The Dirac Delta Function and Distributions

One-line statement: The Dirac delta

δ (x)

is not a function but a distribution — a continuous linear functional on test functions — that "samples" a function at a point:

\int f (x) δ (x) d x = f (0)

. Key result in words: Generalized functions extend classical calculus so that pointwise impossible objects (infinite spikes, derivatives of jumps) become rigorous and differentiable infinitely many times. Where it applies: Partial differential equations (Green's functions, fundamental solutions), Fourier analysis, signal processing, quantum mechanics, probability (point masses), electromagnetism. Who/when: Heuristics by Oliver Heaviside (1890s) and Paul Dirac (1927); rigorous theory by Sergei Sobolev (1936) and Laurent Schwartz (1945–1950, Fields Medal 1950).

The Dirac delta function $δ$ is the prototypical distribution (or generalized function): an object defined not by its values at points but by its action on a space of smooth, rapidly-decaying test functions $φ$ via the pairing $⟨ δ, φ ⟩ = φ (0)$ . No ordinary function has the defining properties $δ (x) = 0$ for $x \neq = 0$ yet $\int δ = 1$ ; the Lebesgue integral of a function vanishing almost everywhere is zero. The theory of distributions, founded rigorously by Laurent Schwartz, resolves this by reconceiving "functions" as functionals, allowing one to differentiate any locally integrable function infinitely often, to make sense of $δ$ , its derivatives, and Fourier transforms of objects that grow at infinity. The surprising fact: in this framework the derivative of the discontinuous Heaviside step is exactly $δ$ , and the second derivative of $∣ x ∣/2$ is again $δ$ — discontinuities and corners "differentiate into" spikes with perfect bookkeeping.

The Core Insight

Stop asking what value a function has at a point. Instead ask what number it produces when integrated against every smooth probe. A distribution is fully determined by these "measurements." The delta is the measurement that reads off $φ (0)$ . This shift — from pointwise evaluation to duality pairing — is the single idea that makes the entire theory rigorous, and it automatically delivers differentiation by transferring derivatives onto the test function via integration by parts.

Formal Statement

Test Function Spaces

Let $Ω \subseteq R^{n}$ be open.

$D (Ω) = C_{c}^{\infty} (Ω)$ : smooth functions with compact support in $Ω$ , with its (LF-)topology of convergence.
$S (R^{n})$ : the Schwartz space of smooth functions $φ$ such that every seminorm

∥ φ ∥_{α, β} = x \in R^{n} sup x^{α} \partial^{β} φ (x) < \infty

is finite, for all multi-indices

α, β

⟳

x^{α} = x_{1}^{α_{1}} \dots x_{n}^{α_{n}}

and

\partial^{β} = \partial_{x_{1}}^{β_{1}} \dots \partial_{x_{n}}^{β_{n}}

; finiteness of all these seminorms is exactly "smooth + faster-than-polynomial decay of all derivatives."

A sequence $φ_{k} \to φ$ in $D$ iff all supports lie in one compact set and all derivatives converge uniformly.

Distributions and the Delta

A distribution is a continuous linear functional $T : D (Ω) \to C$ ; the space is $D^{'} (Ω)$ . Continuity means: there exist $C > 0$ and $N \in N$ (depending on each compact $K$ ) with

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

⟳this is the standard seminorm-bound characterization of continuity on the inductive-limit space

D

; the least valid

N

is the local order of

T

The Dirac delta $δ_{a} \in D^{'} (R^{n})$ at the point $a$ is

⟨ δ_{a}, φ ⟩ = φ (a), δ := δ_{0} .

⟳linearity and continuity (with

N = 0

C = 1

) are immediate from

∣ φ (a) ∣ \leq sup ∣ φ ∣

; thus

δ

is a distribution of order

0

— in fact a Radon measure.

A distribution $T$ is a tempered distribution ( $T \in S^{'}$ ) if it is continuous on $S$ ; these are exactly the distributions admitting a Fourier transform.

Distributional Derivative

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

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

is again a distribution. Every distribution is therefore

C^{\infty}

in the distributional sense.

