Functional Equations and Combinatorial Constructions of Pathological Functions in Real Analysis: Diving into Depths

 

Functional Equations and Combinatorial Constructions of Pathological Functions in Real Analysis

By Arya Dubey                                                                                                                                     Check out the full paper

---

Abstract

The existence of continuous nowhere differentiable (CND) functions remains one of the most striking and counterintuitive phenomena in real analysis. These objects are continuous everywhere but possess no derivative at any point. Once regarded as pathological monsters, they now arise naturally in probability, dynamics, fractal geometry, and harmonic analysis.

This blog post presents two elementary approaches to constructing such functions:

1. A functional-equation method generalizing Takagi’s sawtooth construction, leading to a sharp phase-transition criterion.

2. A combinatorial binary-expansion method using digit manipulations to generate analytic irregularity.

Along the way, we establish results on Hölder regularity for Weierstrass-type series and revisit Hardy’s classical theorem, but framed with elementary tools. This post follows the structure of my paper, while adding exposition, intuition, and commentary.

---

1. Introduction

In 1872, Karl Weierstrass astonished mathematicians by constructing an explicit continuous function that is nowhere differentiable. Before that, analysts assumed continuity usually implied differentiability, or at least differentiability almost everywhere.

Weierstrass’s example shattered this intuition. His function was continuous everywhere, but wild and oscillatory at every scale, admitting no tangent line at any point. These were dubbed “pathological functions,” as they seemed to break the rules.

Over time, they turned out not to be monsters at all:

• They model Brownian motion paths in probability.

• They appear as boundary values of holomorphic functions in complex analysis.

• They arise in dynamical systems as attractors.

• They are examples of fractals with fractional Hausdorff dimension.

This post is devoted to constructing such functions using elementary tools. We will see that pathology often emerges from simple mechanisms like self-similarity and digit expansions.

---

2. Preliminaries

Before constructing new examples, let’s set up definitions and recall the classics.

Definition (Nowhere Differentiable Function).
A function $f : \mathbb{R} \to \mathbb{R}$ is nowhere differentiable if:

1. $f$ is continuous on $\mathbb{R}$, and

2. the derivative $f'(x)$ fails to exist (as a finite real number) for every $x \in \mathbb{R}$.

Two classical constructions:

1. Weierstrass function (1872):

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

with $0 < a < 1$, odd $b \geq 3$, and $ab > 1 + \tfrac{3}{2}\pi$.

2. Takagi function (1903):

$$T(x) = \sum_{n=0}^\infty 2^{-n} \, \tau(2^n x), \quad \tau(x) = \text{dist}(x,\mathbb{Z}).$$

Both are continuous everywhere and nowhere differentiable.

---

3. Functional Equation Approach via Fixed Points

We first adopt a fixed-point framework.

3.1 The Operator Setup

Let $C$ be the Banach space of bounded, continuous, 1-periodic functions on $\mathbb{R}$ with the sup norm.

Given $g \in C$, $0 < \alpha < 1$, and integer $\beta \geq 2$, define the operator:

$$(TF)(x) = g(x) + \alpha F(\beta x).$$

---

Lemma 3.1 (Fixed-Point Existence).
The operator $T$ is a contraction with constant $\alpha$. Thus it admits a unique fixed point

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

with uniform convergence.

Proof.

• The contraction property:

$$\|TF - TH\|_\infty = \alpha \|F-H\|_\infty.$$

• By Banach’s fixed-point theorem, a unique solution exists.

• Iterating the functional equation yields the series. Uniform convergence follows since $\|g\|_\infty < \infty$ and $\sum \alpha^n < \infty$.

---

3.2 Generalized Takagi Functions

Define the generalized Takagi function:

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

where $\tau(x) = \text{dist}(x,\mathbb{Z})$.

---

Theorem 3.2 (Sharp Dichotomy).
For integer $\beta \geq 2$, $0 < \alpha < 1$:

1. If $\alpha\beta < 1$, then $T_{\alpha,\beta}$ is Lipschitz continuous with constant $(1-\alpha\beta)^{-1}$.

2. If $\alpha\beta \geq 1$, then $T_{\alpha,\beta}$ is nowhere differentiable.

Proof idea.

• Case $\alpha\beta < 1$: use the Lipschitz property of $\tau$ to bound differences.

