Theory of Distributions

Every physicist writes $\int \delta(x)\,f(x)\,\mathrm{d}x = f(0)$ , differentiates a step to get a spike, and Fourier-transforms a plane wave into a delta — and each move is, taken literally, nonsense. No ordinary function is zero everywhere except at a point yet integrates to one; a step has no derivative at its jump; a constant has no convergent Fourier integral. The theory of distributions makes all of it rigorous, not by patching the exceptions but by changing what a "function" is allowed to be: a distribution is not a rule that assigns a number to each point of space, but a rule that assigns a number to each smooth test function — the reading a perfect detector would report if you smeared the object against a smooth probe. Once you accept that shift, the whole formal calculus of physics falls out of one idea: every operation on the object — differentiate, translate, scale, Fourier-transform — is instead performed on the smooth test function, using the identity that integration by parts would give if the object were an honest function. Because test functions are infinitely differentiable and decay (or have compact support), the boundary terms that would block those manipulations vanish. Differentiation becomes unconditional: the derivative of the Heaviside step is the Dirac delta, the delta has derivatives of every order, and $1/x$ acquires a principal-value meaning. On the right test space the Fourier transform becomes a clean isomorphism and $\widehat{1}=2\pi\delta$ becomes a theorem. Laurent Schwartz assembled this framework in the mid-1940s; it is now the shared foundation of Green's functions, quantum field theory, and the spectral theory of continuous spectra.

Key takeaways

A distribution is a continuous linear functional on a space of smooth test functions: it acts on functions rather than being evaluated at points, generalizing what "function" means.
The two working test spaces are $\mathcal{D}=C_c^\infty$ (compactly supported) with dual $\mathcal{D}'$ , and the Schwartz space $\mathcal{S}$ (rapidly decreasing) with dual $\mathcal{S}'$ , the tempered distributions.
Every locally integrable $f$ embeds as a regular distribution $T_f(\varphi)=\int f\varphi$ ; the Dirac delta $\langle\delta,\varphi\rangle=\varphi(0)$ is the prototypical singular distribution with no function representative.
Every operation is defined by moving it onto the test function: $\langle T',\varphi\rangle=-\langle T,\varphi'\rangle$ , so every distribution is infinitely differentiable and $H'=\delta$ .
On $\mathcal{S}'$ the Fourier transform $\langle\widehat T,\varphi\rangle=\langle T,\widehat\varphi\rangle$ is an isomorphism, giving rigorous meaning to $\widehat 1=2\pi\delta$ , $\widehat\delta=1$ , and Sokhotski–Plemelj.
The framework grounds delta functions, principal values, Green's functions $LG=\delta$ , and the fundamental solutions behind Coulomb potentials and propagators.
Key idea: stop asking "what is the value of $\delta$ at a point?" and ask "what does $\delta$ do to a smooth test function?" A distribution is defined entirely by its action $\varphi\mapsto\langle T,\varphi\rangle$ , and every operation is defined by transferring it onto $\varphi$ via the identity that integration by parts would give for a genuine function.

The core insight

Stop asking "what is the value of $\delta$ at a point?" Instead ask "what does $\delta$ do to a smooth test function?" A distribution is defined entirely by its action $\varphi\mapsto\langle T,\varphi\rangle$ , and every operation — derivative, Fourier transform, translation, scaling — is defined by moving that operation onto the test function via the identity that would hold by integration by parts for a genuine function.

Where this lives today

Distributions are not a 1950s curiosity — they are the working language of several live research fronts. In quantum field theory the modern rigorous renormalization is Epstein–Glaser causal perturbation theory, which treats products of field distributions using Hörmander's wavefront sets; the "divergences" of Feynman diagrams are precisely ill-defined products of distributions, and controlling their singularity structure is central to constructing QFT on curved spacetimes. In nonlinear PDE theory, distributions and their Sobolev-space refinements are where existence and regularity of weak solutions are proved; the theory of regularity structures, which earned Martin Hairer the 2014 Fields Medal, extends the framework to stochastic PDEs so rough that even the products in the equation must be redefined distributionally, and the finite-element method is built directly on the same weak formulation. And in everyday physics the framework silently underwrites momentum eigenstates, propagators with their $i\varepsilon$ prescriptions, and the rigged Hilbert spaces $\mathcal{S}\subset L^2\subset\mathcal{S}'$ that make Dirac's bra–ket calculus a theorem rather than a convention.