⟳motivated by integration by parts

\int (\partial^{α} f) φ = (- 1)^{∣ α ∣} \int f \partial^{α} φ

, with boundary terms vanishing because

φ

has compact support; the right side is continuous in

φ

since

\partial^{α} : D \to D

is continuous.

Heaviside, Delta, and Localization

Let $H$ be the Heaviside step ( $H (x) = 1$ for $x > 0$ , else $0$ ). Then

H^{'} = δ, δ^{'} (φ) = - φ^{'} (0), x δ (x) = 0, δ (a x) = \frac{1}{∣ a ∣} δ (x) .

⟳each follows from the definitions:

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

;

⟨ x δ, φ ⟩ = ⟨ δ, x φ ⟩ = (x φ) ∣_{0} = 0

; the scaling from change of variables

y = a x

Why It Holds / Derivation

Delta as a Limit of Approximate Identities

Claim. If

η \in L^{1} (R)

with

\int η = 1

and

η_{ε} (x) = ε^{- 1} η (x / ε)

, then

η_{ε} \to δ

D^{'}

ε \to 0^{+}

. Step 1. For

φ \in D

, write the pairing as an integral of the locally integrable function

η_{ε}

⟨ η_{ε}, φ ⟩ = \int_{R} \frac{1}{ε} η (\frac{x}{ε}) φ (x) d x = \int_{R} η (y) φ (ε y) d y .

⟳substitute

y = x / ε

d x = ε d y

; the prefactor

1/ ε

cancels.

Step 2. As

ε \to 0

φ (ε y) \to φ (0)

pointwise, and

∣ η (y) φ (ε y) ∣ \leq ∥ φ ∥_{\infty} ∣ η (y) ∣ \in L^{1}

. By dominated convergence:

ε \to 0 lim ⟨ η_{ε}, φ ⟩ = φ (0) \int_{R} η (y) d y = φ (0) = ⟨ δ, φ ⟩ .

⟳pull the limit inside the integral using the integrable dominator

∥ φ ∥_{\infty} ∣ η ∣

, then use

\int η = 1

Step 3. Since this holds for all

φ \in D

, we have

η_{ε} \to δ

in the weak-

*

topology of

D^{'}

■

The same argument with $η (x) = \frac{1}{π} \frac{1}{1 + x ^{2}}$ (Cauchy/Lorentzian), $η (x) = \frac{1}{2 π} e^{- x^{2} /2}$ (Gaussian), or $\frac{1}{π} \frac{s i n x}{x}$ (Dirichlet, with care) all yield $δ$ .

Fourier Transform of the Delta

With convention $f (ξ) = \int f (x) e^{- i ξ x} d x$ , the Fourier transform on $S^{'}$ is defined by duality $⟨ T, φ ⟩ = ⟨ T, φ ⟩$ . Then

δ = 1, 1 = 2 π δ, δ_{a} (ξ) = e^{- ia ξ} .

⟳

⟨ δ, φ ⟩ = ⟨ δ, φ ⟩ = φ (0) = \int φ

, identifying

δ

with the constant

1

; the second follows by Fourier inversion / duality.

This makes precise the physicists' identity

δ (x) = \frac{1}{2 π} \int_{- \infty}^{\infty} e^{i ξ x} d ξ,

⟳interpret as the inverse transform of

δ = 1

; the divergent integral is meaningful only in

S^{'}

, as a distributional limit of

\frac{1}{2 π} \int_{- R}^{R}

(a Dirichlet kernel).

Paul Adrien Maurice Dirac | 1902–1984

Dirac introduced the delta in his 1927 paper on quantum mechanics and codified it in The Principles of Quantum Mechanics (1930), treating $δ (x)$ as an "improper function" satisfying $\int δ (x) d x = 1$ and $\int f (x) δ (x - a) d x = f (a)$ . He explicitly warned it was "not a quantity which can be generally used in mathematical analysis like an ordinary function, but its use must be confined to certain simple types of expression." Dirac needed it to express the eigenfunction expansion of operators with continuous spectra — the resolution of the identity $\sum_{n} ∣ n ⟩ ⟨ n ∣ + \int ∣ x ⟩ ⟨ x ∣ d x = 1$ — where the orthonormality $⟨ x ∣ x^{'} ⟩ = δ (x - x^{'})$ replaces the Kronecker delta of discrete bases. His pragmatic confidence that the manipulations were sound was vindicated two decades later by Schwartz.

