A fractal is a shape that looks the same at every scale. Zoom in on the coastline of Britain and it stays jagged - bays contain inlets, inlets contain coves, coves contain rocky notches, each level repeating the roughness of the last. Zoom in on a smooth circle and it becomes a straight line. The difference is self-similarity, and it is the defining feature of fractal geometry.

The word itself was coined by Benoît Mandelbrot in 1975, from the Latin fractus (broken, fractured). But the objects had been lurking in mathematics for a century before anyone named them, dismissed as “pathological” curiosities - functions with no derivatives, sets with no sensible dimension. What Mandelbrot recognised was that these were not exceptions to geometric intuition but a more accurate model of the actual world.

Fractal Dimension

The first surprise: fractals do not have integer dimension.

For ordinary geometric objects, dimension behaves simply. A line segment of length $r$ scaled up by a factor of 2 produces two copies - scaling by $s$ produces $s^1$ copies. A square scaled by $s$ gives $s^2$ copies. A cube gives $s^3$. The exponent is the dimension.

For fractals, this exponent is not an integer. Define the Hausdorff dimension (or similarity dimension for self-similar fractals) by:

$$d = \frac{\log N}{\log s}$$

where $N$ is the number of self-similar pieces at scaling factor $s$.

Cantor set. Remove the middle third of $[0,1]$. Remove the middle third of each remaining piece. Repeat forever. What survives has measure zero - but it is uncountable. Scaling by $s = 3$ gives $N = 2$ copies, so $d = \log 2 / \log 3 \approx 0.631$. It is strictly between a point (dimension 0) and a line (dimension 1).

Sierpiński triangle. Divide an equilateral triangle into four smaller equilateral triangles and remove the central one. Repeat on each remaining triangle. Scaling by $s = 2$ gives $N = 3$ copies, so $d = \log 3 / \log 2 \approx 1.585$. More than a curve, less than a plane.

Koch snowflake. Replace the middle third of each line segment with two sides of an equilateral triangle, then iterate. Scaling by $s = 3$ yields $N = 4$ pieces, so $d = \log 4 / \log 3 \approx 1.262$. Its perimeter is infinite (the total length multiplies by $4/3$ at each step), yet it encloses a finite area.

The fractal dimension measures, in a precise sense, how completely a shape fills space. A curve with $d \approx 1.9$ fills the plane almost as thoroughly as a 2-dimensional region.

The Mandelbrot Set

The most famous fractal is defined by an iteration in the complex plane. For each complex number $c$, start at $z_0 = 0$ and iterate:

$$z_{n+1} = z_n^2 + c$$

The Mandelbrot set $\mathcal{M}$ is the set of values $c$ for which this sequence does not escape to infinity:

$$\mathcal{M} = { c \in \mathbb{C} : |z_n| \not\to \infty }.$$

The boundary of $\mathcal{M}$ is infinitely complex - zoom in at any scale and new structures appear, spirals and bulbs and seahorse valleys that echo the overall shape. The Hausdorff dimension of the boundary is exactly 2, despite being a curve.

Points in $\mathcal{M}$ are conventionally coloured black; points outside are coloured by how quickly $|z_n|$ exceeds 2 (a bound guaranteeing escape). The colours are pure convention, added for visibility. The mathematical object is the set itself.

The key theorem is the connectedness of $\mathcal{M}$: it is a single connected piece. This was proved by Douady and Hubbard in 1982. Whether $\mathcal{M}$ is locally connected remains an open problem.

Julia Sets

Closely related: for each fixed $c$, the Julia set $J_c$ is the boundary between the complex numbers $z$ whose orbits under $z \mapsto z^2 + c$ escape and those that do not.

When $c \in \mathcal{M}$, the Julia set is connected and intricate. When $c \notin \mathcal{M}$, it shatters into a Cantor dust - a totally disconnected cloud of points with no interior. This is the deep duality between the Mandelbrot set and Julia sets: $\mathcal{M}$ is precisely the set of parameters for which $J_c$ is connected.

Different values of $c$ produce dramatically different Julia sets. At $c = 0$ the Julia set is the unit circle. At $c = -0.75$ it becomes a dendrite, a branching tree with no interior. At $c = -0.12 + 0.74i$ it produces the “Douady rabbit”, a plane-filling shape with three lobes.

