The Dirac Delta Function

The Dirac delta $\delta(x)$ is the most useful object in mathematical physics that is not, strictly speaking, a function at all. Paul Dirac introduced it in the late 1920s as a compact symbol for an idealized point source — a unit of charge, mass, or probability squeezed onto a single point — and wrote down the impossible-looking recipe $\delta(x)=0$ for $x\neq 0$ , $\delta(0)=\infty$ , with total area $\int\delta\,\mathrm{d}x=1$ . No ordinary Lebesgue-integrable function can meet those conditions: a function vanishing everywhere except one point integrates to zero, not one. The delta is therefore not defined by its pointwise values — those are a fiction — but by what it does inside an integral against a smooth test function. That single operational rule, the sifting property $\int\delta(x-a)f(x)\,\mathrm{d}x=f(a)$ , is complete. From it flow every identity the physicist needs: the scaling law $\delta(ax)=\delta(x)/|a|$ , the root-counting composition formula, the derivative rules, the Fourier integral representation, and the fact that $\delta$ is the identity for convolution. The same rule explains why the delta manufactures Green's functions and why it appears the instant one inverts a Fourier or Laplace transform. Laurent Schwartz's theory of distributions later put the apparatus on rigorous footing: $\delta$ is a continuous linear functional, $\delta[f]=f(0)$ , on a space of well-behaved test functions, where it can be differentiated and transformed as freely as any ordinary function. In practice one computes with the delta as the narrow limit of honest functions — Gaussians, Lorentzians, sinc kernels — that grow tall and thin while pinning their area at one.

Key takeaways

The Dirac delta $\delta(x)$ is not an ordinary function but a generalized function (distribution): it is defined entirely by its action under integration, the sifting property $\int_{-\infty}^{\infty}\delta(x-a)f(x)\,\mathrm{d}x=f(a)$ .
Heuristically it is a unit impulse — zero everywhere except $x=0$ , infinite there, with unit total area $\int_{-\infty}^{\infty}\delta(x)\,\mathrm{d}x=1$ — but that picture is shorthand, never a calculation tool.
It is the narrow limit of nascent delta sequences (Gaussians, Lorentzians, sinc kernels) and the distributional derivative of the Heaviside step, $\delta(x)=H'(x)$ .
Its Fourier transform is the constant 1, $\int\delta(x)e^{-ikx}\,\mathrm{d}x=1$ , which makes it the identity for convolution and the engine of Green's-function methods.
Rigour comes from distribution theory: $\delta$ is a continuous linear functional $\delta[f]=f(0)$ , sidestepping every contradiction of the pointwise picture.
Scaling, composition $\delta(g(x))=\sum_i\delta(x-x_i)/|g'(x_i)|$ , and derivatives all follow from the defining functional, never from manipulating values at a point.
Key idea: the delta function is defined by what it does, not what it equals — $\int\delta(x-a)f(x)\,\mathrm{d}x=f(a)$ — and every identity is derived from this sifting action against smooth test functions, never by reasoning about its pointwise values.

The core insight

The delta function is defined by what it does, not what it equals: $\int\delta(x-a)f(x)\,\mathrm{d}x=f(a)$ . Every identity is derived from this sifting action against smooth test functions, never by reasoning about its pointwise values.

Where this lives today

The delta is the working notation of active physics. In quantum field theory every propagator is a Green's function of a delta source — the Feynman propagator solves $(\Box+m^2)G=-\delta^4(x-x')$ — and lattice-QCD simulations compute exactly these delta-source responses on a discretized spacetime to extract hadron masses. In many-body theory the single-particle Green's function, again defined through a delta, is what angle-resolved photoemission maps out as a spectral function whose quasiparticle peaks are broadened Lorentzians — nascent deltas of finite lifetime. In signal processing the Dirac comb $\sum_n\delta(t-nT)$ models every analog-to-digital converter, and its transform (another comb) is the Nyquist sampling theorem in one line; the same comb in reciprocal space is why an ideal crystal diffracts X-rays into sharp Bragg peaks, with the finite width of a real peak reporting the finite size of the crystallite. And in computational physics it is regularized on purpose: smoothed-particle hydrodynamics and the immersed-boundary method replace point sources with narrow, unit-area kernels — nascent delta sequences made literal. Wherever a point source, an instantaneous kick, or an ideal sample appears, the delta does the work.

