Fractals and Self-Similarity

A fractal is a geometric set exhibiting structure at all scales, typically with a non-integer ("fractal") dimension exceeding its topological dimension. Self-similarity is the defining mechanism: parts of the object resemble the whole under scaling. Key result: many fractals satisfy $N = s^{- D}$ , giving similarity dimension $D = lo g N / lo g (1/ s)$ . Applies across geometry, dynamical systems, signal analysis, physics, and biology. Coined by Benoît Mandelbrot (1975), with roots in Cantor, Koch, Sierpiński, and Hausdorff (1872–1919).

A fractal is a subset of $R^{n}$ whose fractal dimension strictly exceeds its topological dimension, usually built so that it looks (statistically or exactly) the same at every magnification. The notion matters because classical Euclidean geometry — points, lines, smooth surfaces — cannot describe the roughness of coastlines, clouds, turbulence, blood vessels, or strange attractors, while fractal geometry provides quantitative tools (dimensions, scaling exponents) for exactly these objects. The surprising fact: a curve can have infinite length yet enclose finite area, and a set can have dimension like $1.2619$ — a number, not an integer, that genuinely measures how thoroughly the set fills space. Self-similarity converts an impossible "measure-everything" task into a finite recursion. This article develops the definitions, the box-counting and Hausdorff dimensions, iterated function systems, and the worked arithmetic that makes fractals computable.

The Core Insight

If an object is made of $N$ scaled copies of itself, each shrunk by factor $s < 1$ , then mass and length scale by power laws rather than the integer exponents of Euclidean geometry. Demanding consistency of these power laws forces a single number $D = lo g N / lo g (1/ s)$ — the dimension — which need not be an integer. Self-similarity is not decoration; it is the equation that fixes the dimension.

Formal Statement

Self-Similar Set

A nonempty compact set $F \subset R^{n}$ is (exactly) self-similar if it is the union of finitely many contracted copies of itself:

F = i = 1 ⋃ N f_{i} (F),

where each

f_{i} : R^{n} \to R^{n}

is a contracting similarity,

f_{i} (x) = s_{i} R_{i} x + b_{i}

, with ratio

0 < s_{i} < 1

, rotation/reflection matrix

R_{i} \in O (n)

, and translation

b_{i} \in R^{n}

. The family

{f_{1}, \dots, f_{N}}

is an iterated function system (IFS);

F

is its attractor.

⟳

f_{i}

shrinks distances exactly:

∣ f_{i} (x) - f_{i} (y) ∣ = s_{i} ∣ x - y ∣

, so an IFS is just a finite list of "copy-and-shrink" maps.

Similarity Dimension

If the $N$ maps share a common ratio $s$ and overlap negligibly (the open set condition, below), the similarity dimension is the unique $D$ solving the scaling balance

N s^{D} = 1 ⟺ D = \frac{lo g N}{lo g ( 1/ s )} .

For distinct ratios

s_{i}

D

is the unique root of the Moran equation

\sum_{i = 1}^{N} s_{i}^{D} = 1

⟳a measure

μ

with

μ (F) = 1

split into

N

equal pieces gives each piece mass

s^{D}

(mass scales as length

^{D}

); summing

N

pieces back to

1

yields

N s^{D} = 1

Box-Counting Dimension

Let $N_{ε} (F)$ be the number of boxes of side $ε$ in a grid needed to cover $F$ . The box-counting (Minkowski) dimension is

dim_{B} F = ε \to 0^{+} lim \frac{lo g N _{ε} ( F )}{lo g ( 1/ ε )},

when the limit exists; otherwise one takes upper/lower limits

\overline{dim}_{B}, \underline{dim}_{B}

⟳assume

N_{ε} \sim C ε^{- D}

; take logs to get

lo g N_{ε} \approx lo g C + D lo g (1/ ε)

, then divide by

lo g (1/ ε)

and let

ε \to 0

Hausdorff Dimension

For $s \geq 0$ and $δ > 0$ , define the $s$ -dimensional Hausdorff pre-measure

H_{δ}^{s} (F) = in f {i \sum (diam U_{i})^{s} : F \subset i ⋃ U_{i}, diam U_{i} \leq δ},

and the Hausdorff measure

H^{s} (F) = lim_{δ \to 0} H_{δ}^{s} (F)

. The Hausdorff dimension is the threshold

dim_{H} F = in f {s \geq 0 : H^{s} (F) = 0} = sup {s \geq 0 : H^{s} (F) = \infty} .

