Dirac delta function

One way to rigorously capture the notion of the Dirac delta function is to define a measure, called Dirac measure, which accepts a subset A of the real line R as an argument, and returns δ(A) = 1 if 0 ∈ A, and δ(A) = 0 otherwise.[21] If the delta function is conceptualized as modeling an idealized point mass at 0, then δ(A) represents the mass contained in the set A. One may then define the integral against δ as the integral of a function against this mass distribution. Formally, the Lebesgue integral provides the necessary analytic device. The Lebesgue integral with respect to the measure δ satisfies

${\displaystyle \int _{-\infty }^{\infty }f(x)\,\delta (\mathrm {d} x)=f(0)}$

for all continuous compactly supported functions f. The measure δ is not absolutely continuous with respect to the Lebesgue measure—in fact, it is a singular measure. Consequently, the delta measure has no Radon–Nikodym derivative (with respect to Lebesgue measure)—no true function for which the property

${\displaystyle \int _{-\infty }^{\infty }f(x)\,\delta (x)\,\mathrm {d} x=f(0)}$

holds.[22] As a result, the latter notation is a convenient abuse of notation, and not a standard (Riemann or Lebesgue) integral.

As a probability measure on R, the delta measure is characterized by its cumulative distribution function, which is the unit step function.[23]

${\displaystyle H(x)={\begin{cases}1&{\text{if }}x\geq 0\\0&{\text{if }}x<0.\end{cases}}}$

This means that H(x) is the integral of the cumulative indicator function 1_{(−∞, x]} with respect to the measure δ; to wit,

${\displaystyle H(x)=\int _{\mathbf {R} }\mathbf {1} _{(-\infty ,x]}(t)\,\delta (\mathrm {d} t)=\delta (-\infty ,x],}$

the latter being the measure of this interval; more formally, δ((−∞, x]). Thus in particular the integration of the delta function against a continuous function can be properly understood as a Riemann–Stieltjes integral:[24]

${\displaystyle \int _{-\infty }^{\infty }f(x)\,\delta (\mathrm {d} x)=\int _{-\infty }^{\infty }f(x)\,\mathrm {d} H(x).}$

All higher moments of δ are zero. In particular, characteristic function and moment generating function are both equal to one.

In the theory of distributions, a generalized function is considered not a function in itself but only about how it affects other functions when "integrated" against them.[25] In keeping with this philosophy, to define the delta function properly, it is enough to say what the "integral" of the delta function is against a sufficiently "good" test function φ. Test functions are also known as bump functions. If the delta function is already understood as a measure, then the Lebesgue integral of a test function against that measure supplies the necessary integral.

A typical space of test functions consists of all smooth functions on R with compact support that have as many derivatives as required. As a distribution, the Dirac delta is a linear functional on the space of test functions and is defined by[26]

${\displaystyle \delta [\varphi ]=\varphi (0)}$

(1)

for every test function ${\displaystyle \varphi }$.

For δ to be properly a distribution, it must be continuous in a suitable topology on the space of test functions. In general, for a linear functional S on the space of test functions to define a distribution, it is necessary and sufficient that, for every positive integer N there is an integer M_N and a constant C_N such that for every test function φ, one has the inequality[27]

${\displaystyle \left|S[\varphi ]\right|\leq C_{N}\sum _{k=0}^{M_{N}}\sup _{x\in [-N,N]}\left|\varphi ^{(k)}(x)\right|}$

where sup represents the supremum. With the δ distribution, one has such an inequality (with C_N = 1) with M_N = 0 for all N. Thus δ is a distribution of order zero. It is, furthermore, a distribution with compact support (the support being {0}).

The delta distribution can also be defined in several equivalent ways. For instance, it is the distributional derivative of the Heaviside step function. This means that for every test function φ, one has

${\displaystyle \delta [\varphi ]=-\int _{-\infty }^{\infty }\varphi '(x)\,H(x)\,\mathrm {d} x.}$

Intuitively, if integration by parts were permitted, then the latter integral should simplify to

${\displaystyle \int _{-\infty }^{\infty }\varphi (x)\,H'(x)\,\mathrm {d} x=\int _{-\infty }^{\infty }\varphi (x)\,\delta (x)\,\mathrm {d} x,}$

and indeed, a form of integration by parts is permitted for the Stieltjes integral, and in that case, one does have

${\displaystyle -\int _{-\infty }^{\infty }\varphi '(x)\,H(x)\,\mathrm {d} x=\int _{-\infty }^{\infty }\varphi (x)\,\mathrm {d} H(x).}$

In the context of measure theory, the Dirac measure gives rise to distribution by integration. Conversely, equation (1) defines a Daniell integral on the space of all compactly supported continuous functions φ which, by the Riesz representation theorem, can be represented as the Lebesgue integral of φ concerning some Radon measure.