The physical picture

Imagine spreading one coulomb of charge uniformly over a tiny interval of width $\epsilon$ centred on the origin. The charge density on that interval is $1/\epsilon$ — make the interval narrower and the density must climb to keep the total charge fixed at one. As $\epsilon\to 0$ the density spikes toward infinity while the area under it never budges from one. The delta is the idealized end of that process: zero width, infinite height, unit area. It is useful because real point particles, instantaneous hammer-blows, and sharp spectral lines are all well modelled by a narrow, tall bump of fixed area.

Figure

A unit point charge regularized as a box of width ε and height 1/ε — the area stays exactly 1 as ε shrinks (ε = 1.0 gives height 1.0; ε = 0.5 gives height 2.0), and the ε → 0 limit is the Dirac delta δ(x), of zero width, infinite height, unit area

The sifting property makes the delta a continuous analogue of the Kronecker delta. Where the discrete sum $\sum_j\delta_{ij}a_j=a_i$ picks out one component of a vector, the integral $\int\delta(x-a)f(x)\,\mathrm{d}x=f(a)$ picks out one value of a function. The delta is the kernel of the identity operator on function space, exactly as $\delta_{ij}$ is the identity matrix. Read this way, $\delta$ is a measuring device, not a number: it only makes sense once fed a smooth test function and asked to report a single value. Any time you catch yourself asking "what is $\delta(0)$ ?" you have wandered outside the domain where the object is defined — the question is like asking for the colour of the number seven. Many honest functions approach $\delta$ , and the choice of approximant encodes different physics — a Gaussian for diffusion and finite measurement resolution, a Lorentzian for the line shape of a damped resonance, a sinc kernel for a Fourier integral truncated at finite bandwidth. All they share is unit area and vanishing width. The second figure shows the Gaussian family narrowing: the peak climbs as $1/(\epsilon\sqrt{\pi})$ while the area stays locked at one, and in the limit the curve sifts any test function down to its value at the origin.

Figure

The Gaussian nascent-delta sequence δ_ε(x) = e^(−x²/ε²)/(ε√π) at ε = 1.0, 0.5, 0.25 — the peak height 1/(ε√π) climbs through 0.56, 1.13, 2.26 as the width falls, yet every curve has unit area, so the ε → 0 limit sifts any smooth f down to f(0)

The core equations

Everything begins with the sifting property. It is not a theorem about the delta; it is the delta, the one statement from which every other formula is derived by treating $\delta$ as a functional and never as a set of values.

\int_{-\infty}^{\infty}\delta(x-a)\,f(x)\,\mathrm{d}x=f(a) \qquad (1)

where: $\delta(x-a)$ — Dirac delta centred at $x=a$ ; $f(x)$ — smooth test function, continuous at $a$ ; $a$ — the sampling point. Reading: the delta reaches into the integral, samples $f$ at the one point $x=a$ , and discards everything else. The heuristic point-source description is the picture behind the first figure. It is the intuition, and it is where students go wrong: the two conditions are mutually contradictory for any ordinary function and are only ever a mnemonic for the sifting rule above.

\delta(x)=0\ (x\neq 0),\qquad \int_{-\infty}^{\infty}\delta(x)\,\mathrm{d}x=1 \qquad (2)

where: $x$ — the real variable. Reading: "all the area sits at one point of zero width", a statement no genuine function satisfies, which is exactly why the distributional definition is needed. Rescaling the argument stretches the spike horizontally, and unit area forces a compensating change in height. The absolute value is the whole content: area cannot go negative even when $a$ does.

\delta(ax)=\frac{1}{|a|}\,\delta(x) \qquad (3)

where: $a$ — a nonzero real constant. Reading: squeezing the argument by $a$ multiplies the delta's weight by $1/|a|$ ; as a special case $\delta(-x)=\delta(x)$ , so the delta is even. The composition formula generalizes scaling to any smooth $g$ with simple zeros. Near each root $g$ looks linear with slope $g'(x_i)$ , so each root contributes a scaled delta, and the contributions add.

