Distributions and Generalized Functions — Schwartz Theory

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).

A distribution is a continuous linear functional on a space of smooth, compactly supported (or rapidly decaying) test functions. Rather than asking "what is the value of $f$ at a point?", distribution theory asks "what does $f$ do when integrated against every nice test function?" This functional-analytic shift legitimizes operations — pointwise multiplication failures aside — that were used heuristically by physicists for decades, most notably differentiation of non-differentiable functions and the manipulation of the Dirac delta. The framework's most surprising fact is that differentiation becomes a totally unrestricted, continuous operation: every distribution has derivatives of all orders, computed by transferring the derivative onto the test function via integration by parts. This single device unifies weak solutions of PDEs, Green's functions, Fourier transforms of polynomials, and the calculus of singular kernels.

The Core Insight

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

Formal Statement

Test Function Spaces

Let $Ω \subseteq R^{n}$ be open.

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

The Schwartz space is normed by the countable family of seminorms

p_{α, β} (φ) = x \in R^{n} sup x^{α} \partial^{β} φ (x) < \infty, α, β \in N^{n} .

⟳rapid decrease means

x^{α} \partial^{β} φ \to 0

; boundedness of the sup over all multi-indices encodes "faster than any polynomial."

Convergence $φ_{k} \to φ$ in $D (Ω)$ means: there is a fixed compact $K$ containing all $supp φ_{k}$ , and $\partial^{α} φ_{k} \to \partial^{α} φ$ uniformly for every $α$ .

Distributions

A distribution $T \in D^{'} (Ω)$ is a linear functional $T : D (Ω) \to C$ that is continuous: $φ_{k} \to φ$ in $D (Ω)$ implies $⟨ T, φ_{k} ⟩ \to ⟨ T, φ ⟩$ . Equivalently, for each compact $K \subset Ω$ there exist $C > 0$ , $m \in N$ with

∣ ⟨ T, φ ⟩ ∣ \leq C ∣ α ∣ \leq m \sum x \in K sup ∣ \partial^{α} φ (x) ∣, \forall φ with supp φ \subseteq K .

⟳continuity on a Fréchet/LF space is equivalent to a seminorm estimate; the integer

m

(smallest uniform choice) is the local

order of $T$ .

The dual of $S$ is the space of tempered distributions $S^{'} (R^{n})$ — exactly those distributions that admit a Fourier transform.

Regular Distributions and the Embedding

Every $f \in L_{loc}^{1} (Ω)$ defines a distribution $T_{f}$ via

⟨ T_{f}, φ ⟩ = \int_{Ω} f (x) φ (x) d x,

and the map $f \mapsto T_{f}$ is injective (modulo a.e. equality). Thus $L_{loc}^{1} ↪ D^{'}$ , and one writes $f$ for $T_{f}$ .

⟳injectivity follows from the fundamental lemma of calculus of variations — if

\int f φ = 0

for all

φ \in D

then

f = 0

a.e., proved by mollification.

Distributional Derivative

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

⟨ \partial^{α} T, φ ⟩ = (- 1)^{∣ α ∣} ⟨ T, \partial^{α} φ ⟩, φ \in D (Ω) .

⟳integrate by parts

∣ α ∣

times against smooth

f

with compact support; all boundary terms vanish because

φ

has compact support, leaving the sign

(- 1)^{∣ α ∣}

Because $\partial^{α} φ \in D$ , the right side is well-defined for all $T$ and all $α$ : every distribution is infinitely differentiable.

Why It Holds / Derivation

Differentiation is well-defined, linear, and continuous

Step 1 — Linearity. For

φ, ψ \in D

and scalars

a, b

\partial^{α} (a φ + b ψ) = a \partial^{α} φ + b \partial^{α} ψ

, so

⟨ \partial^{α} T, a φ + b ψ ⟩ = (- 1)^{∣ α ∣} ⟨ T, a \partial^{α} φ + b \partial^{α} ψ ⟩