Generally, when the term "Dirac delta function" is used, it is in the sense of distributions rather than measures, the Dirac measure being among several terms for the corresponding notion in measure theory. Some sources may also use the term Dirac delta distribution.

The delta function can be defined in n-dimensional Euclidean space Rⁿ as the measure such that

${\displaystyle \int _{\mathbf {R} ^{n}}f(\mathbf {x} )\,\delta (\mathrm {d} \mathbf {x} )=f(\mathbf {0} )}$

for every compactly supported continuous function f. As a measure, the n-dimensional delta function is the product measure of the 1-dimensional delta functions in each variable separately. Thus, formally, with x = (x₁, x₂, ..., x_n), one has[28]

${\displaystyle \delta (\mathbf {x} )=\delta (x_{1})\,\delta (x_{2})\cdots \delta (x_{n}).}$

(2)

The delta function can also be defined in the sense of distributions exactly as above in the one-dimensional case.[29] However, despite widespread use in engineering contexts, (2) should be manipulated with care, since the product of distributions can only be defined under quite narrow circumstances.[30][31]

The notion of a Dirac measure makes sense on any set.[32] Thus if X is a set, x₀ ∈ X is a marked point, and Σ is any sigma algebra of subsets of X, then the measure defined on sets A ∈ Σ by

${\displaystyle \delta _{x_{0}}(A)={\begin{cases}1&{\text{if }}x_{0}\in A\\0&{\text{if }}x_{0}\notin A\end{cases}}}$

is the delta measure or unit mass concentrated at x₀.

Another common generalization of the delta function is to a differentiable manifold where most of its properties as a distribution can also be exploited because of the differentiable structure. The delta function on a manifold M centered at the point x₀ ∈ M is defined as the following distribution:

${\displaystyle \delta _{x_{0}}[\varphi ]=\varphi (x_{0})}$

(3)

for all compactly supported smooth real-valued functions φ on M.[33] A common special case of this construction is a case in which M is an open set in the Euclidean space Rⁿ.

On a locally compact Hausdorff space X, the Dirac delta measure concentrated at a point x is the Radon measure associated with the Daniell integral (3) on compactly supported continuous functions φ.[34] At this level of generality, calculus as such is no longer possible, however a variety of techniques from abstract analysis are available. For instance, the mapping ${\displaystyle x_{0}\mapsto \delta _{x_{0}}}$ is a continuous embedding of X into the space of finite Radon measures on X, equipped with its vague topology. Moreover, the convex hull of the image of X under this embedding is dense in the space of probability measures on X.[35]

The delta function satisfies the following scaling property for a non-zero scalar α:[36]

${\displaystyle \int _{-\infty }^{\infty }\delta (\alpha x)\,\mathrm {d} x=\int _{-\infty }^{\infty }\delta (u)\,{\frac {\mathrm {d} u}{|\alpha |}}={\frac {1}{|\alpha |}}}$

and so

${\displaystyle \delta (\alpha x)={\frac {\delta (x)}{|\alpha |}}.}$

(4)

Scaling property proof:

$${\displaystyle {\begin{aligned}\delta (\alpha x)&=\lim _{b\to 0}{\frac {1}{|b|{\sqrt {\pi }}}}e^{-(\alpha x/b)^{2}}\qquad {\text{since }}b{\text{ is a dummy variable, we set }}b=\alpha c\\&=\lim _{c\to 0}{\frac {1}{|\alpha c|{\sqrt {\pi }}}}e^{-(\alpha x/(\alpha c))^{2}}\\&=\lim _{c\to 0}{\frac {1}{|\alpha |}}{\frac {1}{|c|{\sqrt {\pi }}}}e^{-(x/c)^{2}}={\frac {1}{|\alpha |}}\delta (x)\end{aligned}}}$$

In this proof, the delta function representation as the limit of the sequence of zero-centered normal distributions ${\displaystyle \delta (x)=\lim _{b\to 0}{\frac {1}{|b|{\sqrt {\pi }}}}e^{-(x/b)^{2}}}$ is used. This proof can be made by using other delta function representations as the limits of sequences of functions, as long as these are even functions.

In particular, the delta function is an even distribution (symmetry), in the sense that

${\displaystyle \delta (-x)=\delta (x)}$

which is homogeneous of degree −1.