⟳

H^{s} (F)

as a function of

s

jumps from

\infty

0

at exactly one value; that crossover defines

dim_{H}

. Infima over all small-diameter covers make it finer than box-counting.

Existence/Uniqueness of the Attractor — Hutchinson 1981

Let ${f_{1}, \dots, f_{N}}$ be contractions on a complete metric space. Define the Hutchinson operator $F (A) = ⋃_{i = 1}^{N} f_{i} (A)$ on nonempty compact sets with the Hausdorff metric $d_{H}$ . Then $F$ is a contraction, and there exists a unique nonempty compact $F$ with $F (F) = F$ ; moreover $F^{k} (A_{0}) \to F$ for any starting compact $A_{0}$ .

Moran / Open Set Condition

If the IFS satisfies the open set condition — there is a nonempty bounded open set $V$ with $⋃_{i} f_{i} (V) \subseteq V$ and the $f_{i} (V)$ pairwise disjoint — then the attractor's Hausdorff and box dimensions coincide with the similarity dimension:

dim_{H} F = dim_{B} F = D, i = 1 \sum N s_{i}^{D} = 1.

⟳OSC prevents copies from overlapping enough to lose dimension; one then proves

0 < H^{D} (F) < \infty

via a mass-distribution argument, pinning

dim_{H} = D

Why It Holds / Derivation

Hutchinson Operator is a Contraction

Step 1. Equip the space

K

of nonempty compact subsets of

R^{n}

with the Hausdorff metric

d_{H} (A, B) = max {a \in A sup b \in B in f ∣ a - b ∣, b \in B sup a \in A in f ∣ a - b ∣} .

⟳

d_{H}

measures the worst-case distance you must travel from one set to reach the other, symmetrized.

Step 2. For a single similarity

f_{i}

with ratio

s_{i}

d_{H} (f_{i} (A), f_{i} (B)) \leq s_{i} d_{H} (A, B)

, since

f_{i}

scales all distances by

s_{i}

. Step 3. Use the union inequality

d_{H} (⋃_{i} A_{i}, ⋃_{i} B_{i}) \leq max_{i} d_{H} (A_{i}, B_{i})

. Hence

d_{H} (F (A), F (B)) \leq i max d_{H} (f_{i} (A), f_{i} (B)) \leq (i max s_{i}) d_{H} (A, B) .

⟳combine Steps 2–3; the contraction constant is

s_{m a x} = max_{i} s_{i} < 1

Step 4. Since

s_{m a x} < 1

F

is a contraction on the complete metric space

(K, d_{H})

. By the Banach Fixed Point Theorem, a unique fixed point

F = F (F)

exists and iterates converge to it geometrically.

■

Cantor Set Dimension via Self-Similarity

Step 1. The middle-thirds Cantor set

C

is the attractor of

f_{1} (x) = x /3

f_{2} (x) = x /3 + 2/3

. So

N = 2

s = 1/3

. Step 2. These maps send

[0, 1]

[0, 1/3]

and

[2/3, 1]

, disjoint open intervals — the OSC holds with

V = (0, 1)

. Step 3. Apply equation [2]:

D = lo g 2/ lo g 3 \approx 0.6309

dim_{H} C = \frac{lo g 2}{lo g 3} = 0.63092975 \dots

⟳substitute

N = 2, 1/ s = 3

into

D = lo g N / lo g (1/ s)

Step 4. Cross-check by box-counting: at scale

ε = 3^{- k}

, the

k

-th stage has

2^{k}

intervals, so

N_{ε} = 2^{k}

. Then

lo g N_{ε} / lo g (1/ ε) = k lo g 2/ (k lo g 3) = lo g 2/ lo g 3

, confirming [3]=[9].

■

Benoît B. Mandelbrot | 1924–2010

Polish-born French-American mathematician at IBM Research and later Yale. Mandelbrot synthesized scattered 19th- and early-20th-century "pathological" curves into a coherent geometry of roughness, coining fractal (from Latin fractus, "broken") in 1975 and publishing The Fractal Geometry of Nature (1982). His 1967 paper "How Long Is the Coast of Britain?" showed measured length depends on ruler size via a power law, making fractal dimension empirically meaningful. The Mandelbrot set, a parameter-space portrait of quadratic dynamics, bears his name and became the icon of computer-age mathematics.

