The Pit of Fractal Integrals

The Pit of Fractal Integrals: Measuring the Unmeasurable with Takagi’s Monster

There are functions that refuse to be smooth. They wiggle, they bend, they squirm, and the closer you look, the worse they behave. In real analysis, such creatures live everywhere — but a few have become infamous for how innocently they present themselves before revealing their pathological depths.

One such monster is the Takagi function, sometimes called the “blancmange function.” On first sight, it looks like a child’s zig-zag doodle. Defined by a simple infinite sum of tent functions, it has the unnerving property of being continuous everywhere, but differentiable nowhere.

When I first met it, I thought: “Fine, I know Weierstrass did this already. This is just another example of pathological analysis.”

But then I asked a simple question:

I(\alpha) = \int_0^1 f_\alpha(x) \, dx,

where

f_\alpha(x) = \sum_{n=0}^\infty \alpha^n \, \phi(2^n x), \qquad \phi(x) = \min_{k\in\mathbb{Z}} |x-k|.

That is: the Takagi–Landsberg function with parameter $0 < \alpha < 1$ .

It looked like a straightforward problem: just compute the area under the fractal. Surely it should be simple — after all, $\phi$ is just a triangle wave.

That thought was the rope that pulled me straight into the pit.

Step 1: First Impressions

At a glance:

Each $\phi(2^n x)$ is a sawtooth with frequency $2^n$ and amplitude at most $1/2$ .
The factor $\alpha^n$ damps the terms, so the series converges uniformly.
So $f_\alpha$ is continuous. Good.

If $\alpha < 1/2$ , the higher frequencies shrink rapidly. The graph looks smooth-ish.
If $\alpha \to 1$ , the graph grows spiky and chaotic, clearly fractal-like.

So maybe the integral would interpolate smoothly between “easy” and “wild.”

I was wrong.

Step 2: The Naive Attack

The integral looks linear. Why not swap sum and integral?

I(\alpha) = \sum_{n=0}^\infty \alpha^n \int_0^1 \phi(2^n x) \, dx.

That seems safe: the series converges absolutely.

Now, $\int_0^1 \phi(2^n x) \, dx$ should be the average value of the triangle wave. Since $\phi$ is periodic with mean $1/4$ , scaling doesn’t change the mean.

So each term contributes $1/4$ .

I(\alpha) = \sum_{n=0}^\infty \alpha^n \cdot \frac{1}{4}.

That’s just a geometric series:

I(\alpha) = \frac{1}{4} \cdot \frac{1}{1-\alpha}.

I felt triumphant. Easy. Done.

Then I realized: this was only the beginning.

Step 3: Suspicion Creeps In

The formula $\tfrac{1}{4(1-\alpha)}$ is elegant, but too simple.

Was it really capturing the whole story?

I checked:

For $\alpha = 0$ , we get $I(0) = 1/4$ . Correct: it’s just one triangle wave.
For $\alpha = 1/2$ , we get $I(1/2) = 1/2$ . Numerical checks confirmed it.
For $\alpha = 0.9$ , we get $I(0.9) = 2.5$ . Plausible, since the function grows tall.

So maybe the formula was correct.

But integrals don’t just give numbers — they tell stories. What about variations, moments, energies?

I tried to compute

J(\alpha) = \int_0^1 f_\alpha(x)^2 \, dx.

This was the beginning of the real pit.

Step 4: Squaring the Monster

Expanding:

J(\alpha) = \int_0^1 \left( \sum_{m=0}^\infty \alpha^m \phi(2^m x) \right) \left( \sum_{n=0}^\infty \alpha^n \phi(2^n x) \right) dx.

Swapping sums:

J(\alpha) = \sum_{m,n \geq 0} \alpha^{m+n} \int_0^1 \phi(2^m x)\phi(2^n x) \, dx.

Now things were messy.

The cross terms $\int \phi(2^m x)\phi(2^n x)$ aren’t trivial. They encode correlations between different scales of the fractal.

I thought: maybe they vanish when $m \neq n$ . That would make life easy.

Wrong. They don’t vanish. They overlap, resonate, and produce subtle correlations.

Dead end #1.

Step 5: Fourier Analysis to the Rescue?

I remembered: triangle waves have simple Fourier series.

\phi(x) = \frac{1}{4} - \frac{2}{\pi^2} \sum_{k=1}^\infty \frac{\cos(2\pi k x)}{k^2}.

