Distributions and Generalized Functions

Distributions make singular sources calculable through their action on smooth probes. Name the test space first, move derivatives onto the test function, and never multiply singular objects unless a theorem or a declared regularization defines the product. This record collects the identities and stop conditions most often needed in gravity and field theory.

Helpful background. Calculus and Linear Algebra for Relativity supplies multivariable differentiation, Jacobians, and integration by parts. PDEs and Green Functions for Gravitation supplies operator domains, boundary data, and causal support. The analysis preparation module places these tools within weak convergence and low-regularity geometry.

Choose the distributional object

Calculation	Start here	Full treatment or application
Weak derivative or singular limit	Declare the test space and topology	Analysis for Mathematical Relativity
Retarded or advanced solution	Declare the operator, measure, domain, and support condition	Retarded and Advanced Green Functions
Material interface or non-null shell	Fix the defining function, orientation, surface measure, and traces	Darmois–Israel Junction Conditions
Worldline source	Pair the tensor source with a smooth test tensor on a fixed background	Point Particles and Dust
Sourced black-hole perturbation	Integrate the exact radial equation across the source	Sourced Master Equations and Junction Conditions
Nonlinear singular curvature	Establish the metric regularity class or a declared generalized-function construction	Distributional Curvature

Test spaces, derivatives, and deltas

Let $\Omega\subset\mathbb R^n$ be open. The standard test space is

$$\mathcal D(\Omega)=C_c^\infty(\Omega),$$

with convergence $\varphi_j\to0$ meaning that all supports lie in one compact subset of $\Omega$ and every derivative converges uniformly there. A distribution is a continuous linear functional

$$T:\mathcal D(\Omega)\longrightarrow\mathbb C, \qquad \varphi\longmapsto\langle T,\varphi\rangle.$$

Every $f\in L^1_{\mathrm{loc}}(\Omega)$ defines $\langle T_f,\varphi\rangle=\int_\Omega f\varphi\,d^nx$. Functions equal almost everywhere define the same distribution, so point values are not generally meaningful. See Hörmander 2003, pp. 5–53 and NIST DLMF § 1.16(i), Eqs. (1.16.1)–(1.16.3).

Distributional convergence means $\langle T_j,\varphi\rangle\to\langle T,\varphi\rangle$ for every $\varphi\in\mathcal D(\Omega)$. It implies neither pointwise convergence nor convergence of nonlinear expressions. The support of $T$ is the complement of the largest open set on which every such pairing vanishes; this, rather than a pointwise graph, localizes the singular source.

The derivative is defined by duality:

$$\langle\partial^\alpha T,\varphi\rangle =(-1)^{\lvert\alpha\rvert} \langle T,\partial^\alpha\varphi\rangle.$$

Compact support removes the boundary term. Multiplication by a smooth function $a$ is also defined:

$$\langle aT,\varphi\rangle=\langle T,a\varphi\rangle.$$

The signs and product rule are NIST DLMF § 1.16(i)–(ii), Eqs. (1.16.6), (1.16.8), and (1.16.9). Smoothness of $a$ on a neighborhood of $\operatorname{supp}T$ is essential; assigning a value to a discontinuous coefficient at the singular support does not define a product Hörmander 2003, pp. 54–86.

Composition is also conditional. A smooth map may pull back a distribution only when its normal covectors avoid the distribution’s wavefront set. The familiar simple-zero formula for $\delta(F)$ below is a special transverse case, not permission to substitute an arbitrary nonsmooth $F$ or a smooth map tangent to the singular directions.

For Fourier analysis, the Schwartz space $\mathcal S(\mathbb R^n)$ and its dual $\mathcal S'$ give tempered distributions. Fourier transformation preserves $\mathcal S'$, but a general element of $\mathcal D'$ need not be tempered: the test space is mathematical data, not notation.

On an oriented manifold, scalar distributions act on compactly supported smooth top forms and tensor distributions on smooth dual-tensor densities. A smooth metric supplies the volume form. With a smooth connection,

$$\langle\nabla_\mu T^{\mu\nu},X_\nu\rangle =-\langle T^{\mu\nu},\nabla_\mu X_\nu\rangle.$$

This is safe because the connection coefficients are smooth multipliers. A distributional metric or connection instead requires a separate product analysis Steinbauer and Vickers 2006, § 2, arXiv PDF pp. 4–9.

