Karl Weierstrass

Karl Weierstrass (1815–1897) was a German mathematician known as the "father of modern analysis." He rigorously refounded calculus on the precise (ε, δ) definition of limits, discovered a continuous function that is nowhere differentiable, and developed the theory of analytic functions via power series. His work appears throughout real and complex analysis, approximation theory, and the calculus of variations. Active primarily in Berlin from the 1850s through 1890s.

Karl Theodor Wilhelm Weierstrass was a German mathematician whose lifelong project was the arithmetization of analysis — the elimination of vague geometric and intuitive arguments from calculus in favor of rigorous definitions built from inequalities involving real numbers. He gave the modern (ε, δ) formulation of limits and continuity, established the theory of Uniform Convergence, and built complex analysis from the foundation of Power Series. His most famous single result is the construction of a function that is continuous everywhere but differentiable nowhere — a discovery that overturned the widespread belief that continuity implied differentiability except at isolated points. Remarkably, Weierstrass spent his early career as an obscure secondary-school teacher, publishing almost nothing, before a single 1854 paper catapulted him to the front rank of European mathematics at age 39.

The Core Insight

Analysis can — and must — be made rigorous by reducing all statements about limits, continuity, and convergence to precise inequalities between real numbers, banishing appeals to motion, geometry, or infinitesimals. Weierstrass showed both the power of this program (clean theorems) and its necessity (pathological counterexamples that intuition cannot reach).

Formal Statement

Weierstrass is associated with several foundational definitions and theorems. We collect the central ones.

Limit of a Function (ε–δ)

A function $f$ has limit $L$ at a point $a$ , written $lim_{x \to a} f (x) = L$ , if for every $ε > 0$ there exists $δ > 0$ such that

0 < ∣ x - a ∣ < δ ⟹ ∣ f (x) - L ∣ < ε .

⟳This is a definition, not a derived result; the structure "for all ε there exists δ" formalizes the intuitive phrase "f(x) gets arbitrarily close to L as x approaches a."

Here $ε$ (epsilon) bounds the allowed error in the output, $δ$ (delta) bounds the input distance, $a$ is the limit point, and $L$ the candidate limit value.

Continuity

A function $f$ is continuous at $a$ if

\forall ε > 0 \exists δ > 0 : ∣ x - a ∣ < δ \Rightarrow ∣ f (x) - f (a) ∣ < ε .

⟳Specialize Equation [1] to the case L = f(a) and drop the punctured-neighborhood requirement (x = a trivially satisfies the consequent).

The symbols carry the same meaning as in [1]; continuity is the special case where the limit equals the function's actual value.

Weierstrass Approximation Theorem (1885)

Let $f : [a, b] \to R$ be continuous. Then for every $ε > 0$ there exists a polynomial $p$ such that

x \in [a, b] sup ∣ f (x) - p (x) ∣ < ε .

⟳Polynomials are dense in C([a,b]) under the uniform norm; one constructive proof convolves f with the Landau kernel

(1 - t^{2})^{n}

, normalized.

That is, every continuous function on a closed bounded interval is the uniform limit of polynomials.

Bolzano–Weierstrass Theorem

Every bounded sequence $(x_{n})$ in $R^{d}$ has a convergent subsequence:

(x_{n}) bounded ⟹ \exists (x_{n_{k}}) such that x_{n_{k}} \to x^{*} \in R^{d} .

⟳Repeated bisection of an interval containing infinitely many terms produces nested intervals shrinking to a point; the nested interval theorem yields the limit.

Extreme Value Theorem

A continuous function $f$ on a compact set $K \subset R^{d}$ attains its maximum and minimum:

\exists x_{m i n}, x_{m a x} \in K : f (x_{m i n}) \leq f (x) \leq f (x_{m a x}) \forall x \in K .

⟳f(K) is bounded (else a subsequence escapes to infinity, contradicting Bolzano–Weierstrass); the supremum is approached by a sequence whose convergent subsequence limit attains it by continuity.

The Weierstrass Function

Weierstrass's Nowhere-Differentiable Function (1872)

Define

W (x) = n = 0 \sum \infty a^{n} cos (b^{n} π x),

where $0 < a < 1$ , $b$ is an odd integer, and

ab > 1 + \frac{3}{2} π .

⟳The factor a^n < 1 with Σaⁿ < ∞ gives uniform convergence (Weierstrass M-test), hence continuity; condition [7] forces the difference quotients to oscillate unboundedly so no derivative can exist.

Then $W$ is continuous on all of $R$ but differentiable at no point. Here $a$ controls the amplitude decay, $b$ controls the frequency growth, and the product condition $ab > 1 + \frac{3}{2} π$ ensures roughness wins over smoothness.

The Weierstrass function W(x) shown at increasing magnification — self-similar roughness at every scale

Why It Holds

Sketch: Continuity of W(x)

We verify that the series in [6] defines a continuous function.

Each term $f_{n} (x) = a^{n} cos (b^{n} π x)$ is continuous on $R$ .