\delta(g(x))=\sum_{i}\frac{\delta(x-x_i)}{|g'(x_i)|} \qquad (4)

where: $g(x)$ — smooth function with simple zeros at the points $x_i$ ; $x_i$ — the roots, $g(x_i)=0$ with $g'(x_i)\neq 0$ . Reading: the delta of a function fires once at each root, weighted by the reciprocal of the local slope. The delta can be differentiated as a distribution. There is no pointwise derivative to speak of; instead $\delta'$ is defined by moving the derivative onto the test function through integration by parts.

\int_{-\infty}^{\infty}\delta^{(n)}(x-a)\,f(x)\,\mathrm{d}x=(-1)^{n}f^{(n)}(a) \qquad (5)

where: $\delta^{(n)}$ — the $n$ -th distributional derivative of the delta; $f^{(n)}(a)$ — the $n$ -th derivative of the test function at $a$ . Reading: each derivative that lands on the delta is repaid as a derivative on $f$ , and each transfer flips the sign — in particular $\int\delta'(x-a)f(x)\,\mathrm{d}x=-f'(a)$ . Two algebraic companions follow the same way: $x\,\delta(x)=0$ and $x\,\delta'(x)=-\delta(x)$ . The Fourier integral representation is the delta's second face and the reason it dominates transform theory. It says the delta is the inverse Fourier transform of the constant 1 — an equal-weight superposition of every frequency.

\delta(x)=\frac{1}{2\pi}\int_{-\infty}^{\infty}e^{ikx}\,\mathrm{d}k \qquad (6)

where: $k$ — the wavenumber (frequency) variable. Reading: perfect localization in $x$ demands complete delocalization in $k$ — the extreme case of the position–wavenumber uncertainty relation — and the transform pair $\int\delta(x-a)e^{-ikx}\,\mathrm{d}x=e^{-ika}$ carries all frequencies at unit amplitude. Finally, the delta is the identity for convolution and, in Schwartz's language, a continuous linear functional. These two lines are the operational and the rigorous statement of the same fact.

(f*\delta)(x)=\int_{-\infty}^{\infty}f(y)\,\delta(x-y)\,\mathrm{d}y=f(x),\qquad \delta[\,f\,]=f(0),\ f\in\mathcal{S}(\mathbb{R}) \qquad (7)

where: $f*\delta$ — convolution of $f$ with the delta; $\mathcal{S}(\mathbb{R})$ — the Schwartz space of rapidly decreasing smooth test functions; $\delta[\,\cdot\,]$ — a continuous linear functional. Reading: convolving with $\delta$ returns the function unchanged, and the physicist's integral notation is nothing but the action of the functional $\delta[f]=f(0)$ .

Derivation AThe composition formula for a function with simple zeros

The goal is to prove that $\delta(g(x))$ resolves into a sum of scaled deltas, one at each simple root of $g$ , weighted by the reciprocal slope $1/|g'(x_i)|$ . This is the identity that lets a delta whose argument is a whole function — an on-shell condition $\delta(p^2-m^2)$ , a dispersion relation, an energy shell — be reduced to ordinary point deltas one can integrate. The strategy is the only one the delta permits: never touch pointwise values, but integrate against an arbitrary smooth test function $f$ , localize the support to neighbourhoods of the zeros, change variables so $g$ becomes the integration variable, and read off the weight.

Step 1. Integrate against a test function and split the support. Because $\delta(g(x))$ is supported only where $g(x)=0$ , write $\int\delta(g(x))f(x)\,\mathrm{d}x=\sum_i\int_{I_i}\delta(g(x))f(x)\,\mathrm{d}x$ , where each $I_i$ is a small interval containing exactly one simple root $x_i$ . Why: away from the roots $g$ is nonzero, so $\delta(g(x))$ contributes nothing there; all the action is concentrated near the zeros, and simple roots are isolated so the intervals do not overlap.