• Case $\alpha\beta \geq 1$: analyze the difference quotient:

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

Using the functional equation,

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

Iterating shows sequences $h \to 0$ exist where $Q(x,h) \to \infty$. Thus no derivative exists.

---

Remark 3.3.
The boundary case $\alpha\beta = 1$ is a genuine phase transition: below it, the function is smooth enough to be Lipschitz; at or above it, self-similarity produces unbounded oscillations.

---

4. Hölder Regularity of Weierstrass-Type Series

Consider trigonometric series of the form:

$$W_{a,b}(x) = \sum_{n=0}^\infty a^n \cos(2\pi b^n x), \quad 0 < a < 1, \, b \geq 2.$$

Proposition 4.1.
$W_{a,b}$ is Hölder continuous with exponent

$$\gamma = -\frac{\log a}{\log b}.$$

Proof sketch.
Split the series at scale $m$ with $b^{-(m+1)} < |h| \leq b^{-m}$.

• For $n \leq m$: estimate by $ab < 1$.

• For $n > m$: tail sum bounded by $O(|h|^\gamma)$.

Thus $|W_{a,b}(x+h)-W_{a,b}(x)| \lesssim |h|^\gamma$. Optimality follows from oscillations.

---

Theorem 4.2 (Hardy, 1916).
If $ab \geq 1$, then $W_{a,b}$ is nowhere differentiable.

Hardy’s proof uses Fourier analysis, but the above gives intuition: beyond the threshold, oscillations dominate.

---

5. Combinatorial Construction via Binary Expansions

We now turn to a digit-expansion method.

5.1 Setup

Every $x \in [0,1]$ has a binary expansion:

$$x = \sum_{n=1}^\infty d_n(x) 2^{-n}, \quad d_n(x) \in \{0,1\}.$$

For dyadic rationals, pick the terminating expansion.

---

5.2 Digit-Weighted Functions

For $p > 1$, define:

$$f_p(x) = \sum_{n=1}^\infty \frac{d_n(x)}{n^p}.$$

Proposition 5.3.
For $p > 1$, $f_p$ converges uniformly and is continuous.

Proof.
Bounded by $\sum n^{-p} < \infty$. Uniform convergence follows by the Weierstrass M-test.

---

5.3 Key Lemma

Lemma 5.4 (Isolated Digit Flip).
For any $x$ and integer $m \geq 1$, there exists $y = x+h$ with $0<|h|\leq 2^{-m}$ such that:

• $d_n(y) = d_n(x)$ for $1 \leq n < m$.

• $d_m(y) = 1 - d_m(x)$.

This shows we can flip a single binary digit while keeping earlier ones unchanged.

---

5.4 Nowhere Differentiability

Theorem 5.5.
For every $p > 1$, the function $f_p$ is nowhere differentiable.

Proof.
Take $x$ and construct $y_m$ as in Lemma 5.4. Then:

$$f_p(y_m) - f_p(x) = \frac{\pm 1}{m^p} + \text{(small tail)}.$$

Since $|h_m| \leq 2^{-m}$,

$$\Bigg|\frac{f_p(y_m)-f_p(x)}{h_m}\Bigg| \gtrsim \frac{2^m}{m^p}.$$

As $m \to \infty$, this diverges. Thus no derivative exists anywhere.

---

Remark 5.6.
This construction highlights how symbolic dynamics (digit manipulations) can produce analytic irregularity.

---

6. Comparative Analysis and Extensions

Comparison of methods:

• Functional equation approach: emphasizes self-similarity and scaling.

• Combinatorial approach: emphasizes number-theoretic structure of binary expansions.

Both give elementary alternatives to Fourier analysis.

Further directions:

1. Higher dimensions ($\mathbb{R}^n \to \mathbb{R}$).

2. Fractal geometry (Hausdorff dimensions of graphs).

3. Probabilistic variants (random coefficients).

4. Applications to stochastic processes and dynamics.

---

7. Conclusion

We have developed two rigorous, elementary approaches to constructing continuous nowhere differentiable functions:

• A sharp dichotomy for generalized Takagi functions ($\alpha\beta < 1$ vs. $\alpha\beta \geq 1$).

• A digit-expansion construction with full proof of nowhere differentiability.

These illustrate that pathological behavior often arises from simple, transparent mechanisms: self-similarity and digit complexity.

Far from being exotic monsters, CND functions are now understood as natural and central objects, bridging analysis, probability, geometry, and dynamics.

---