lin 22

    Example: Fixed-point Iteration

    032)( 2 xxxf

    32)(1 xxg

    2

    3)(2

    xxg

  • 7/29/2019 MNchap1

    23/46

    Numerical Methods Wen-Chieh Lin 23

    Example: Fixed-point Iteration (cont.)

    032)( 2 xxxf

    2

    3)(

    2

    3

    xxg

    Diver

    ge!

  • 7/29/2019 MNchap1

    24/46

    Numerical Methods Wen-Chieh Lin 24

    Convergence Rate

    For general iterative methods, define error atiteration n by

    en = xn R

    where xn is approximate solution and R is true

    solution

    For methods that maintain interval known to

    contain solution, rather than specific

    approximate value for solution, take error to be

    length of interval containing solution

  • 7/29/2019 MNchap1

    25/46

    Numerical Methods Wen-Chieh Lin 25

    Convergence Rate (cont.)

    Sequence converges with rate rif

    for some finite nonzero constant C Some cases of interest

    r=1: linear (C1: superlinear

    r=2: quadratic

    Ce

    er

    n

    n

    n

    1lim

  • 7/29/2019 MNchap1

    26/46

    Numerical Methods Wen-Chieh Lin 26

    Convergence Rate of Bisection

    Length of interval containing solution reduced

    by half at each iteration

    Linearly convergent

    r = 1 C = 0.5

  • 7/29/2019 MNchap1

    27/46

    Numerical Methods Wen-Chieh Lin 27

    Convergence of Fixed-point Iteration

    IfR = g(R) and |g(R)| < 1, then there is aninterval containing R such that iteration

    xn+1 = g(xn)

    converges to R if started within that interval

    If |g(R)| > 1, then iterative scheme diverges

  • 7/29/2019 MNchap1

    28/46

    Numerical Methods Wen-Chieh Lin 28

    Proof of Convergent Condition

    nnn ege )('1

    )()()(1 nnn xgRgxgRxR

    n

    n

    nn xR

    xR

    xgRgxR

    )()(1

    ))(('1 nnn xRgxR

    Mean Value Theorem

    ),( nn xR

    Ce

    er

    n

    n

    n

    1lim

    Fixed-point iteration is linearly convergent

    )(' RgC

  • 7/29/2019 MNchap1

    29/46

    Numerical Methods Wen-Chieh Lin 29

    Convergence of Newtons Method

    Represent Newtons method in fixed-pointiteration form

    Condition for convergence

    )()('

    )(1 n

    n

    nnn xg

    xf

    xfxx

    1)]('[

    )(")()('

    2

    xf

    xfxfxg

  • 7/29/2019 MNchap1

    30/46

    Numerical Methods Wen-Chieh Lin 30

    Convergence rate of Newtons method

    )()(1 nn xgRgxR 2)()(")()(')()( 2nnn xRgxRRgRgxg

    0)('

    )]('[

    )(")()('

    0)(

    2 Rg

    xf

    xfxfxg

    Rf

    2)()(")()( 2nn xRgRgxg

    2)(")()(2

    11 nnnn egxgRgxRe

    Ce

    er

    n

    n

    n

    1lim

    Newtons method isquadratically convergent!

    Recall that

  • 7/29/2019 MNchap1

    31/46

    Numerical Methods Wen-Chieh Lin 31

    Question from the last class

    Is the initial solution important for theconvergence of fixed-point iteration?

  • 7/29/2019 MNchap1

    32/46

    Numerical Methods Wen-Chieh Lin 32

    Answer to the convergence problem of

    fixed-point Iteration

    |g(R)|1, the fixed-point iteration will diverge

    even if the initial condition is very close to a rootsince the iteration will eventually reach the region

    causing divergence

    Initial solution is important but less critical

    The algorithm may not converge if the initial

    solution is far from the true solution Recall the conditions that Newtons method does

    not converge?

  • 7/29/2019 MNchap1

    33/46

    Numerical Methods Wen-Chieh Lin 33

    Example: Newtons Method for Finding

    Complex Roots

    f(x) = x3 + 2x2 x + 5

  • 7/29/2019 MNchap1

    34/46

    Numerical Methods Wen-Chieh Lin 34

    Start Newtons method with a complex value

    ix 10 ixf 52)( 0 ixf 103)(' 0

    i

    i

    ii

    xf

    xfxx 04587.1486238.0

    103

    521

    )('

    )(

    0

    001

    52)( 23 xxxxf 143)(' 2 xxxf

    ixf

    xfxx 23665.1448139.0

    )('

    )(

    1

    112

  • 7/29/2019 MNchap1

    35/46

    Numerical Methods Wen-Chieh Lin 35

    Newtons Methods for Multiple Roots

    Quadratically convergent for simple root,

    Linearly convergent for multiple roots as

    0)(')]('[

    )(")()(' 0)(

    2 Rg

    xf

    xfxfxg

    Rf

    0

    0

    )]('[

    )(")()('

    2

    Rf

    RfRfRg

  • 7/29/2019 MNchap1

    36/46

    Numerical Methods Wen-Chieh Lin 36

    Remedies for Multiple Roots with

    Newtons method

    Iff(x) has a root of multiplicity kat x=R, wecan factor out (x-R)k from f(x) to get

    With a slightly modified Newtons method

    It can be proved that and

    Newtons method still converges quadratically

    )()()( xQRxxfk

    0)(' Rgk

    )()('

    )(1 nk

    n

    nnn xg

    xf

    xfkxx

  • 7/29/2019 MNchap1

    37/46

    Numerical Methods Wen-Chieh Lin 37

    Remedies for Multiple Roots with

    Newtons method (cont.)

