The Problem Pit: The Pit of the Generalized Takagi Function

 

The Pit of the Generalized Takagi Function: Smooth Nowhere

There are monsters in mathematics that don’t roar right away. They don’t shout “I am a fractal.” They start quietly, with innocent definitions, and lure you into believing they are tame. The Takagi function is one of them.

It’s defined by the infinite series

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

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

The ingredients are harmless: sawteeth, scaling, some coefficients. You can plot the first few partial sums, and what you get looks like a jagged mountain range. But it’s continuous. It seems innocent enough.

I thought: Surely this is just a quirky continuous function. Maybe not smooth everywhere, but nothing too wild.

That’s when the pit opened.

---

Step 1: The False Sense of Security

At first, I remembered that the classical Takagi function (with $\alpha=\tfrac{1}{2}, \beta=2$) is known to be continuous everywhere but differentiable nowhere. That already feels paradoxical: a function that behaves like a “line” with no tangent line anywhere.

But the generalized version adds a parameter dance: $\alpha, \beta$ control how much scaling and shrinking happens. There’s a trichotomy of behavior:

• Subcritical ($\alpha\beta < 1$): Lipschitz continuous. The graph is jagged but still relatively well-behaved.

• Supercritical ($\alpha\beta > 1$): wild — continuous but nowhere differentiable, with difference quotients exploding like a power law.

• Critical ($\alpha\beta = 1$): the cliff edge. What happens right here?

I thought: maybe at the critical point, the function balances. Perhaps it’s differentiable “almost everywhere”? Or maybe it’s “piecewise smooth”?

I was wrong.

---

Step 2: Naive Differentiation (The Dead End)

I tried to formally differentiate:

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

Differentiate term by term? Disaster. Each $\tau(\beta^n x)$ is piecewise linear but has kinks at every rational with denominator $\beta^n$. The kinks accumulate densely. The derivative series diverges almost everywhere.

Okay, maybe consider the difference quotient:

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

If $\alpha\beta < 1$, this stays bounded — so the function is Lipschitz. If $\alpha\beta > 1$, it blows up like a power of $|h|^{-1}$.

But at $\alpha\beta=1$… my estimates collapsed. Sometimes it looked bounded, sometimes unbounded. I couldn’t tell which way it would tip.

Dead end.

---

Step 3: Wrong Guesses and Half-Truths

I guessed: maybe the critical case is differentiable almost everywhere. After all, “critical” suggests borderline behavior.

So I tried small experiments: plotting difference quotients for tiny $h$. Instead of stabilizing, they oscillated. Sometimes they grew, sometimes they shrank. No clear pattern.

Next guess: maybe it’s differentiable at rationals but not irrationals? Or vice versa? But calculations around rationals with denominator powers of $\beta$ gave chaotic blowups.

Another wrong turn.

---

Step 4: The Reveal — Logarithmic Explosion

The actual truth is stranger:

• At the critical threshold ($\alpha\beta=1$), the function is continuous but nowhere differentiable.

• The difference quotient grows like

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

  for some constant $C > 0$.

That means the function doesn’t explode as violently as in the supercritical case, but it still refuses to have a tangent anywhere. The growth is just slower — logarithmic rather than power-law.

It’s an in-between monster: smoother than chaos, rougher than order.

---

Step 5: Fractal Features

Here’s where the pit deepens:

• Self-similarity. The function satisfies a functional equation

  $$T_{\alpha,\beta}(x) = \tau(x) + \alpha T_{\alpha,\beta}(\beta x),$$

  which means zooming in reproduces scaled versions of itself. It’s fractal in nature.

• Combinatorics of digits. Whether the difference quotient explodes at a given scale depends on the base-$\beta$ expansion of $x$. Runs of identical digits correspond to “bad” scales where spikes happen.

• Nowhere differentiability. No matter where you look, the kinks are too dense, the oscillations too sharp. Smoothness never emerges.

Visually, the graph is like a coastline: zoom in and it never straightens out.

---

Step 6: The Abyss Below — Open Questions

Even after proving nowhere differentiability at the critical threshold, the pit doesn’t end. Deeper questions remain:

1. Optimal constants. Exactly how large is the logarithmic growth constant $C$? How does it depend on $\beta$?

2. Fractal dimension. What is the Hausdorff dimension of the graph at criticality? Is it different from subcritical and supercritical regimes?

3. Multifractal spectrum. How are the pointwise Hölder exponents distributed?

4. Randomized analogues. If we randomize coefficients or phases, does the same logarithmic failure persist?

5. Higher dimensions. What happens if we define Takagi-like monsters on $\mathbb{R}^d$?

These are not exercises. They are live research questions.

---

Step 7: Climbing Out

What started as a playful-looking sawtooth series turned out to be an abyss. The function is continuous — so you can draw its graph without lifting your pen — but nowhere differentiable, so you can’t find a tangent line anywhere. At the critical threshold, the failure is delicate, balanced exactly on a logarithmic edge.

I went in expecting a quirky toy. I came out realizing I had stumbled onto a razor’s edge fractal.

That’s the heart of The Problem Pit: the fall from “I can do this” to “this is stranger than I imagined.”

---