An equation $f(x) = 0$ can always be written as $g(x) = x$. Fixed-point iteration can be applied to approximate the fixed number $r = g(r)$.

After selecting an initial guess $x_0$, the fixed-point iteration is $x_{n+1} = g(x_n)$. That is,

\begin{equation*} x_0 = \textrm{initial guess} \\ x_1 = g(x_0) \\ x_2 = g(x_1) \\ x_3 = g(x_2) \\ \vdots \end{equation*}The algorithm is repeated until either a testing condition $e_{i+1} = \left|\:x_{i+1} - x_{i}\:\right| < \alpha$, where $\alpha$ is some tolerance limit is met, or until a fixed number of iterations $N$ is reached. Note that the method may or may not converge. Note that the choice of $g(x)$ is in general not unique.

As a simple introductory example we use the fixed-point iteration to solve the equation

$$\frac{1}{2}\sin(x) - x + 1 = 0.$$The most natural is to use $g(x)= \sin(x)/2 + 1$ and then use the fixed-point iteration to solve $g(x)=x$.

In [5]:

```
import math
x = 1 # initial guess
N = 10 # iterations
for i in range(1, N):
x = math.sin(x) + 1
print("x%i:\t%.5f"%(i, x))
```

For the iteration scheme to return a fixed point $r$, it needs to converge. A criterion for convergence is that the error, $e_i = \left|\:r - x_i \:\right|$, decreases for each iteration step. This means that the change from $x_i$ to $x_{i+1}$ also has to decrease for each step. The convergence criterion is explained in the following theorem [1]:

Theorem 1Assume that $g$ is continuously differentiable, that $g(r) = r$, and that $S = \left|\:g'(r)\:\right| < 1$. Then fixed-point iteration converges linearly with rate $S$ to the fixed point $r$ for initial guesses sufficiently close to $r$.

Note that in the example above we have $|g'(x)|=\cos(x)/2 < 1$ which means that the method will converge for all initial guesses.

The convergence can in fact be of higher order. This is explained in the following theorem [2]:

Theorem 2Assume that $g$ is $p$ times continuously differentiable and that the fixed-iteration converges to $r$ for some initial guess $x_0$. If $g'(r)=g''(r)=\cdots=g^{(p-1)}(r)=0$ and $g^{(p-1)}(r)=0$, then the order of convergence is $p$.

The convergence is said to be of order $p$ when $\lim_{i\to\infty}e_{i+1}/e_{i}^p = \text{constant}$.

**Excercise:** Prove theorem 1 using the mean value theorem and the theorem 2 using Taylor's theorem.

As mentioned in the introduction, the choice of $g(x)$ is far from unique. We will now use fixed-point iteration to estimate the square root of a real and positive number $a$. That is, we want to solve $f(x) = x^2-a^2=0$. Two natural choices for $g(x)$ are $g_1(x)=a/x$ and $g_2(x)=x$. However, none of these converges since $|g_1'(\sqrt a)| = |g_2'(\sqrt a)| =1$. We can choose the mean as $g(x)$:

$$g(x) = \frac{1}{2}\left(\frac{a}{x} + x\right).$$In this case we have

$$\left|\:g'(x)\:\right|_{x=\sqrt a}=\left|\:\frac{1}{2}(1-\frac{a}{x^2})\:\right|_{x=\sqrt a} = 0 < 1$$The theorem above thus implies that method converges for some initial guess $x_0$. The method is in fact globally convergent due to the convexity of $f(x)$.

In [1]:

```
a = 0.07 # square of root
x = 1 # initial guess
N = 10 # iterations
for i in range(1, N):
x = (a/x + x)/2
print("x%i:\t%.5f"%(i, x))
```

This method can also be derived from Newton's method (see the section on Newton's method below). The method was however first used by the Babylonian people long before Newton [2].

Underneath follows an implementation of the Fixed-Point Iteration with a testing condition $\left|\:x_{i+1} - x_i\:\right| < \alpha$. There are other stopping criteria that may be relevant such as the backward error $ \left|\:f(x_a)- f(0)\right|$. We need an additional stopping criteria in case convergence fails. We therefore stop the computation if the a given number of iterations $N$ is reached.