is linear in

φ

. Step 2 — The functional is continuous. If

φ_{k} \to 0

D

, then

\partial^{α} φ_{k} \to 0

D

(same compact support, uniform convergence of all derivatives). By continuity of

T

⟨ T, \partial^{α} φ_{k} ⟩ \to 0

, hence

⟨ \partial^{α} T, φ_{k} ⟩ \to 0

. So

\partial^{α} T \in D^{'}

. Step 3 — Consistency with classical derivatives. If

f \in C^{1}

, integration by parts gives

\int (\partial_{j} f) φ d x = - \int f (\partial_{j} φ) d x,

with no boundary term since

supp φ

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^{α} T

is sequentially continuous on

D^{'}

(weak-* topology), and mixed partials commute,

\partial_{i} \partial_{j} T = \partial_{j} \partial_{i} T

, because they commute on

φ

■

Paul Dirac | 1902–1984

Dirac introduced the "delta function" $δ (x)$ in his 1930 Principles of Quantum Mechanics as an object satisfying $\int δ (x) f (x) d x = f (0)$ and $δ (x) = 0$ 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: $δ$ is precisely the evaluation functional $⟨ δ, φ ⟩ = φ (0)$ , a distribution of order $0$ . Dirac's formal calculus thereby became a theorem-rich theory.

Building Intuition

Think of a distribution as a measuring device rather than a curve. You never "see" $δ$ ; you only see what it reports when you feed it a smooth test signal $φ$ — namely $φ (0)$ . The genius of test functions is that they are infinitely smooth, so any roughness in your object is laundered through integration by parts onto $φ$ , 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 $H^{'} = δ$ distributionally because $⟨ H^{'}, φ ⟩ = - \int_{0}^{\infty} φ^{'} d x = φ (0)$ . The "infinite spike" of intuition is encoded in the single rule "evaluate at $0$ ."

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

<svg viewBox="0 0 640 360" xmlns="http://www.w3.org/2000/svg">
<rect width="640" height="360" fill="#0a0e17"/>
<line x1="40" y1="300" x2="600" y2="300" stroke="#94a3b8" stroke-width="1.5"/>
<line x1="320" y1="40" x2="320" y2="320" stroke="#94a3b8" stroke-width="1.5"/>
<text x="595" y="320" fill="#e2e8f0" font-size="13">x</text>
<!-- Heaviside -->
<polyline points="40,260 320,260 320,120 600,120" fill="none" stroke="#94a3b8" stroke-width="2" stroke-dasharray="4 3"/>
<text x="450" y="110" fill="#e2e8f0" font-size="13">H(x)  (Heaviside)</text>
<text x="120" y="252" fill="#e2e8f0" font-size="13">0</text>
<text x="450" y="138" fill="#e2e8f0" font-size="13">1</text>
<!-- delta spike -->
<line x1="320" y1="300" x2="320" y2="60" stroke="#d4a853" stroke-width="3"/>
<polygon points="320,50 313,68 327,68" fill="#d4a853"/>
<text x="330" y="60" fill="#d4a853" font-size="14">δ = H′</text>
<!-- mollifier bump approximating delta -->
<path d="M 240 300 Q 320 90 400 300" fill="none" stroke="#d4a853" stroke-width="2" stroke-dasharray="2 4"/>
<text x="405" y="220" fill="#e2e8f0" font-size="12">ρₑ * δ  (mollified)</text>
<text x="40" y="345" fill="#94a3b8" font-size="11">Distributional derivative of a jump is a point mass at the jump location.</text>
</svg>

Worked Examples

Derivative of the Heaviside function

Given:

H (x) = 1

for

x > 0

H (x) = 0

for

x < 0

, a regular distribution in

D^{'} (R)

. Compute

H^{'}

. Step 1. By the derivative rule,

⟨ H^{'}, φ ⟩ = - ⟨ H, φ^{'} ⟩ = - \int_{- \infty}^{\infty} H (x) φ^{'} (x) d x = - \int_{0}^{\infty} φ^{'} (x) d x .

⟳apply equation [4] with

∣ α ∣ = 1

, then use the explicit form of

H

to cut the integral at

0

Step 2. Since

φ

has compact support,

φ (\infty) = 0

, so

- \int_{0}^{\infty} φ^{'} d x = - [φ (\infty) - φ (0)] = φ (0)

. Step 3. Therefore

⟨ H^{'}, φ ⟩ = φ (0) = ⟨ δ, φ ⟩

for every

φ

H^{'} = δ

(dimensionless;

δ

has units of

1/ x

x

carries units).

Fourier transform of the Dirac delta and of 1

Given: Work in

S^{'} (R)

with

φ (ξ) = \int φ (x) e^{- 2 π i x ξ} d x

. The Fourier transform of a tempered distribution is defined by

⟨ T, φ ⟩ = ⟨ T, φ ⟩

. Step 1 — Transform of $δ$ .

