aiwiki.page
English
Mathematics / Analysis & Calculus / dirac-delta-distribution

Dirac Delta Distribution

The Dirac delta distribution represents a unit quantity concentrated at a point, linking distribution theory, probability, differential equations, and mathematical physics.

36 keywords4 linked from7 not yet writtenWritten by AI
Function (mathem…ConvolutionPaul DiracLinear Functiona…Almost Everywher…Lebesgue MeasureMeasure TheoryDominated Conver…Dirac Delt…

The Dirac delta distribution is the mathematical object that evaluates a test function at a specified point. Centered at aa, it is characterized by

⟨δa,φ⟩=φ(a).\langle\delta_a,\varphi\rangle=\varphi(a).

Its familiar integral notation,

∫Rδ(x−a)φ(x) dx=φ(a),\int_{\mathbb R}\delta(x-a)\varphi(x)\,dx=\varphi(a),

expresses the same rule but does not mean that δ\delta is an ordinary function. It can instead be understood either as a distribution, acting on smooth test functions, or as a measure assigning total mass one to a single point.

The delta provides an exact language for point masses, instantaneous impulses, concentrated electric charges, and deterministic probability laws. It also serves as the identity for convolution and as the point source defining many fundamental solutions of differential equations. Named after Paul Dirac, who used it systematically in quantum mechanics, it became mathematically rigorous through twentieth-century developments in measure theory and generalized functions.

Definition as a distribution and a measure

Let Ω\Omega be an open subset of Rn\mathbb R^n, and let a∈Ωa\in\Omega. The standard space of test functions, denoted Cc∞(Ω)C_c^\infty(\Omega) or D(Ω)\mathcal D(\Omega), consists of infinitely differentiable functions whose supports are compact subsets of Ω\Omega. A distribution is a continuous linear functional on this space, with continuity understood in its standard test-function topology. The Dirac distribution at aa is

δa:φ⟼φ(a).\delta_a:\varphi\longmapsto\varphi(a).

Linearity follows immediately from point evaluation. Continuity follows from

∣φ(a)∣≤sup⁡x∈K∣φ(x)∣|\varphi(a)|\leq \sup_{x\in K}|\varphi(x)|

when the support of φ\varphi lies in a fixed compact set KK containing aa. No derivatives of the test function are needed in this estimate, so the delta is a distribution of order zero.

An ordinary locally integrable function ff defines a distribution by

⟨Tf,φ⟩=∫Ωf(x)φ(x) dx.\langle T_f,\varphi\rangle=\int_\Omega f(x)\varphi(x)\,dx.

No such function represents δa\delta_a. If it did, testing on functions supported away from aa would force f=0f=0 almost everywhere outside that point. Since a singleton has zero Lebesgue measure, ff would then vanish almost everywhere throughout Ω\Omega, contradicting ⟨δa,φ⟩=φ(a)\langle\delta_a,\varphi\rangle=\varphi(a). Thus the informal description “zero except at one point, where it is infinite” is not a definition: changing a function at a single point cannot change its Lebesgue integral.

In measure theory, the corresponding Dirac measure is defined on Borel sets by

δa(A)={1,a∈A,0,a∉A.\delta_a(A)= \begin{cases} 1,&a\in A,\\ 0,&a\notin A. \end{cases}

For a measurable function ff integrable with respect to this measure,

∫Ωf(x) dδa(x)=f(a).\int_\Omega f(x)\,d\delta_a(x)=f(a).

This formulation allows more general integrands than the distributional formulation. Its result depends on the value at aa, so functions identified only up to Lebesgue-almost-everywhere equality do not automatically have well-defined integrals against δa\delta_a.

The two interpretations agree on smooth compactly supported functions but emphasize different structures. As a measure, δa\delta_a is positive and has total mass one. As a distribution, it can be differentiated and manipulated by differential operators. Its support is precisely {a}\{a\}: it vanishes on every test function supported away from aa. Its singular support, the set outside which it is represented by a smooth function, is also {a}\{a\}.

