Exploring Fractals

Generating and exloring fractal structures with Python and Numba.

Exploring Fractals

Generating and exloring fractal structures with Python and Numba.

I recently rewatched a video from Arthur C. Clarke which is where I first saw what fractals are. I saw this video while I was in high school and having rewatched it now I wondered if I can generate the Mandelbrot set and explore some parts of it myself. I did a quick search online and I discovered that Jean Francois Puget had already done that and achieved great results. Learning from his work I decided to do some computational experiments.

For this post I will use the acceleration provided by numba as in the original post by Jean Francois Puget. In a further post he developed a parallel version of the code so you can check it out, but for my purposes now numba will be enough.


Fractals are abstract objects that are infinitely complicated and exhibit similar patterns at increasingly smaller scales. What this means is that we can keep zooming in the fractal until the original picture is bigger than the visible universe and still see patterns. For this post I want to explore the Mandelbrot set and produce the fractal from it. However, before delving in the Mandelbrot set we need to look at the more general picture.

The Mandelbrot set is made by iteration using the simple formula \(z_{n+1} = z_n^2 + c\) for different values of $c$ in the complex plane. Then we have two possible paths: the iteration will converge or diverge. Points on the complex plane $\mathbb{C}$ for which $z_n$ is bounded form the Mandelbrot set. Points which are not bounded do not belong to the Mandelbrot set and to get artistic pictures are usually coloured based on how fast they change. Each point $c$ then specifies the geometric structure of the corresponding Julia set.

Julia sets are formed by performing a repeated iteration of a holomorphic function and in this post we will investigate polynomial functions of the form $z_{n+1} = z_{n}^2 + c$. The set has an interesting property of chaotic behaviour which produces drastically different result from a small perturbation. The Julia set also has a complementary set known as the Fatou set which has the property that all values behave similarly under the repeated iteration. Now if a point $c$ is in the Mandelbrot set the Julia set is called connected, if it is not then it is known as Cantor/Fatou dust.

We can now visualise what we mean by the above by using some Python code and the Scipy library.

Mandelbrot Set

The Mandelbrot set is generated using the iterative formula below.

\[z_{n+1} = z_{n}^{2} + c\]

To get the set we start from $z_{0} = 0$ for a given $c$ coordinate. Expressing it mathematically, for a point to belong to the Mandelbrot set we need the limit to be $\leq 2$. The precise reason is given below.

\[c \in M \iff \lim_{n \to \infty} \sup|z_{n+1}| \leq 2\]

When I first saw the mathematical description I could not understand why was the number 2 written as the upper boundary to determine if a point belongs to the Mandelbrot set or not. I decided to dig further and I found the mathematical proof. I found the explanation in Alun Williams' website and you can read it there.

Lets define the function for the Mandelbrot set.

In [1]:
import numpy as np
from numba import jit

def mandelbrot(c, maxiter):
    z = c
    for n in range(maxiter):
        if abs(z) > 2:
            return n
        z = z*z + c
    return 0

The holomorphic function can then be called to generate the Mandelbrot set. I will take this opportunity to generalise the set generation and plot functions so we can use them with the Julia set as well.

In [2]:
def generate_set(set_function, xmin, xmax, ymin, ymax, width, height, maxiter):
    r1 = np.linspace(xmin, xmax, width)
    r2 = np.linspace(ymin, ymax, height)
    n3 = np.empty((width,height))
    for i in range(width):
        for j in range(height):
            n3[i, j] = set_function(r1[i] + 1j * r2[j], maxiter)
    return (r1, r2, n3)

Having generated the set we need a way to visualise it. We can do this with Matplotlib. As states above the algorithm to display the fractals is explained here.

In [3]:
from matplotlib import pyplot as plt
from matplotlib import colors
from matplotlib import patches
from IPython.display import set_matplotlib_formats
%matplotlib inline

def generate_image(set_function, xmin, xmax, ymin, ymax, width=10, height=10, maxiter=256):
    dpi = 72
    img_width = dpi * width
    img_height = dpi * height
    x,y,z = generate_set(set_function, xmin, xmax, ymin, ymax, img_width, img_height, maxiter)
    fig, ax = plt.subplots(figsize=(width, height), dpi=72)
    ticks = np.arange(0, img_width, 3*dpi)
    x_ticks = xmin + (xmax-xmin)*ticks/img_width
    plt.xticks(ticks, x_ticks)
    y_ticks = ymin + (ymax-ymin)*ticks/img_width
    plt.yticks(ticks, y_ticks)
    ax.imshow(z.T, origin='lower')

We can then visualise the Mandelbrot set.

In [4]:
generate_image(mandelbrot, -2.0, 0.5, -1.25, 1.25)


The "main" part of the Mandelbrot is called the main cardioid and the set contains small copies of itself connected to the main cardioid. This continues and we get structures which resemble hairs, not surprisingly these features are called Mandelbrot hair.

Other explorers have discovered interesting anchor points. For example lets set sail to $-0.761574-i0.0847596$ as shown on maps generated by Paul Bourke. We will centre at this point and zoom $\times 10$ to start with. Note that the original generated image is of size $2.5\times2.5$ and we need a function to generate the size of the image for the given zoom.

