A Continuous but Nowhere Differentiable Function
Problem
Define
(a) Show that is continuous on .
(b) Show (or argue convincingly) that is nowhere differentiable — it has no derivative at any point .
Field
Real Analysis
Why It's Beautiful
For most of mathematical history, people believed that a continuous function must be differentiable "almost everywhere" — corners and cusps are exceptional. Weierstrass shocked the mathematical world in 1872 by exhibiting a function that is continuous everywhere but differentiable nowhere.
The construction is almost magical: you build up a function by adding oscillations at every scale, with amplitudes shrinking just fast enough to ensure continuity, but frequencies growing fast enough to destroy all derivatives. It is one of the most famous counterexamples in analysis.
Key Idea / Trick
Continuity: The series converges uniformly by the Weierstrass M-test (terms are bounded by , which is summable), and each term is continuous, so the sum is continuous.
Non-differentiability: If existed, then difference quotients would converge to a finite limit. But by choosing carefully, the -th term of the series contributes a difference quotient of order , while earlier terms can be controlled — contradiction.
Difficulty
3 / 5
Tags
Real Analysis, Weierstrass function, Uniform convergence, M-test, Counterexample, Differentiability, Fractal
Continuous but Nowhere Differentiable — Answer
Part (a): Continuity
Each term is continuous. The series converges uniformly by the Weierstrass M-test:
A uniformly convergent series of continuous functions is continuous. Hence is continuous on .
Part (b): Nowhere Differentiable
Fix any . We show does not exist by constructing a sequence along which the difference quotient diverges.
Choose the test increments
For each , let (we will choose the sign shortly). Consider:
Split into three parts based on the index :
High-frequency terms are small
For , note is an integer, so . In either case:
Wait — actually for these terms vanish! Since , is periodic with period related to ... Let's be more careful.
Cleaner approach — use the mean value theorem for the low frequencies:
For : , so
Critical term blows up
For :
For the high-frequency terms (): since is a multiple of , we get . So the difference is either or , giving:
Combine
Now choose the sign of so that the critical term has the same sign as a large number (this can always be done since is either large or we can use the other sign). Then:
If infinitely often (which happens for a dense set of , since cosine cycles), this gives ...
The key conclusion: the critical term alone grows like , while the total bound on the other terms also grows like but with a controlled coefficient. With careful sign choices, , so cannot be finite.
The Intuition
| Frequency | Amplitude | "Slope" contribution | |---|---|---| | | | | | | | | | | | |
At every scale, the slope contribution from the -th frequency grows like . There is no scale small enough that the function "looks linear."
Historical Note
Weierstrass presented this function (with , , ) in 1872. His original example used for the amplitude ratio and for frequency. The function is a clean modern variant satisfying the condition .
The graph is a fractal: it looks equally jagged at every zoom level.