Step 2. Change variable to $u=g(x)$ on each interval. Since the root is simple, $g'(x_i)\neq 0$ , so $g$ is strictly monotone on $I_i$ and the substitution $u=g(x)$ , $\mathrm{d}u=g'(x)\,\mathrm{d}x$ , is one-to-one. As $x$ sweeps $I_i$ , $u$ sweeps an interval around $0$ . Why: monotonicity is what makes the change of variable legitimate; a double root would fail here and needs separate, more delicate treatment.

Step 3. Apply sifting in the new variable. The integral becomes $\int\delta(u)\,f(x(u))\,\frac{\mathrm{d}u}{|g'(x(u))|}$ , and sifting at $u=0$ selects $x=x_i$ , giving the single contribution $f(x_i)/|g'(x_i)|$ . The absolute value appears because reversing the orientation for $g'<0$ flips the limits, and restoring them costs a sign that the modulus absorbs. Why: this is the one place the direction of the substitution matters, and taking the modulus is exactly what keeps the delta's weight positive.

Step 4. Sum over the roots. Adding the contributions gives $\int\delta(g(x))f(x)\,\mathrm{d}x=\sum_i f(x_i)/|g'(x_i)|=\int f(x)\sum_i\delta(x-x_i)/|g'(x_i)|\,\mathrm{d}x$ . Since $f$ was arbitrary, the two functionals are equal, which is the boxed identity. Why: matching against every test function is precisely what "equal as distributions" means — it is the rigorous replacement for comparing values. *Two quick checks: the special case $g(x)=ax$ has the single root $x=0$ with $|g'|=|a|$ , reproducing $\delta(ax)=\delta(x)/|a|$ ; and $g(x)=x^2-a^2$ gives roots $\pm a$ with $|g'|=2a$ , hence $\delta(x^2-a^2)=[\delta(x-a)+\delta(x+a)]/2a$ .*

Go deeper · Derivation B · 1 minThe Fourier integral representation of the delta

Derivation BThe Fourier integral representation of the delta

The goal is to establish $\delta(x)=\frac{1}{2\pi}\int_{-\infty}^{\infty}e^{ikx}\,\mathrm{d}k$ , an integral that does not converge in the ordinary sense and only makes sense as a distribution. This representation is the hinge of the transform machinery: it states the Fourier inversion theorem, makes the delta the completeness relation of the plane-wave basis, and is why momentum eigenstates satisfy $\langle p|p'\rangle\propto\delta(p-p')$ . The honest route is to insert a convergence factor $e^{-\epsilon|k|}$ that tames the integral, evaluate it exactly to a Lorentzian, and show the Lorentzian is itself a nascent delta sequence as the regulator is removed.

Step 1. Regularize the divergent integral. Multiply the integrand by $e^{-\epsilon|k|}$ with $\epsilon>0$ and split the range at $k=0$ : $\frac{1}{2\pi}\int_{-\infty}^{\infty}e^{ikx-\epsilon|k|}\,\mathrm{d}k=\frac{1}{2\pi}\left(\frac{1}{\epsilon-ix}+\frac{1}{\epsilon+ix}\right)$ . Why: without the regulator the integral is meaningless; the factor $e^{-\epsilon|k|}$ makes each half converge while approaching 1 pointwise as $\epsilon\to 0$ , so it changes nothing in the limit.

Step 2. Combine into a Lorentzian. Adding the two complex fractions over a common denominator gives $\frac{1}{2\pi}\cdot\frac{2\epsilon}{\epsilon^2+x^2}=\frac{1}{\pi}\frac{\epsilon}{x^2+\epsilon^2}$ . This is a Lorentzian of width $\epsilon$ , peak height $1/(\pi\epsilon)$ , and — crucially — unit area for every $\epsilon$ , since $\frac{1}{\pi}\int\frac{\epsilon\,\mathrm{d}x}{x^2+\epsilon^2}=\frac{1}{\pi}[\arctan(x/\epsilon)]_{-\infty}^{\infty}=1$ . Why: the exact evaluation turns an abstract regulated integral into a concrete, area-normalized bump whose limit we already understand.