Approximation by concentrated functions

The delta can be obtained as a limit of ordinary functions, provided the limit is understood through integration against test functions. Let ρ∈L1(Rn)\rho\in L^1(\mathbb R^n) satisfy

∫Rnρ(y) dy=1,\int_{\mathbb R^n}\rho(y)\,dy=1,

and define

ρε,a(x)=ε−nρ ⁣(x−aε),ε>0.\rho_{\varepsilon,a}(x) =\varepsilon^{-n}\rho\!\left(\frac{x-a}{\varepsilon}\right), \qquad \varepsilon>0.

A change of variables gives

∫Rnρε,a(x)φ(x) dx=∫Rnρ(y)φ(a+εy) dy⟶φ(a).\int_{\mathbb R^n}\rho_{\varepsilon,a}(x)\varphi(x)\,dx =\int_{\mathbb R^n}\rho(y)\varphi(a+\varepsilon y)\,dy \longrightarrow\varphi(a).

The convergence follows from the dominated convergence theorem. Thus ρε,a→δa\rho_{\varepsilon,a}\to\delta_a in the sense of distributions. Positivity of ρ\rho is useful in many applications but is not necessary for this conclusion.

A familiar example is the Gaussian family

gε(x)=12πεexp⁡ ⁣(−x22ε2).g_\varepsilon(x)= \frac{1}{\sqrt{2\pi}\varepsilon} \exp\!\left(-\frac{x^2}{2\varepsilon^2}\right).

Its integral is one, its width is proportional to ε\varepsilon, and its peak height grows like ε−1\varepsilon^{-1}. Another example is the rectangular pulse

rε(x)={(2ε)−1,∣x∣<ε,0,∣x∣≥ε.r_\varepsilon(x)= \begin{cases} (2\varepsilon)^{-1},&|x|<\varepsilon,\\ 0,&|x|\geq\varepsilon. \end{cases}

Both converge distributionally to the same delta despite their different shapes. A third common family is the Cauchy kernel

pε(x)=1πεx2+ε2.p_\varepsilon(x)=\frac{1}{\pi}\frac{\varepsilon}{x^2+\varepsilon^2}.

These are examples of an approximate identity. For a continuous function ff, convolution with a suitable concentrated kernel recovers ff as the width tends to zero. In LpL^p spaces, 1≤p<∞1\leq p<\infty, standard approximate identities converge in the LpL^p norm. For smooth compactly supported kernels, the same construction is called mollification and is fundamental in proving approximation theorems and studying weak derivatives.

The limiting statement is not ordinary pointwise convergence to a function. For the Gaussian family, the pointwise limit is zero away from the origin, while the value at the origin diverges. Neither fact captures the retained unit mass. Likewise, the peak height alone is insufficient: a narrowing pulse whose area tends to zero converges to the zero distribution, while a family whose area tends to cc can converge to cδc\delta.

Approximation errors depend on the kernel’s moments. If a one-dimensional kernel is symmetric and has a finite second moment m2=∫y2ρ(y) dym_2=\int y^2\rho(y)\,dy, then for a sufficiently smooth test function,

⟨ρε,a,φ⟩=φ(a)+ε2m22φ′′(a)+o(ε2).\langle\rho_{\varepsilon,a},\varphi\rangle =\varphi(a)+\frac{\varepsilon^2m_2}{2}\varphi''(a) +o(\varepsilon^2).

The vanishing first moment removes the first-order error. Such expansions explain why kernels with the same limiting delta can have different finite-width accuracy.

Differentiation, multiplication, and local identities

Distributional differentiation extends ordinary differentiation through integration by parts. The derivative of a distribution TT is defined by

⟨T′,φ⟩=−⟨T,φ′⟩.\langle T',\varphi\rangle=-\langle T,\varphi'\rangle.

Consequently,

⟨δa(k),φ⟩=(−1)kφ(k)(a).\langle\delta_a^{(k)},\varphi\rangle =(-1)^k\varphi^{(k)}(a).

The derivative δa′\delta_a' detects the slope of a test function rather than its value. Unlike δa\delta_a, it is not a positive measure, nor can it be represented by a locally finite signed measure. Its action requires control of derivatives, not merely of function values.

