Distributions and Generalized Functions — Schwartz Theory

⚡ Key Takeaways

Distributions (generalized functions) extend the notion of function so that objects like the Dirac delta, derivatives of discontinuous functions, and divergent integrals acquire rigorous meaning. Key result: every locally integrable function and every "infinitely singular" object can be represented as a continuous linear functional on a space of test functions, and every distribution is infinitely differentiable. Where it applies: PDEs, Fourier analysis, quantum field theory, signal processing, spectral theory. Who/when: Laurent Schwartz, Théorie des distributions (1945–1950).

💡 The Core Insight

Replace the pointwise object $f(x)$ by the action $\varphi \mapsto \int f\varphi$ . Since test functions $\varphi$ are infinitely smooth, any "bad behavior" of $f$ — jumps, spikes, growth — is absorbed by the smoothness of $\varphi$ after integration by parts. Differentiation is defined by pushing derivatives onto the test function, and because $\varphi$ tolerates all derivatives, so does every distribution. The price: distributions cannot in general be multiplied, and they have no pointwise values.

DefinitionTest Function Spaces

Let $\Omega \subseteq \mathbb{R}^n$ be open.

$\mathcal{D}(\Omega) = C_c^\infty(\Omega)$ : smooth functions with compact support in $\Omega$ .
$\mathcal{S}(\mathbb{R}^n)$ (Schwartz space): smooth functions of rapid decrease, with all derivatives decaying faster than any polynomial.
$\mathcal{E}(\Omega) = C^\infty(\Omega)$ : all smooth functions.

The Schwartz space is normed by the countable family of seminorms

1

\quad p_{\alpha,\beta}(\varphi) = \sup_{x\in\mathbb{R}^n}\,\big| x^\alpha\, \partial^\beta \varphi(x)\big| < \infty,\qquad \alpha,\beta \in \mathbb{N}^n.

Derivation hint: rapid decrease means $x^\alpha\partial^\beta\varphi \to 0$ ; boundedness of the sup over all multi-indices encodes "faster than any polynomial." Convergence $\varphi_k \to \varphi$ in $\mathcal{D}(\Omega)$ means: there is a fixed compact $K$ containing all $\mathrm{supp}\,\varphi_k$ , and $\partial^\alpha\varphi_k \to \partial^\alpha\varphi$ uniformly for every $\alpha$ .

DefinitionDistributions

A distribution $T \in \mathcal{D}'(\Omega)$ is a linear functional $T : \mathcal{D}(\Omega) \to \mathbb{C}$ that is continuous: $\varphi_k \to \varphi$ in $\mathcal{D}(\Omega)$ implies $\langle T,\varphi_k\rangle \to \langle T,\varphi\rangle$ . Equivalently, for each compact $K \subset \Omega$ there exist $C>0$ , $m\in\mathbb{N}$ with

2

\quad |\langle T,\varphi\rangle| \le C \sum_{|\alpha|\le m} \sup_{x\in K}\,|\partial^\alpha \varphi(x)|,\qquad \forall\,\varphi \text{ with } \mathrm{supp}\,\varphi \subseteq K.

Derivation hint: continuity on a Fréchet/LF space is equivalent to a seminorm estimate; the integer $m$ (smallest uniform choice) is the local <strong>order</strong> of $T$ . The dual of $\mathcal{S}$ is the space of tempered distributions $\mathcal{S}'(\mathbb{R}^n)$ — exactly those distributions that admit a Fourier transform.

TheoremRegular Distributions and the Embedding

Every $f \in L^1_{\mathrm{loc}}(\Omega)$ defines a distribution $T_f$ via

3

\quad \langle T_f,\varphi\rangle = \int_\Omega f(x)\,\varphi(x)\,dx,

and the map $f \mapsto T_f$ is injective (modulo a.e. equality). Thus $L^1_{\mathrm{loc}} \hookrightarrow \mathcal{D}'$ , and one writes $f$ for $T_f$ . Derivation hint: injectivity follows from the fundamental lemma of calculus of variations — if $\int f\varphi = 0$ for all $\varphi\in\mathcal{D}$ then $f=0$ a.e., proved by mollification.