Step 3. Remove the regulator. As $\epsilon\to 0^+$ the Lorentzian grows tall and narrow with fixed unit area, so $\lim_{\epsilon\to 0^+}\frac{1}{\pi}\frac{\epsilon}{x^2+\epsilon^2}=\delta(x)$ in the distributional sense, establishing equation (6). Why: this is precisely a nascent delta sequence — the defining behaviour of tall, thin, unit-area families — so the regulated integral converges to the delta, not to any ordinary function. *A second route to the same result: truncating the integral at finite bandwidth gives $\frac{1}{2\pi}\int_{-K}^{K}e^{ikx}\,\mathrm{d}k=\frac{\sin(Kx)}{\pi x}$ , the sinc (Dirichlet) kernel of unit area, and $\lim_{K\to\infty}\int\frac{\sin(Kx)}{\pi x}f(x)\,\mathrm{d}x=f(0)$ using $\int_{-\infty}^{\infty}(\sin u)/u\,\mathrm{d}u=\pi$ — confirming the delta by a second route.*

Go deeper · Derivation C · 1 minGreen's function as the response to a delta source

Derivation CGreen's function as the response to a delta source

The goal is to show why solving one equation with a delta on the right, $L_x\,G(x,x')=\delta(x-x')$ , hands you the solution of the inhomogeneous problem $L_x\,u(x)=s(x)$ for any forcing $s$ , through one convolution $u(x)=\int G(x,x')s(x')\,\mathrm{d}x'$ . This is the most consequential use of the delta in physics: it converts a differential equation into an integral, and it underlies electrostatic potentials, quantum field propagators, and the impulse response of every linear system. The argument rests entirely on the sifting property and the linearity of $L$ .

Step 1. Define the Green's function. Let $G(x,x')$ solve $L_x\,G(x,x')=\delta(x-x')$ , subject to the problem's boundary conditions. Physically $G$ is the response measured at $x$ to a unit point source placed at $x'$ . Why: the delta is the mathematical stand-in for a point source, so $G$ is literally "the field of a point source", the elementary building block of any extended source.

Step 2. Decompose the source into point sources. By sifting, any forcing can be written $s(x)=\int_{-\infty}^{\infty}\delta(x-x')\,s(x')\,\mathrm{d}x'$ — a continuous superposition of point sources of strength $s(x')$ . Why: this is the sifting property read backwards; it expresses an arbitrary function as a weighted pile of deltas, which is exactly the set of sources we already know how to respond to.

Step 3. Superpose the responses. Because $L$ is linear, the response to the sum of sources is the sum of the responses, so $u(x)=\int_{-\infty}^{\infty}G(x,x')\,s(x')\,\mathrm{d}x'$ . Why: linearity is the license to add solutions; each point source $\delta(x-x')s(x')$ produces $G(x,x')s(x')$ , and integrating collects them.

Step 4. Verify by applying the operator. Act with $L_x$ under the integral: $L_x u(x)=\int (L_x G(x,x'))s(x')\,\mathrm{d}x'=\int\delta(x-x')s(x')\,\mathrm{d}x'=s(x)$ . The delta's sifting property closes the loop and recovers $L u=s$ . Why: the final sift is what proves the construction actually solves the equation, not merely resembles a solution. *Two concrete examples: for $L=\mathrm{d}^2/\mathrm{d}x^2$ on the line, $G(x,x')=\tfrac12|x-x'|$ gives $G'=\tfrac12\,\mathrm{sgn}(x-x')$ and $G''=\delta(x-x')$ ; in three dimensions $\nabla^2(-1/4\pi|\mathbf{r}-\mathbf{r}'|)=\delta^3(\mathbf{r}-\mathbf{r}')$ , the Coulomb Green's function.*

Limiting cases

The delta is the common narrow limit of many unit-area families, and each family is the correct model for a different physical regularization. The table collects the standard four.