Iterated Function Systems

Many fractals are more cleanly described by iterated function systems (IFS). An IFS is a finite collection of contractions ${f_1, \ldots, f_k}$ on a complete metric space, typically $\mathbb{R}^2$. The attractor is the unique compact set $A$ satisfying:

$$A = f_1(A) \cup f_2(A) \cup \cdots \cup f_k(A).$$

Existence and uniqueness follow from Banach’s fixed-point theorem applied to the Hausdorff metric on compact sets.

The Sierpiński triangle is the attractor of three contractions, each mapping the plane toward one vertex of the triangle with ratio $1/2$:

$$f_1(x) = \tfrac{1}{2}x, \quad f_2(x) = \tfrac{1}{2}x + \begin{pmatrix}1/2 \ 0\end{pmatrix}, \quad f_3(x) = \tfrac{1}{2}x + \begin{pmatrix}1/4 \ \sqrt{3}/4\end{pmatrix}.$$

The chaos game for the Sierpiński triangle: start anywhere, randomly pick one of the three functions, apply it, repeat. After a few hundred iterations the points trace the attractor. The fractal emerges from pure randomness - not drawn, but collapsed to.

The Barnsley fern is produced by four affine contractions chosen to mimic the self-similar structure of a fern leaf. Each contraction maps the plane to a differently scaled and rotated copy of the whole. The result is indistinguishable from a photograph of a real fern, a fact that says something deep about how biological growth works.

Strange Attractors

In dynamical systems, chaotic attractors often have fractal geometry. The Lorenz attractor, discovered by Edward Lorenz in 1963 while modelling atmospheric convection, is a bounded region in 3D that trajectories never leave and never revisit. It has a fractal cross-section - slice it with a plane and you see a Cantor-set structure.

The system is:

$$\dot{x} = \sigma(y - x), \quad \dot{y} = x(\rho - z) - y, \quad \dot{z} = xy - \beta z$$

with parameters $\sigma = 10$, $\rho = 28$, $\beta = 8/3$. Nearby trajectories diverge exponentially - this is the “butterfly effect” - yet all remain on the same fractal attractor.

The attractor’s Lyapunov dimension (Kaplan-Yorke formula) is approximately 2.06. It is two-dimensional almost everywhere, barely more than a surface, yet folded and stretched so densely that it fills a region of space.

Why Fractals Appear in Nature

The Koch snowflake is a mathematical toy, but coastlines, mountains, trees, river networks, bronchial passages, and clouds all display statistical self-similarity over several orders of magnitude. Mandelbrot showed that the measured “length” of the British coastline depends on the ruler size - shorter rulers find more detail, and the length diverges. The coastline’s fractal dimension is approximately 1.25.

This is not coincidence. Processes governed by simple local rules - erosion, diffusion-limited aggregation, branching growth - produce fractal structure at every scale, because the same physical forces act at every scale. Fractal dimension provides a single number that captures geometric complexity the way roughness would informally: the British coastline ($d \approx 1.25$) is more irregular than South Africa’s ($d \approx 1.02$), less than Norway’s fjords ($d \approx 1.52$).

The lung maximises surface area for gas exchange within a fixed volume. A bronchial tree that branched without fractal scaling would reach the alveolar scale in far fewer generations, wasting volume. Fractal branching with ratio approximately $2^{-1/3}$ at each generation (Murray’s law) packs surface area into the available space as efficiently as possible.

Examples

Compression. Fractal image compression (Barnsley, Jacquin, 1980s-90s) represents an image as the attractor of an IFS. The compression ratio can be very high because the IFS description is compact. An image of a fern can be stored in four affine transformations rather than pixel arrays. The decompression iterates the IFS until convergence.

Computer graphics. Procedural terrain generation uses fractal subdivision: start with a coarse mesh, recursively subdivide triangles and displace midpoints by a random amount proportional to $s^{-H}$ where $H$ is the Hurst exponent controlling roughness. At $H = 1$ you get smooth hills; at $H = 0.5$ you get rocky mountains. The landscapes in early 3D games were literally Koch curves with noise.

Signal processing. $1/f$ noise (pink noise) has power spectral density $S(f) \propto 1/f^\alpha$. Music, heartbeat intervals, neuron firing rates, and many other natural signals have $\alpha \approx 1$. This is the spectral signature of fractal time series: scale-invariant fluctuations at all timescales. White noise ($\alpha = 0$) has no structure; Brownian motion ($\alpha = 2$) is too smooth. Fractional Brownian motion interpolates between them with Hurst exponent $H = (3-\alpha)/2$.

Read Next: