Tempered Distributions and Fourier Calculus

The Fourier transform of a singular object is defined by what it does to rapidly decreasing probes, not by an oscillatory integral evaluated point by point. This viewpoint turns delta functions, polynomially growing fields, and momentum-space Green kernels into members of one stable space: the tempered distributions $\mathcal S'(\mathbb R^d)$. It also makes the $i0$ in a propagator part of the distribution itself rather than a disposable decoration.

Required background. Test-Function Spaces, Distributions, Support, and Convergence supplies test-space duality, continuity, support, and distributional convergence.

Helpful background. Fourier Series, Fourier Transforms, and Plancherel Theory supplies the function-level transform, inversion, and norm identities.

Readers who need a focused review can use Fourier Distributions and Green Functions or the mathematics readiness diagnostic.

Schwartz functions are the right Fourier probes

A smooth function $\varphi$ on $\mathbb R^d$ belongs to the Schwartz space $\mathcal S(\mathbb R^d)$ when every polynomially weighted derivative is bounded:

$$q_{\alpha\beta}(\varphi) = \sup_{x\in\mathbb R^d} \left|x^\alpha\partial^\beta\varphi(x)\right| <\infty \qquad \text{for all multi-indices }\alpha,\beta.$$

These seminorms define the Schwartz topology. A sequence $\varphi_n\to\varphi$ in $\mathcal S$ precisely when every $q_{\alpha\beta}(\varphi_n-\varphi)$ tends to zero. Thus convergence controls all derivatives and every inverse-power tail at once. The coordinates and weights in this definition are Euclidean even when $\mathbb R^d$ later labels Minkowski momentum space; the Lorentzian metric enters algebraic expressions such as $p^2$, not the test-space topology.

Compactly supported smooth functions sit continuously and densely inside Schwartz space:

$$\mathcal D(\mathbb R^d) \lhook\joinrel\longrightarrow \mathcal S(\mathbb R^d).$$

The inclusion is strict. A Gaussian is Schwartz but not compactly supported. The decisive feature is that $\mathcal S$ is invariant under the Fourier transform. Its dual is consequently a distribution space closed under Fourier transformation while still controlling polynomial growth at infinity. For the structural theory, see Hörmander 2003, §§ 7.1–7.3; for the density, continuity criteria, and distributional Fourier extension, see Hunter 2014, Definitions 5.51, 5.55, and 5.63 and Theorem 5.61, PDF.

Tempered distributions control growth through continuity

A tempered distribution is a continuous complex-linear functional on $\mathcal S$:

$$\mathcal S'(\mathbb R^d) = \mathcal L_{\mathrm{cont}} \bigl(\mathcal S(\mathbb R^d),\mathbb C\bigr).$$

Here the pairing is linear, not sesquilinear. Continuity means that for each $u\in\mathcal S'$ there are a constant $C$, an integer $N$, and finitely many Schwartz seminorms such that

$$\left|\langle u,\varphi\rangle\right| \leq C \sum_{|\alpha|,|\beta|\leq N} q_{\alpha\beta}(\varphi).$$

Restricting a tempered distribution to the dense subspace $\mathcal D$ gives an ordinary distribution. The restriction map is injective, so one usually writes

$$\mathcal S'(\mathbb R^d) \subset \mathcal D'(\mathbb R^d).$$

The reverse inclusion is false: an arbitrary distribution may grow too quickly at infinity for a Schwartz probe to control it. Temperedness is a global growth condition, not a statement about local singularity. The delta distribution and all its derivatives are tempered even though they are singular, whereas the regular distribution represented by $e^{|x|^2}$ is not tempered.

A useful sufficient condition for a locally integrable function $f$ is

$$\int_{\mathbb R^d} |f(x)|(1+|x|)^{-N}\,\mathrm d^d x <\infty$$

for some $N$. In particular, every locally integrable function of at most polynomial growth defines a tempered distribution by integration. This criterion is not necessary. On $\mathbb R$,

$$e^x\cos(e^x) = \frac{\mathrm d}{\mathrm dx}\sin(e^x)$$

defines a tempered distribution because it is the distributional derivative of a bounded function, despite its exponentially large pointwise oscillations as $x\to+\infty$. Cancellation can therefore make a distribution tempered even when an absolute-growth estimate fails.

Polynomials, plane waves, finite measures of polynomial growth, and derivatives of all these objects are standard members of $\mathcal S'$. This closure under differentiation is one reason the space is suited to linear field equations.

Smooth multiplication has a growth condition

If $m\in C^\infty(\mathbb R^d)$ and every derivative of $m$ is polynomially bounded, then multiplication by $m$ maps $\mathcal S$ continuously to itself. Such a function is often called a Schwartz multiplier. It acts on tempered distributions by

$$\langle mu,\varphi\rangle = \langle u,m\varphi\rangle.$$

Polynomials and bounded smooth functions with polynomially bounded derivatives are safe examples. An arbitrary smooth function need not be: rapid growth can carry $m\varphi$ out of $\mathcal S$. Smoothness alone, which is sufficient for multiplying an element of $\mathcal D'$, is not sufficient for preserving $\mathcal S'$.