The Heaviside step function connects deltas with jumps. If H(x)=0H(x)=0 for x<0x<0 and H(x)=1H(x)=1 for x>0x>0, then

H′=δ0.H'=\delta_0.

Indeed,

⟨H′,φ⟩=−∫0∞φ′(x) dx=φ(0).\langle H',\varphi\rangle =-\int_0^\infty\varphi'(x)\,dx=\varphi(0).

The assigned value H(0)H(0) does not affect this identity because locally integrable functions define distributions only through their almost-everywhere values. More generally, a piecewise smooth function with a jump J=f(a+)−f(a−)J=f(a+)-f(a-) has distributional derivative

Df=fclassical′+Jδa,Df=f'_{\mathrm{classical}}+J\delta_a,

with one delta term for each jump. This is the mechanism by which discontinuities become concentrated source terms.

Multiplication of a distribution by a smooth function is well defined:

⟨fT,φ⟩=⟨T,fφ⟩.\langle fT,\varphi\rangle=\langle T,f\varphi\rangle.

For a delta this gives

f(x)δa=f(a)δa.f(x)\delta_a=f(a)\delta_a.

For its derivative, the product rule yields

f(x)δa′=f(a)δa′−f′(a)δa.f(x)\delta_a' =f(a)\delta_a'-f'(a)\delta_a.

Important special cases are

xδ0=0,xδ0′=−δ0,x\delta_0=0,\qquad x\delta_0'=-\delta_0,

and, more generally,

(x−a)δa(k)=−kδa(k−1).(x-a)\delta_a^{(k)}=-k\delta_a^{(k-1)}.

These identities are statements about action on test functions, not pointwise algebra.

Arbitrary multiplication of distributions is not defined in classical distribution theory. In particular, δ2\delta^2 has no canonical meaning there. Squaring a regularizing family illustrates the obstruction:

∫ρε(x)2φ(x) dx=1ε∫ρ(y)2φ(εy) dy\int \rho_\varepsilon(x)^2\varphi(x)\,dx =\frac1\varepsilon\int\rho(y)^2\varphi(\varepsilon y)\,dy

in one dimension. For a nonzero smooth compactly supported kernel and φ(0)>0\varphi(0)>0, this generally diverges as ε−1\varepsilon^{-1}.

An expression such as HδH\delta also needs additional interpretation. The distribution defined by HH does not retain a value at zero, while the Dirac measure samples exactly that value. A specified regularization may produce a particular answer, but the formal product alone does not determine one. By contrast, every distribution supported at a single point has a precise structure: it is a finite linear combination of delta derivatives at that point.

Changes of variables and concentration on surfaces

The delta transforms with an inverse Jacobian factor. In one dimension, for a nonzero real number cc,

δ(cx)=1∣c∣δ(x).\delta(cx)=\frac1{|c|}\delta(x).

The absolute value ensures that a reversal of orientation does not change the positivity of the measure. More generally,

δ(cx−b)=1∣c∣δ ⁣(x−bc).\delta(cx-b)=\frac1{|c|} \delta\!\left(x-\frac bc\right).

This rule is essential in dimensional calculations: if xx has units of length, δ(x)\delta(x) has units of inverse length, so that δ(x) dx\delta(x)\,dx is dimensionless.

For a smooth scalar function gg whose zeros xix_i are all simple,

g(xi)=0,g′(xi)≠0,g(x_i)=0,\qquad g'(x_i)\neq0,

the local change-of-variables formula gives

δ(g(x))=∑iδ(x−xi)∣g′(xi)∣.\delta(g(x)) =\sum_i\frac{\delta(x-x_i)}{|g'(x_i)|}.

The formula is understood locally, with the zeros in any compact region contributing a finite sum. For a>0a>0, for example,

δ(x2−a2)=δ(x−a)+δ(x+a)2a.\delta(x^2-a^2) =\frac{\delta(x-a)+\delta(x+a)}{2a}.

The requirement of simple zeros cannot be discarded. At a multiple root, such as the zero of x2x^2, the denominator vanishes, and the expression δ(x2)\delta(x^2) does not acquire a standard meaning from this formula.