    In practice, we dont know kin advance

    Remedies

    Try and error

    Deflate f(x)f(x)/(x-s) where s is an approximate R

    Be warned that an indeterminate form at x=R is

    created

  • 7/29/2019 MNchap1

    38/46

    Numerical Methods Wen-Chieh Lin 38

    Systems of Nonlinear Equations

    Solving systems of nonlinear equations ismuch more difficult than scalar case because

    Wide variety of behavior is possible, so

    determining existence and number of solutions or

    good starting guess is much more complex In general, there is no simple way to guarantee

    convergence to desired solution or to bracket

    solution to produce absolutely safe method

    Computational overhead increases rapidly withdimension of problem

  • 7/29/2019 MNchap1

    39/46

    Numerical Methods Wen-Chieh Lin 39

    Example: Systems in Two Dimensions

    From Michael T. Heath

  • 7/29/2019 MNchap1

    40/46

    Numerical Methods Wen-Chieh Lin 40

    Newtons Method

    In n dimensions, Newtons method has form

    where J(x) is Jacobian matrix off

    In practice, we do not explicitly invert J(xn),but instead solve linear system

    for Newton step sn, then take as next iterate

    )()( 1 nnn1n xfxJxx

    )()( nnn xfsxJ

    nn1n sxx

  • 7/29/2019 MNchap1

    41/46

    Numerical Methods Wen-Chieh Lin 41

    Example: Newtons Method

    01

    4

    ),(

    ),(

    )(2

    2

    2

    2

    1

    212

    211

    1

    xe

    xx

    xxf

    xxfxxf

    1

    22)(

    1

    21

    2

    2

    1

    2

    2

    1

    1

    1

    xe

    xx

    xf

    xf

    x

    f

    x

    f

    xJ

    T7.110 x

    0183.0

    11.0)( 0xf

    17183.2

    4.32

    )( 0xJ

    )()( 000 xfsxJ

    0183.0

    11.0

    17183.2

    4.320s

    T]7298.10043.1[00 sxx1

  • 7/29/2019 MNchap1

    42/46

    Numerical Methods Wen-Chieh Lin 42

    Example: Newtons Method

    01

    4

    ),(

    ),(

    )(2

    2

    2

    2

    1

    212

    211

    1

    xe

    xx

    xxf

    xxfxxf

    1

    22)(

    1

    21

    2

    2

    1

    2

    2

    1

    1

    1

    xe

    xx

    xf

    xf

    x

    f

    x

    f

    xJ

    T7298.10043.11 x

    49651.1

    42653.8)( 1

    e

    exf

    17300.2

    4596.30086.2

    )( 1xJ

    )()( 111 xfsxJ

    49651.1

    42653.8

    17300.2

    4596.30086.21

    e

    es

    T]729637.1004169.1[112 sxx

  • 7/29/2019 MNchap1

    43/46

    Numerical Methods Wen-Chieh Lin 43

    Fixed-Point Iteration

    Fixed-point problem for is to find

    vector x such that

    Corresponding fixed-point iteration is

    Converges if

    starts close enough to solution

    )(xgx

    )(1 nn xgx

    nnRR :g

    1))(( RJ

    A

    A

    R

    J

    matrixaofseigenvaluethe

    ofnormcomplexmaximum:)(

    solutiontrue:

    matrixJacobian:

  • 7/29/2019 MNchap1

    44/46

    Numerical Methods Wen-Chieh Lin 44

    Fixed-Point Iteration (cont.)

    Convergence rate is normally linear, with

    constant C given by

    If then convergence rate is at least

    quadratic, e.g., Newtons method

    nnRR :g

    0)( RJ

    ))(( RJ

  • 7/29/2019 MNchap1

    45/46

    Numerical Methods Wen-Chieh Lin 45

    Example: Fixed-point Iteration

    014

    ),(

    ),()(

    2

    2

    2

    2

    1

    212

    2111

    xe

    xxxxf

    xxfxxf

    2

    1

    2

    2

    1

    4

    )1ln(

    x

    x

    x

    x T

    7.110 x

    1.00000000000000 -1.73205080756888

    1 .00505253874238 -1.72912388057290

    1 .00398063482304 -1.72974647995027

    1 .00420874039761 -1.72961406264780

    1 .00416023020985 -1.72964222660783

  • 7/29/2019 MNchap1

    46/46

    Numerical Methods Wen-Chieh Lin 47

    Next Monday

    Introduction to Matlab/Octave

    Matlab offers a student version with discounted

    price

    Octave is a shareware that has similar environmentas Matlab!

    You are encouraged to bring your laptop with

    Octave or Matlab installed next Monday