Plugging this into $f_\alpha$ gave:

f_\alpha(x) = \sum_{n=0}^\infty \alpha^n \left( \frac{1}{4} - \frac{2}{\pi^2} \sum_{k=1}^\infty \frac{\cos(2\pi k \cdot 2^n x)}{k^2} \right).

Suddenly the function looked like a Fourier fractal: a weighted sum of cosines with exponentially growing frequencies.

I thought this would help compute $J(\alpha)$ . Parseval’s theorem relates integrals of squares to Fourier coefficients.

But the sum was chaotic: frequencies doubled at each step. Orthogonality didn’t cleanly cancel them.

Dead end #2.

Step 6: Self-Similarity as a Tool

The Takagi function is self-similar. That suggested recursion.

Indeed, $f_\alpha(x)$ satisfies a functional equation:

f_\alpha(x) = \phi(x) + \alpha f_\alpha(2x).

That’s gold: it encodes the fractal scaling.

So maybe I could derive an equation for $J(\alpha)$ .

Squaring both sides and integrating:

J(\alpha) = \int_0^1 \phi(x)^2 \, dx + 2\alpha \int_0^1 \phi(x)f_\alpha(2x)\, dx + \alpha^2 \int_0^1 f_\alpha(2x)^2 \, dx.

Now the integrals mix scales again.

But change of variables helped: $\int_0^1 f_\alpha(2x)^2 dx = \frac{1}{2} \int_0^2 f_\alpha(u)^2 du = \frac{1}{2}(J(\alpha) + J(\alpha)) = J(\alpha)$ .

So the last term simplified.

Still, the middle term was ugly.

Progress, but the walls of the pit grew steeper.

Step 7: Numerical Experiments

At this point, I abandoned analysis for a while and just computed $J(\alpha)$ numerically.

For $\alpha = 0.5$ , I got roughly 0.333...
For $\alpha = 0.9$ , I got ~7.8.

The growth was rapid. It looked like $J(\alpha)$ diverged as $\alpha \to 1$ .

That made sense: the function gets rougher and rougher, the oscillations stack, the energy blows up.

But could I prove this rigorously?

Step 8: The Fractal Dimension Angle

Another detour: what if I didn’t care about the exact integral, but about the scaling behavior?

The Takagi–Landsberg function’s graph has Hausdorff dimension > 1 when $\alpha$ is large.

So the question “what is $\int f_\alpha$ ” morphs into “how does the fractal geometry of the graph control integrals of its powers?”

This was a whole new pit: geometric measure theory.

I realized the integral wasn’t just a number. It was a window into dimension.

Step 9: Connection to Probability

Then I found another surprising link:

f_\alpha(x) = \mathbb{E}\left[ \sum_{n=0}^\infty \alpha^n \left| x - \frac{K}{2^n} \right| \right],

where $K$ is a random integer with certain distribution.

So $f_\alpha(x)$ could be seen as an expected distance to random dyadic rationals.

The integral became an expectation over expectations. A probabilistic object in disguise.

I didn’t solve it this way, but it gave a new perspective: the Takagi monster is a random walk over scales.

Step 10: Where the Pit Opens Wider

Every attempt at a closed form for $J(\alpha)$ dragged me into heavier machinery:

Harmonic analysis: to manage cross terms.
Fractal measure theory: to interpret divergence.
Probability: to reinterpret the function as a stochastic process.
Dynamical systems: since $f_\alpha(x)$ relates to iterating the tent map.

What began as “just compute an integral” became a tour of half of modern analysis.

Step 11: What We Actually Know

After hours in the pit, here’s what emerged:

The first moment $I(\alpha)$ is simple: $I(\alpha) = \tfrac{1}{4(1-\alpha)}$ .
The second moment $J(\alpha)$ is messy, tied to correlations across scales. Explicit formulas exist, but they involve zeta values and infinite series.
As $\alpha \to 1$ , $J(\alpha)$ diverges, reflecting the function’s growing roughness.
The integrals encode fractal dimension information.

So the integral that looked harmless was actually a barometer for fractality itself.

Step 12: Climbing Out

By the end, I realized:

The area under a fractal is sometimes easy.
The energy under a fractal is chaos.
Simple sums of triangle waves can encode dimensions, measures, even randomness.
Integrals, instead of taming the function, reveal its wildness.

I went in expecting to “just integrate a function.” I came out tangled in Fourier series, Hausdorff dimension, and stochastic processes.

That’s the essence of The Problem Pit: beneath the simplest problem lies the whole of analysis, waiting to pull you down.

Arya’s Algorithms

Search This Blog