Fractals and Self-Similarity

⚡ Key Takeaways

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 $D = \log N / \log(1/s)$ . 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).

💡 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 $D = \log N/\log(1/s)$ — the dimension — which need not be an integer. Self-similarity is not decoration; it is the equation that fixes the dimension.

DefinitionSelf-Similar Set

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

1

\quad F = \bigcup_{i=1}^{N} f_i(F),

where each $f_i:\mathbb{R}^n \to \mathbb{R}^n$ 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 $b_i \in \mathbb{R}^n$ . The family $\{f_1,\dots,f_N\}$ is an iterated function system (IFS); $F$ is its attractor. Derivation hint: $f_i$ 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.

DefinitionSimilarity 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

2

\quad N s^{D} = 1 \quad\Longleftrightarrow\quad D = \frac{\log N}{\log(1/s)}.

For distinct ratios $s_i$ , $D$ is the unique root of the Moran equation $\sum_{i=1}^N s_i^{D} = 1$ . Derivation hint: a measure $\mu$ with $\mu(F)=1$ split into $N$ equal pieces gives each piece mass $s^D$ (mass scales as length $^D$ ); summing $N$ pieces back to $1$ yields $Ns^D=1$ .

DefinitionBox-Counting Dimension

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

3

\quad \dim_B F = \lim_{\varepsilon \to 0^+} \frac{\log N_\varepsilon(F)}{\log(1/\varepsilon)},

when the limit exists; otherwise one takes upper/lower limits $\overline{\dim}_B,\underline{\dim}_B$ . Derivation hint: assume $N_\varepsilon \sim C\varepsilon^{-D}$ ; take logs to get $\log N_\varepsilon \approx \log C + D\log(1/\varepsilon)$ , then divide by $\log(1/\varepsilon)$ and let $\varepsilon\to0$ .

DefinitionHausdorff Dimension

For $s\ge 0$ and $\delta>0$ , define the $s$ -dimensional Hausdorff pre-measure

4

\quad \mathcal{H}^s_\delta(F) = \inf\Big\{ \sum_{i} (\operatorname{diam} U_i)^s : F \subset \bigcup_i U_i,\ \operatorname{diam} U_i \le \delta \Big\},

and the Hausdorff measure $\mathcal{H}^s(F) = \lim_{\delta\to 0}\mathcal{H}^s_\delta(F)$ . The Hausdorff dimension is the threshold

5

\quad \dim_H F = \inf\{ s \ge 0 : \mathcal{H}^s(F) = 0\} = \sup\{ s \ge 0 : \mathcal{H}^s(F) = \infty\}.

Derivation hint: $\mathcal H^s(F)$ as a function of $s$ jumps from $\infty$ to $0$ at exactly one value; that crossover defines $\dim_H$ . Infima over all small-diameter covers make it finer than box-counting.

TheoremExistence/Uniqueness of the Attractor — Hutchinson 1981

Let $\{f_1,\dots,f_N\}$ be contractions on a complete metric space. Define the Hutchinson operator $\mathcal{F}(A) = \bigcup_{i=1}^N f_i(A)$ on nonempty compact sets with the Hausdorff metric $d_H$ . Then $\mathcal{F}$ is a contraction, and there exists a unique nonempty compact $F$ with $\mathcal{F}(F)=F$ ; moreover $\mathcal{F}^k(A_0) \to F$ for any starting compact $A_0$ .

TheoremMoran / Open Set Condition

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

6

\quad \dim_H F = \dim_B F = D, \qquad \sum_{i=1}^N s_i^{D} = 1.

Derivation hint: OSC prevents copies from overlapping enough to lose dimension; one then proves $0<\mathcal H^D(F)<\infty$ via a mass-distribution argument, pinning $\dim_H=D$ .

ProofHutchinson Operator is a Contraction

Step 1. Equip the space $\mathcal{K}$ of nonempty compact subsets of $\mathbb{R}^n$ with the Hausdorff metric

7

\quad d_H(A,B) = \max\Big\{\, \sup_{a\in A}\inf_{b\in B}|a-b|,\ \sup_{b\in B}\inf_{a\in A}|a-b| \,\Big\}.

Derivation hint: $d_H$ 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 $s_i$ , $d_H(f_i(A), f_i(B)) \le s_i\, d_H(A,B)$ , since $f_i$ scales all distances by $s_i$ .

Step 3. Use the union inequality $d_H\big(\bigcup_i A_i, \bigcup_i B_i\big) \le \max_i d_H(A_i,B_i)$ . Hence

8

\quad d_H(\mathcal F(A), \mathcal F(B)) \le \max_i d_H(f_i(A),f_i(B)) \le \Big(\max_i s_i\Big)\, d_H(A,B).

