(Robinson, 2014)
TABLE OF CONTENTS
1. Introduction
3. The Distributional Derivative: A Formal Trick
4. Pathologies and the Limits of Generalization
5. Constructivist and Finitist Concerns
6. Conclusion
The essay below is being published in two installments; this one, the second and final installment, contains sections 4-6.
But you can also download and read or share a .pdf of the complete text of this essay by scrolling down to the bottom of this post and clicking on the Download tab.
4. Pathologies and the Limits of Generalization
To clarify the critique of distribution theory, we define a function f: R → R (or on a subset of R) as pathological if it exhibits behavior that challenges classical differentiation, particularly in the context of distributional derivatives. A function is pathological if it satisfies one or more of the following properties.
Nowhere Differentiable: The function is continuous but not differentiable at any point due to rapid oscillations or fractal-like behavior. Example: the Weierstrass function, W(x) = Σ (a^n * cos(b^n * π * x)), n from 0 to ∞, with 0 < a < 1, b a positive odd integer, and a * b > 1 + (3/2) * π, is continuous but nowhere differentiable due to high-frequency oscillations.
Non-Locally Integrable: The function is so erratic that it is not locally integrable, preventing it from defining a distribution. Example: The Dirichlet function, χ_Q(x) = 1 if x is rational and 0 otherwise, is discontinuous everywhere and not locally integrable, as it oscillates too wildly to have a well-defined integral over any interval.
Singular with Non-Trivial Measure Properties: The function is continuous but has singular behavior, such as a derivative that is zero almost everywhere yet increases, or is defined on a set with zero measure but uncountable points. Example: The Cantor function, c: [0,1] → [0,1], is continuous, non-decreasing, and constant on intervals removed during the Cantor set construction, with derivative zero almost everywhere, yet increases from 0 to 1.
Defined on Fractal or Irregular Domains: The function is defined on a fractal or self-similar set, leading to irregular behavior. Example: The Cantor function, tied to the Cantor set (uncountable, zero measure, nowhere dense), exhibits fractal-like properties that defy classical differentiation.
Infinitely Spiky or Oscillatory: The function has increasingly frequent spikes or oscillations, making classical derivatives undefined or meaningless. Example: A function f(x) = Σ (1/n^2) * exp(-n^2 * (x – q_n)^2), summed over n, with q_n dense points (e.g., rationals), has a distributional derivative that is a sum of delta functions at dense points, which is physically uninterpretable.
These properties highlight why pathological functions challenge distribution theory. Their distributional derivatives, when they exist, are abstract and lack pointwise meaning, while some functions (e.g., Dirichlet) cannot even define distributions, exposing the theory’s reliance on local integrability.
Let’s look at these mathematical pathologies in more detail.
1. Nowhere-Differentiable Functions
Here are some examples and classes of functions that throw a wrench in the optimism of distribution theory.
Weierstrass function: Continuous everywhere, differentiable nowhere. In distribution theory, it does have a generalized derivative—but this derivative is no longer a function in the classical sense. It’s a distribution, which may correspond to wild oscillations with no pointwise meaning.
The graph of such a “derivative” cannot be plotted in the usual sense. So what is this object, really? Let’s look at this “function” in more detail to illustrate the limitations of distributional derivatives.
The Weierstrass function is a pathological function that is continuous everywhere but differentiable nowhere. Define W(x) = Σ (a^n * cos(b^n * π * x)), summed from n = 0 to infinity, where 0 < a < 1, b is a positive odd integer, and a * b > 1 + (3/2) * π. This ensures W(x) is nowhere differentiable due to its infinite sum of high-frequency oscillations.
In distribution theory, the derivative of W(x) is defined as a distribution acting on a smooth test function φ. The distributional derivative is W'[φ] = -∫ (W(x) * φ'(x)) dx, where the integral is over the real line. Since W(x) is continuous, this integral is well-defined, and W’ is a tempered distribution. However, W’ is not a function; it cannot be evaluated at any point and reflects the rapid oscillations of W(x) due to terms like b^n * π. This lacks the classical meaning of a derivative as a local rate of change.