⟨ δ, φ ⟩ = ⟨ δ, φ ⟩ = φ (0) = \int_{- \infty}^{\infty} φ (x) d x = ⟨ 1, φ ⟩ .

⟳

δ

evaluates at

0

, and

φ (0)

is the total integral of

φ

, which is the action of the constant distribution

1

Hence $δ = 1$ .

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

⟨ 1, φ ⟩ = ⟨ 1, φ ⟩ = \int φ (ξ) d ξ = φ (0) = ⟨ δ, φ ⟩,

⟳the integral of

φ

over all

ξ

recovers

φ (0)

by the inversion formula evaluated at the origin.

so $1 = δ$ . 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, $ξ$ is Hz; $δ (ξ)$ carries units of seconds.

Common Misconceptions

Distributions cannot be multiplied in general. There is no associative product on

D^{'}

extending pointwise multiplication. The notorious obstruction (Schwartz impossibility theorem) shows

δ \cdot δ

, or even

δ (x) \cdot x^{- 1}

, is undefined. One can multiply a distribution by a smooth function (

a \in C^{\infty}

gives

⟨ a T, φ ⟩ = ⟨ T, a φ ⟩

), and convolve under support conditions, but a full algebra requires Colombeau algebras at the cost of breaking some identities. Distributions have no pointwise values. Writing "

δ (0) = \infty

" is meaningless. Likewise "

δ (x) = 0

for

x \neq = 0

" is only the statement that

δ

vanishes on test functions supported away from

0

(its support is

{0}

). The object lives entirely in its pairing with test functions.

The square of the delta has no meaning

Suppose a continuous, associative product $\cdot$ extended pointwise multiplication and the Leibniz rule, with $x \cdot δ = 0$ (true, since $⟨ x δ, φ ⟩ = ⟨ δ, x φ ⟩ = 0 \cdot φ (0) = 0$ ). Combine with $x \cdot pv \frac{1}{x} = 1$ . Associativity would force

(δ \cdot x) \cdot pv \frac{1}{x} = 0, δ \cdot (x \cdot pv \frac{1}{x}) = δ \cdot 1 = δ,

0 = δ

— a contradiction. No such product exists. This is the concrete face of the Schwartz impossibility theorem.

Connections

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 $P (\partial) E = δ$ , 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.

Real-World Applications

Quantum mechanics — continuous spectra. Position and momentum eigenstates $∣ x ⟩$ , $∣ p ⟩$ are not in $L^{2}$ ; they are tempered distributions handled rigorously by rigged Hilbert spaces (Gelfand triples $S \subset L^{2} \subset S^{'}$ ).
Quantum field theory. Quantum fields are operator-valued distributions; the Wightman axioms phrase correlation functions as distributions, and propagators (Feynman, retarded) are fundamental solutions of wave/Klein–Gordon operators.
Signal processing. The Dirac comb $\sum_{n} δ (x - n)$ models ideal sampling; its self-Fourier property yields the Poisson summation formula and underlies the Nyquist–Shannon sampling theorem.
Electromagnetism / mechanics. Point charges, line charges, and surface densities are distributions; $\nabla^{2} (1/4 π r) = - δ$ expresses the Coulomb/Newton potential as a fundamental solution of the Laplacian.
Solid mechanics. Concentrated point loads and moments on beams produce distributional terms in the Euler–Bernoulli equation; deflection is the convolution with the Green's function.
Numerical PDEs. The finite element method discretizes the weak (distributional) formulation of boundary value problems; well-posedness rests on Sobolev/distribution theory.
Probability. Characteristic functions and the densities of singular or discrete random variables (e.g. point masses) are tempered distributions; the Lévy continuity theorem lives in $S^{'}$ .

Historical Development

1822 — Fourier manipulates series and "impulse" kernels heuristically; the Dirichlet kernel foreshadows approximations to $δ$ .
1894 — Heaviside's operational calculus computes with $H^{'}$ and operators $p = d / d t$ without rigorous foundation, drawing fierce criticism.
1926–1930 — Dirac systematizes $δ (x)$ 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 $D$ and $S$ 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.

Subtleties & Limits

The central limitation — no canonical multiplication of distributions — remains a live frontier. Colombeau algebras (1980s) construct differential algebras containing $D^{'}$ 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 $Φ_{3}^{4}$ . A fully canonical, universally applicable nonlinear theory of generalized functions does not exist; each application selects its own consistent enlargement.