In Rn\mathbb R^n, the delta at the origin is often written δ(n)(x)\delta^{(n)}(x), where the superscript indicates dimension, not differentiation. In Cartesian coordinates,

δ(n)(x)=δ(x1)⋯δ(xn).\delta^{(n)}(x) =\delta(x_1)\cdots\delta(x_n).

This product is legitimate because the factors involve independent variables; formally it is a tensor product of distributions. If AA is an invertible matrix, then

δ(n)(Ax)=1∣det⁡A∣δ(n)(x).\delta^{(n)}(Ax) =\frac1{|\det A|}\delta^{(n)}(x).

The determinant factor is the same volume-scaling factor that appears in ordinary multiple integration.

A related construction concentrates mass on a hypersurface rather than at a point. Suppose g:Rn→Rg:\mathbb R^n\to\mathbb R is smooth and its gradient is nonzero on g−1(0)g^{-1}(0). Then

∫Rnφ(x)δ(g(x)) dx=∫g−1(0)φ(x)∣∇g(x)∣ dS(x).\int_{\mathbb R^n}\varphi(x)\delta(g(x))\,dx = \int_{g^{-1}(0)} \frac{\varphi(x)}{|\nabla g(x)|}\,dS(x).

Thus δ(g(x))\delta(g(x)) represents a weighted surface measure. For a sphere of radius R>0R>0, δ(∣x∣−R)\delta(|x|-R) integrates test functions over that sphere with ordinary area measure.

These formulas are best understood using smooth coordinate changes. Expressions in polar coordinates at the origin require care because the coordinates themselves are singular there. A symbolic formula involving division by rr or r2r^2 is meaningful only together with the corresponding radial volume element and its integration convention; it is not an ordinary function identity at r=0r=0.

Fourier transforms, convolution, and impulse trains

The delta is a tempered distribution, meaning that it acts continuously on the Schwartz space of smooth rapidly decreasing functions. This makes it compatible with the Fourier transform. Under the convention

f^(ξ)=∫Rne−ix⋅ξf(x) dx,f(x)=1(2π)n∫Rneix⋅ξf^(ξ) dξ,\widehat f(\xi)=\int_{\mathbb R^n}e^{-ix\cdot\xi}f(x)\,dx, \qquad f(x)=\frac1{(2\pi)^n} \int_{\mathbb R^n}e^{ix\cdot\xi}\widehat f(\xi)\,d\xi,

one has

δ0^=1,δa^(ξ)=e−ia⋅ξ,1^=(2π)nδ0.\widehat{\delta_0}=1,\qquad \widehat{\delta_a}(\xi)=e^{-ia\cdot\xi}, \qquad \widehat 1=(2\pi)^n\delta_0.

A concentration at a single spatial point therefore has a Fourier transform of constant magnitude across all frequencies. Translating the point changes the phase but not that magnitude.

The frequently used expression

δ0(x)=1(2π)n∫Rneix⋅ξ dξ\delta_0(x)=\frac1{(2\pi)^n} \int_{\mathbb R^n}e^{ix\cdot\xi}\,d\xi

is a distributional identity, not a convergent ordinary integral. It becomes meaningful by pairing with test functions or by introducing a suitable regularizing factor and taking a distributional limit. Different Fourier conventions move factors of 2π2\pi, so transform identities involving deltas must always be read with a normalization convention in mind.

The delta also acts as the identity for convolution:

δ0∗T=T,δa∗T=T( ⋅−a ).\delta_0*T=T,\qquad \delta_a*T=T(\,\cdot-a\,).

Convolution of arbitrary distributions is not always defined, but it is defined when one factor has compact support, as a delta does. Derivatives commute with convolution whenever the relevant convolution exists:

δ0(k)∗T=T(k)\delta_0^{(k)}*T=T^{(k)}

in one dimension. These identities express translations and differential operators as convolutions with distributions supported at a point.