Now, apply W(x) to a physical problem: the one-dimensional wave equation, u_tt = c^2 * u_xx, where u_tt is the second time derivative and u_xx is the second spatial derivative. Set initial conditions u(x,0) = W(x) and u_t(x,0) = 0 (zero initial velocity). The solution is u(x,t) = (1/2) * [W(x + ct) + W(x – ct)], and its time derivative is u_t(x,t) = (1/2) * [W'(x + ct) – W'(x – ct)]. Both exist as distributions, but u(x,t) is nowhere differentiable in x for all t, and u_t(x,t) is a distribution with wild oscillations. This makes it impossible to interpret as a physical wave, such as the displacement of a vibrating string, as it lacks measurable local behavior.
Contrast this with a classical smooth case. If u(x,0) = sin(x), the solution is u(x,t) = (1/2) * [sin(x + ct) + sin(x – ct)], which is smooth and represents a propagating wave with measurable velocity u_t(x,t) = (c/2) * [cos(x + ct) – cos(x – ct)]. This is physically interpretable, unlike the Weierstrass case.
Classical numerical methods, like finite differences, fail for W(x). The difference quotient [W(x + h) – W(x)] / h oscillates uncontrollably as h approaches 0 due to high-frequency terms, and no limit exists. Distribution theory avoids this by redefining the derivative as a functional, but this comes at the cost of losing physical interpretability. Thus, while the distributional derivative of W(x) exists, it does not resolve the problem of differentiating pathological functions; it merely shifts the definition to an abstract framework.
2. Infinitely Spiky Functions
Imagine a function that’s continuous but has spikes of increasing frequency and decreasing width, for example, a sum of narrow Gaussians or triangles with divergent second moments.
These functions can be made smooth yet so erratic that their classical derivatives diverge. In such cases, the “distributional derivative” may exist but might be almost meaningless physically, or it might diverge to infinity on dense subsets.
Infinitely spiky functions are pathological, combining continuity with infinite, sharp peaks of increasing frequency and decreasing width. Consider f(x) = Σ (1/n^2) * exp(-n^2 * (x – q_n)^2), n from 1 to ∞, where q_n are rationals in [0,1]. Each term creates a spike at q_n with height 1/n^2 and width ~1/n, making f continuous but erratic due to dense spikes.
Classically, f(x) has no derivative, as finite differences [f(x + h) – f(x)]/h oscillate wildly near each q_n. In distribution theory, the derivative is f'[φ] = -∫ f(x) * φ'(x) dx, or formally f'(x) = Σ (-2 * (x – q_n) * exp(-n^2 * (x – q_n)^2)). This distribution, a sum of delta-like terms at dense points, lacks pointwise meaning.
Apply f(x) to the wave equation u_tt = c^2 * u_xx, with u(x,0) = f(x), u_t(x,0) = 0. The solution u(x,t) = (1/2) * [f(x + ct) + f(x – ct)] has velocity u_t(x,t) = (1/2) * [f'(x + ct) – f'(x – ct)], a distribution with dense spikes, uninterpretable as physical motion. Contrast with u(x,0) = sin(x), where u_t(x,t) = (c/2) * [cos(x + ct) – cos(x – ct)] is measurable.
The distributional derivative exists but isn’t a rate of change, supporting the claim that distribution theory redefines differentiation abstractly. Finitists object to its non-constructive reliance on infinite sums, questioning its philosophical validity. This shows distribution theory sidesteps the challenge of spiky functions, formalizing an algebraic construct rather than resolving their pathological nature.
3. Functions with Dense Discontinuities
Take the indicator function of the rationals (Dirichlet function). It’s discontinuous everywhere. This function is not even integrable, let alone differentiable. It has no distributional derivative, because it’s not a locally integrable function, so it can’t define a distribution in the standard sense.
The Dirichlet function, χ_Q(x) = 1 (x rational), 0 (x irrational), is listed as pathological because it is not locally integrable, preventing it from defining a distribution. In the Lebesgue sense, χ_Q(x) is not integrable on any interval [a, b] due to its everywhere discontinuous nature, as it oscillates between 0 and 1 on dense sets of rationals and irrationals. This means ∫_K χ_Q(x) dx is undefined for any compact set K, so χ_Q(x) cannot define a distribution via T[φ] = ∫ χ_Q(x) φ(x) dx, where φ is a test function.
Some modern integration theories, like the Henstock-Kurzweil integral, can assign an integral of 0 to χ_Q(x) over [0, 1], as the rationals have measure zero.