DefinitionDistributional Derivative

For any $T\in\mathcal{D}'(\Omega)$ and multi-index $\alpha$ , the derivative $\partial^\alpha T$ is the distribution

4

\quad \langle \partial^\alpha T,\varphi\rangle = (-1)^{|\alpha|}\,\langle T,\partial^\alpha\varphi\rangle,\qquad \varphi\in\mathcal{D}(\Omega).

Derivation hint: integrate by parts $|\alpha|$ times against smooth $f$ with compact support; all boundary terms vanish because $\varphi$ has compact support, leaving the sign $(-1)^{|\alpha|}$ . Because $\partial^\alpha\varphi\in\mathcal{D}$ , the right side is well-defined for all $T$ and all $\alpha$ : every distribution is infinitely differentiable.

ProofDifferentiation is well-defined, linear, and continuous

Step 1 — Linearity. For $\varphi,\psi\in\mathcal{D}$ and scalars $a,b$ , $\partial^\alpha(a\varphi+b\psi)=a\,\partial^\alpha\varphi+b\,\partial^\alpha\psi$ , so $\langle\partial^\alpha T, a\varphi+b\psi\rangle = (-1)^{|\alpha|}\langle T, a\partial^\alpha\varphi+b\partial^\alpha\psi\rangle$ is linear in $\varphi$ .

Step 2 — The functional is continuous. If $\varphi_k\to 0$ in $\mathcal{D}$ , then $\partial^\alpha\varphi_k\to 0$ in $\mathcal{D}$ (same compact support, uniform convergence of all derivatives). By continuity of $T$ , $\langle T,\partial^\alpha\varphi_k\rangle\to 0$ , hence $\langle\partial^\alpha T,\varphi_k\rangle\to 0$ . So $\partial^\alpha T\in\mathcal{D}'$ .

Step 3 — Consistency with classical derivatives. If $f\in C^1$ , integration by parts gives

\int (\partial_j f)\varphi\,dx = -\int f\,(\partial_j\varphi)\,dx,

with no boundary term since $\mathrm{supp}\,\varphi$ is compact. Thus $\partial_j T_f = T_{\partial_j f}$ : the distributional derivative coincides with the classical one when the latter exists.

Step 4 — Continuity in $T$ and commutativity. The map $T\mapsto\partial^\alpha T$ is sequentially continuous on $\mathcal{D}'$ (weak-* topology), and mixed partials commute, $\partial_i\partial_j T = \partial_j\partial_i T$ , because they commute on $\varphi$ . $\blacksquare$

∎

PD

Paul Dirac

1902–1984

Dirac introduced the "delta function" $\delta(x)$ in his 1930 Principles of Quantum Mechanics as an object satisfying $\int\delta(x)f(x)\,dx=f(0)$ and $\delta(x)=0$ for $x\neq 0$ — a contradiction for ordinary functions. Physicists computed with it freely and successfully (continuous spectra, completeness relations, Green's functions). Schwartz's later achievement was to show this and far more is mathematically rigorous: $\delta$ is precisely the evaluation functional $\langle\delta,\varphi\rangle=\varphi(0)$ , a distribution of order $0$ . Dirac's formal calculus thereby became a theorem-rich theory.

🧠 Intuition

Think of a distribution as a measuring device rather than a curve. You never "see" $\delta$ ; you only see what it reports when you feed it a smooth test signal $\varphi$ — namely $\varphi(0)$ . The genius of test functions is that they are infinitely smooth, so any roughness in your object is laundered through integration by parts onto $\varphi$ , which can absorb arbitrarily many derivatives. This is why differentiation never fails: differentiating the device just means handing the derivative to the test function. The Heaviside step $H$ has no classical derivative at $0$ , yet $H' = \delta$ distributionally because $\langle H',\varphi\rangle = -\int_0^\infty\varphi'\,dx = \varphi(0)$ . The "infinite spike" of intuition is encoded in the single rule "evaluate at $0$ ."