Building Intuition

Think of dimension as "how fast content grows when you zoom out." Double the ruler size: a line covers 2× as much ( $2^{1}$ ), a square 4× ( $2^{2}$ ), a cube 8× ( $2^{3}$ ). A fractal sits between: the Koch curve grows by $4$ when scaled by $3$ , so its exponent is $lo g 4/ lo g 3 \approx 1.26$ — "more than a line, less than a plane." Self-similarity is the promise that this growth rate is the same at every scale, which is exactly what lets a single equation [2] capture an infinitely detailed object. The Cantor set is "dust" too sparse to be a line ( $dim < 1$ ); the Koch curve is "too wiggly" to be a line ( $dim > 1$ ). Both are quantifying the gap between topology (how connected) and metric content (how much room it fills).

Koch curve generator: replacing one segment by four, each scaled by 1/3

Worked Examples

Sierpiński Triangle — dimension and stage-3 count

Given. Start with a filled equilateral triangle of side

L = 1

. At each stage remove the central (inverted) sub-triangle, leaving

N = 3

copies each scaled by

s = 1/2

. Step 1 — Dimension. Apply [2]:

D = \frac{lo g 3}{lo g 2} = \frac{1.0986}{0.6931} = 1.5849 \dots

⟳

N = 3

1/ s = 2

, plug into

lo g N / lo g (1/ s)

Step 2 — Number of triangles at stage 3. Each stage triples the count: stage

k

has

3^{k}

triangles. For

k = 3

3^{3} = 27

. Step 3 — Side length at stage 3.

L_{3} = (1/2)^{3} \cdot 1 = 1/8

. Step 4 — Surviving area at stage 3. Each stage keeps

3/4

of the area:

A_{3} = (3/4)^{3} A_{0} = 0.4219 A_{0}

, with

A_{0} = \frac{3}{4}

for unit side. So

A_{3} = 0.4219 \times 0.4330 = 0.1827

D = 1.585, 27 triangles, side = 0.125, A_{3} = 0.183 (unit^{2})

As $k \to \infty$ , $A_{k} = (3/4)^{k} A_{0} \to 0$ : the limit set has zero area yet dimension $1.585$ .

Coastline measurement — empirical fractal dimension

Given. A coastline measured with ruler length

ε

yields total length

L (ε)

. Data: with

ε = 200

km,

L = 2400

km; with

ε = 50

km,

L = 3800

km. Step 1 — Model. Richardson's law:

L (ε) = C ε^{1 - D}

, so number of ruler steps is

N (ε) = L / ε = C ε^{- D}

. Step 2 — Take the ratio (eliminate $C$ ).

\frac{L ( ε _{2} )}{L ( ε _{1} )} = (\frac{ε _{2}}{ε _{1}})^{1 - D} .

⟳divide

L (ε_{2}) = C ε_{2}^{1 - D}

L (ε_{1}) = C ε_{1}^{1 - D}

;

C

cancels.

Step 3 — Solve for $D$ . Plug numbers:

\frac{3800}{2400} = (\frac{50}{200})^{1 - D} \Rightarrow 1.5833 = (0.25)^{1 - D} .

Take logs:

1 - D = \frac{ln 1.5833}{ln 0.25} = \frac{0.4595}{- 1.3863} = - 0.3315

. Step 4 — Answer.

D = 1 + 0.3315

D \approx 1.33

This matches typical rugged coastlines (Britain's west coast $\approx 1.25$ , Norway's fjords $\approx 1.5$ ). The longer length at finer rulers ( $3800 > 2400$ ) is the signature of $D > 1$ .

Common Misconceptions

"Fractal dimension is always non-integer." False — fractal dimension can be an integer. A space-filling Peano curve has

dim_{H} = 2

; Brownian motion paths have dimension

2

in the plane. The defining feature is

dim_{H} > dim_{topological}

, not non-integrality. Conversely, non-integer box dimension does not by itself prove self-similarity — many fractals (coastlines, attractors) are only statistically self-similar, with no exact IFS. "Box and Hausdorff dimensions are interchangeable." Always

dim_{H} F \leq \underline{dim}_{B} F \leq \overline{dim}_{B} F

, and the inequality can be strict. The set

{0} \cup {1/ n : n \in N}

has

dim_{H} = 0

but

dim_{B} = 1/2

. Box dimension is not countably stable; Hausdorff dimension is.