The site Fourier convention

For $\varphi\in\mathcal S(\mathbb R^d)$, define

$$\widetilde\varphi(p) = \mathcal F\varphi(p) = \int_{\mathbb R^d} \mathrm d^d x\, e^{+ip\cdot x}\varphi(x),$$

with inverse

$$\varphi(x) = \mathcal F^{-1}\widetilde\varphi(x) = \int_{\mathbb R^d} \frac{\mathrm d^d p}{(2\pi)^d}\, e^{-ip\cdot x}\widetilde\varphi(p).$$

In spacetime notation, $p\cdot x=p_\mu x^\mu$. Component formulas therefore use the dual coordinate pair $(x^\mu,p_\mu)$; this convention prevents an extra apparent sign when spatial contravariant components are introduced.

The map $\mathcal F:\mathcal S\to\mathcal S$ is a continuous linear isomorphism. With this normalization,

$$\mathcal F^2\varphi(x) = (2\pi)^d\varphi(-x).$$

For $f,\psi\in\mathcal S$, Fubini’s theorem gives

$$\begin{aligned} \langle\mathcal F u_f,\psi\rangle &= \int\mathrm d^d p\, \widetilde f(p)\psi(p) \\ &= \int\mathrm d^d x\, f(x)\, \mathcal F\psi(x). \end{aligned}$$

This identity supplies the extension to every tempered distribution:

$$\boxed{ \langle\mathcal F u,\psi\rangle = \langle u,\mathcal F\psi\rangle, \qquad u\in\mathcal S',\ \psi\in\mathcal S. }$$

Because $\mathcal F$ is an automorphism of the test space, this definition makes it an automorphism of $\mathcal S'$ as well. Its inverse is fixed by the same reflection and normalization:

$$\mathcal F^{-1}u(x) = \frac{1}{(2\pi)^d} \bigl(\mathcal F u\bigr)(-x)$$

in distribution notation. No oscillatory integral for $u$ is required. This is also why Fourier transformation is not defined on every element of $\mathcal D'$: the Fourier transform of a compactly supported test function is Schwartz but generally not compactly supported, so an arbitrary distribution cannot act on it.

Differentiation and multiplication exchange roles

Integration by parts on $\mathcal S$, followed by duality, gives

$$\mathcal F(\partial_\mu u) = -ip_\mu\mathcal Fu, \qquad \mathcal F(x^\mu u) = -i\frac{\partial}{\partial p_\mu}\mathcal Fu.$$ $$\boxed{ \mathcal F(\partial^\alpha u) = (-ip)^\alpha\mathcal F u, \qquad \mathcal F(x^\alpha u) = (-i\partial_p)^\alpha\mathcal F u. }$$

The minus signs follow from the positive phase $e^{+ip\cdot x}$. They are not portable from a source using $e^{-ip\cdot x}$ without translation.

Several normalization checks are immediate:

$$\begin{aligned} \mathcal F\delta_0&=1, & \mathcal F1&=(2\pi)^d\delta_0, \\ \mathcal F\delta_a&=e^{+ip\cdot a}, & \mathcal F(e^{-ik\cdot x}) &=(2\pi)^d\delta^{(d)}(p-k). \end{aligned}$$

For example,

$$\langle\mathcal F\delta_a,\psi\rangle = \langle\delta_a,\mathcal F\psi\rangle = \int\mathrm d^d p\, e^{+ip\cdot a}\psi(p),$$

so the third identity is an equality of tempered distributions rather than a pointwise Fourier integral.

Linear constant-coefficient differential equations become algebraic multiplier equations. If $P(\partial)=\sum_\alpha c_\alpha\partial^\alpha$, then

$$P(\partial)u=f \quad\Longleftrightarrow\quad P(-ip)\widetilde u(p)=\widetilde f(p).$$

The apparent algebraic division by $P(-ip)$ is precisely where singular kernels and boundary prescriptions enter.

Product and convolution formulas have domains

For Schwartz functions,

$$\mathcal F(f*g) = (\mathcal Ff)(\mathcal Fg), \qquad \mathcal F(fg) = \frac{1}{(2\pi)^d} \bigl(\mathcal Ff*\mathcal Fg\bigr).$$

Useful versions survive when only one factor is singular. If $u\in\mathcal S'$ and $\varphi\in\mathcal S$, the convolution

$$(u*\varphi)(x) = \left\langle u,\, y\longmapsto\varphi(x-y) \right\rangle$$

is a smooth function of at most polynomial growth, and

$$\mathcal F(u*\varphi) = (\mathcal Fu)(\mathcal F\varphi).$$

Likewise the product $u\varphi$ is tempered, and

$$\mathcal F(u\varphi) = \frac{1}{(2\pi)^d} \bigl(\mathcal Fu*\mathcal F\varphi\bigr).$$

Compact support of one factor permits further well-defined convolution cases, but compact support alone does not define the product of two distributions. These formulas therefore define neither $uv$ nor $u*v$ for two arbitrary tempered distributions. The undefined expression has merely moved from one representation to another. The local criterion and extension problem are developed in Products, Scaling Degree, and Distribution Extensions.

Boundary values define singular momentum kernels

For $\varepsilon>0$, the function $(s\pm i\varepsilon)^{-1}$ is smooth on the real line and grows slowly. Its limit as $\varepsilon\downarrow0$ exists in $\mathcal S'(\mathbb R)$. Separating real and imaginary parts,

$$\frac{1}{s\pm i\varepsilon} = \frac{s}{s^2+\varepsilon^2} \mp i\frac{\varepsilon}{s^2+\varepsilon^2},$$

shows that the first term tends to the Cauchy principal value and the second is an approximate delta:

$$\boxed{ \frac{1}{s\pm i0} = \operatorname{PV}\frac{1}{s} \mp i\pi\delta(s). }$$

The notation abbreviates a distributional boundary value:

$$\left\langle\frac{1}{s\pm i0},\psi\right\rangle = \lim_{\varepsilon\downarrow0} \int_{\mathbb R} \frac{\psi(s)}{s\pm i\varepsilon}\,\mathrm ds.$$

Dyatlov 2022, § 5.2.3 and Exercise 5.4(c), PDF constructs the principal value and proves these boundary-value identities.

For $m>0$ and the $(+---)$ quadratic form

$$q(p) = p^2-m^2 = (p^0)^2-|\mathbf p|^2-m^2,$$

the mass shell $q^{-1}(0)$ is a regular hypersurface. The corresponding boundary value is

$$\frac{1}{p^2-m^2+i0} = \operatorname{PV}\frac{1}{p^2-m^2} -i\pi\delta(p^2-m^2).$$

Indeed, $\mathrm dq$ cannot vanish on $q=0$: its vanishing would force $p=0$, where $q=-m^2\neq0$. This regularity is what licenses pulling the one-dimensional boundary value back through $q$.

Writing $E_{\mathbf p}=\sqrt{|\mathbf p|^2+m^2}$ makes the shell term explicit:

$$\begin{aligned} \left\langle \delta(p^2-m^2),\psi \right\rangle = \int_{\mathbb R^{d-1}} \frac{\mathrm d^{d-1}\mathbf p}{2E_{\mathbf p}} \Bigl[ \psi(E_{\mathbf p},\mathbf p) + \psi(-E_{\mathbf p},\mathbf p) \Bigr]. \end{aligned}$$

This is the regular-level-set delta formula of Dyatlov 2022, Proposition 10.12, PDF, applied to the two roots $p^0=\pm E_{\mathbf p}$. The principal-value part and the on-shell delta part are both tempered. Removing the $i0$, or reversing its sign, changes the distribution.

First QFT application: the Feynman inverse

The free massive scalar Feynman kernel in momentum space is

$$\widetilde\Delta_F(p) = \frac{i}{p^2-m^2+i0} \in\mathcal S'(\mathbb R^d).$$

Tong 2006–2007, § 2.7.1, Eq. (2.174) uses this $e^{-ip\cdot x}$ kernel with the same metric and $i0$ conventions.

Its position-space inverse transform is the distribution

$$\Delta_F(x) = \int_{\mathbb R^d} \frac{\mathrm d^d p}{(2\pi)^d}\, \frac{i\,e^{-ip\cdot x}}{p^2-m^2+i0}.$$

The displayed integral means $\mathcal F^{-1}\widetilde\Delta_F$, not a pointwise improper integral. The boundary-value identity resolves its momentum-space content:

$$\frac{i}{p^2-m^2+i0} = i\,\operatorname{PV}\frac{1}{p^2-m^2} + \pi\delta(p^2-m^2).$$

It is also a distributional inverse of the Klein–Gordon operator with the site’s normalization. Since

$$(\Box+m^2)e^{-ip\cdot x} = -(p^2-m^2)e^{-ip\cdot x}$$

and multiplication by the polynomial $p^2-m^2$ is defined, with

$$(p^2-m^2) \frac{1}{p^2-m^2+i0} = 1$$

as a distribution, Fourier inversion gives

$$\boxed{ (\Box+m^2)\Delta_F(x) = -i\delta^{(d)}(x). }$$

This sign and normalization agree with Tong 2006–2007, § 2.7.2, Eq. (2.175). The same free Feynman kernel and Klein–Gordon contact term are treated in Schwartz 2014, §§ 6.2 and 24.1.