Case	Result	Meaning
Gaussian	$\delta(x)=\lim_{\epsilon\to 0^+}\frac{1}{\epsilon\sqrt{\pi}}\,e^{-x^2/\epsilon^2}$	Smooth and infinitely differentiable; the model for diffusion, heat spreading, and finite measurement resolution.
Lorentzian	$\delta(x)=\lim_{\epsilon\to 0^+}\frac{1}{\pi}\frac{\epsilon}{x^2+\epsilon^2}$	Heavy-tailed, full width at half maximum $2\epsilon$ ; the natural line shape of a damped oscillator or a decaying state (resonance width $\epsilon$ ).
Sinc / Dirichlet	$\delta(x)=\lim_{K\to\infty}\frac{\sin(Kx)}{\pi x}$	Oscillatory, not positive; arises whenever a Fourier inversion is truncated at finite bandwidth $K$ , and drives Fourier-series convergence.
Heaviside derivative	$\delta(x)=\frac{\mathrm{d}}{\mathrm{d}x}H(x),\ H(x)=\begin{cases}0&x<0\\1&x>0\end{cases}$	In the distributional sense: differentiating a unit step produces a unit impulse exactly at the jump.

Common trapTreating δ as an ordinary function with a value at a point

The single most common error is to manipulate $\delta(x)$ as though it had genuine pointwise values — to write $\delta(0)=\infty$ and compute with that infinity, or to form the product $\delta(x)^2$ of two deltas at the same point. Both are meaningless: $\delta$ is a distribution defined only through its action on test functions, and $\delta(x)^2$ has no distributional meaning at all. Symbolic identities such as $\delta(ax)=\delta(x)/|a|$ , $x\,\delta'(x)=-\delta(x)$ , or the composition rule are true only as statements about integrals against a test function, and the only safe way to check one is to integrate both sides against an arbitrary smooth $f$ and confirm they agree. Whenever a calculation seems to require the number $\delta(0)$ , it has left the domain where the delta is defined, and a nascent-delta sequence should be reinstated until the offending quantity cancels or is exposed as genuinely divergent.

Worked examples

1 · The composition formula with numbers (confidence builder)

Evaluate $I=\int_{-\infty}^{\infty}\delta(x^2-4)\,(x^2+1)\,\mathrm{d}x$ using the composition rule.

Step 1. Locate the roots and their slopes. Set $g(x)=x^2-4$ ; it vanishes at $x=\pm 2$ , and $g'(x)=2x$ gives $|g'(2)|=|g'(-2)|=4$ . Both roots are simple, so the composition formula applies with no special handling. The two roots are $x=\pm 2$ , each carrying weight $1/4$ .

Step 2. Resolve the delta. By equation (4), $\delta(x^2-4)=\frac{1}{4}\big[\delta(x-2)+\delta(x+2)\big]$ . The single delta of a quadratic has become two point deltas, one at each root, each softened by the local slope $4$ .

Step 3. Sift the test function. With $f(x)=x^2+1$ , sifting at each root gives $f(2)=f(-2)=5$ , so $I=\frac{1}{4}\big[f(2)+f(-2)\big]=\frac{1}{4}(5+5)=\mathbf{2.5}$ . The integral collapses to summing $f$ over the roots and dividing by the slope — no antiderivative, no limit, just the sifting rule applied twice. I = 2.5.

2 · An impulsive force and the momentum it delivers

A mass $m=0.5\ \mathrm{kg}$ sits at rest on a frictionless track. At $t=t_0$ it is struck by a force $F(t)=I_0\,\delta(t-t_0)$ with impulse strength $I_0=3\ \mathrm{N\,s}$ . Find the momentum and speed it acquires, and comment on the acceleration.

Step 1. Integrate the force to get the impulse. Newton's second law gives the momentum change as $\Delta p=\int F(t)\,\mathrm{d}t=\int I_0\,\delta(t-t_0)\,\mathrm{d}t=I_0=\mathbf{3\ kg\,m/s}$ by sifting the constant $1$ . The delta delivers the whole impulse instantly, independent of the shape of any real, finite-duration blow of the same area.

Step 2. Convert to a velocity. Starting from rest, $v=\Delta p/m=3/0.5=\mathbf{6\ m/s}$ . The mass leaves the interaction moving at $6$ metres per second; the speed depends only on the impulse and the mass, never on how the impulse was distributed in time.