Derivation hint: combine Steps 2–3; the contraction constant is $s_{\max}=\max_i s_i<1$ .

Step 4. Since $s_{\max}<1$ , $\mathcal F$ is a contraction on the complete metric space $(\mathcal K, d_H)$ . By the Banach Fixed Point Theorem, a unique fixed point $F=\mathcal F(F)$ exists and iterates converge to it geometrically. $\blacksquare$

∎

ProofCantor 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 $[0,1]$ to $[0,1/3]$ and $[2/3,1]$ , disjoint open intervals — the OSC holds with $V=(0,1)$ .

Step 3. Apply equation [2]: $D = \log 2/\log 3 \approx 0.6309$ .

9

\quad \dim_H C = \frac{\log 2}{\log 3} = 0.63092975\ldots

Derivation hint: substitute $N=2,\ 1/s=3$ into $D=\log N/\log(1/s)$ .

Step 4. Cross-check by box-counting: at scale $\varepsilon = 3^{-k}$ , the $k$ -th stage has $2^k$ intervals, so $N_\varepsilon = 2^k$ . Then $\log N_\varepsilon/\log(1/\varepsilon) = k\log 2/(k\log 3) = \log 2/\log 3$ , confirming [3]=[9]. $\blacksquare$

∎

BB

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.

🧠 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 $\log 4/\log 3 \approx 1.26$ — "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).

Figure

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

ExampleSierpiński 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]:

10

\quad D = \frac{\log 3}{\log 2} = \frac{1.0986}{0.6931} = 1.5849\ldots

Derivation hint: $N=3$ , $1/s=2$ , plug into $\log N/\log(1/s)$ .

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 $A_0=\frac{\sqrt3}{4}$ for unit side. So $A_3 = 0.4219 \times 0.4330 = 0.1827$ .

\boxed{D = 1.585,\quad 27\text{ triangles},\quad \text{side }=0.125,\quad A_3 = 0.183 \text{ (unit}^2)}

As $k\to\infty$ , $A_k = (3/4)^k A_0 \to 0$ : the limit set has zero area yet dimension $1.585$ .

ExampleCoastline measurement — empirical fractal dimension

Given. A coastline measured with ruler length $\varepsilon$ yields total length $L(\varepsilon)$ . Data: with $\varepsilon = 200$ km, $L = 2400$ km; with $\varepsilon = 50$ km, $L = 3800$ km.

Step 1 — Model. Richardson's law: $L(\varepsilon) = C\,\varepsilon^{1-D}$ , so number of ruler steps is $N(\varepsilon)=L/\varepsilon = C\varepsilon^{-D}$ .

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

11

\quad \frac{L(\varepsilon_2)}{L(\varepsilon_1)} = \left(\frac{\varepsilon_2}{\varepsilon_1}\right)^{1-D}.

Derivation hint: divide $L(\varepsilon_2)=C\varepsilon_2^{1-D}$ by $L(\varepsilon_1)=C\varepsilon_1^{1-D}$ ; $C$ cancels.

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

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

Take logs: $1-D = \dfrac{\ln 1.5833}{\ln 0.25} = \dfrac{0.4595}{-1.3863} = -0.3315$ .

Step 4 — Answer. $D = 1 + 0.3315$ .

\boxed{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 mistake

"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 $\dim_H > \dim_{\text{topological}}$ , 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.

⚠️ Common mistake

"Box and Hausdorff dimensions are interchangeable." Always $\dim_H F \le \underline{\dim}_B F \le \overline{\dim}_B F$ , and the inequality can be strict. The set $\{0\}\cup\{1/n : n\in\mathbb N\}$ has $\dim_H = 0$ but $\dim_B = 1/2$ . Box dimension is not countably stable; Hausdorff dimension is.

CounterexampleA countable set with positive box dimension

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

12

\quad \dim_B F = \tfrac{1}{2}.

Derivation hint: near $0$ the points $1/n$ cluster; covering them at scale $\varepsilon$ needs about $\varepsilon^{-1/2}$ boxes because consecutive gaps $1/n - 1/(n+1)\sim n^{-2}$ exceed $\varepsilon$ only up to $n\sim\varepsilon^{-1/2}$ . This single example demolishes the assumption that "small/countable" forces both dimensions to agree.

connection

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.

🔬 Where this appears

⏳ History

- 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).

open-problem

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, $\lambda$ -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 $f(\alpha)$ — 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.

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also