This calculation determines the inverse equation but does not by itself derive time ordering or compare Feynman, retarded, and advanced boundary conditions. Those physical statements belong to Scalar Propagators, Ordered Correlators, and Sources.

Operation checklist and stop conditions

Before applying Fourier calculus to a singular expression, check:

1. Test space. Has the object been shown to lie in $\mathcal S'$? Local distributional existence in $\mathcal D'$ does not control growth at infinity.
2. Convention. Record the phase and every $(2\pi)^d$ factor before importing derivative, inversion, or delta formulas.
3. Multiplier. A smooth multiplier must preserve $\mathcal S$; smoothness without polynomial derivative bounds is not enough.
4. Product or convolution. Identify a Schwartz factor, compact support, or another theorem that makes the operation legitimate. Do not multiply two singular kernels formally.
5. Boundary prescription. Treat $+i0$, $-i0$, and principal value as different distributions. An algebraic denominator alone does not select an inverse across its zero set.
6. Equation check. Apply the differential operator and verify the resulting delta normalization and sign in the declared convention.

Stop when one of these data is missing. A formal Fourier integral can hide both an undefined operation and a wrong boundary condition.

Exercises

1. In the convention of this page, compute $\mathcal F(\partial_\mu\delta_a)$.

   Check

   Differentiation becomes multiplication by the covariant component $-ip_\mu$, while $\mathcal F\delta_a=e^{ip\cdot a}$. Therefore

   $$\mathcal F(\partial_\mu\delta_a) = -ip_\mu e^{ip\cdot a}.$$

   Pairing both sides with a Schwartz function also follows directly from $\langle\partial_\mu\delta_a,\varphi\rangle =-\partial_\mu\varphi(a)$.

2. Show that the constant function $1$ is tempered and recover its Fourier transform.

   Check

   Every Schwartz function is integrable, and its integral is bounded by finitely many Schwartz seminorms, so $\langle1,\varphi\rangle=\int\varphi$ is continuous. For $\psi\in\mathcal S$,

   $$\langle\mathcal F1,\psi\rangle = \int\mathrm d^d x\,\mathcal F\psi(x) = (2\pi)^d\psi(0).$$

   Hence $\mathcal F1=(2\pi)^d\delta_0$.

3. Find the difference between the two boundary values $(s+i0)^{-1}$ and $(s-i0)^{-1}$.

   Check

   Their principal-value parts cancel, while the delta parts add:

   $$\frac{1}{s+i0} - \frac{1}{s-i0} = -2\pi i\,\delta(s).$$

   Thus changing the sign of $i0$ changes a term supported exactly at the singular set.

4. Apply $\Box+m^2$ to the inverse transform of $-i/(p^2-m^2+i0)$.

   Check

   The Klein–Gordon operator contributes $-(p^2-m^2)$. Multiplication by the proposed kernel gives $+i$, whose inverse Fourier transform is $+i\delta^{(d)}(x)$. Reversing the numerator therefore reverses the Green-equation source.

References

• Semyon Dyatlov, Lecture Notes for 18.155: Differential Analysis, §§ 5.2.3 and 11.1–11.3, PDF, MIT, 2022. This is the teaching source for Schwartz seminorms, tempered distributions, dual Fourier transformation, the calculus identities, safe Schwartz convolution, principal value, and boundary values. Dyatlov uses a negative forward phase; the formulas here have been translated to the site’s positive forward phase.
• Lars Hörmander, The Analysis of Linear Partial Differential Operators I: Distribution Theory and Fourier Analysis, 2nd ed., §§ 7.1–7.3, Springer, 2003. This is the structural reference for tempered distributions and Fourier analysis.
• John K. Hunter, *Notes on Partial Differential Equations, Chapter 5: The Fourier Transform, PDF, Definitions 5.51, 5.55, and 5.63; Examples 5.56–5.58; and Theorem 5.61, University of California, Davis, revised 2014. This supplies the density of $\mathcal D$ in $\mathcal S$, continuity criteria and counterexamples for tempered distributions, and the distributional Fourier extension. Hunter uses a normalized negative-phase transform; both its phase and normalization have been converted here.
• Matthew D. Schwartz, Quantum Field Theory and the Standard Model, §§ 6.2 and 24.1, Cambridge University Press, 2014. Book record. This is the QFT source for the free Feynman kernel, its Klein–Gordon equation, and the delta contribution obtained from the $i0$ prescription.
• David Tong, Quantum Field Theory, § 2.7: Propagators, Cambridge Part III lecture notes, University of Cambridge, 2006–2007. Equations (2.174)–(2.175) fix the $e^{-ip\cdot x}$ Feynman kernel and the source $-i\delta^{(d)}$ in the $(+---)$ convention used here.