Figure

The Heaviside step, its distributional derivative δ, and a mollifier approximation

```svg

```

ExampleDerivative of the Heaviside function

Given: $H(x)=1$ for $x>0$ , $H(x)=0$ for $x<0$ , a regular distribution in $\mathcal{D}'(\mathbb{R})$ . Compute $H'$ .

Step 1. By the derivative rule,

5

\quad \langle H',\varphi\rangle = -\langle H,\varphi'\rangle = -\int_{-\infty}^{\infty} H(x)\,\varphi'(x)\,dx = -\int_0^\infty \varphi'(x)\,dx.

Derivation hint: apply equation [4] with $|\alpha|=1$ , then use the explicit form of $H$ to cut the integral at $0$ .

Step 2. Since $\varphi$ has compact support, $\varphi(\infty)=0$ , so $-\int_0^\infty\varphi'\,dx = -[\varphi(\infty)-\varphi(0)] = \varphi(0)$ .

Step 3. Therefore $\langle H',\varphi\rangle = \varphi(0) = \langle\delta,\varphi\rangle$ for every $\varphi$ .

\boxed{\,H' = \delta\,}$$ (dimensionless; $\delta$ has units of $1/x$ if $x$ carries units).

ExampleFourier transform of the Dirac delta and of 1

Given: Work in $\mathcal{S}'(\mathbb{R})$ with $\widehat{\varphi}(\xi)=\int \varphi(x)e^{-2\pi i x\xi}\,dx$ . The Fourier transform of a tempered distribution is defined by $\langle\widehat{T},\varphi\rangle=\langle T,\widehat{\varphi}\rangle$ .

Step 1 — Transform of $\delta$ .

6

\quad \langle\widehat{\delta},\varphi\rangle = \langle\delta,\widehat{\varphi}\rangle = \widehat{\varphi}(0) = \int_{-\infty}^\infty \varphi(x)\,dx = \langle 1,\varphi\rangle.

Derivation hint: $\delta$ evaluates at $0$ , and $\widehat{\varphi}(0)$ is the total integral of $\varphi$ , which is the action of the constant distribution $1$ . Hence $\boxed{\widehat{\delta}=1}$ .

Step 2 — Transform of $1$ . Using Fourier inversion / duality,

7

\quad \langle\widehat{1},\varphi\rangle = \langle 1,\widehat{\varphi}\rangle = \int \widehat{\varphi}(\xi)\,d\xi = \varphi(0) = \langle\delta,\varphi\rangle,

Derivation hint: the integral of $\widehat\varphi$ over all $\xi$ recovers $\varphi(0)$ by the inversion formula evaluated at the origin. so $\boxed{\widehat{1}=\delta}$ . The constant function (not in $L^1$ , $L^2$ ) has a perfectly well-defined transform as a tempered distribution: a point mass at frequency $0$ . Units: if $x$ is seconds, $\xi$ is Hz; $\delta(\xi)$ carries units of seconds.

⚠️ Common mistake

Distributions cannot be multiplied in general. There is no associative product on $\mathcal{D}'$ extending pointwise multiplication. The notorious obstruction (Schwartz impossibility theorem) shows $\delta\cdot\delta$ , or even $\delta(x)\cdot x^{-1}$ , is undefined. One can multiply a distribution by a smooth function ( $a\in C^\infty$ gives $\langle aT,\varphi\rangle=\langle T,a\varphi\rangle$ ), and convolve under support conditions, but a full algebra requires Colombeau algebras at the cost of breaking some identities.