The distributional product of δ with x is equal to zero:

${\displaystyle x\,\delta (x)=0.}$

Conversely, if xf(x) = xg(x), where f and g are distributions, then

${\displaystyle f(x)=g(x)+c\delta (x)}$

for some constant c.[37]

The integral of the time-delayed Dirac delta is[38]

${\displaystyle \int _{-\infty }^{\infty }f(t)\,\delta (t-T)\,\mathrm {d} t=f(T).}$

This is sometimes referred to as the shifting property[39] or the sampling property.[40] The delta function is said to "shift out" the value at t = T.[41]

It follows that the effect of convolving a function f(t) with the time-delayed Dirac delta ${\displaystyle \delta _{T}(t)=\delta (t-T)}$ is to time-delay f(t) by the same amount:

${\displaystyle {\begin{aligned}(f*\delta _{T})(t)\ &{\stackrel {\mathrm {def} }{=}}\ \int _{-\infty }^{\infty }f(\tau )\,\delta (t-T-\tau )\,\mathrm {d} \tau \\&=\int _{-\infty }^{\infty }f(\tau )\,\delta (\tau -(t-T))\,\mathrm {d} \tau \qquad {\text{since}}~\delta (-x)=\delta (x)~~{\text{by (4)}}\\&=f(t-T).\end{aligned}}}$

This holds under the precise condition that f be a tempered distribution (see the discussion of the Fourier transform below). As a special case, for instance, we have the identity (understood in the distribution sense)

${\displaystyle \int _{-\infty }^{\infty }\delta (\xi -x)\delta (x-\eta )\,\mathrm {d} x=\delta (\eta -\xi ).}$

More generally, the delta distribution may be composed with a smooth function g(x) in such a way that the familiar change of variables formula holds, that

${\displaystyle \int _{\mathbb {R} }\delta {\bigl (}g(x){\bigr )}f{\bigl (}g(x){\bigr )}\left|g'(x)\right|\mathrm {d} x=\int _{g(\mathbb {R} )}\delta (u)\,f(u)\,\mathrm {d} u}$

provided that g is a continuously differentiable function with g′ nowhere zero.[42] That is, there is a unique way to assign meaning to the distribution ${\displaystyle \delta \circ g}$ so that this identity holds for all compactly supported test functions f. Therefore, the domain must be broken up to exclude the g′ = 0 point. This distribution satisfies δ(g(x)) = 0 if g is nowhere zero, and otherwise if g has a real root at x₀, then

${\displaystyle \delta (g(x))={\frac {\delta (x-x_{0})}{|g'(x_{0})|}}.}$

It is natural therefore to define the composition δ(g(x)) for continuously differentiable functions g by

${\displaystyle \delta (g(x))=\sum _{i}{\frac {\delta (x-x_{i})}{|g'(x_{i})|}}}$

where the sum extends over all roots (i.e., all the different ones) of g(x), which are assumed to be simple. Thus, for example

${\displaystyle \delta \left(x^{2}-\alpha ^{2}\right)={\frac {1}{2|\alpha |}}{\Big [}\delta \left(x+\alpha \right)+\delta \left(x-\alpha \right){\Big ]}.}$

In the integral form, the generalized scaling property may be written as

${\displaystyle \int _{-\infty }^{\infty }f(x)\,\delta (g(x))\,\mathrm {d} x=\sum _{i}{\frac {f(x_{i})}{|g'(x_{i})|}}.}$

The delta distribution in an n-dimensional space satisfies the following scaling property instead,

${\displaystyle \delta (\alpha \mathbf {x} )=|\alpha |^{-n}\delta (\mathbf {x} )~,}$

so that δ is a homogeneous distribution of degree −n.

Under any reflection or rotation ρ, the delta function is invariant,

${\displaystyle \delta (\rho \mathbf {x} )=\delta (\mathbf {x} )~.}$

As in the one-variable case, it is possible to define the composition of δ with a bi-Lipschitz function[43] g: Rⁿ → Rⁿ uniquely so that the identity

${\displaystyle \int _{\mathbb {R} ^{n}}\delta (g(\mathbf {x} ))\,f(g(\mathbf {x} ))\left|\det g'(\mathbf {x} )\right|\mathrm {d} \mathbf {x} =\int _{g(\mathbb {R} ^{n})}\delta (\mathbf {u} )f(\mathbf {u} )\,\mathrm {d} \mathbf {u} }$

for all compactly supported functions f.