The Dirac distribution at $a$ is defined by

$$\langle\delta_a,\varphi\rangle=\varphi(a), \qquad \langle\partial^\alpha\delta_a,\varphi\rangle =(-1)^{\lvert\alpha\rvert}(\partial^\alpha\varphi)(a).$$

The following one-dimensional identities are distributional equalities:

Identity	Meaning or check
$H'=\delta_0$	A unit jump contributes a delta derivative.
$(\operatorname{sgn}x)'=2\delta_0$	The jump has size two.
$(\lvert x\rvert)''=2\delta_0$	The first derivative jumps from $-1$ to $+1$.
$f\delta_a=f(a)\delta_a$ for smooth $f$	Only the value of a smooth multiplier on the support survives.
$\delta(F(x))=\sum_i\delta(x-x_i)/\lvert F'(x_i)\rvert$	The zeros $x_i$ must be simple.

These formulas are NIST DLMF § 1.16(iii)–(iv), Eqs. (1.16.10)–(1.16.18). The zeros in the $\delta(F)$ identity must be simple; the general pullback requires its own transversality hypothesis Hörmander 2003, pp. 133–157.

The invariant spacetime delta is fixed by

$$\int_M f(x)\,\delta_g(x,x')\,dV_g(x)=f(x'), \qquad \delta_g(x,x')=\frac{\delta^{(4)}(x-x')}{\sqrt{-g(x)}}.$$

The numerator is a coordinate delta density; the quotient is paired with $dV_g=\sqrt{-g}\,d^4x$. This prevents a missing Jacobian under coordinate changes Poisson, Pound, and Vega 2011, § 13.1, Eqs. (13.1)–(13.3).

If a coordinate expression is written with $d^4x$ instead, the density weight must remain explicit. Mixing the coordinate delta with the invariant volume form counts the determinant twice; pairing the scalar $\delta_g$ with bare coordinate measure omits it.

Principal values and finite parts

The Cauchy principal value is a different distribution:

$$\left\langle\operatorname{pv}\frac1x,\varphi\right\rangle =\lim_{\varepsilon\downarrow0} \left( \int_{-\infty}^{-\varepsilon}\frac{\varphi(x)}x\,dx +\int_{\varepsilon}^{\infty}\frac{\varphi(x)}x\,dx \right).$$

It obeys

$$x\operatorname{pv}\frac1x=1, \qquad \frac{d}{dx}\log\lvert x\rvert =\operatorname{pv}\frac1x.$$

The boundary values from the upper or lower complex half-plane are

$$\frac{1}{x\pm i0} =\operatorname{pv}\frac1x\mp i\pi\delta_0.$$

The sign follows from the limit of $(x\pm i\varepsilon)^{-1}$. A principal value also remembers its symmetric limiting variable; changing the prescription can change local terms. See Duistermaat and Kolk 2010, pp. 33–58 and 177–220 and NIST DLMF § 1.16(viii), Eqs. (1.16.45)–(1.16.49).

Under a nonlinear change of variable, a symmetric deletion in the old coordinate need not remain symmetric in the new one. Therefore record the defining variable and limiting rule alongside any principal-value identity used in a coordinate or mode calculation.

A Hadamard finite part is not merely another name for a principal value: the symmetric integral of $1/x^2$ still diverges. One extension subtracts the divergence explicitly:

$$\left\langle\operatorname{Pf}\frac1{x^2},\varphi\right\rangle =\lim_{\varepsilon\downarrow0} \left( \int_{\lvert x\rvert\gt\varepsilon}\frac{\varphi(x)}{x^2}\,dx -\frac{2\varphi(0)}{\varepsilon} \right), \qquad \frac{d}{dx}\operatorname{pv}\frac1x =-\operatorname{Pf}\frac1{x^2}.$$

The subtractions, coordinate, scale, and extension prescription are part of the definition Bhattacharyya 2012, § 1.4, pp. 43–46, Eqs. (1.4.11)–(1.4.19).

Jumps and hypersurface sources

Let a smooth function $F$ define a regular non-null hypersurface $\Sigma=\{F=0\}$, with $dF\ne0$ there. Set

$$N=\sqrt{\left\lvert g^{\mu\nu}\nabla_\mu F\nabla_\nu F\right\rvert}, \qquad n_\mu=\frac{\nabla_\mu F}{N},$$

where $n_\mu$ points toward increasing $F$. The surface delta is characterized by

$$\int_M f\,\delta_\Sigma\,dV_g =\int_\Sigma f\,d\Sigma, \qquad \delta_\Sigma=N\,\delta(F).$$