However, this integrability is context-dependent and does not imply local integrability in the sense required for distributions, which relies on Lebesgue integration for consistency across all test functions. Thus, the Dirichlet function remains outside the scope of distribution theory, reinforcing its pathological nature and the limitations of distributional derivatives.
4. Lacunary Functions and Fractal Structures
Functions defined on fractal domains or exhibiting “holes” in their domain (like the Cantor function) often have pathological behavior. The Cantor function has derivative zero almost everywhere, but increases from 0 to 1—this undermines the classical intuition of derivatives as “rate of change.”
The Cantor set is a compact, perfect, uncountable subset of [0,1] with zero Lebesgue measure. It is constructed by iteratively removing the open middle third of intervals:
Start with [0,1].
Remove (1/3, 2/3), leaving [0,1/3] ∪ [2/3,1].
Repeat for each remaining interval, removing middle thirds ad infinitum.
The Cantor set is the limit of this process.
Properties:
Self-similarity: A fractal with self-similar structure.
Zero measure: Uncountable yet has Lebesgue measure zero.
Nowhere dense: No interior points, not dense in any interval.
Perfect set: Every point is a limit point, with no isolated points.
Pathological Nature: The Cantor set defies intuition by being uncountable yet measure-zero, unlike the real numbers.
Cantor Function Definition: The Cantor function c: [0,1] → [0,1] is a continuous, non-decreasing function defined as follows:
For x in the Cantor set, express x in base-3 (ternary) with digits 0 or 2 (no 1s).
Replace 2s with 1s and interpret as a base-2 number.
On intervals removed during Cantor set construction, the function is constant.
Properties:
Continuity: Continuous everywhere.
Monotonicity: Non-decreasing.
Zero derivative almost everywhere: Constant on removed intervals (total measure 1), so the derivative is zero almost everywhere.
Singular function: Continuous but not absolutely continuous; its derivative does not integrate to its increment.
Pathological Nature: Continuous and non-decreasing yet has zero derivative almost everywhere, contradicting expectations of increasing functions.
5. Constructivist and Finitist Concerns
From a constructivist or finitist standpoint, this entire enterprise is suspect. The notion that an object with no tangible local behavior can be said to have a derivative at all strains credibility. For those who take seriously the requirement that mathematics describe computable, observable, or at least finitely representable phenomena, the distributional derivative is an abstraction too far. It does not describe a process or transformation of real quantities; it defines an algebraic shadow, often built on infinite or non-constructive assumptions.
The Physical Disconnect
Physicists often accept distributional derivatives as tools because they are operationally useful. But their physical meaning is nebulous at best. When the Dirac delta is interpreted as the “derivative” of a step function, what exactly is changing? Where is the motion, the transition, the force? These concepts evaporate when reduced to functionals on test spaces. The promise of continuity between mathematics and the physical world—a principle historically central to calculus—is lost.
6. Conclusion
The theory of distributions, while elegant and powerful in certain domains, does not ultimately resolve the philosophical or interpretive difficulties posed by pathological functions. It sidesteps them. By redefining differentiation in a way that no longer demands any local structure, it evacuates the concept of its original meaning. For finitists, constructivists, and skeptical physicists and mathematicians, the derivative of a nowhere-differentiable function is not a triumph of modern analysis, but a cautionary tale in the seduction of formalism. The claim that such functions “have” derivatives must be understood not as a clarification, but as a profound shift in what we mean by mathematical truth. If differentiation no longer tracks change, and if mathematical objects no longer model anything observable, then we are not saving calculus—we are embalming it. The derivative, like the Emperor’s new clothes, is admired only because no one dares to ask what it is.
REFERENCES
(Halperin and Schwartz, 1952). Halperin, I. and Schwartz, L. Introduction to the Theory of Distributions. Toronto CA: Univ. of Toronto Press.
(Robinson, 2014). Robinson, M. “Math 601: Distribution Theory.” YouTube. Available online at URL = <https://www.youtube.com/watch?v=jBI5RM3xc7k>.
(von Neumann, 1955). Von Neumann, J. Mathematical Foundations of Quantum Mechanics. Princeton NJ: Princeton University Press.
Against Professional Philosophy is a sub-project of the online mega-project Philosophy Without Borders, which is home-based on Patreon here.
Please consider becoming a patron!