The physical picture

No real instrument reports the value of a field at a mathematical point. A photodiode integrates over its active area, a microphone over a weighted window of the pressure wave. What it returns is $\int T(x)\,\varphi(x)\,\mathrm{d}x$ , where $\varphi$ is the instrument's smooth response profile. A distribution abstracts what every possible smooth detector would read: two are equal precisely when every test function gives the same reading. This is why the delta is legitimate even though it is not a function — it is the idealized detector of infinite resolution, $\langle\delta,\varphi\rangle=\varphi(0)$ .

Figure

A distribution reads a test function φ by pairing. A regular T_f returns the overlap integral ∫ f φ dx, the shaded area; the singular δ_a ignores all of φ except its value φ(a) at the point a, here φ(0) = 1.0

The engine of the whole subject is integration by parts. For smooth $f$ and compactly supported $\varphi$ , $\int f'\varphi = -\int f\varphi'$ , with no boundary term because $\varphi$ vanishes at the ends. We define the derivative of any distribution by the right-hand side — and now a step has a derivative, and that derivative has derivatives, forever, promoting the delta to an object with a full differential calculus. Where does the delta come from? As a limit of ordinary functions: ever-narrower bumps of fixed area, tested against $\varphi$ , converge to $\varphi(0)$ . The second figure shows the canonical nascent delta, the Lorentzian family underlying the Sokhotski–Plemelj formula.

Figure

Nascent delta from Lorentzians ε/(x²+ε²): each encloses area π, and as ε shrinks 1.0 → 0.5 → 0.25 the peak height 1/ε grows 1 → 2 → 4 while the width collapses, so the family converges to π·δ(x)

Finally, the choice of test space is not a technicality — it fixes the theory. Demand more of your test functions (smoothness and rapid decay, the Schwartz space $\mathcal{S}$ ) and you admit fewer functionals, but you gain the property physics needs most: those functionals are exactly the ones whose Fourier transform stays in the same world — the tempered distributions $\mathcal{S}'$ , where plane waves and polynomials finally get honest transforms.

The core equations

The starting point is the test space and its convergence, since a distribution must be continuous with respect to it. Sequences of test functions must share one compact support and converge uniformly together with all their derivatives.

\mathcal{D}(\Omega)=C_c^\infty(\Omega),\qquad \varphi_n\to\varphi \iff \begin{cases}\operatorname{supp}\varphi_n\subset K\text{ fixed},\\ \partial^\alpha\varphi_n\to\partial^\alpha\varphi \text{ uniformly }\forall\alpha.\end{cases} \qquad (1)

where: $\Omega$ — open subset of $\mathbb{R}^n$ ; $C_c^\infty$ — infinitely differentiable functions of compact support; $K$ — one compact set containing every support; $\alpha$ — a multi-index. Reading it: convergence in $\mathcal{D}$ is uniform in every derivative at once — far stronger than pointwise, keeping the dual usefully small. A distribution is a continuous linear functional on that space; the pairing $\langle T,\varphi\rangle$ is what $T$ returns when read by the probe $\varphi$ .

T:\mathcal{D}(\Omega)\to\mathbb{C}\ \text{linear},\qquad \varphi_n\to\varphi\ \Rightarrow\ \langle T,\varphi_n\rangle\to\langle T,\varphi\rangle. \qquad (2)

where: $T\in\mathcal{D}'(\Omega)$ — a distribution, a continuous linear functional on $\mathcal{D}$ ; $\langle T,\varphi\rangle$ — its action on $\varphi$ . Reading it: continuity is equivalent to a local seminorm bound $|\langle T,\varphi\rangle|\le C_K\sum_{|\alpha|\le m}\sup_K|\partial^\alpha\varphi|$ , whose least $m$ is the order of $T$ ( $\delta$ has order $0$ , $\delta^{(k)}$ order $k$ ). Two objects anchor the theory: the regular distributions from ordinary functions, and the singular delta that no function can represent.