A countable set with positive box dimension

Let $F = {0} \cup {1/ n : n \geq 1}$ . Being countable, $dim_{H} F = 0$ (each point has measure zero and Hausdorff dimension is countably stable). Yet a covering-count estimate gives

dim_{B} F = \frac{1}{2} .

⟳near

0

the points

1/ n

cluster; covering them at scale

ε

needs about

ε^{- 1/2}

boxes because consecutive gaps

1/ n - 1/ (n + 1) \sim n^{- 2}

exceed

ε

only up to

n \sim ε^{- 1/2}

This single example demolishes the assumption that "small/countable" forces both dimensions to agree.

Connections

Fractal geometry sits at the crossroads of several undergraduate fields. Dimension generalizes the Topological Dimension of point-set topology, while Hausdorff Measure extends Lebesgue Measure to fractional exponents. The IFS framework is a direct application of the Banach Fixed Point Theorem in the complete metric space of compact sets under the Hausdorff Metric. Dynamically, fractals appear as Strange Attractors and Julia Sets in Chaos Theory, where sensitivity to initial conditions produces self-similar invariant sets. Statistically self-similar fractals connect to Power Law distributions, Brownian Motion, and Scale Invariance in Renormalization Group physics.

Real-World Applications

Antenna design — Cohen's fractal antennas (Sierpiński, Koch geometries) pack multi-band resonance into compact area; used in mobile phones and RFID tags.
Image compression — Fractal image compression (Barnsley, Jacquin) encodes images as IFS fixed points, exploiting self-similarity between image blocks.
Medical diagnostics — fractal dimension of retinal vasculature, tumor boundaries, and EEG/heart-rate signals serves as a biomarker; lower dimension can indicate pathology.
Computer graphics — terrain and cloud generation via fractional Brownian motion and the diamond–square algorithm in films and games (e.g. Star Trek II Genesis sequence).
Geophysics & hydrology — river network branching, fault roughness, and porosity follow fractal scaling, informing reservoir and earthquake models.
Finance — Mandelbrot's multifractal models of asset returns capture fat tails and volatility clustering missed by Gaussian models.
Materials & turbulence — aggregate growth (DLA), polymer conformations, and the inertial range of turbulent flows obey power-law (fractal) scaling laws.

Historical Development

1872 — Karl Weierstrass exhibits a continuous, nowhere-differentiable function, a "monster" curve shocking the analytic establishment.
1883 — Georg Cantor constructs the middle-thirds set, an uncountable measure-zero "dust."
1890 — Giuseppe Peano builds a continuous space-filling curve, dimension $1$ mapping onto dimension $2$ .
1904 — Helge von Koch describes his snowflake curve: continuous, nowhere differentiable, infinite length, finite enclosed area.
1915–16 — Wacław Sierpiński introduces his triangle (gasket) and carpet.
1918–19 — Felix Hausdorff defines fractional-dimensional measure; Constantin Carathéodory's earlier outer-measure ideas underpin it.
1967 — Mandelbrot's Science paper "How Long Is the Coast of Britain?" links Richardson's empirical data to fractional dimension.
1975 — Mandelbrot coins the term fractal.
1981 — John Hutchinson formalizes iterated function systems and proves attractor existence/uniqueness; Michael Barnsley later develops the "collage theorem" (1988).

Subtleties & Limits

Despite computational ubiquity, deep questions remain. Computing Hausdorff dimension exactly is hard when the open set condition fails (overlapping IFS): the dimension of self-similar measures with overlaps (e.g. Bernoulli convolutions, $λ$ -expansions) was open for decades — Hochman (2014) and Shmerkin resolved major cases, but a full classification is incomplete. The dimension of the Mandelbrot set boundary is exactly $2$ (Shishikura, 1991), yet whether it has positive area is still open. Multifractal formalism — describing sets where the scaling exponent varies pointwise via a spectrum $f (α)$ — lacks a complete rigorous theory for general measures. And whether physical "fractals" truly scale across infinitely many decades (versus a finite scaling range cut off by molecular or planetary limits) is a perpetual modeling caveat: nature's fractals are bounded; mathematics' are not.