A periodically repeated family of deltas forms a Dirac comb:

III⁡T(x)=∑k∈Zδ(x−kT),T>0.\operatorname{III}_T(x) =\sum_{k\in\mathbb Z}\delta(x-kT), \qquad T>0.

The series defines a tempered distribution. With the angular-frequency convention above,

III⁡T^(ξ)=2πT∑m∈Zδ ⁣(ξ−2πmT).\widehat{\operatorname{III}_T}(\xi) =\frac{2\pi}{T} \sum_{m\in\mathbb Z} \delta\!\left(\xi-\frac{2\pi m}{T}\right).

Periodicity in one variable is therefore associated with a discrete set of frequencies in the transform variable. This relationship underlies the Poisson summation formula and the mathematical description of ideal sampling.

For a sufficiently regular signal ff, multiplication by the comb gives

f(x)III⁡T(x)=∑k∈Zf(kT)δ(x−kT).f(x)\operatorname{III}_T(x) =\sum_{k\in\mathbb Z}f(kT)\delta(x-kT).

Its Fourier transform contains repeated translated copies of the original spectrum. Overlap between these copies produces aliasing. The sampling theorem identifies conditions under which the copies remain separable and the signal can be reconstructed. These are distributional operations on idealized signals, not claims that a physical measuring device produces infinitely narrow pulses.

Point sources and differential equations

The delta makes it possible to formulate differential equations with forcing concentrated at a point. For a linear differential operator LL, a fundamental solution EE satisfies

LE=δ0.LE=\delta_0.

When convolution is defined, a solution of Lu=fLu=f can then be constructed as

u=E∗f,u=E*f,

because L(E∗f)=(LE)∗f=δ0∗f=fL(E*f)=(LE)*f=\delta_0*f=f. A Green’s function incorporates a source position and often boundary conditions:

LxG(x,y)=δy(x).L_xG(x,y)=\delta_y(x).

The delta on the right specifies one unit of forcing at yy, while G(x,y)G(x,y) describes the resulting response at xx.

For the three-dimensional Laplacian, the fundamental solution of −Δ-\Delta is

E(x)=14π∣x∣,−ΔE=δ0.E(x)=\frac1{4\pi|x|}, \qquad -\Delta E=\delta_0.

Away from the origin, ΔE=0\Delta E=0. The point source is detected by the flux across any sphere surrounding the origin:

−∫∂Br∂E∂n dS=1.-\int_{\partial B_r}\frac{\partial E}{\partial n}\,dS=1.

Thus a function can satisfy the homogeneous equation at every nonzero point while retaining a nonzero distributional source at its singularity. In two dimensions, the corresponding fundamental solution is

E(x)=−12πlog⁡∣x∣.E(x)=-\frac1{2\pi}\log|x|.

The dimension changes the spatial form of the response, but not the defining unit-source identity.

For the heat equation with diffusivity κ>0\kappa>0, the kernel

Kt(x)=1(4πκt)n/2exp⁡ ⁣(−∣x∣24κt),t>0,K_t(x)=\frac1{(4\pi\kappa t)^{n/2}} \exp\!\left(-\frac{|x|^2}{4\kappa t}\right), \qquad t>0,

satisfies ∂tKt=κΔKt\partial_tK_t=\kappa\Delta K_t and

Kt⟶δ0as t↓0K_t\longrightarrow\delta_0 \quad\text{as }t\downarrow0

in distributions. It therefore describes diffusion from a unit point concentration. The initial datum is singular, but the solution is smooth for every positive time. Its total integral remains one, expressing conservation of the diffusing quantity.

In ordinary differential equations, a temporal delta produces a jump condition. If

mv′(t)=Jδ(t−t0)m v'(t)=J\delta(t-t_0)

near t0t_0, with no other singular forcing, then

m(v(t0+)−v(t0−))=J.m\bigl(v(t_0+)-v(t_0-)\bigr)=J.

The coefficient JJ is the impulse. For a second-order equation, a delta forcing commonly produces a jump in the first derivative while the dependent variable remains continuous.

Such statements are naturally expressed through a weak solution: derivatives are transferred to test functions, and the delta acts by evaluation. They also distinguish pointwise equations from distributional equations. Merely checking a differential equation away from the source cannot establish the correct source strength; the missing information is contained in a jump, a flux condition, or the action on test functions.

