The Problem Pit: The Pit of the Generalized Takagi Function

 

The Pit of the Generalized Takagi Function: Smooth Nowhere

Some monsters in mathematics announce themselves immediately. The Mandelbrot set, with its spirals and lightning bolts, screams “fractal.” The Cantor set, built by removing the middle third forever, is obviously bizarre.

But the Takagi function whispers.

It starts quietly: just sawtooth waves stacked on top of each other. It looks like a child’s drawing of mountains. Yet when I tried to understand its smoothness, I fell into a pit — and I didn’t find my way out easily.

---

Step 1: Meeting the Function

The generalized Takagi function is defined as

$$T_{\alpha,\beta}(x) = \sum_{n=0}^\infty \alpha^n \, \tau(\beta^n x),$$

where $\tau(x) = \text{dist}(x, \mathbb{Z})$ is the sawtooth function — the distance from $x$ to the nearest integer.

• Each $\tau(\beta^n x)$ is a little triangular wave, oscillating faster as $n$ increases.

• The coefficients $\alpha^n$ dampen the size of the waves.

At first sight: this looks perfectly harmless.

I thought: surely this is continuous, and maybe even differentiable except at some rationals.

That was the first wrong step.

---

Step 2: The Trichotomy (A Rope Into the Pit)

The behavior depends on the product $\alpha \beta$:

• If $\alpha \beta < 1$: The damping is strong. The series is Lipschitz continuous. Smooth enough.

• If $\alpha \beta > 1$: The oscillations outrun the damping. The function is continuous but nowhere differentiable, with difference quotients blowing up like a power law.

• If $\alpha \beta = 1$: The cliff. The balance point. What happens here?

I naively thought: maybe this is the “nice” case — a perfect balance, maybe even differentiable almost everywhere.

But the cliff was sharper than I expected.

---

Step 3: My Naive Differentiation

I tried to differentiate term by term:

$$\frac{d}{dx} T_{\alpha,\beta}(x) = \sum_{n=0}^\infty \alpha^n \frac{d}{dx}\tau(\beta^n x).$$

But here’s the problem:

• Each $\tau(\beta^n x)$ is piecewise linear with slope $\pm \beta^n$.

• At integers and half-integers (scaled by $\beta^{-n}$), there are kinks.

As $n$ grows, the kinks get denser. In the limit, every neighborhood of every $x$ contains infinitely many kinks.

So the derivative series doesn’t converge anywhere.

Dead end.

---

Step 4: The Difference Quotient Trap

Okay, maybe I could analyze the difference quotient:

$$Q(x,h) = \frac{T_{\alpha,\beta}(x+h) - T_{\alpha,\beta}(x)}{h}.$$

For subcritical ($\alpha \beta < 1$), estimates show $|Q(x,h)|$ is bounded, so the function is Lipschitz.

For supercritical ($\alpha \beta > 1$), difference quotients grow like $|h|^{-\gamma}$. No hope of differentiability.

But at the critical case ($\alpha \beta = 1$), I ran into fog.

Sometimes $Q(x,h)$ seemed to stay small. Sometimes it grew. I thought: maybe it depends on $x$. Maybe rationals behave differently than irrationals? Maybe “almost everywhere” it’s fine?

I plotted some difference quotients numerically. For $x=0.1, h=10^{-k}$, I got values bouncing unpredictably. For rationals with denominators powers of $\beta$, they spiked.

My guesses crumbled.

---

Step 5: Trying Specific Examples

Take the classical Takagi function: $\alpha=\tfrac{1}{2}, \beta=2$.

So we’re at the critical case: $\alpha \beta = 1$.

Try $x=0.25$. Compute

$$T_{1/2,2}(x) = \sum_{n=0}^\infty 2^{-n-1} \tau(2^n x).$$

• For $n=0$: $\tau(0.25) = 0.25.$ Contribution: $0.25$.

• For $n=1$: $\tau(0.5) = 0.5.$ Contribution: $0.25$.

• For $n=2$: $\tau(1) = 0.$ Contribution: 0.

• For $n=3$: $\tau(2) = 0.$ Contribution: 0.

• For $n=4$: $\tau(4) = 0.$ Contribution: 0.

So partial sums: $0.25 + 0.25 = 0.5$.

Seems tame. But now look at the difference quotient around $x=0.25$:

Take $h=0.01$.

$$Q(0.25,0.01) = \frac{T(0.26) - T(0.25)}{0.01}.$$

Numerical experiments show it’s around 4. Shrink $h$ to 0.001 → value jumps to ~9. Shrink further → it keeps growing. Not exploding like $1/h$, but creeping up like $\log(1/h)$.

The tame face hides logarithmic teeth.

---

Step 6: The Logarithmic Monster

The actual theorem says:

At $\alpha \beta = 1$, the Takagi function is continuous but nowhere differentiable. And the difference quotient satisfies

$$|Q(x,h)| \geq C |\log h|$$

along infinitely many scales.

That’s the secret: the growth is logarithmic. Slower than supercritical chaos, but enough to kill every chance of differentiability.

It’s not smooth anywhere.

---

Step 7: Digit Combinatorics — Why It Fails

Why logarithmic? The answer lies in base-$\beta$ expansions.

Write $x$ in base $\beta$. Whether a difference quotient at scale $h = \beta^{-m}$ grows depends on whether the digits line up to create “bad scales.”

• If you hit long runs of the same digit, you get spikes.

• These spikes are rare, but they happen infinitely often.

So no matter which $x$ you pick, infinitely many “bad scales” ruin differentiability.

It’s like trying to walk across a floor with random nails sticking up. You’ll always step on one eventually.

---

Step 8: Comparing to Other Monsters

At this point, I tried to compare:

• Weierstrass function:

  $$W(x) = \sum_{n=0}^\infty a^n \cos(b^n \pi x).$$

  Its differentiability depends on $ab$. At the critical line, it’s power-law bad. The Takagi function is different — its oscillations are piecewise linear, so at criticality we get logarithmic failure instead of power-law.

• Cantor function: continuous but flat almost everywhere. Opposite problem: too differentiable (derivative 0 almost everywhere). Takagi is never differentiable.

The Takagi sits in between, carving out its own pit.

---

Step 9: The Open Abyss

Even after understanding the logarithmic failure, the pit doesn’t end.

• Exact constants: What is the best possible $C$ in $|Q(x,h)| \geq C \log(1/h)$? Does it depend delicately on $\beta$?

• Fractal dimension: What is the Hausdorff dimension of the graph at criticality? Does the logarithmic failure affect it?

• Multifractal analysis: How do the Hölder exponents distribute across points?

• Randomized versions: If we add randomness to coefficients, does the critical logarithmic behavior survive?

• Higher dimensions: What happens if we define Takagi-like functions on $\mathbb{R}^d$?

The function refuses to give a complete picture.

---

Step 10: Climbing Out (Barely)

By the end of my journey, I knew this:

• The generalized Takagi function at $\alpha\beta=1$ is continuous but nowhere differentiable.

• Its difference quotients grow logarithmically.

• Its graph is a fractal coast — never smooth, always jagged.

• It hides combinatorics of base-$\beta$ digits, and connects to questions still open in fractal geometry.

I went in expecting a mountain of sawteeth. I came out knowing I had been staring into a razor’s edge fractal.

That’s The Problem Pit: what seems playful becomes unfathomably strange the deeper you dig.

---