Consequently,

$$\nabla_\mu H(F)=n_\mu\delta_\Sigma.$$

If $u^\pm$ are smooth up to $\Sigma$ from their respective sides, choose smooth extensions of both fields across a collar of $\Sigma$ and set $[u]=u^+\rvert_\Sigma-u^-\rvert_\Sigma$. Then

$$u=u^-+(u^+-u^-)H(F)$$

has derivative

$$\nabla_\mu u =(1-H)\nabla_\mu u^-+H\nabla_\mu u^+ +[u]n_\mu\delta_\Sigma.$$

The result is independent of the chosen collar extensions. A field jump gives a delta in its first derivative; a jump only in the first derivative gives a delta in the second. Nonlinear equations may instead create undefined $\delta_\Sigma^2$ terms.

Rescaling the defining function by a positive smooth factor changes $\delta(F)$ and $N$ inversely, leaving $\delta_\Sigma$ unchanged. Reversing the sign of $F$ reverses $n_\mu$ and exchanges the plus and minus traces; the products $[u]n_\mu$ and $[J^\mu]n_\mu$ remain invariant when both choices are changed consistently.

For a piecewise smooth current,

$$\nabla_\mu J^\mu =(\nabla_\mu J^\mu)_{\mathrm{reg}} +[J^\mu]n_\mu\delta_\Sigma.$$

Thus source-free conservation requires $[J^\mu]n_\mu=0$; a surface current adds its tangential divergence.

This condition is a pillbox statement, not merely a pointwise equation on either side. A nonzero normal-flux jump is precisely the coefficient of the surface-supported source and cannot be found by discarding $\Sigma$ before differentiating.

For a timelike thin shell in general relativity, take a spacelike unit normal $n^\mu$ with $n^2=+1$, directed from the minus side to the plus side, and define

$$K_{ab}=e_a{}^\mu e_b{}^\nu\nabla_\mu n_\nu, \qquad [K_{ab}]=K^+_{ab}-K^-_{ab}.$$

Continuity of the induced metric removes ill-defined squared deltas, while the jump in extrinsic curvature produces the surface stress tensor

$$[K_{ab}]-h_{ab}[K]=-8\pi G\,S_{ab}.$$

Reverse the normal and both $K^\pm_{ab}$ and the jump reverse. A source using the opposite $K_{ab}$ sign needs the corresponding translation. Null shells require a separate transverse structure because no unit normal exists. See Poisson 2004, §§ 3.7 and 3.11, pp. 84–112 and, in the convention above, Israel 1966, p. 7, Eq. (19).

Worldline sources

For a point mass on a timelike worldline $z(\tau)$ in a fixed smooth background, the rigorous definition is its pairing with a compactly supported smooth test tensor:

$$\left\langle T^{\mu\nu},X_{\mu\nu}\right\rangle =m\int u^\mu u^\nu X_{\mu\nu}(z(\tau))\,d\tau.$$

Symbolically,

$$T^{\mu\nu}(x) =m\int u^\mu u^\nu \delta_g(x,z(\tau))\,d\tau.$$

Testing its divergence against $X_\nu$ and integrating along the worldline gives

$$\left\langle\nabla_\mu T^{\mu\nu},X_\nu\right\rangle =m\int \frac{Du^\nu}{d\tau}X_\nu(z(\tau))\,d\tau,$$

after the endpoint term vanishes. Distributional conservation therefore implies geodesic motion for a free monopole source in the stated smooth background.

The endpoint assumption matters: a worldline segment ending inside the test function’s support contributes an endpoint distribution. Likewise, a varying monopole weight contributes a term proportional to $dm/d\tau$; conservation yields the displayed geodesic conclusion only for the stated free, constant-mass source.

This does not construct an exact point-particle spacetime. It is valid on a fixed smooth geometry and as a linearized source; nonlinear curvature contains products of weak derivatives. See Poisson, Pound, and Vega 2011, §§ 13.1 and 19.1, especially Eqs. (19.3)–(19.4). The Geroch–Traschen regular-metric class admits codimension-one concentrated curvature but excludes generic string and point sources within that class Geroch and Traschen 1987, p. 1022, Theorem 1.