Using the coarea formula from geometric measure theory, one can also define the composition of the delta function with a submersion from one Euclidean space to another one of different dimension; the result is a type of current. In the special case of a continuously differentiable function g : Rⁿ → R such that the gradient of g is nowhere zero, the following identity holds[44]

${\displaystyle \int _{\mathbb {R} ^{n}}f(\mathbf {x} )\,\delta (g(\mathbf {x} ))\,\mathrm {d} \mathbf {x} =\int _{g^{-1}(0)}{\frac {f(\mathbf {x} )}{|\mathbf {\nabla } g|}}\,\mathrm {d} \sigma (\mathbf {x} )}$

where the integral on the right is over g⁻¹(0), the (n − 1)-dimensional surface defined by g(x) = 0 with respect to the Minkowski content measure. This is known as a simple layer integral.

More generally, if S is a smooth hypersurface of Rⁿ, then we can associate to S the distribution that integrates any compactly supported smooth function g over S:

${\displaystyle \delta _{S}[g]=\int _{S}g(\mathbf {s} )\,\mathrm {d} \sigma (\mathbf {s} )}$

where σ is the hypersurface measure associated to S. This generalization is associated with the potential theory of simple layer potentials on S. If D is a domain in Rⁿ with smooth boundary S, then δ_S is equal to the normal derivative of the indicator function of D in the distribution sense,

${\displaystyle -\int _{\mathbb {R} ^{n}}g(\mathbf {x} )\,{\frac {\partial 1_{D}(\mathbf {x} )}{\partial n}}\,\mathrm {d} \mathbf {x} =\int _{S}\,g(\mathbf {s} )\,\mathrm {d} \sigma (\mathbf {s} ),}$

where n is the outward normal.[45][46] For a proof, see e.g. the article on the surface delta function.

More generally, on an open set ${\displaystyle U}$ in the ${\displaystyle n}$-dimensional Euclidean space ${\displaystyle \mathbb {R} ^{n}}$, the Dirac delta distribution centered at a point ${\displaystyle a\in U}$ is defined by[54]

${\displaystyle \delta _{a}[\varphi ]=\varphi (a)}$

for all ${\displaystyle \varphi \in C_{c}^{\infty }(U)}$, the space of all smooth functions with compact support on ${\displaystyle U}$. If ${\displaystyle \alpha =(\alpha _{1},\ldots ,\alpha _{n})}$ is any multi-index with ${\displaystyle |\alpha |=\alpha _{1}+\cdots +\alpha _{n}}$ and ${\displaystyle \partial ^{\alpha }}$ denotes the associated mixed partial derivative operator, then the ${\displaystyle \alpha }$-th derivative ${\displaystyle \partial ^{\alpha }\delta _{a}}$ of ${\displaystyle \delta _{a}}$ is given by[54]

${\displaystyle \left\langle \partial ^{\alpha }\delta _{a},\,\varphi \right\rangle =(-1)^{|\alpha |}\left\langle \delta _{a},\partial ^{\alpha }\varphi \right\rangle =(-1)^{|\alpha |}\partial ^{\alpha }\varphi (x){\Big |}_{x=a}\quad {\text{ for all }}\varphi \in C_{c}^{\infty }(U).}$

That is, the ${\displaystyle \alpha }$-th derivative of ${\displaystyle \delta _{a}}$ is the distribution whose value on any test function ${\displaystyle \varphi }$ is the ${\displaystyle \alpha }$-th derivative of ${\displaystyle \varphi }$ at ${\displaystyle a}$ (with the appropriate positive or negative sign).

The first partial derivatives of the delta function are thought of as double layers along the coordinate planes. More generally, the normal derivative of a simple layer supported on a surface is a double layer supported on that surface and represents a laminar magnetic monopole. Higher derivatives of the delta function are known in physics as multipoles.

Higher derivatives enter into mathematics naturally as the building blocks for the complete structure of distributions with point support. If ${\displaystyle S}$ is any distribution on ${\displaystyle U}$ supported on the set ${\displaystyle \{a\}}$ consisting of a single point, then there is an integer ${\displaystyle m}$ and coefficients ${\displaystyle c_{\alpha }}$ such that[54][55]

${\displaystyle S=\sum _{|\alpha |\leq m}c_{\alpha }\partial ^{\alpha }\delta _{a}.}$

Typically a nascent delta function η_ε can be constructed in the following manner. Let η be an absolutely integrable function on R of total integral 1, and define

${\displaystyle \eta _{\varepsilon }(x)=\varepsilon ^{-1}\eta \left({\frac {x}{\varepsilon }}\right).}$

In n dimensions, one uses instead the scaling

${\displaystyle \eta _{\varepsilon }(x)=\varepsilon ^{-n}\eta \left({\frac {x}{\varepsilon }}\right).}$

Then a simple change of variables shows that η_ε also has integral 1. One may show that (5) holds for all continuous compactly supported functions f,[56] and so η_ε converges weakly to δ in the sense of measures.