\langle T_f,\varphi\rangle=\int_{\mathbb{R}^n} f\,\varphi\,\mathrm{d}x\quad(f\in L^1_{\mathrm{loc}}),\qquad \langle\delta_a,\varphi\rangle=\varphi(a),\qquad \langle\delta_a^{(k)},\varphi\rangle=(-1)^k\varphi^{(k)}(a). \qquad (3)

where: $T_f$ — the regular distribution of a locally integrable $f$ ; $\delta_a$ — the point mass at $a$ ; $\varphi^{(k)}$ — the $k$ -th derivative of the probe. Reading it: $f\mapsto T_f$ is injective, so genuine functions sit inside $\mathcal{D}'$ without loss, while $\delta_a$ has no $f$ behind it. The central construction is the distributional derivative, defined by throwing the derivative onto the test function with a sign — making every distribution infinitely differentiable.

\langle \partial^\alpha T,\varphi\rangle=(-1)^{|\alpha|}\langle T,\partial^\alpha\varphi\rangle. \qquad (4)

where: $\partial^\alpha T$ — the distributional derivative, always well defined; $(-1)^{|\alpha|}$ — the sign from $|\alpha|$ integrations by parts. Reading it: no difference quotient and no smoothness of $T$ is needed — the derivative is defined as the functional on the right, valid since $\partial^\alpha\varphi$ is still a probe. Two further operations transfer onto $\varphi$ the same way: multiplication by a smooth $g\in C^\infty$ is $\langle gT,\varphi\rangle=\langle T,g\varphi\rangle$ (the multiplier must be smooth so $g\varphi$ stays a test function — the shadow of $\delta^2$ having no meaning), and translation is $\langle \tau_a T,\varphi\rangle=\langle T,\tau_{-a}\varphi\rangle$ . To Fourier-transform, restrict to the Schwartz space, whose seminorms control every derivative against every polynomial weight; its dual $\mathcal{S}'$ is closed under the transform.

\mathcal{S}(\mathbb{R}^n)=\Big\{\varphi\in C^\infty:\ \sup_x |x^\beta\partial^\alpha\varphi(x)|<\infty\ \ \forall\alpha,\beta\Big\},\qquad \langle \widehat{T},\varphi\rangle=\langle T,\widehat{\varphi}\rangle\ \ (T\in\mathcal{S}'). \qquad (5)

where: $\mathcal{S}$ — the rapidly decreasing smooth functions; $x^\beta$ — a monomial weight enforcing polynomial decay; $\widehat T$ — the transform of the tempered distribution. Reading it: the transform is an isomorphism $\mathcal{S}\to\mathcal{S}$ , so $\widehat\varphi$ is again a legal probe — the reason it lives on $\mathcal{S}'$ . Two named identities package the payoff — a convolution structure making $\delta$ a unit, and a boundary-value formula separating a principal value from a delta.

(T*\varphi)(x)=\langle T,\ \tau_x\check\varphi\rangle,\quad \delta * T=T;\qquad \frac{1}{x\mp i0}=\mathrm{p.v.}\frac1x\pm i\pi\,\delta(x). \qquad (6)

where: $T*\varphi$ — convolution, smooth when $\varphi\in\mathcal{D}$ ; $\check\varphi(y)=\varphi(-y)$ — reflection; $\tfrac{1}{x\mp i0}$ — boundary value of $1/(x\mp i\varepsilon)$ as $\varepsilon\to0^+$ ; $\mathrm{p.v.}$ — Cauchy principal value. Reading it: $\delta$ is the convolution identity, so " $LG=\delta$ " makes $G$ an inverse of $L$ , and Sokhotski–Plemelj underlies every $i\varepsilon$ prescription.

Derivation AThe distributional derivative and why the derivative of a step is a delta

The goal is to show that the sign rule of equation (4) is not arbitrary but forced — it is the unique definition that keeps agreeing with ordinary calculus wherever ordinary calculus applies — and then to run it on the Heaviside step to produce the most quoted identity of the subject, $H'=\delta$ . We derive the definition from integration by parts, apply it to $H$ , and confirm the general jump rule, so that "differentiating a discontinuity gives a spike at the jump" becomes a theorem rather than a slogan.

Step 1. For smooth $f$ and compactly supported $\varphi$ , integrate by parts: $\int_{\mathbb{R}} f'\,\varphi\,\mathrm{d}x = [f\varphi] - \int_{\mathbb{R}} f\,\varphi'\,\mathrm{d}x = -\int_{\mathbb{R}} f\,\varphi'\,\mathrm{d}x$ . Why: the boundary term $[f\varphi]$ vanishes because $\varphi$ has compact support — the property that makes the construction consistent.

