Scribd
Upload
35 views

MAT 461/561: 10.3 Fixed-Point Iteration: Announcements

This document discusses fixed-point iteration, a method for solving nonlinear equations of the form x = g(x). It provides examples of functions that do and do not converge under fixed-point iteration. For convergence, g(x) must be a contraction mapping with derivative |g'(x)| < 1. The examples illustrate how rewriting the equation can sometimes improve convergence if |g'(x*)| is close to 1. Taylor expansion is used to analyze the error at each iteration, which is typically linear but can be quadratic if g'(x*) = 0.

Uploaded by

Debisa
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
35 views

MAT 461/561: 10.3 Fixed-Point Iteration: Announcements

This document discusses fixed-point iteration, a method for solving nonlinear equations of the form x = g(x). It provides examples of functions that do and do not converge under fixed-point iteration. For convergence, g(x) must be a contraction mapping with derivative |g'(x)| < 1. The examples illustrate how rewriting the equation can sometimes improve convergence if |g'(x*)| is close to 1. Taylor expansion is used to analyze the error at each iteration, which is typically linear but can be quadratic if g'(x*) = 0.

Uploaded by

Debisa
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

MAT 461/561: 10.

3 Fixed-Point Iteration

James V. Lambers

March 4, 2020

Announcements
Get some homework done dangit!

Homework Questions
3.2.4: In showing L−1 is lower triangular, can show that the solution of the system

Lxj = ej ,

where X = L−1 and ej is the jth column of I, has elements x1j , x2j ,. . .,xj−1,j = 0. This would
mean that xij = 0 for i < j, meaning X is lower triangular
No need to repeat the whole proof for upper triangular after doing lower triangular: use the fact
that if A is nonsingular, then (A−1 )T = (AT )−1 = A−T
3.4.3: use loglog to plot vector of n-values against vector of condition numbers, use cond(A) to
compute 2-norm condition number of each A. Submit plot and code used to generate it
For example if using n = 10p : use a loop for p to create each matrix and get its condition number,
store in vectors: let conds be the vector of condition numbers, do conds(i)=cond(A)

Fixed-Point Iteration
To solve a nonlinear equation f (x) = 0, can rewrite (in many ways) in the form

x = g(x)

Example: g(x) = x + λf (x) where λ 6= 0. This leads to the algorithm fixed-point iteration.

Solution by Successive Substitution


The algorithm:

Choose initial guess x(0)


for k = 0, 1, 2, . . .
x(k+1) = g(x(k) )
Use x(k) and x(k+1) to check for convergence
end

1
Examples:
• f (x) = x − cos x leading to x = cos x. Converged, but slooooooowly
• f (x) = x2 − x + 3/16, or x = x2 + 3/16. Solutions: x∗ = 1/4, 3/4. Converges to 1/4 unless
x > 3/4, diverges!
• f (x) = x+ln x, or x = − ln x. By IVT, solution x∗ ∈ (0.5, 0.6). However, fixed-point iteration
with initial guess 0.55 diverges. Instead: from −x = ln x, exponentiate: e−x = x. This time
it converges.
• f (x) = x − cos x − 2, or x = cos x + 2. Initial guess x(0) = 2. Iterates bounce back and forth,
convergence very slow. Instead try:
x + g(x)
x=
2
This iteration converges rapidly. Why?

Convergence Analysis
Existence of a fixed point (solution) x∗ : Brouwer’s Fixed-Point Theorem states that g(x) has a
fixed point in [a, b] if g maps [a, b] into [a, b]. Proof: use IVT.
Uniqueness: if g has a fixed point in [a, b] and satisfies a Lipschitz condition on [a, b]:
|g(x) − g(y)| ≤ L|x − y|, x, y ∈ [a, b]
where L < 1 is a Lipschitz constant, then g is a contraction and the fixed point x∗ is unique,
and Fixed-Point iteration will converge for any x(0) ∈ [a, b]. (Contraction Mapping Theorem)
Fixed-Point Theorem: By the Mean Value Theorem,

g(x) − g(y) 0
x − y = |g (c)|, c between x, y

If |g 0 (x)| ≤ L < 1 on [a, b], then g is a contraction and L is a Lipschitz constant.


What about the examples:
• x = cos x: g(x) = cos x, g 0 (x) = − sin x, so is a contraction on an interval [0, 1]. But g 0 (x∗ ) is
not very small, so convergence is slow
• x = x2 + 3/16, g 0 (x) = 2x. g 0 (1/4) = 1/2, g 0 (3/4) = 3/2. Because g 0 (3/4) > 1, that fixed
point is unstable so Fixed-Point iteration will not converge to it.
• x = − ln x: g 0 (x) = −1/x, so |g 0 (x)| > 1 on [0.5, 0.6]. But for g(x) = e−x , g 0 (x) = −e−x . Here
we used this rule:
d −1 1
[f (x)] = 0 −1 .
dx f (f (x))
Note that x = g(x) is equivalent to x = g −1 (x).
• x = cos x + 2: g 0 (x) = − sin x, g 0 (x∗ ) ≈ 1 thus convergence is slow. For
x + g(x)
h(x) = ,
2
h0 (x) = (1 − sin x)/2, so h0 (x∗ ) ≈ 0.

2
Taylor expansion of g to analyze error:
1
x(k+1) − x∗ = g(x(k) ) − g(x∗ ) = g 0 (x∗ )(x(k) − x∗ ) + g 00 (x∗ )(x(k) − x∗ )2 + · · ·
2
Normally, convergence is linear, with asymptotic error constant |g 0 (x∗ )|. But if g 0 (x∗ ) = 0, conver-
gence is quadratic with asymptotic error constant |g 00 (x∗ )/2|.

Relaxation

You might also like