The η_ε constructed in this way are known as an approximation to the identity.[57] This terminology is because the space L¹(R) of absolutely integrable functions is closed under the operation of convolution of functions: f ∗ g ∈ L¹(R) whenever f and g are in L¹(R). However, there is no identity in L¹(R) for the convolution product: no element h such that f ∗ h = f for all f. Nevertheless, the sequence η_ε does approximate such an identity in the sense that

${\displaystyle f*\eta _{\varepsilon }\to f\quad {\text{as }}\varepsilon \to 0.}$

This limit holds in the sense of mean convergence (convergence in L¹). Further conditions on the η_ε, for instance that it be a mollifier associated to a compactly supported function,[58] are needed to ensure pointwise convergence almost everywhere.

If the initial η = η₁ is itself smooth and compactly supported then the sequence is called a mollifier. The standard mollifier is obtained by choosing η to be a suitably normalized bump function, for instance

${\displaystyle \eta (x)={\begin{cases}e^{-{\frac {1}{1-|x|^{2}}}}&{\text{if }}|x|<1\\0&{\text{if }}|x|\geq 1.\end{cases}}}$

In some situations such as numerical analysis, a piecewise linear approximation to the identity is desirable. This can be obtained by taking η₁ to be a hat function. With this choice of η₁, one has

${\displaystyle \eta _{\varepsilon }(x)=\varepsilon ^{-1}\max \left(1-\left|{\frac {x}{\varepsilon }}\right|,0\right)}$

which are all continuous and compactly supported, although not smooth and so not a mollifier.

In the context of probability theory, it is natural to impose the additional condition that the initial η₁ in an approximation to the identity should be positive, as such a function then represents a probability distribution. Convolution with a probability distribution is sometimes favorable because it does not result in overshoot or undershoot, as the output is a convex combination of the input values, and thus falls between the maximum and minimum of the input function. Taking η₁ to be any probability distribution at all, and letting η_ε(x) = η₁(x/ε)/ε as above will give rise to an approximation to the identity. In general this converges more rapidly to a delta function if, in addition, η has mean 0 and has small higher moments. For instance, if η₁ is the uniform distribution on [−1/2, 1/2], also known as the rectangular function, then:[59]

${\displaystyle \eta _{\varepsilon }(x)={\frac {1}{\varepsilon }}\operatorname {rect} \left({\frac {x}{\varepsilon }}\right)={\begin{cases}{\frac {1}{\varepsilon }},&-{\frac {\varepsilon }{2}}<x<{\frac {\varepsilon }{2}},\\0,&{\text{otherwise}}.\end{cases}}}$

Another example is with the Wigner semicircle distribution

${\displaystyle \eta _{\varepsilon }(x)={\begin{cases}{\frac {2}{\pi \varepsilon ^{2}}}{\sqrt {\varepsilon ^{2}-x^{2}}},&-\varepsilon <x<\varepsilon ,\\0,&{\text{otherwise}}.\end{cases}}}$

This is continuous and compactly supported, but not a mollifier because it is not smooth.

Nascent delta functions often arise as convolution semigroups.[60] This amounts to the further constraint that the convolution of η_ε with η_δ must satisfy

${\displaystyle \eta _{\varepsilon }*\eta _{\delta }=\eta _{\varepsilon +\delta }}$

for all ε, δ > 0. Convolution semigroups in L¹ that form a nascent delta function are always an approximation to the identity in the above sense, however the semigroup condition is quite a strong restriction.

In practice, semigroups approximating the delta function arise as fundamental solutions or Green's functions to physically motivated elliptic or parabolic partial differential equations. In the context of applied mathematics, semigroups arise as the output of a linear time-invariant system. Abstractly, if A is a linear operator acting on functions of x, then a convolution semigroup arises by solving the initial value problem

${\displaystyle {\begin{cases}{\dfrac {\partial }{\partial t}}\eta (t,x)=A\eta (t,x),\quad t>0\\[5pt]\displaystyle \lim _{t\to 0^{+}}\eta (t,x)=\delta (x)\end{cases}}}$

in which the limit is as usual understood in the weak sense. Setting η_ε(x) = η(ε, x) gives the associated nascent delta function.

Some examples of physically important convolution semigroups arising from such a fundamental solution include the following.