Bound each term: $∣ a^{n} cos (b^{n} π x) ∣ \leq a^{n} =: M_{n}$ , independent of $x$ .

Since $0 < a < 1$ , the series $\sum_{n = 0}^{\infty} M_{n} = \sum a^{n} = \frac{1}{1 - a} < \infty$ converges.

By the Weierstrass M-test, the series $\sum f_{n}$ converges uniformly on $R$ .

A uniform limit of continuous functions is continuous. Therefore $W$ is continuous everywhere. $■$

Sketch: Non-differentiability of W(x)

The full argument is delicate; here is the strategy.

Fix any point $x_{0}$ . For a chosen integer $m$ , pick a point $x_{m}$ near $x_{0}$ of the form $\frac{α _{m} + 1/2}{b ^{m}}$ (so $b^{m} x_{m}$ lands at a half-integer where cosine behaves predictably).

Split the difference quotient $\frac{W ( x _{m} ) - W ( x _{0} )}{x _{m} - x _{0}}$ into the sum over $n < m$ and $n \geq m$ .

The low-frequency part ( $n < m$ ) is bounded by a geometric-type estimate of size $\leq \frac{π ( ab ) ^{m}}{ab - 1}$ in magnitude (a controlled term).

The high-frequency part ( $n \geq m$ ) contributes a term of magnitude at least $\frac{2}{3} (ab)^{m}$ , growing without bound.

Condition $ab > 1 + \frac{3}{2} π$ guarantees the divergent part dominates the bounded part, so the difference quotient $\to \infty$ in absolute value as $m \to \infty$ . No finite derivative exists. $■$

Karl Weierstrass | 1815–1897

Born in Ostenfelde, Westphalia, Weierstrass was pushed by his father into studying law and finance at Bonn — a subject he loathed, devoting himself instead to beer, fencing, and self-taught mathematics. He left without a degree, then trained as a secondary-school teacher and spent over a decade in remote Prussian towns (Deutsch-Krone, Braunsberg) teaching not only mathematics but gymnastics, botany, geography, and penmanship by day, while researching abelian functions by night. His 1854 paper Zur Theorie der Abelschen Functionen in Crelle's Journal stunned the mathematical world; Königsberg granted him an honorary doctorate, and he soon moved to Berlin, where his lectures became legendary throughout Europe. A famously meticulous, gentle teacher, he influenced a generation including Cantor, Schwarz, Mittag-Leffler, Frobenius, and Sofia Kovalevskaya — whom he mentored and championed when women were barred from universities.

Building Intuition

Picture trying to "draw a tangent line" to the Weierstrass function. At any point you zoom in expecting the curve to flatten into a straight line — that's what differentiability means. But each magnification reveals wiggles of comparable steepness, because the high-frequency terms $cos (b^{n} π x)$ never become negligible: as $a^{n}$ shrinks the amplitude, $b^{n}$ amplifies the slope, and condition $ab > 1$ guarantees slope wins. The curve is self-similar roughness all the way down, a fractal before fractals had a name. The deeper lesson of Weierstrass's whole career is captured here: human geometric intuition about smooth curves is simply wrong about the general continuous function, and only the discipline of inequalities can tell us what is true.

Worked Examples

Verifying continuity with ε–δ

Given:

f (x) = 3 x + 1

. Prove

f

is continuous at

a = 2

using Equation [2]. Steps:

We want: given $ε > 0$ , find $δ$ with $∣ x - 2∣ < δ \Rightarrow ∣ f (x) - f (2) ∣ < ε$ .
Compute $∣ f (x) - f (2) ∣ = ∣ (3 x + 1) - (7) ∣ = ∣3 x - 6∣ = 3∣ x - 2∣$ .
We need $3∣ x - 2∣ < ε$ , i.e. $∣ x - 2∣ < ε /3$ .
Choose $δ = ε /3$ . Then $∣ x - 2∣ < δ$ gives $∣ f (x) - f (2) ∣ = 3∣ x - 2∣ < 3 \cdot \frac{ε}{3} = ε$ . ✓

Answer:

δ = ε /3

works;

f

is continuous at 2. (For

ε = 0.06

, take

δ = 0.02

Checking the roughness condition

Given: A Weierstrass function with

a = 0.5

b = 7

(odd). Does it satisfy [7] and so define a nowhere-differentiable function? Steps:

Compute the product: $ab = 0.5 \times 7 = 3.5$ .
Compute the threshold: $1 + \frac{3}{2} π = 1 + 1.5 \times 3.14159 = 1 + 4.712 = 5.712$ .
Compare: $ab = 3.5$ vs $5.712$ . Since $3.5 < 5.712$ , condition [7] fails.
Try instead $a = 0.9$ , $b = 7$ : $ab = 6.3 > 5.712$ . ✓ And $0 < a < 1$ , $b$ odd. ✓

Answer:

a = 0.5, b = 7 fails Weierstrass’s sufficient condition; a = 0.9, b = 7 succeeds.