Step 2. The right side $-\langle T_f,\varphi'\rangle$ uses only the action of $T_f$ on $\varphi'$ , never any smoothness of $f$ , so promote it to a definition: $\langle T',\varphi\rangle := -\langle T,\varphi'\rangle$ , inductively $\langle \partial^\alpha T,\varphi\rangle=(-1)^{|\alpha|}\langle T,\partial^\alpha\varphi\rangle$ . Why: the right side is always defined, so every distribution becomes infinitely differentiable, and it agrees with the classical derivative when $f$ is differentiable.

Step 3. Apply it to the Heaviside step: $\langle H',\varphi\rangle = -\int_{-\infty}^{\infty} H\,\varphi'\,\mathrm{d}x = -\int_{0}^{\infty}\varphi'\,\mathrm{d}x$ . Why: $H$ is zero for $x<0$ , so only the half-line $x>0$ contributes.

Step 4. Evaluate: $-\int_0^\infty\varphi' = -\big[\varphi\big]_0^\infty = -\big(0-\varphi(0)\big)=\varphi(0)=\langle\delta,\varphi\rangle$ for every $\varphi$ , hence $H'=\delta$ . Why: compact support gives $\varphi(\infty)=0$ , leaving the delta's defining action; a jump of size $J$ at $x_0$ contributes $J\,\delta_{x_0}$ on top of the classical derivative. *This is the general jump formula $f' = \{f'\} + \sum J_i\,\delta_{x_i}$ , here with $J=1$ at $x_0=0$ : a jump of size $J$ at any point contributes $J\,\delta_{x_0}$ on top of the classical derivative.*

Go deeper · Derivation B · 1 minFourier transform of the constant 1 gives 2π times the delta

Derivation BFourier transform of the constant 1 gives 2π times the delta

The aim is to turn the physicist's completeness relation $\int e^{ikx}\,\mathrm{d}k = 2\pi\,\delta(x)$ , which diverges as a literal integral, into a rigorous statement about tempered distributions, and to extract its dual $\widehat\delta=1$ and the shifted plane-wave rule as free corollaries. We fix a Fourier convention, apply the duality definition, and recognize the inversion theorem hiding inside the integral — the identity that legitimizes momentum eigenstates throughout physics.

Step 1. Use the convention $\widehat{\varphi}(k)=\int e^{-ikx}\varphi\,\mathrm{d}x$ with inverse $\varphi(x)=\frac{1}{2\pi}\int e^{ikx}\widehat\varphi\,\mathrm{d}k$ . Since $1$ is a tempered distribution, $\langle\widehat 1,\varphi\rangle=\langle 1,\widehat\varphi\rangle=\int_{-\infty}^{\infty}\widehat\varphi(k)\,\mathrm{d}k$ . Why: the transform is defined by moving it onto the probe, and $1$ acts by integration — so everything reduces to integrating $\widehat\varphi$ .

Step 2. Read the inverse transform at $x=0$ : $\varphi(0)=\frac{1}{2\pi}\int_{-\infty}^{\infty}\widehat\varphi(k)\,\mathrm{d}k$ . Why: the integral we need is exactly $2\pi$ times the inversion formula at the origin — recognition, not computation.

Step 3. Therefore $\int\widehat\varphi\,\mathrm{d}k = 2\pi\,\varphi(0)$ , so $\langle\widehat 1,\varphi\rangle = 2\pi\,\varphi(0) = \langle 2\pi\delta,\varphi\rangle$ for every $\varphi$ , giving $\widehat 1 = 2\pi\,\delta$ . Why: equality of functionals on all test functions is equality of distributions, the only equality available.

Step 4. Dually $\langle\widehat\delta,\varphi\rangle=\langle\delta,\widehat\varphi\rangle=\widehat\varphi(0)=\int\varphi\,\mathrm{d}x=\langle 1,\varphi\rangle$ , so $\widehat\delta=1$ ; and modulation gives $\widehat{e^{ik_0 x}}=2\pi\,\delta(k-k_0)$ . Why: the transform pairs constants with points — the content of "a plane wave is a single momentum," and the origin of $\langle k|k'\rangle=2\pi\delta(k-k')$ . *As a consistency check: applying the Fourier transform twice returns $2\pi$ times the reflection operator, and $\widehat\delta=1$ together with $\widehat 1=2\pi\delta$ are mutually dual under this convention.*