⚠️ Common mistake

Distributions have no pointwise values. Writing " $\delta(0)=\infty$ " is meaningless. Likewise " $\delta(x)=0$ for $x\neq0$ " is only the statement that $\delta$ vanishes on test functions supported away from $0$ (its support is $\{0\}$ ). The object lives entirely in its pairing with test functions.

CounterexampleThe square of the delta has no meaning

Suppose a continuous, associative product $\cdot$ extended pointwise multiplication and the Leibniz rule, with $x\cdot\delta=0$ (true, since $\langle x\delta,\varphi\rangle=\langle\delta,x\varphi\rangle=0\cdot\varphi(0)=0$ ). Combine with $x\cdot\mathrm{pv}\tfrac1x=1$ . Associativity would force

\big(\delta\cdot x\big)\cdot \mathrm{pv}\tfrac1x = 0,\qquad \delta\cdot\big(x\cdot \mathrm{pv}\tfrac1x\big)=\delta\cdot 1=\delta,

so $0=\delta$ — a contradiction. No such product exists. This is the concrete face of the Schwartz impossibility theorem.

connection

Distribution theory is the connective tissue of modern analysis. The distributional Fourier transform turns differentiation into multiplication, making it the natural setting for Fourier Analysis and constant-coefficient PDEs. Weak derivatives of distributions underpin Sobolev Spaces, where one demands the derivative distribution be representable by an $L^p$ function. Fundamental solutions (Green's functions) are distributional solutions of $P(\partial)E=\delta$ , linking to Partial Differential Equations and the Malgrange–Ehrenpreis theorem. Tempered distributions are the natural carriers of the Spectral Theory of Self-Adjoint Operators via generalized eigenfunctions (rigged Hilbert spaces / Gelfand triples). The support and singular-support apparatus generalizes into the Wavefront Set and microlocal analysis.

🔬 Where this appears

⏳ History

- 1822 — Fourier manipulates series and "impulse" kernels heuristically; the Dirichlet kernel foreshadows approximations to

\delta

.

- 1894 — Heaviside's operational calculus computes with

H'

and operators

p=d/dt

without rigorous foundation, drawing fierce criticism.

- 1926–1930 — Dirac systematizes

\delta(x)

in quantum mechanics, declaring its defining sifting property despite its non-functional nature.

- 1935–1936 — Sobolev introduces generalized derivatives and functionals on test functions to study hyperbolic PDEs, anticipating much of the theory.

- 1936 — Bochner studies Fourier transforms of functions of polynomial growth, a precursor to tempered distributions.

- 1945–1948 — Laurent Schwartz unifies these threads, defining distributions as dual spaces of

\mathcal{D}

and

\mathcal{S}

and proving the structural theorems.

- 1950–1951 — Schwartz publishes Théorie des distributions (two volumes); the Fourier transform, kernel theorem, and convolution structure are established.

- 1950 — Schwartz awarded the Fields Medal, largely for this work.

- 1955–1965 — Gelfand–Shilov generalize spaces; Hörmander, Sato, and others develop microlocal analysis and hyperfunctions, extending the reach to analytic singularities and wavefront sets.

open-problem

The central limitation — no canonical multiplication of distributions — remains a live frontier. Colombeau algebras (1980s) construct differential algebras containing $\mathcal{D}'$ in which products exist, but at the price of nonstandard equality and loss of some classical identities; choosing the "right" framework for nonlinear PDEs with singular data is unsettled. The renormalization problem in QFT is precisely the failure to multiply distributions (products of propagators at coincident points), addressed by Epstein–Glaser causal perturbation theory and, recently, Hairer's regularity structures and Gubinelli's paracontrolled calculus (Fields Medal 2014), which give meaning to products in stochastic PDEs like KPZ and $\Phi^4_3$ . A fully canonical, universally applicable nonlinear theory of generalized functions does not exist; each application selects its own consistent enlargement.

Distributions and Generalized Functions — Schwartz Theory

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also