(Note: later work by Hardy weakened [7] to merely

ab \geq 1

Common Misconceptions

"A continuous function is differentiable except possibly at a few corner points." This was the prevailing 19th-century belief (Ampère even attempted to "prove" it). Weierstrass's function [6] is a permanent refutation: continuity is a strictly weaker condition than differentiability, and the "typical" continuous function is in fact nowhere differentiable — in the sense of Baire category and Wiener measure, almost all continuous functions are nowhere differentiable. "Uniform convergence is the same as pointwise convergence." It is not, and the distinction is central to all of Weierstrass's foundations. Pointwise convergence only requires the speed of convergence to be finite at each point; uniform convergence (the M-test in step 4 of the proof above) requires a single

N

to work for all points at once. Only uniform convergence transmits continuity, integrability, and (with care) differentiability from terms to limit.

The sequence $f_{n} (x) = x^{n}$ on $[0, 1]$ converges pointwise to the discontinuous function that is $0$ on $[0, 1)$ and $1$ at $x = 1$ . Each $f_{n}$ is continuous, yet the pointwise limit is not — precisely because the convergence is not uniform near $x = 1$ . This shows why Weierstrass insisted on uniform convergence: without it, the continuity of the terms tells you nothing about the limit.

Connections

Weierstrass's program touches nearly all of modern analysis. The ε–δ definitions [1], [2] underlie Real Analysis and the rigorous theory of Limits and Continuity. The M-test and uniform convergence connect to Series of Functions and Fourier Series. His power-series approach to complex functions is the foundation of Complex Analysis alongside the contour-integral approach of Cauchy and the geometric approach of Riemann. The Bolzano–Weierstrass theorem [4] is the sequential face of Compactness, and the Extreme Value Theorem [5] is a cornerstone of Optimization. His rigor in the Calculus of Variations (the Weierstrass E-function and sufficient conditions) closed gaps left by Euler and Lagrange.

Real-World Applications

Weierstrass's results, though foundational rather than applied, enable enormous practical machinery:

Numerical analysis & CAD — the Approximation Theorem [3] guarantees polynomials (and the constructive Bernstein polynomials) can approximate any continuous shape, underlying Bézier curves used in fonts and computer-aided design.

Signal processing — uniform convergence justifies term-by-term manipulation of Fourier and wavelet expansions used in audio/image compression (JPEG, MP3).

Fractal modeling — the Weierstrass function and its relatives (Weierstrass–Mandelbrot function) model rough natural phenomena: coastlines, fracture surfaces, and turbulent profiles.

Financial mathematics — Brownian-motion paths are, like $W (x)$ , continuous but nowhere differentiable; Weierstrass's example was the historical precursor to the rigorous treatment of stochastic processes underlying the Black–Scholes model.

Optimization & machine learning — the Extreme Value Theorem [5] guarantees that continuous loss functions attain minima on compact parameter sets, justifying the existence of optimal solutions.

Antenna and rough-surface scattering — fractal Weierstrass profiles model electromagnetic scattering from natural terrain.

Approximation in engineering control — the Stone–Weierstrass generalization underpins universal approximation results for neural networks.

Historical Development

1815 — Karl Weierstrass born 31 October in Ostenfelde, Westphalia (Kingdom of Prussia).
1834–1838 — Studies law/finance at Bonn against his will; leaves without a degree.
1841–1854 — Works as a schoolteacher in Deutsch-Krone and Braunsberg, researching abelian functions in isolation; develops much of his analysis privately.
1854 — Publishes Zur Theorie der Abelschen Functionen in Crelle's Journal; Königsberg awards an honorary doctorate. Sudden fame.
1856 — Appointed to the Industry Institute and then the University of Berlin; becomes a leading figure of the "Berlin school."
1861 — Construction of an early continuous nowhere-differentiable function (he had examples before public presentation).
1872 — Presents his nowhere-differentiable function [6] to the Berlin Academy (published by du Bois-Reymond in 1875), shocking the mathematical community.
1870s — Mentors Sofia Kovalevskaya, securing her path to a doctorate (1874) despite institutional exclusion of women.
1885 — Proves the Weierstrass Approximation Theorem [3].
1897 — Dies 19 February in Berlin of pneumonia, having shaped the foundations of analysis for the entire 20th century.

Subtleties & Limits

Weierstrass gave the sufficient condition $ab > 1 + \frac{3}{2} π$ for non-differentiability; G. H. Hardy (1916) sharpened this to the essentially optimal $ab \geq 1$ (with $a < 1$ ). Deeper questions about these functions remain active: the exact Hausdorff dimension of the graph of the classical Weierstrass function was only rigorously determined to equal $2 + lo g_{b} a$ (the "box-counting" value) in full generality by Shen (2018), after decades of partial results. More broadly, the precise multifractal and local-regularity structure of Weierstrass-type and related fractal functions — and their connections to dynamical systems and number theory — continues to generate research.