Go deeper · Derivation C · 1 minSokhotski–Plemelj: the boundary value splits into a principal value and a delta

Derivation CSokhotski–Plemelj: the boundary value splits into a principal value and a delta

The purpose is to derive the relation $\tfrac{1}{x\mp i0}=\mathrm{p.v.}\tfrac1x\pm i\pi\delta(x)$ that governs every $i\varepsilon$ prescription in scattering theory, Green's functions, and dispersion relations, by regularizing the kernel $1/x$ off the real axis and taking the limit under the pairing. We split the kernel into real and imaginary parts and recognize each limit as a distribution we already understand — a principal value and a nascent delta of area $\pi$ .

Step 1. For $\varepsilon>0$ , $1/(x-i\varepsilon)$ is smooth and bounded; rationalize it as $\frac{1}{x-i\varepsilon}=\frac{x}{x^2+\varepsilon^2}+i\,\frac{\varepsilon}{x^2+\varepsilon^2}$ . Why: separating real and imaginary parts isolates the two pieces that survive the limit differently.

Step 2. The imaginary part is the Lorentzian nascent delta: $\int \frac{\varepsilon}{x^2+\varepsilon^2}\,\mathrm{d}x = \pi$ for every $\varepsilon$ while the mass concentrates at $0$ , so $\lim_{\varepsilon\to0^+}\int \frac{\varepsilon}{x^2+\varepsilon^2}\,\varphi\,\mathrm{d}x = \pi\,\varphi(0)$ . Why: fixed area $\pi$ plus shrinking width is convergence to $\pi\delta$ .

Step 3. The real part has an odd kernel that symmetrically excises the singularity: $\lim_{\varepsilon\to0^+}\int \frac{x}{x^2+\varepsilon^2}\,\varphi\,\mathrm{d}x = \mathrm{p.v.}\!\int\frac{\varphi}{x}\,\mathrm{d}x$ . Why: odd symmetry cancels the divergences on either side of $x=0$ , leaving the principal value.

Step 4. Combine: $\frac{1}{x-i0}=\mathrm{p.v.}\frac1x+i\pi\,\delta(x)$ , and $\varepsilon\to-\varepsilon$ flips the delta's sign, so $\frac{1}{x+i0}=\mathrm{p.v.}\frac1x-i\pi\,\delta(x)$ . Why: subtracting the two gives the Plemelj jump $\frac{1}{x-i0}-\frac{1}{x+i0}=2\pi i\,\delta(x)$ , which puts on-shell deltas into propagators. *As a consistency check: subtracting the two boundary values gives the Plemelj jump $2\pi i\,\delta$ , and the whole derivation rests on the Lorentzian normalization $\int\varepsilon/(x^2+\varepsilon^2)\,\mathrm{d}x=\pi$ .*

Limiting cases

The framework reproduces ordinary analysis exactly wherever it applies and extends smoothly into the singular territory where it breaks down.

Case	Result	Meaning
Regular distribution	$\langle T_f,\varphi\rangle=\int f\,\varphi\,\mathrm{d}x$	any $f\in L^1_{\mathrm{loc}}$ embeds; if $f\in C^1$ its distributional and classical derivatives coincide
Dirac measure (singular)	$\langle\delta_a,\varphi\rangle=\varphi(a)$	no locally integrable function represents it; the prototypical singular distribution, a point mass of order 0
Heaviside step	$H'=\delta$ in $\mathcal{D}'(\mathbb{R})$	classically $H'=0$ almost everywhere, but the distributional derivative captures the jump exactly
Principal value	$\langle \mathrm{p.v.}\tfrac1x,\varphi\rangle=\lim_{\varepsilon\to0^+}\int_{\|x\|>\varepsilon}\tfrac{\varphi}{x}\,\mathrm{d}x$	$1/x$ is not locally integrable, yet the symmetric limit defines a tempered distribution
Scaling of the delta	$\delta(ax)=\tfrac{1}{\|a\|}\delta(x)$	the delta carries a density weight; $\delta(a\mathbf{x})=\|a\|^{-n}\delta(\mathbf{x})$ in $n$ D