Building Intuition

Think of a distribution as an idealized measuring device. A locally integrable function $f$ "measures" a test signal $φ$ by computing the weighted average $\int f φ$ . The delta is the perfect point sensor: it reports the value of the signal exactly at one point, ignoring everything else. You can never build such a sensor from an ordinary function (any real average smears over a region), but you can build a sequence of ever-sharper sensors — taller, narrower bumps of unit area — whose readings converge to the ideal one. The distribution is the limit of the readings, not of the graphs.

For the derivative, integration by parts is the whole story: you can't differentiate a spike pointwise, but you can hand the derivative to the smooth probe instead. $δ^{'}$ is a "dipole sensor" that reports $- φ^{'} (0)$ — the negative slope at the origin — which is why a derivative of a spike behaves like a balanced $+ / -$ pair infinitely close together.

Approximate identities: Gaussian bumps $\eta_\varepsilon$ sharpening to the delta spike as $\varepsilon\to0$, each of unit area

Worked Examples

Charge density of a point charge and its potential

Given: A point charge

q = 2 \times 1 0^{- 9} C

at the origin in 3D. Model its density as

ρ (x) = q δ^{(3)} (x)

. Step 1 — Total charge. Integrate the density:

Q = \int_{R^{3}} ρ d^{3} x = q \int δ^{(3)} (x) d^{3} x = q .

⟳

⟨ δ^{(3)}, 1 ⟩

via the

n = 3

sampling property at

x = 0

with

φ \equiv 1

on the support.

Step 2 — Poisson equation. The electrostatic potential satisfies

- ε_{0} \nabla^{2} ϕ = ρ

. The fundamental solution of the Laplacian in 3D is known:

\nabla^{2} (\frac{- 1}{4 π ∣ x ∣}) = δ^{(3)} (x) .

⟳

1/ (4 π r)

is the Green's function of

- \nabla^{2}

; verify by integrating

\nabla^{2} (1/ r) = 0

off the origin plus Gauss's flux

\oint \nabla (1/ r) \cdot d S = - 4 π

Step 3 — Solve. Hence

ϕ (x) = \frac{q}{4 π ε _{0} ∣ x ∣}

. Numerical answer at $r = 1 m$ (

ε_{0} = 8.854 \times 1 0^{- 12} F/m

ϕ = \frac{2 \times 1 0 ^{- 9}}{4 π ( 8.854 \times 1 0 ^{- 12} ) ( 1 )} \approx 17.98 V

The delta cleanly produces the Coulomb potential as the Green's function evaluation.

Derivative of a jump: distributional bookkeeping

Given:

f (x) = {x^{2} 3 x x < 1 x \geq 1

, which has a jump of size

3 - 1 = 2

x = 1

(since

lim_{x \to 1^{-}} = 1

lim_{x \to 1^{+}} = 3

). Step 1 — Classical derivative away from the jump.

f_{cl}^{'} (x) = 2 x

for

x < 1

and

= 3

for

x > 1

. Step 2 — Jump-derivative formula. For a piecewise

C^{1}

function with jumps

[f]_{x_{j}} = f (x_{j}^{+}) - f (x_{j}^{-})

f^{'} = {f^{'}}_{cl} + j \sum [f]_{x_{j}} δ_{x_{j}} .