Fundamental solutions and causal choices

A fundamental solution satisfies

$$P E=\delta.$$

A Green kernel adds domain, boundary, or support information:

$$P_xG(x,x')=\delta_g(x,x').$$

The equation alone does not usually select $G$. In $\mathbb R^3$,

$$-\Delta\frac{1}{4\pi r}=\delta^{(3)}(\mathbf x).$$

The normalization follows from a small-sphere flux. With $\Box=-\partial_t^2+\nabla^2$ and $P=-\Box$ in $3+1$ Minkowski spacetime,

$$G_{\mathrm{ret}}(t,\mathbf x) =\frac{\theta(t)}{4\pi r}\,\delta(t-r), \qquad P G_{\mathrm{ret}}=\delta(t)\delta^{(3)}(\mathbf x).$$

Its support is on the future light cone. Curvature or another dimension can produce an interior tail, so this Huygens property does not define “retarded” generally Duistermaat and Kolk 2010, pp. 115–152 and 271–285.

Adding any homogeneous solution preserves $PG=\delta$ while changing boundary or support behavior. Retarded, advanced, Feynman, Euclidean, radiative, and boundary-value kernels are therefore different inverses or parametrices even when they share the same local differential expression.

For a normally hyperbolic operator on a globally hyperbolic spacetime, advanced and retarded Green operators are unique with the corresponding support. Timelike boundaries, failure of global hyperbolicity, gauge constraints, zero modes, or non-self-adjoint boundary conditions require more data Bär, Ginoux, and Pfäffle 2007, Chapter 3, pp. 67–101.

For $\delta(r-r_0)$ or $\delta'(r-r_0)$ sources, integrate the exact differential equation across $r_0$. A copied unit-leading-coefficient jump can miss its measure, Jacobian, or normalization.

Products, wavefront sets, and regularization

Ordinary distribution theory is linear. Its product rules have a sharp boundary:

Proposed operation	Status
$aT$ with $a\in C^\infty$	Always defined.
Product of locally integrable functions whose ordinary product is locally integrable	Defines a regular distribution.
Product of distributions with disjoint singular supports	Defined by localization.
$uv$ satisfying Hörmander’s wavefront condition	Defined canonically within that criterion.
$H\delta$, $\delta^2$, or an arbitrary product in $\mathcal D'$	Not canonically defined.

The wavefront set records singular points and covector directions. A sufficient condition for $uv$ to exist is

$$(x,\xi)\in\operatorname{WF}(u) \quad\Longrightarrow\quad (x,-\xi)\notin\operatorname{WF}(v).$$

This condition is sufficient, not necessary. See Brouder, Dang, and Hélein 2014, §§ 2–3, especially Theorem 13, arXiv PDF pp. 5–12.

Disjoint singular supports are a simpler sufficient case, but overlapping singular supports alone do not decide the product: the covector directions matter. Conversely, failure of Hörmander’s criterion says that this canonical construction is unavailable; it does not prove that no problem-specific extension can be chosen. Any such extension must expose its local ambiguity.

A mollifier $\rho_\varepsilon(x)=\varepsilon^{-n}\rho(x/\varepsilon)$ with $\int\rho=1$ gives $T*\rho_\varepsilon\to T$ in $\mathcal D'$, but products need not converge independently of $\rho$. In one dimension,

$$\left\langle\rho_\varepsilon^2,\varphi\right\rangle =\frac{\varphi(0)}{\varepsilon} \int\rho(y)^2\,dy+O(1),$$

so a delta approximation has no finite distributional square without a subtraction whose scale and allowed local terms are extra input.

Colombeau algebras retain smooth representatives and permit nonlinear products. The relation $U\approx T$ means that $U$ is associated with $T$, not equal to it in the algebra. Even when generalized functions are associated with distributions, not arbitrary products preserve that association; addition, differentiation, and smooth multiplication do Steinbauer and Vickers 2006, §§ 2–3, arXiv PDF pp. 4–12.

In particular, $U\approx0$ does not imply $U^2\approx0$. A result obtained after smoothing must therefore state the embedding, admissible mollifiers, coordinate or geometric construction, and which observable is independent of those choices.

Gravity application map

Problem	Distributional object	Data that make it well defined	Stop condition
Material interface or shock	$H(F)$ and $\delta_\Sigma$	Defining function, orientation, measure, traces, flux law	No simple zero or no controlled traces
Timelike thin shell	$S_{ab}\delta_\Sigma$	Continuous induced metric, normal and $K_{ab}$ convention, two smooth sides	Discontinuous metric or unexamined $\delta^2$ terms
Point source in perturbation theory	Worldline delta $\delta_g(x,z)$	Smooth background, proper-time measure, tensor pairing, gauge and field operator	Treating the source as an exact nonlinear point-mass metric
Retarded perturbation	Green kernel $G_{\mathrm{ret}}$	Operator sign, measure, causal domain, initial/boundary conditions	Calling a fundamental solution “retarded” without support data
Curved-spacetime QFT	Two-point and time-ordered distributions	State, wavefront condition, extension or renormalization prescription	Multiplying coincident propagators pointwise
Rough or impulsive geometry	Distributional or nonlinear generalized curvature	Metric regularity class, differentiable structure, smoothing/embedding, invariant comparison	Assuming weak convergence survives nonlinear curvature