The heat kernel

The heat kernel, defined by

${\displaystyle \eta _{\varepsilon }(x)={\frac {1}{\sqrt {2\pi \varepsilon }}}\mathrm {e} ^{-{\frac {x^{2}}{2\varepsilon }}}}$

represents the temperature in an infinite wire at time t > 0, if a unit of heat energy is stored at the origin of the wire at time t = 0. This semigroup evolves according to the one-dimensional heat equation:

${\displaystyle {\frac {\partial u}{\partial t}}={\frac {1}{2}}{\frac {\partial ^{2}u}{\partial x^{2}}}.}$

In probability theory, η_ε(x) is a normal distribution of variance ε and mean 0. It represents the probability density at time t = ε of the position of a particle starting at the origin following a standard Brownian motion. In this context, the semigroup condition is then an expression of the Markov property of Brownian motion.

In higher-dimensional Euclidean space Rⁿ, the heat kernel is

${\displaystyle \eta _{\varepsilon }={\frac {1}{(2\pi \varepsilon )^{n/2}}}\mathrm {e} ^{-{\frac {x\cdot x}{2\varepsilon }}},}$

and has the same physical interpretation, mutatis mutandis. It also represents a nascent delta function in the sense that η_ε → δ in the distribution sense as ε → 0.

The Poisson kernel

The Poisson kernel

${\displaystyle \eta _{\varepsilon }(x)={\frac {1}{\pi }}\mathrm {Im} \left\{{\frac {1}{x-\mathrm {i} \varepsilon }}\right\}={\frac {1}{\pi }}{\frac {\varepsilon }{\varepsilon ^{2}+x^{2}}}={\frac {1}{2\pi }}\int _{-\infty }^{\infty }\mathrm {e} ^{\mathrm {i} \xi x-|\varepsilon \xi |}\,d\xi }$

is the fundamental solution of the Laplace equation in the upper half-plane.[61] It represents the electrostatic potential in a semi-infinite plate whose potential along the edge is held at fixed at the delta function. The Poisson kernel is also closely related to the Cauchy distribution and Epanechnikov and Gaussian kernel functions.[62] This semigroup evolves according to the equation

${\displaystyle {\frac {\partial u}{\partial t}}=-\left(-{\frac {\partial ^{2}}{\partial x^{2}}}\right)^{\frac {1}{2}}u(t,x)}$

where the operator is rigorously defined as the Fourier multiplier

${\displaystyle {\mathcal {F}}\left[\left(-{\frac {\partial ^{2}}{\partial x^{2}}}\right)^{\frac {1}{2}}f\right](\xi )=|2\pi \xi |{\mathcal {F}}f(\xi ).}$

In areas of physics such as wave propagation and wave mechanics, the equations involved are hyperbolic and so may have more singular solutions. As a result, the nascent delta functions that arise as fundamental solutions of the associated Cauchy problems are generally oscillatory integrals. An example, which comes from a solution of the Euler–Tricomi equation of transonic gas dynamics,[63] is the rescaled Airy function

${\displaystyle \varepsilon ^{-1/3}\operatorname {Ai} \left(x\varepsilon ^{-1/3}\right).}$

Although using the Fourier transform, it is easy to see that this generates a semigroup in some sense—it is not absolutely integrable and so cannot define a semigroup in the above strong sense. Many nascent delta functions constructed as oscillatory integrals only converge in the sense of distributions (an example is the Dirichlet kernel below), rather than in the sense of measures.

Another example is the Cauchy problem for the wave equation in R¹⁺¹:[64]

${\displaystyle {\begin{aligned}c^{-2}{\frac {\partial ^{2}u}{\partial t^{2}}}-\Delta u&=0\\u=0,\quad {\frac {\partial u}{\partial t}}=\delta &\qquad {\text{for }}t=0.\end{aligned}}}$

The solution u represents the displacement from equilibrium of an infinite elastic string, with an initial disturbance at the origin.

Other approximations to the identity of this kind include the sinc function (used widely in electronics and telecommunications)

${\displaystyle \eta _{\varepsilon }(x)={\frac {1}{\pi x}}\sin \left({\frac {x}{\varepsilon }}\right)={\frac {1}{2\pi }}\int _{-{\frac {1}{\varepsilon }}}^{\frac {1}{\varepsilon }}\cos(kx)\,dk}$

and the Bessel function

${\displaystyle \eta _{\varepsilon }(x)={\frac {1}{\varepsilon }}J_{\frac {1}{\varepsilon }}\left({\frac {x+1}{\varepsilon }}\right).}$

One approach to the study of a linear partial differential equation

${\displaystyle L[u]=f,}$

where L is a differential operator on Rⁿ, is to seek first a fundamental solution, which is a solution of the equation

${\displaystyle L[u]=\delta .}$

When L is particularly simple, this problem can often be resolved using the Fourier transform directly (as in the case of the Poisson kernel and heat kernel already mentioned). For more complicated operators, it is sometimes easier first to consider an equation of the form

${\displaystyle L[u]=h}$

where h is a plane wave function, meaning that it has the form

${\displaystyle h=h(x\cdot \xi )}$

for some vector ξ. Such an equation can be resolved (if the coefficients of L are analytic functions) by the Cauchy–Kovalevskaya theorem or (if the coefficients of L are constant) by quadrature. So, if the delta function can be decomposed into plane waves, then one can in principle solve linear partial differential equations.