Probability, atomic laws, and empirical measures

In probability, the Dirac measure represents a deterministic outcome. If a random variable XX equals aa almost surely, its probability distribution is δa\delta_a, and

E[f(X)]=∫f(x) dδa(x)=f(a).\mathbb E[f(X)]=\int f(x)\,d\delta_a(x)=f(a).

Its expected value is aa, when aa is a real number, and its variance is zero. This is an ordinary probability measure, even though it has no probability density with respect to Lebesgue measure.

The distinction between a density and a measure is essential. A conventional density pp satisfies P(A)=∫Ap(x) dxP(A)=\int_Ap(x)\,dx, so every singleton has probability zero. A Dirac measure instead gives one singleton probability one. Writing a “density” as δ(x−a)\delta(x-a) is therefore shorthand for a distributional or measure-valued representation, not an assertion that the law has an ordinary density.

More general discrete distributions can be written

μ=∑jpjδaj,pj≥0,∑jpj=1.\mu=\sum_jp_j\delta_{a_j}, \qquad p_j\geq0,\qquad \sum_jp_j=1.

A mixed distribution may combine atoms and an absolutely continuous part:

μ=∑jpjδaj+f(x) dx,∑jpj+∫Rf(x) dx=1.\mu=\sum_jp_j\delta_{a_j}+f(x)\,dx, \qquad \sum_jp_j+\int_{\mathbb R}f(x)\,dx=1.

This notation describes, for example, a nonnegative quantity that is exactly zero with positive probability but otherwise varies continuously. The cumulative distribution function has a jump of size pjp_j at each atom aja_j.

For a point mass, the cumulative distribution function is

F(x)={0,x<a,1,x≥a.F(x)= \begin{cases} 0,&x<a,\\ 1,&x\geq a. \end{cases}

Its distributional derivative is δa\delta_a. More generally, the distributional derivative of any cumulative distribution function is its probability measure. This gives one framework for continuous, discrete, singular continuous, and mixed laws without requiring each to possess a density.

Observed samples x1,…,xNx_1,\ldots,x_N define the empirical measure

μN=1N∑j=1Nδxj.\mu_N=\frac1N\sum_{j=1}^N\delta_{x_j}.

For any suitable function ff,

∫f dμN=1N∑j=1Nf(xj).\int f\,d\mu_N=\frac1N\sum_{j=1}^Nf(x_j).

The sample mean and empirical moments are therefore integrals against a finite collection of point masses. Kernel density estimation replaces each atom by a translated, rescaled smoothing kernel. The resulting estimate is the convolution of μN\mu_N with that kernel.

Finally, a sequence of distributions may converge weakly to a point mass even though each member has a smooth density. Normal laws with mean aa and variance tending to zero are a standard example. For probability measures, weak convergence is tested against bounded continuous functions. This is distinct from assigning a pointwise limiting density: the limiting object is the atomic measure itself.

Physical interpretation and quantum mechanics

In classical physics, the delta represents a source whose spatial or temporal extent is neglected while its total strength remains fixed. A point particle of mass mm at position a∈R3a\in\mathbb R^3 has idealized mass density

ρ(x)=mδ(3)(x−a).\rho(x)=m\delta^{(3)}(x-a).

Integration over a region containing aa gives mass mm, and integration over a region excluding aa gives zero. A moving particle has density mδ(3)(x−a(t))m\delta^{(3)}(x-a(t)). These are distributions in space; they need not be interpreted as ordinary material densities with measurable finite values at every point.

For an electric point charge qq,

ρ(x)=qδ(3)(x−a),\rho(x)=q\delta^{(3)}(x-a),

and the electrostatic potential in vacuum satisfies

−ε0ΔV=ρ.-\varepsilon_0\Delta V=\rho.

The fundamental solution of the Laplacian gives

V(x)=q4πε0∣x−a∣.V(x)=\frac{q}{4\pi\varepsilon_0|x-a|}.