In the Geroch–Traschen class, $g$, $g^{-1}\in L^\infty_{\mathrm{loc}}$ and weak first metric derivatives lie in $L^2_{\mathrm{loc}}$, making curvature distributional. This is sufficient, not exhaustive: special Kerr–Schild, impulsive, conical, or Colombeau constructions require their own covariance and regularization-independence checks Steinbauer and Vickers 2006, §§ 2 and 4–8.

A fast validity check

Before accepting a singular-source calculation, answer these questions in order:

1. What is the test space, and is the object a scalar, density, tensor distribution, or current?
2. Which measure defines the delta, and what Jacobian appears under the chosen coordinates?
3. Which derivatives are weak, and what jump terms appear after integration by parts?
4. Does every multiplier have enough regularity on the singular support?
5. If two distributions are multiplied, which theorem, wavefront criterion, or generalized-function algebra defines the product?
6. Which topology defines the singular limit, and do the nonlinear observables converge in that topology?
7. Which boundary, causal, gauge, or regularization choice selects the solution?
8. Do conservation, dimensions, support, a pillbox or small-sphere integral, and a smooth limiting case agree?

Pointwise equations away from the singular set cannot determine the coefficient of a delta supported on that set.

Common pitfalls

Treating $\delta$ as an infinite function. Evaluate the full expression against a test function, including its measure and Jacobian.

Using $f\delta=f(0)\delta$ for nonsmooth $f$. The multiplier must be smooth near the support; $H\delta$ has no canonical value in ordinary distribution theory.

Passing a weak limit through a nonlinear map. From $u_\varepsilon\to u$ in $\mathcal D'$ one cannot infer $u_\varepsilon^2\to u^2$; prove stronger convergence or specify and test a product construction.

Forgetting that a Green function includes a choice. $PG=\delta$ fixes neither support nor boundary behavior; state the causal or boundary prescription.

References

• Bär, Christian, Nicolas Ginoux, and Frank Pfäffle. Wave Equations on Lorentzian Manifolds and Quantization. ESI Lectures in Mathematics and Physics. Zürich: European Mathematical Society, 2007. DOI. Open preprint.
• Bhattacharyya, Pulin Kumar. Distributions: Generalized Functions with Applications in Sobolev Spaces. Berlin and Boston: De Gruyter, 2012. DOI.
• Brouder, Christian, Nguyen Viet Dang, and Frédéric Hélein. “A Smooth Introduction to the Wavefront Set.” Journal of Physics A: Mathematical and Theoretical 47 (2014): 443001. DOI. Open PDF.
• Duistermaat, J. J., and J. A. C. Kolk. Distributions: Theory and Applications. Translated by J. P. van Braam Houckgeest. Boston: Birkhäuser, 2010. DOI.
• Geroch, Robert, and Jennie Traschen. “Strings and Other Distributional Sources in General Relativity.” Physical Review D 36, no. 4 (1987): 1017–1031. DOI.
• Hörmander, Lars. The Analysis of Linear Partial Differential Operators I: Distribution Theory and Fourier Analysis. 2nd ed. Classics in Mathematics. Berlin: Springer, 2003. DOI.
• Israel, Werner. “Singular Hypersurfaces and Thin Shells in General Relativity.” Il Nuovo Cimento B 44, no. 1 (1966): 1–14; erratum 48, no. 2 (1967): 463. Article DOI. Erratum DOI.
• National Institute of Standards and Technology. Digital Library of Mathematical Functions, § 1.16, “Distributions.” Version 1.2.7, 15 June 2026. DLMF.
• Poisson, Eric. A Relativist’s Toolkit: The Mathematics of Black-Hole Mechanics. Cambridge: Cambridge University Press, 2004. DOI.
• Poisson, Eric, Adam Pound, and Ian Vega. “The Motion of Point Particles in Curved Spacetime.” Living Reviews in Relativity 14 (2011): 7. DOI. Open full text.
• Steinbauer, Roland, and James A. Vickers. “The Use of Generalised Functions and Distributions in General Relativity.” Classical and Quantum Gravity 23, no. 10 (2006): R91–R114. DOI. Open PDF.