Such a decomposition of the delta function into plane waves was part of a general technique first introduced essentially by Johann Radon, and then developed in this form by Fritz John (1955).[65] Choose k so that n + k is an even integer, and for a real number s, put

${\displaystyle g(s)=\operatorname {Re} \left[{\frac {-s^{k}\log(-is)}{k!(2\pi i)^{n}}}\right]={\begin{cases}{\frac {|s|^{k}}{4k!(2\pi i)^{n-1}}}&n{\text{ odd}}\\[5pt]-{\frac {|s|^{k}\log |s|}{k!(2\pi i)^{n}}}&n{\text{ even.}}\end{cases}}}$

Then δ is obtained by applying a power of the Laplacian to the integral with respect to the unit sphere measure dω of g(x · ξ) for ξ in the unit sphere Sⁿ⁻¹:

${\displaystyle \delta (x)=\Delta _{x}^{(n+k)/2}\int _{S^{n-1}}g(x\cdot \xi )\,d\omega _{\xi }.}$

The Laplacian here is interpreted as a weak derivative, so that this equation is taken to mean that, for any test function φ,

${\displaystyle \varphi (x)=\int _{\mathbf {R} ^{n}}\varphi (y)\,dy\,\Delta _{x}^{\frac {n+k}{2}}\int _{S^{n-1}}g((x-y)\cdot \xi )\,d\omega _{\xi }.}$

The result follows from the formula for the Newtonian potential (the fundamental solution of Poisson's equation). This is essentially a form of the inversion formula for the Radon transform because it recovers the value of φ(x) from its integrals over hyperplanes. For instance, if n is odd and k = 1, then the integral on the right hand side is

${\displaystyle {\begin{aligned}&c_{n}\Delta _{x}^{\frac {n+1}{2}}\iint _{S^{n-1}}\varphi (y)|(y-x)\cdot \xi |\,d\omega _{\xi }\,dy\\[5pt]={}&c_{n}\Delta _{x}^{(n+1)/2}\int _{S^{n-1}}\,d\omega _{\xi }\int _{-\infty }^{\infty }|p|R\varphi (\xi ,p+x\cdot \xi )\,dp\end{aligned}}}$

where Rφ(ξ, p) is the Radon transform of φ:

${\displaystyle R\varphi (\xi ,p)=\int _{x\cdot \xi =p}f(x)\,d^{n-1}x.}$

An alternative equivalent expression of the plane wave decomposition, from Gelfand & Shilov (1966–1968, I, §3.10), is

${\displaystyle \delta (x)={\frac {(n-1)!}{(2\pi i)^{n}}}\int _{S^{n-1}}(x\cdot \xi )^{-n}\,d\omega _{\xi }}$

for n even, and

${\displaystyle \delta (x)={\frac {1}{2(2\pi i)^{n-1}}}\int _{S^{n-1}}\delta ^{(n-1)}(x\cdot \xi )\,d\omega _{\xi }}$

for n odd.

In the study of Fourier series, a major question consists of determining whether and in what sense the Fourier series associated with a periodic function converges to the function. The nth partial sum of the Fourier series of a function f of period 2π is defined by convolution (on the interval [−π,π]) with the Dirichlet kernel:

${\displaystyle D_{N}(x)=\sum _{n=-N}^{N}e^{inx}={\frac {\sin \left(\left(N+{\frac {1}{2}}\right)x\right)}{\sin(x/2)}}.}$

Thus,

${\displaystyle s_{N}(f)(x)=D_{N}*f(x)=\sum _{n=-N}^{N}a_{n}e^{inx}}$

where

${\displaystyle a_{n}={\frac {1}{2\pi }}\int _{-\pi }^{\pi }f(y)e^{-iny}\,dy.}$

A fundamental result of elementary Fourier series states that the Dirichlet kernel tends to the a multiple of the delta function as N → ∞. This is interpreted in the distribution sense, that

${\displaystyle s_{N}(f)(0)=\int _{\mathbf {R} }D_{N}(x)f(x)\,dx\to 2\pi f(0)}$

for every compactly supported smooth function f. Thus, formally one has

${\displaystyle \delta (x)={\frac {1}{2\pi }}\sum _{n=-\infty }^{\infty }e^{inx}}$

on the interval [−π,π].

