Kolmogorov’s Pathological Function and the Limits of Fourier Series Convergence

“One could just as well provide Kolmogorov’s function.”

Kolmogorov is one of the greatest minds of the 20th century: he built the axiomatic foundation of probability theory, laid the groundwork for the theory of Markov processes, advanced chaos theory, and built one of the largest mathematical schools in history. But he burst onto the mathematical stage as a student — with an elegant and audacious “pathology” that exposed the limits of the most celebrated tool in signal analysis: the Fourier series.

What Fourier Series Convergence Is Supposed to Look Like

The idea seems straightforward: take a function f, compute its harmonics, assemble its “orchestra” of sines and cosines, gradually add more and more harmonics, and expect the partial sums to converge ever more precisely to the original curve. In plain terms, the more pure tones you mix together, the closer the result should sound to the original. In engineering, Fourier series appear wherever signals exist — from smart speakers to MRI machines to quantum mechanics equations in frequency space.

The Surprise from a 19-Year-Old: Breaking Fourier Series Convergence

In 1922, Kolmogorov constructed a perfectly “well-behaved” function. It has finite area under the curve over one period — that is, it is integrable in the L¹ sense. But when you faithfully apply the Fourier method and keep adding harmonics, the partial sums fail to converge to the original function at almost every point.

In 1926, he struck harder: he found a function from the same class whose Fourier expansion diverges at every single point.

What does such a function look like? It is not a monster bristling with infinite spikes — it is more of an “uneven” function, with bursts and sharp transitions. Just enough irregularity to cause the summation to produce large dips and overshoots.

What Exactly Did Kolmogorov Break?

He shattered the expectation that integrability alone is sufficient for stable reconstruction from harmonics — in other words, for guaranteed convergence of Fourier series partial sums to the original function.

The problem lies in how standard summation works: through what is called the Dirichlet kernel. Do not be alarmed — it is simply a set of weights used to blend the harmonics, a kind of “room acoustics” for the summation. And this acoustics has an unpleasant property: its total “volume” grows as log N with the number of harmonics N. The growth is slow but relentless. Eventually, even small oscillations in the original function begin to be amplified many times over in the sum. Instead of an increasingly accurate copy, you get a trembling sawtooth with amplified ripples.

Why This Matters Beyond Mathematics

The Fourier series is a universal tool. In the real world, we often want to decompose a complex signal into understandable components and then reassemble it. Kolmogorov showed that this universality does not hold everywhere. There exist data sets where “honest” frequency summation, without additional smoothness, is destined to produce instability. The implication is clear: either demand more regularity from the data, or change the method of summation.

How Fourier Series Convergence Was Restored

As early as 1915, Luzin had intuited the right direction: one should work not with L¹ functions (finite area under the curve) but with L² functions (finite energy — meaning the square is integrable). In 1966, Lennart Carleson proved that for f ∈ L², the partial sums do indeed converge to f almost everywhere. In 1968, Richard Hunt extended this result to all L^p spaces with p > 1.

If a function has finite “energy” (its square is integrable), the harmonics reassemble into the original. If only the integral of the absolute value is finite — there are no guarantees.

The counterexample forced a reconstruction of technique. More flexible summation methods emerged (Fejér, Pólya), maximal operators were developed, and the whole edifice of modern harmonic analysis was built — with its gentler “lenses” that smooth the summation process and restore stability. Paradoxically, it was precisely the counterexample that revealed how to make Fourier methods reliable.

The Line Kolmogorov Drew: A Lesson for Signal Processing

Kolmogorov drew a red line: “integrable” does not mean “recoverable from harmonics.” Carleson completed the picture: “finite energy — passage permitted.” For mathematics, this clarified which infinite sums can be assembled “honestly” and where a different ordering or smoothing is required. For the practical world, it explains why signal and data processing almost always applies windowing, apodization, and averaging — so that frequency methods are not only mathematically beautiful, but stable in practice.