⟳integrate

⟨ f^{'}, φ ⟩ = - \int f φ^{'}

by parts on each smooth piece; the boundary terms collect into

\sum_{j} [f]_{x_{j}} φ (x_{j})

, which is the action of

[f]_{x_{j}} δ_{x_{j}}

Step 3 — Assemble. Here the single jump is

[f]_{1} = 2

f^{'} (x) = (2 x 1_{x < 1} + 3 1_{x > 1}) + 2 δ (x - 1)

Check by re-integrating: pairing

f^{'}

against a

φ

with

φ (1) = 5

, the delta contributes

2 \cdot 5 = 10

⟨ f^{'}, φ ⟩

, exactly the missing boundary term a naive classical derivative would drop.

Common Misconceptions

$δ$ is not a function. Writing "

δ (0) = \infty

" is meaningless;

δ

has no pointwise values. It only makes sense inside a pairing

⟨ δ, φ ⟩

or under an integral sign understood as that pairing. Products of distributions are generally undefined. There is no associative product making

D^{'}

an algebra (Schwartz's impossibility result). In particular

δ^{2}

and

δ (x) H (x)

are ill-defined; attempting

δ \cdot δ

leads to contradictions. Only special products like (smooth function)

\times

(distribution) are legitimate. $δ (x) x = 0$ but you cannot "divide" to conclude $δ (x) = 0$ . Division is not generally available; the equation

x T = 0

has general solution

T = c δ

, not

T = 0

. Convergence is weak- $*$ , not pointwise.

η_{ε} \to δ

D^{'}

even though

η_{ε} (x) \to 0

for every fixed

x \neq = 0

and

η_{ε} (0) \to \infty

. The graphs do not converge to anything sensible; only the integrals against test functions converge.

Non-uniqueness in $xT=0$

Suppose a student "solves" $x T (x) = 0$ by dividing: $T = 0/ x = 0$ . But take $T = 5 δ$ . Then for any $φ \in D$ ,

⟨ x \cdot 5 δ, φ ⟩ = ⟨ 5 δ, x φ ⟩ = 5 (x φ)_{x = 0} = 0,

⟳multiplication by the smooth function

x

is transferred onto

φ

, and

x φ

vanishes at

0

regardless of

φ

so $x \cdot 5 δ = 0$ as a distribution, yet $5 δ \neq = 0$ . Thus the "solution" $T = 0$ is not unique; the kernel of "multiply by $x$ " is the line $C δ$ . This is exactly why naive cancellation fails in distribution theory.

Connections

The delta sits at the crossroads of several theories. As a measure, it is the unit point mass $δ_{a}$ in Measure Theory — a Radon measure, hence a distribution of order zero; the Lebesgue Integral against a continuous function reproduces $φ (a)$ . Through the Fourier Transform, $δ = 1$ links localization in space to spreading in frequency, the cleanest instance of the uncertainty principle. The delta is the identity for convolution, $f * δ = f$ , which is the engine behind Green's Functions and Fundamental Solutions: solving $Lu = f$ reduces to finding $E$ with $L E = δ$ , then $u = E * f$ . In Sobolev Spaces $H^{s}$ , $δ \in H^{s} (R^{n})$ exactly when $s < - n /2$ , quantifying how singular it is. Finally, in Functional Analysis, $D^{'}$ is the strong dual of a nuclear Fréchet/LF space, and the delta is the evaluation functional whose representability (or lack thereof) characterizes Reproducing Kernel Hilbert Spaces.

Real-World Applications

Green's functions in electromagnetism. The retarded potential of a point source uses $\nabla^{2} G - c^{- 2} \partial_{t}^{2} G = - δ^{(3)} (x) δ (t)$ ; the solution underlies antenna theory and the Liénard–Wiechert potentials.

Quantum mechanics — continuous spectra. Position eigenstates obey $⟨ x ∣ x^{'} ⟩ = δ (x - x^{'})$ ; momentum-position completeness and the propagator $K (x, t; x^{'}, 0)$ as a kernel rest on delta normalization.

Signal processing — sampling. The Dirac comb $III_{T} (t) = \sum_{n} δ (t - n T)$ models ideal sampling; its self-dual Fourier transform (another comb) yields the Nyquist–Shannon sampling theorem and aliasing analysis.