In [14]:

```
def FPI(g, x, maxit=100, alpha=1e-5):
""" A function using fixed-point iteration to compute
an approximate solution to the fixed point problem g(r) = r.
Arguments:
g callable function, g(x)
x float. The initial guess
alpha float. Error tolerance
maxit int. Maximum number of iterations
Returns:
array of each fixed-point approximation
"""
result = [x]
x_next = g(x)
i = 0 # Counter
e = abs(x_next - x)
while e > alpha :
if i > maxit:
print("No convergence.")
break
x = x_next
x_next = g(x_next)
result.append(x)
if x - x_next > e:
print("Divergence.")
break
e = abs(x_next - x)
i += 1
return result
```

We want to solve the equation,

\begin{equation*} f(x) = x^4 + x - 1 = 0 %\label{eq:example} \end{equation*}using fixed point iteration. First, we rewrite the equation to the form, $x = g(x)$. This particular equation may be rewritten in three ways:

$$g_1(x) = x = 1 - x^4,$$$$g_2(x) = x = \sqrt[4]{1 - x},$$$$g_3(x) = x = \frac{1 + x^4}{1 + 2x^3}.$$Each of these has its own convergence rate. One of the solutions is clearly somewhere between 0 and 1, so we will use $x_0=0.5$ as initial guess.

In [13]:

```
x_0 = 0.5
N = 100
alpha = 1e-5
def g1(x): return 1 - x**4
def g2(x): return (1 - x)**(1/4)
def g3(x): return (1 + 3*x**4)/(1 + 4*x**3)
def print_result(result):
for i in range(len(result)):
print("x%i:\t%.5f"%(i, result[i]))
```

In [10]:

```
result_g1 = FPI(g1, x_0, N, alpha)
print_result(result_g1)
```

In [11]:

```
result_g2 = FPI(g2, x_0, N, alpha)
print_result(result_g2)
```

In [12]:

```
result_g3 = FPI(g3, x_0, N, alpha)
print_result(result_g3)
```

The method does not converge for $g_1$, but both $g_2$ and $g_3$ converges towards the fixed number $0.72449$, which is a solution of the equation $f(x) = 0$. The method converges much faster for $g_3$ than for $g_2$.

This result can be verified by calculating the convergence rates $S = \left|\: g'(r) \:\right|$:

\begin{equation*} \left|\:g_1'(0.72449)\:\right| \approx 1.521 > 1, \\ \left|\:g_2'(0.72449)\:\right| \approx 0.657 < 1, \\ \left|\:g_3'(0.72449)\:\right| \approx 0.000 < 1. \\ \end{equation*}Theorem 1 tells us that the method will converge for $g_2$ and $g_3$ for some inital guess. Moreover, from theorem 2 we know that the method converges linearly for $g_2$ and the order of convergence for $g_3$ is two.

**Exercise:** Verify that the convergence for $g_2$ is linear and that the convergence of $g_3$ is of second order by using the numerical results above.

The Newton-Rhapson method is a special case of fixed-point iteration where the convergence rate is zero - the fastest possible. The method estimates the root of a differentiable function $f(x)$ by iteratively calculating the expression

\begin{equation*} x_{n+1} = x_n -\frac{f(x_n)}{f'(x_n)}. \end{equation*}By using theorem 2, we can show that Newton's method is of second order (if $f'(r)\neq 0$). If $f''(r)=0$, then the method has a third order convergence.

This Newton iteration may be rewritten as a fixed point iteration

\begin{equation*} x_{n+1} = g(x_n), \: n = 1, 2 ,3, ... \end{equation*}where

\begin{equation*} r = g(r) = r - \frac{f(r)}{f'(r)} %\label{eq:newton} \end{equation*}Provided that the iteration converges to a fixed point $r$ of $g$. From this we obtain

$$ \frac{f(r)}{f'(r)} = 0 \: \Rightarrow \: f(r) = 0$$and thus the fixed point $r$ is a root of $f$.

The Newton's method is further discussed in the module Root Finding - Newton-Rhapson Method.

We have now used the fixed-point iteration method to solve the equation $f(x) = 0$ by rewriting it as a fixed-point problem. This is a simple method and which is easy to implement, but unlike the Bisection Method it only converges if the initial guess is sufficiently close to the root $r$. It is in general only locally convergent. However, the convergence rate may, or may not, be faster than that of the Bisection Method, which is 1/2.

[1] Sauer, T.: Numerical Analysis international edition, second edition, Pearson 2014

[2] Gautschi, W.: Numerical Analysis, second edition, Birkhäuser 2012 (https://doi.org/10.1007/978-0-8176-8259-0)