Step 3. Interpret the acceleration. During the strike the acceleration is $a(t)=F/m=(I_0/m)\,\delta(t-t_0)$ , itself a delta — formally infinite, yet perfectly integrable, with $\int a\,\mathrm{d}t=6\ \mathrm{m/s}$ finite. This is the idealization at work: an infinite acceleration acting for zero time still transfers a finite, well-defined velocity. Δp = 3 kg·m/s, v = 6 m/s, with a delta-function acceleration of finite time integral.

Physics of the portrait

The delta function is one of the rare cases where a physicist's bold, "illegal" notation forced mathematicians to build an entirely new theory to justify it — and the theory, distributions, turned out to be one of the great unifying frameworks of twentieth-century analysis.

Origins · 1 minPaul Dirac and how this was discovered

Paul Dirac

1902–1984 · the delta function and the foundations of quantum mechanics

Paul Adrien Maurice Dirac, trained first as an electrical engineer in Bristol and then as a mathematician, brought an engineer's willingness to compute with objects before anyone could say what they were. In his 1927 paper on the quantum theory of radiation and the 1930 masterwork The Principles of Quantum Mechanics, he introduced the "improper function" $\delta(x)$ to handle continuous spectra the way the Kronecker delta handles discrete ones — so that eigenstates could be normalized, $\langle x|x'\rangle=\delta(x-x')$ , and completeness $\int|x\rangle\langle x|\,\mathrm{d}x=\mathbb{1}$ written down. He knew it was not a function; he called it a convenient notation and noted any rigorous statement could be rephrased with limits. Mathematicians were uneasy for two decades until Laurent Schwartz's theory of distributions (1945–1950) gave $\delta$ and all its derivatives a rigorous home as continuous linear functionals — vindicating Dirac's instinct. Dirac shared the 1933 Nobel Prize with Schrödinger; his delta, invented as a bookkeeping symbol, is now inseparable from Green's functions, quantum field propagators, and signal theory. It is the model example of physics running ahead of rigour and rigour catching up to enshrine it.

History

1822

— Fourier's integral theorem implicitly encodes the sifting identity in its double integral

1893

— Heaviside's operational calculus manipulates impulsive functions freely, without justification

1927

— Dirac introduces δ(x) to normalize continuous spectra in quantum mechanics

1936

— Sobolev defines generalized functions through weak derivatives for wave equations

1950

— Schwartz's théorie des distributions puts δ and its derivatives on rigorous footing

Going further

Texts. Stone & Goldbart, Mathematics for Physics (CUP), chapters on distributions and Green's functions · Arfken & Weber, Mathematical Methods for Physicists, 7th ed., §1.11 (the Dirac delta) · Hassani, Mathematical Physics, 2nd ed., Ch. 6–7 (generalized functions) · Byron & Fuller, Mathematics of Classical and Quantum Physics, §5.4 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., §13.1 · Lighthill, Introduction to Fourier Analysis and Generalised Functions · Schwartz, Théorie des distributions. Key numerical anchors used throughout this article: the composition formula $\delta(x^2-a^2)=[\delta(x-a)+\delta(x+a)]/2a$ and the worked value $\int\delta(x^2-4)(x^2+1)\,\mathrm{d}x=2.5$ ; the Lorentzian regularization $\frac{1}{2\pi}\int e^{ikx-\epsilon|k|}\,\mathrm{d}k=\frac{1}{\pi}\epsilon/(x^2+\epsilon^2)$ of unit area; the Dirichlet-kernel limit using $\int(\sin u)/u\,\mathrm{d}u=\pi$ ; the Gaussian peak heights $1/(\epsilon\sqrt{\pi})=0.56,\,1.13,\,2.26$ at $\epsilon=1.0,\,0.5,\,0.25$ ; the impulse result $\Delta p=I_0=3\ \mathrm{N\,s}$ , $v=6\ \mathrm{m/s}$ ; and $G''=\delta$ for $G=\tfrac12|x-x'|$ with $\nabla^2(-1/4\pi r)=\delta^3(\mathbf{r})$ . All constants ( $\pi$ , $2\pi$ , $4\pi$ , $\sqrt{\pi}$ ) are standard values.

Prerequisites. The Fourier Transform · Fourier Series and Periodic Functions · Theory of Distributions. Leads to. Tempered Distributions and Fourier Transforms · The Convolution Theorem · Green's Functions via Integral Transforms · Solving Differential Equations by Transform Methods.