Despite this, the result does not hold for all compactly supported continuous functions: that is D_N does not converge weakly in the sense of measures. The lack of convergence of the Fourier series has led to the introduction of a variety of summability methods to produce convergence. The method of Cesàro summation leads to the Fejér kernel[66]

${\displaystyle F_{N}(x)={\frac {1}{N}}\sum _{n=0}^{N-1}D_{n}(x)={\frac {1}{N}}\left({\frac {\sin {\frac {Nx}{2}}}{\sin {\frac {x}{2}}}}\right)^{2}.}$

The Fejér kernels tend to the delta function in a stronger sense that[67]

${\displaystyle \int _{\mathbf {R} }F_{N}(x)f(x)\,dx\to 2\pi f(0)}$

for every compactly supported continuous function f. The implication is that the Fourier series of any continuous function is Cesàro summable to the value of the function at every point.

The Dirac delta distribution is a densely defined unbounded linear functional on the Hilbert space L² of square-integrable functions. Indeed, smooth compactly supported functions are dense in L², and the action of the delta distribution on such functions is well-defined. In many applications, it is possible to identify subspaces of L² and to give a stronger topology on which the delta function defines a bounded linear functional.

Sobolev spaces

The Sobolev embedding theorem for Sobolev spaces on the real line R implies that any square-integrable function f such that

${\displaystyle \|f\|_{H^{1}}^{2}=\int _{-\infty }^{\infty }|{\widehat {f}}(\xi )|^{2}(1+|\xi |^{2})\,d\xi <\infty }$

is automatically continuous, and satisfies in particular

${\displaystyle \delta [f]=|f(0)|<C\|f\|_{H^{1}}.}$

Thus δ is a bounded linear functional on the Sobolev space H¹. Equivalently δ is an element of the continuous dual space H⁻¹ of H¹. More generally, in n dimensions, one has δ ∈ H^−s(Rⁿ) provided s > n / 2.

In complex analysis, the delta function enters via Cauchy's integral formula, which asserts that if D is a domain in the complex plane with smooth boundary, then

${\displaystyle f(z)={\frac {1}{2\pi i}}\oint _{\partial D}{\frac {f(\zeta )\,d\zeta }{\zeta -z}},\quad z\in D}$

for all holomorphic functions f in D that are continuous on the closure of D. As a result, the delta function δ_z is represented in this class of holomorphic functions by the Cauchy integral:

${\displaystyle \delta _{z}[f]=f(z)={\frac {1}{2\pi i}}\oint _{\partial D}{\frac {f(\zeta )\,d\zeta }{\zeta -z}}.}$

Moreover, let H²(∂D) be the Hardy space consisting of the closure in L²(∂D) of all holomorphic functions in D continuous up to the boundary of D. Then functions in H²(∂D) uniquely extend to holomorphic functions in D, and the Cauchy integral formula continues to hold. In particular for z ∈ D, the delta function δ_z is a continuous linear functional on H²(∂D). This is a special case of the situation in several complex variables in which, for smooth domains D, the Szegő kernel plays the role of the Cauchy integral.[68]

Given a complete orthonormal basis set of functions {φ_n} in a separable Hilbert space, for example, the normalized eigenvectors of a compact self-adjoint operator, any vector f can be expressed as

${\displaystyle f=\sum _{n=1}^{\infty }\alpha _{n}\varphi _{n}.}$

The coefficients {α_n} are found as

${\displaystyle \alpha _{n}=\langle \varphi _{n},f\rangle ,}$

which may be represented by the notation:

${\displaystyle \alpha _{n}=\varphi _{n}^{\dagger }f,}$

a form of the bra–ket notation of Dirac.[69] Adopting this notation, the expansion of f takes the dyadic form:[70]

${\displaystyle f=\sum _{n=1}^{\infty }\varphi _{n}\left(\varphi _{n}^{\dagger }f\right).}$

Letting I denote the identity operator on the Hilbert space, the expression

${\displaystyle I=\sum _{n=1}^{\infty }\varphi _{n}\varphi _{n}^{\dagger },}$

is called a resolution of the identity. When the Hilbert space is the space L²(D) of square-integrable functions on a domain D, the quantity:

${\displaystyle \varphi _{n}\varphi _{n}^{\dagger },}$

is an integral operator, and the expression for f can be rewritten

${\displaystyle f(x)=\sum _{n=1}^{\infty }\int _{D}\,\left(\varphi _{n}(x)\varphi _{n}^{*}(\xi )\right)f(\xi )\,d\xi .}$