Common trapthe misconception that distributions can be multiplied like functions

There is no associative product on distributions that extends pointwise multiplication — the objects $\delta^2$ , $\delta\cdot H$ , and $\delta\cdot\mathrm{p.v.}(1/x)$ are all undefined, and this is not a fixable gap but a theorem (Schwartz's impossibility result). You may multiply by a genuinely smooth function, because then $g\varphi$ stays a test function; you may not multiply by another distribution, a discontinuous function, or itself. This is exactly the origin of the ultraviolet divergences of quantum field theory, where the naive product $\phi(x)^2$ at a single point is ill-defined and must be tamed by renormalization (rigorously, by Hörmander's wavefront-set conditions). A related trap: pointwise convergence of functions does not imply distributional convergence, and "the value of a distribution at a point" is meaningless unless it is regular.

Worked examples

1 · The nascent delta, with real numbers (confidence builder)

Consider the Lorentzian family $L_\varepsilon(x)=\dfrac{\varepsilon}{x^2+\varepsilon^2}$ . Show it has $\varepsilon$ -independent area, compute its peak heights for $\varepsilon=1,\,0.5,\,0.25$ , and show that as a distribution $L_\varepsilon\to\pi\delta$ .

Step 1. Area: $\displaystyle\int_{-\infty}^{\infty}\frac{\varepsilon}{x^2+\varepsilon^2}\,\mathrm{d}x=\Big[\arctan\frac{x}{\varepsilon}\Big]_{-\infty}^{\infty}=\boldsymbol{\pi}$ , locked at $\pi$ no matter how narrow the peak.

Step 2. Peak height at $x=0$ is $L_\varepsilon(0)=1/\varepsilon$ : $\varepsilon=1\Rightarrow\mathbf{1}$ , $\varepsilon=0.5\Rightarrow\mathbf{2}$ , $\varepsilon=0.25\Rightarrow\mathbf{4}$ . Halving $\varepsilon$ doubles the peak and halves the width at fixed area — the nascent-delta signature.

Step 3. Substitute $x=\varepsilon u$ : $\displaystyle\int \frac{\varepsilon}{x^2+\varepsilon^2}\,\varphi\,\mathrm{d}x=\int\frac{\varphi(\varepsilon u)}{u^2+1}\,\mathrm{d}u\to\varphi(0)\int\frac{\mathrm{d}u}{u^2+1}=\pi\,\varphi(0)$ as $\varepsilon\to0^+$ .

Step 4. Hence $\lim_{\varepsilon\to0^+}\langle L_\varepsilon,\varphi\rangle=\pi\,\varphi(0)=\langle\pi\delta,\varphi\rangle$ for every test function. Result: $L_\varepsilon\to\pi\delta$ distributionally, and $\tfrac{1}{\pi}L_\varepsilon\to\delta$ . This is exactly the imaginary-part limit of the Sokhotski–Plemelj derivation — a Lorentzian lineshape acts as a delta once its width drops below any probe's resolution.

2 · The Laplacian of 1/r is minus 4π times the delta (numeric)

Show that in three dimensions $-\nabla^2\!\left(\dfrac{1}{4\pi r}\right)=\delta^{(3)}(\mathbf{x})$ , the identity that makes $1/(4\pi r)$ the Green's function of the Laplacian and grounds the Coulomb and Newtonian potentials.

Step 1. Away from the origin $\nabla^2(1/r)=0$ ( $1/r$ is harmonic), so as a distribution $\nabla^2(1/r)$ is supported at the single point $r=0$ .

Step 2. Integrate over a ball $B_R$ and convert to a surface flux by the divergence theorem: $\int_{B_R}\nabla^2\frac1r\,\mathrm{d}V=\oint_{\partial B_R}\nabla\frac1r\cdot\mathrm{d}\mathbf{S}$ . With $\nabla(1/r)=-\hat{\mathbf{r}}/r^2$ over a sphere of area $4\pi R^2$ , the flux is $-\frac{1}{R^2}(4\pi R^2)=\boldsymbol{-4\pi}$ .

Step 3. The result $-4\pi$ is independent of $R$ and the whole contribution sits at the origin, so as a distribution $\nabla^2(1/r)=-4\pi\,\delta^{(3)}(\mathbf{x})$ . Dividing by $-4\pi$ , $-\nabla^2\!\left(\dfrac{1}{4\pi r}\right)=\delta^{(3)}(\mathbf{x})$ .