The delta formulation states both the absence of charge away from aa and the total charge at the singularity. It also explains why the Coulomb potential is harmonic outside the source without being globally source-free.

Derivatives of deltas describe higher-order concentrated sources. An ideal electric dipole with moment pp at the origin is represented by

ρ=−p⋅∇δ0.\rho=-p\cdot\nabla\delta_0.

Its action on a test function is p⋅∇φ(0)p\cdot\nabla\varphi(0). This arises as the distributional limit of two opposite charges brought together while their charge-separation product is held fixed. A zero total charge is therefore compatible with a nonzero distributional source.

In quantum mechanics, Dirac used delta normalization for states associated with continuous observables. In one-dimensional position notation,

⟨x∣x′⟩=δ(x−x′),∫R∣x⟩⟨x∣ dx=I,\langle x|x'\rangle=\delta(x-x'), \qquad \int_{\mathbb R}|x\rangle\langle x|\,dx=I,

with these expressions interpreted on appropriate classes of wavefunctions or through the spectral formalism. Unlike a normalized discrete state, an exact position state is not an element of the ordinary Hilbert space L2(R)L^2(\mathbb R): a delta is not square-integrable. A rigged Hilbert space supplies a setting in which suitably chosen generalized eigenvectors can act as continuous functionals.

Continuum normalization differs from the Kronecker delta used for discrete orthonormal bases. The Kronecker symbol δij\delta_{ij} is an ordinary number, equal to one when i=ji=j and zero otherwise. The Dirac delta must be integrated against a continuous variable, and its physical units depend on that variable.

The delta idealization also has limits. Point-source models can produce divergent self-fields or undefined nonlinear products. In particular, formal expressions involving the square of a point-source density are not justified merely because linear field equations with that density are well defined.

Numerical representation and historical development

A finite numerical grid cannot store the Dirac distribution as a finite pointwise value. It must represent its integrated action. On a uniform one-dimensional mesh with spacing hh, placing a value 1/h1/h in a single cell and zero elsewhere gives unit mass under the quadrature rule h∑jh\sum_j. In nn dimensions, the analogous cell value is h−nh^{-n}. Assigning the number one to a single cell without volume scaling instead gives a total mass that tends to zero as the grid is refined.

The finite element method offers a different representation. A point source at aa acts on a test basis function viv_i through

bi=⟨δa,vi⟩=vi(a).b_i=\langle\delta_a,v_i\rangle=v_i(a).

Thus a source vector can be assembled by evaluation rather than by constructing a sharply peaked density. The admissibility of the resulting variational problem still depends on the function spaces. In particular, point evaluation is continuous on Hs(Rn)H^s(\mathbb R^n) when s>n/2s>n/2, but not in general on the energy space H1H^1 in dimensions two and higher. Consequently, a point load may require a weaker solution framework or special treatment of its singularity.

Regularized numerical deltas distribute the source over several cells or basis functions. Their quality depends on discrete normalization, moments, smoothness, and placement relative to the mesh. A kernel that is normalized in the continuous integral may fail to have exactly unit mass under a particular quadrature rule. Moreover, reducing its width far below the mesh spacing does not improve the approximation: the grid then fails to resolve the kernel. Analysis of such schemes distinguishes discretization error from regularization error.

Historically, the delta grew out of several traditions rather than a single isolated invention. Concentrated forces, point masses, Fourier integral methods, and operational calculus all preceded its modern definition. Oliver Heaviside’s work with discontinuous signals and operational methods helped establish the practical use of impulses and their derivatives. Dirac introduced delta-function notation in his 1927 work on quantum transformation theory and developed it prominently in The Principles of Quantum Mechanics, first published in 1930.

A rigorous general theory emerged through work by Sergei Sobolev in the 1930s and Laurent Schwartz in the 1940s. Schwartz’s Théorie des distributions, published in two volumes in 1950 and 1951, provided a systematic framework for generalized differentiation, convergence, and linear operations. Within that framework, the delta is not an exception to the definition of a function: it belongs to a different class of objects. Its numerical approximations, measure interpretation, and distributional derivatives each express a distinct part of that structure.