In [5]:
def zoom_image(xc, yc, size):
    x1 = xc - size
    x2 = xc + size
    y1 = yc - size
    y2 = yc + size
    return x1, x2, y1, y2

We can than choose our point of interest and zoom into it.

In [6]:
xc = -0.761574
yc = -0.0847596
size = 2.5 / 10

x1, x2, y1, y2 = zoom_image(xc, yc, size)
generate_image(mandelbrot, x1, x2, y1, y2)


Great, everything is looking fine. We can see the primary continental mu-atom and a mu-atom of period 2, if you want an explanation of the names go here. The place in the middle where all the interesting features occur is know as the seahorse valley and we are going to explore the southern part. Lets zoom $\times 500$ and explore the seahorses.

In [7]:
size = 2.5 / 500

x1, x2, y1, y2 = zoom_image(xc, yc, size)
generate_image(mandelbrot, x1, x2, y1, y2)


We can see why mathematicians are referring to these patters as seahorses. Also notice that more patterns are starting to emerge as we go deeper in the set. Why stop now, $\times 1000$ it is.

In [8]:
size = 2.5 / 1000

x1, x2, y1, y2 = zoom_image(xc, yc, size)
generate_image(mandelbrot, x1, x2, y1, y2)


I will test the ability of our code to cope with even more decimals places and zoom to $\times 15625$. We can then compare with the image obtained from Paul Bourke.

In [9]:
size = 2.5 / 15625

x1, x2, y1, y2 = zoom_image(xc, yc, size)
generate_image(mandelbrot, x1, x2, y1, y2)


Okay, lets go wild and do a $5\times10^6$ zoom.

In [10]:
size = 2.5 / (5 * (10**6))

x1, x2, y1, y2 = zoom_image(xc, yc, size)
generate_image(mandelbrot, x1, x2, y1, y2)


Whoops, it seems that we have reached the limit of the colouring algorithm. This is an issue for another post, where I am planning to program the problem in C++ and use different generation and colouring algorithms.

Julia Set

To generate a Julia set we need a holomorphic function. In other words we need a continuously differentiable complex-valued function with one or more complex variable. Lets start with something random... the first thing that comes to mind.

\[f(z) = z^3 + c\]

The iteration is then performed for:

\[z_{n+1} = z_{n}^{3} + c\]

We can then get the points that tend to infinity, those that do not (Fatou set) and the boundary between them (Julia set) for a complex number $k$ if we set $c=k$ and start with $z_{0}$ as the coordinates of a point. Lets explore $c = -0.1 + i0.65$ which is in the Mandelbrot set and we would get a connected Julia set. The way I have implemented this is to define a holomorphic function with a given complex parameter $c$, the seemingly weird use of $c$ as an argument to define $z$ and then be redefined is just for convenience so I can use the same set generation function as for the Mandelbrot set.

In [11]:
def holomorphic_julia(c, maxiter):
    z = c
    c = complex(-0.1, 0.65)
    for n in range(maxiter):
        if abs(z) > 10:
            return n
        z = z*z + c
    return 0

We can then show our results and we get a weird looking bulge.

In [12]:
generate_image(holomorphic_julia, -1.5, 1.5, -1.5, 1.5)


Lets examine the fractal nature of this thing we have produced. We would focus in the middle of our bulge.

In [13]:
xc = 0.0
yc = 0.0
size = 3.0 / 10

x1, x2, y1, y2 = zoom_image(xc, yc, size)
generate_image(holomorphic_julia, x1, x2, y1, y2)


We can see that we have generated a fractal since zooming in reveals similar patterns of infinity complexity. We can also see the Julia set as the boundary between the Fatou sets.

Now we can pick a point which is outside of the Mandelbrot set. Such a point is $c = -2.5 - i2.5$.

In [14]:
def holomorphic_candor_dust(c, maxiter):
    z = c
    c = complex(-2.5, -2.5)
    for n in range(maxiter):
        if abs(z) > 10:
            return n
        z = z*z + c
    return 0

generate_image(holomorphic_candor_dust, -4.0, 4.0, -4.0, 4.0)


This is known as Cantor/Fatou dust and the calculation oscillates between the points on the plot which is very different from the previous Julia set. As can be seen the resultant behaviour is very much dependent on whether $c$ is in the Mandelbrot set or not. We can also investigate different complex parameters. For example lets see a $c = -0.78 + 0.1i$ which is in the Mandelbrot set and the function $z_{n+1} = z_{n}^{2} + c$ as shown here.

In [15]:
def holomorphic_julia_spinning_eyes(c, maxiter):
    z = c
    c = complex(-0.78, 0.1)
    for n in range(maxiter):
        if abs(z) > 10:
            return n
        z = z*z + c
    return 0

generate_image(holomorphic_julia_spinning_eyes,  -1.5, 1.5, -1.5, 1.5)


Thank you for reading!!! I hope this was a fun post.

Nikola Stoyanov
PhD Researcher

My research interests include distributed robotics, mobile computing and programmable matter.


comments powered by Disqus