Step 4. Read as a Green's function: with $L=-\nabla^2$ and $G=1/(4\pi r)$ , $LG=\delta$ , so the solution of $-\nabla^2 u=\rho/\varepsilon_0$ is $u=G*(\rho/\varepsilon_0)=\dfrac{1}{4\pi\varepsilon_0}\displaystyle\int\frac{\rho(\mathbf{x}')}{|\mathbf{x}-\mathbf{x}'|}\,\mathrm{d}^3x'$ . The Coulomb potential is the fundamental solution of the Laplacian, verified distributionally — and the same duality algebra gives $x\,\delta=0$ and $x\,\delta'=-\delta$ , solving equations like $xT=1$ (solution $T=\mathrm{p.v.}(1/x)+C\delta$ ).

Physics of the portrait

The theory has a rare, clean origin: one person, in the mid-1940s, fused a scattered collection of physicists' tricks and PDE-theorists' workarounds into one framework — and won a Fields Medal for it.

Origins · 1 minLaurent Schwartz and how this was discovered

Laurent Schwartz

1915–2002, French mathematician, Fields Medal 1950 · founder of the theory of distributions

Laurent Schwartz created the theory of distributions in Grenoble in 1944–1945, unifying a century of formal manipulations that had never had a rigorous home. Heaviside had computed with the step function's "derivative" in the 1890s to the fury of contemporary analysts; Dirac had introduced $\delta(x)$ in 1927 as an indispensable but illegitimate tool of quantum mechanics; Sobolev had defined generalized derivatives and weak solutions of PDEs in the 1930s; Hadamard had extracted "finite parts" of divergent integrals. Schwartz's decisive move was to stop trying to be a generalized function and instead define one by its action on smooth test functions — a continuous linear functional on $C_c^\infty$ — so that differentiation, Fourier transformation, and convolution all became automatic by duality. He published the two-volume Théorie des distributions in 1950–51, and in 1950 became the first French mathematician to receive the Fields Medal, cited for exactly this work; he later proved the kernel theorem that bears his name. A lifelong political activist and passionate lepidopterist with over 20,000 butterflies, Schwartz taught for decades at the École Polytechnique. The framework he built is now so basic that the delta "function" is taught without apology — the apology having been supplied by him.

History

1893

— Heaviside's operational calculus manipulates the step and its derivative formally

1927

— Dirac introduces the delta function δ(x) in the principles of quantum mechanics

1936

— Sobolev defines generalized derivatives and weak solutions of partial differential equations

1945

— Schwartz defines distributions as continuous linear functionals on test functions

1950

— Schwartz publishes Théorie des distributions and receives the Fields Medal

Going further

Texts. Stone and Goldbart, Mathematics for Physics, ch. 8 (distributions and special functions) · Arfken, Weber and Harris, Mathematical Methods for Physicists, 7th ed., ch. 1 and 20 · Hassani, Mathematical Physics, ch. on generalized functions · Riley, Hobson and Bence, Mathematical Methods for Physics and Engineering, 3rd ed., ch. 13 · Reed and Simon, Methods of Modern Mathematical Physics, Vol. I, ch. V · Hörmander, The Analysis of Linear Partial Differential Operators I, ch. I–II · L. Schwartz, Théorie des distributions (1950–51). The key anchor identities developed in this article are: (1) $H'=\delta$ by integration by parts; (2) $\widehat 1=2\pi\delta$ and $\widehat\delta=1$ under the convention $\widehat\varphi(k)=\int e^{-ikx}\varphi\,\mathrm{d}x$ ; (3) Sokhotski–Plemelj via the Lorentzian limit $\int\varepsilon/(x^2+\varepsilon^2)\,\mathrm{d}x=\pi$ ; (4) $\nabla^2(1/r)=-4\pi\delta^{(3)}$ by the divergence theorem; (5) the algebra identities $x\delta=0$ and $x\delta'=-\delta$ . The seminorm continuity criterion and the order of $\delta^{(k)}$ follow from the standard definition of distributional order.

Prerequisites. The Dirac Delta Function · The Fourier Transform · Properties of the Fourier Transform · Lebesgue Integration. Leads to. Tempered Distributions and Fourier Transforms · Green's Functions via Integral Transforms · Solving Differential Equations by Transform Methods · The Convolution Theorem.