Probability and statistics. A point mass $δ_{a}$ represents a deterministic random variable; mixed distributions (e.g. censored data, zero-inflated models) are sums of a density and weighted deltas.

Structural and mechanical engineering. A point load on a beam is $w (x) = P δ (x - x_{0})$ ; the Euler–Bernoulli equation $E I u^{''''} = w$ integrates this into the deflection curve and influence functions.

Control theory / LTI systems. The impulse response $h (t)$ is the system's response to $δ (t)$ ; convolution $y = h * u$ characterizes every linear time-invariant system, and the transfer function is $h$ .

Fracture and crack mechanics / fluid dynamics. Vortex sheets, point vortices, and dislocation lines are modeled by deltas supported on lower-dimensional sets (single-layer and double-layer potentials).

Historical Development

1822 — Fourier: In Théorie analytique de la chaleur, Fourier writes the inversion formula $f (x) = \frac{1}{2 π} \iint f (α) e^{i ξ (x - α)} d α d ξ$ , implicitly containing the delta as the inner integral.
1828 — Green / 1830s — Poisson: Green's functions for potential theory introduce source terms that are, in retrospect, delta-supported.
1880s–1890s — Heaviside: Operational calculus uses the step function $H$ and its "derivative" (the impulse) freely; ridiculed by contemporaries for lack of rigor but spectacularly effective.
1894 — Kirchhoff / 1900s — physicists use impulsive idealizations in optics and mechanics.
1927 — Dirac: Introduces $δ$ explicitly in The physical interpretation of the quantum dynamics (Proc. Roy. Soc.); systematized in his 1930 textbook.
1936 — Sobolev: Defines generalized derivatives and "fonctionnelles" to solve hyperbolic PDEs — the first rigorous distribution-like theory, with what we now call Sobolev spaces.
1944–1950 — Schwartz: Laurent Schwartz develops the full théorie des distributions, unifying duality, Fourier transform, convolution, and the kernel theorem; publishes Théorie des distributions (1950–1951).
1950 — Fields Medal awarded to Schwartz, largely for distributions.
1955–1961 — Gelfand–Shilov, Hörmander: Generalized Functions (multivolume) and Hörmander's wavefront-set / microlocal refinement extend the theory; Mikio Sato's hyperfunctions (1959) give an alternative sheaf-theoretic foundation.

Subtleties & Limits

Multiplication of distributions. Schwartz's impossibility theorem shows no associative, commutative product on

D^{'}

extends pointwise multiplication of continuous functions while keeping

1

as unit and

δ

behaving consistently. Yet physics (renormalization in QFT, nonlinear PDEs with rough data) demands products like

δ^{2}

H δ

. Partial resolutions include Colombeau algebras (sequences modulo negligibles, giving an associative differential algebra containing

D^{'}

, but with a delicate "association" relation), Hörmander's product valid when wavefront sets do not collide (microlocal condition

WF (u) + WF (v) \neq ∋ 0

), and paraproducts / rough path & regularity structures (Martin Hairer's 2014 Fields Medal work) for stochastic PDEs. Determining the "right" canonical product remains domain-dependent: there is no universal answer, only context-specific algebras and renormalization schemes that fix the ambiguous finite parts. The frontier question — a fully canonical, physically faithful nonlinear theory of singular distributions — stays open.

The Dirac Delta Function and Distributions

The Core Insight

Formal Statement

Test Function Spaces

Distributions and the Delta

Distributional Derivative

Heaviside, Delta, and Localization

Why It Holds / Derivation

Delta as a Limit of Approximate Identities

Fourier Transform of the Delta

Paul Adrien Maurice Dirac | 1902–1984

Building Intuition

Approximate identities: Gaussian bumps $\eta_\varepsilon$ sharpening to the delta spike as $\varepsilon\to0$, each of unit area

Worked Examples

Charge density of a point charge and its potential

Derivative of a jump: distributional bookkeeping

Common Misconceptions

Non-uniqueness in $xT=0$

Connections

Real-World Applications

Historical Development

Subtleties & Limits

See Also

Before this

Where this leads