Ask the tutor: Ask me to prove any delta identity by testing it against a smooth function, or to build the Green's function of a differential operator you name from a single delta source.

Is the Dirac delta really a function?

No — and this is the whole point. An ordinary function assigns a number to each input, and no such assignment can be zero everywhere except at one point yet integrate to one, because changing a function at a single point never changes its integral. The delta is instead a distribution, or generalized function: an object defined by the single number it returns when integrated against a smooth test function, namely $\delta[f]=f(0)$ . The pointwise picture " $\delta(0)=\infty$ , zero elsewhere" is a useful mnemonic for that behaviour, but it is not a definition and cannot be computed with directly. Everything rigorous about the delta is a statement about integrals, not about values.

What is a test function, and why does the delta need one?

A test function is a very well-behaved function — smooth (infinitely differentiable) and either of compact support or rapidly decreasing at infinity (the Schwartz space). The delta is defined only by what it does to such functions: $\delta[f]=f(0)$ . It needs them because a distribution is fundamentally a rule for producing numbers from functions, not a function in its own right, and the good behaviour of test functions is what makes the rule sensible — you can differentiate the delta any number of times precisely because the derivative is shovelled onto the test function, which can always take it. Physically, the test function is the smooth probe — a detector response, a smearing kernel — through which any idealized point object is actually observed.

Why does the scaling law δ(ax) = δ(x)/|a| carry an absolute value?

Because area, not sign, is what the delta conserves. Substituting $u=ax$ in $\int\delta(ax)f(x)\,\mathrm{d}x$ produces a factor $1/a$ from $\mathrm{d}x=\mathrm{d}u/a$ , but when $a<0$ the substitution also reverses the limits of integration, and swapping them back contributes a second minus sign. The two sign effects combine so that only the magnitude $|a|$ survives. Intuitively, squeezing the horizontal axis by a factor $a$ makes the spike narrower, and to keep the total area at one its height — and hence its weight $1/|a|$ — must scale by the same magnitude regardless of whether $a$ flips the axis. A useful consequence is $\delta(-x)=\delta(x)$ : the delta is even.

How can you differentiate something that is not even a function?

By passing the derivative to the test function. The distributional derivative $\delta'$ is defined by the rule $\int\delta'(x)f(x)\,\mathrm{d}x=-\int\delta(x)f'(x)\,\mathrm{d}x=-f'(0)$ , which is just integration by parts with the boundary term dropped because test functions vanish at infinity. Nothing is ever differentiated pointwise; the operation is moved onto $f$ , where it is harmless. The same trick iterates: $\delta^{(n)}[f]=(-1)^n f^{(n)}(0)$ . This is why distributions are so powerful — every distribution is infinitely differentiable, so objects like the derivative of a step function or the second derivative of $|x|$ suddenly make perfect sense.

What does the Fourier representation δ(x) = (1/2π)∫e^{ikx}dk actually mean?

It means the delta is an equal-weight superposition of every plane wave. Taken literally the integral does not converge, so it is a distributional statement: integrate both sides against a test function and the right side reproduces $f(0)$ , the Fourier inversion theorem in disguise. Physically it encodes the extreme of the uncertainty principle — perfect localization in position requires every wavenumber in equal measure. It is also the completeness relation of the plane-wave basis, which is why continuous-spectrum eigenstates are normalized to a delta, $\langle x|x'\rangle=\delta(x-x')$ , rather than a Kronecker delta.

Why is the delta the identity for convolution but not for multiplication?

Because they are different operations. Convolving with the delta, $(f*\delta)(x)=\int f(y)\delta(x-y)\,\mathrm{d}y=f(x)$ , follows straight from sifting and returns $f$ unchanged, so $\delta$ is the convolution unit — exactly why the impulse response characterizes a linear system completely. Ordinary pointwise multiplication instead gives $f(x)\delta(x)=f(0)\delta(x)$ , which collapses $f$ to a single number times the delta and is not $f$ ; the multiplicative identity is instead the constant function $1$ . Keeping the two straight matters when moving between a signal and its spectrum, where one becomes the other.