Ask the tutor: ask me to run any distributional identity — $H'=\delta$ , $\widehat 1=2\pi\delta$ , Sokhotski–Plemelj, or $\nabla^2(1/r)=-4\pi\delta$ — line by line, or to show why $\delta^2$ has no meaning.

Is the Dirac delta really not a function?

Correct, and this is a theorem, not a convention. If $\delta$ were locally integrable it would be zero almost everywhere (it vanishes off the origin), and any such function integrates to zero against every test function — contradicting $\langle\delta,\varphi\rangle=\varphi(0)$ . So no ordinary function reproduces the delta's readings. What it is, precisely, is a continuous linear functional: the rule $\varphi\mapsto\varphi(0)$ . Physicists write $\delta(x)$ inside integrals because, read as the pairing $\int\delta\,\varphi=\varphi(0)$ , it always gives correct answers — but that integral sign symbolizes the functional, not a literal Lebesgue integral.

Why must test functions be infinitely differentiable?

Because you want to differentiate a distribution arbitrarily many times, and each derivative on the distribution is paid for by one on the test function. If test functions were only $C^k$ , then $\delta^{(k+1)}$ would have nothing to act on — its definition $\langle\delta^{(k+1)},\varphi\rangle=(-1)^{k+1}\varphi^{(k+1)}(0)$ needs a $(k+1)$ -th derivative of $\varphi$ . Demanding $C^\infty$ once and for all guarantees derivatives of every order. Compact support (in $\mathcal{D}$ ) or rapid decay (in $\mathcal{S}$ ) plays the complementary role: it kills the boundary terms in integration by parts.

What is the difference between D′ and S′, and when does it matter?

$\mathcal{D}'$ is the dual of $\mathcal{D}=C_c^\infty$ ; $\mathcal{S}'$ is the dual of the Schwartz space $\mathcal{S}$ . Because $\mathcal{S}$ is a larger test space than $\mathcal{D}$ (it drops compact support), its dual $\mathcal{S}'$ is smaller: every tempered distribution is a distribution, but not conversely. The distinction matters when you Fourier-transform: the transform maps $\mathcal{S}$ onto itself, so it acts cleanly on $\mathcal{S}'$ , but it does not preserve $\mathcal{D}$ , so there is no transform on all of $\mathcal{D}'$ . The price of admission to $\mathcal{S}'$ is at most polynomial growth: $e^{x^2}$ is a distribution in $\mathcal{D}'(\mathbb{R})$ but is not tempered.

Why can't you just multiply two distributions?

Schwartz proved there is no associative product on distributions that restricts to ordinary multiplication and obeys the Leibniz rule. The obstruction is concrete: using $x\,\delta=0$ and $x\cdot\mathrm{p.v.}(1/x)=1$ , associativity would force $0=(x\delta)\cdot\mathrm{p.v.}\tfrac1x=\delta\cdot(x\cdot\mathrm{p.v.}\tfrac1x)=\delta$ , a contradiction. You can always multiply by a smooth function, and you can sometimes multiply distributions whose singularities do not collide — Hörmander's wavefront-set condition, the modern language of QFT renormalization.

Does the distributional derivative ever disagree with the ordinary one?

For a $C^1$ function they agree exactly: the distributional derivative of $T_f$ is $T_{f'}$ . They part company only where $f$ fails to be classically differentiable. At a jump the distributional derivative adds a delta of weight equal to the jump — the content of $H'=\delta$ — while the classical one does not exist there (for $|x|$ , the classical derivative is $\mathrm{sgn}(x)$ and the distributional second derivative is $2\delta$ ). So the distributional derivative never contradicts the classical one where the latter exists; it extends it, filling in the singular contributions it misses.

Why do physicists get away with treating δ as an ordinary function?

Because almost every manipulation physicists perform on the delta is secretly a legal distributional operation. Writing $\int\delta f\,\mathrm{d}x=f(0)$ is the pairing $\langle\delta,f\rangle$ ; $\delta(ax)=\delta(x)/|a|$ is the scaling rule; differentiating to get $\delta'$ is the duality derivative. The formal rules coincide with the theorems, because Schwartz built the theory to reproduce them. Trouble appears only at the forbidden operations — multiplying two deltas, squaring a field at a point, evaluating pointwise — which is precisely where physics itself runs into ultraviolet divergences. The rigor does not change the answers; it flags which moves are safe.