Hyperbolic Equations and Causal Propagators

Hyperbolicity supplies a well-posed Cauchy problem and a finite domain of dependence only under appropriate hypotheses on the operator, data surface, and function spaces. Within that setting, a future-supported inverse is the retarded Green operator, while a past-supported inverse is the advanced Green operator. Their difference is the causal propagator: it solves the homogeneous equation and carries Cauchy or commutator data, so it is not itself an inverse. A Feynman-type inverse makes a different choice, selecting frequency and pole boundary values rather than future or past cone support.

Required background. Fundamental Solutions and Green Operators supplies distributional kernels and typed inverses; Symbols, Characteristics, and PDE Type supplies characteristic covectors and cones.

Helpful background. Contour Deformation, Pinches, and Causal Prescriptions supplies the contour argument behind the pole placements used below.

Hyperbolicity controls propagation, not the inverse

Work first on time-oriented Minkowski space $\mathbb R^{1+n}$ with the site’s $(+,-,\ldots,-)$ metric, and let

$$P=\Box+m^2 = \partial_t^2-\nabla^2+m^2, \qquad m\geq0.$$

For a set $A$, its causal future and past are

$$\begin{aligned} J^+(A) &= \left\{ x:\text{some }a\in A \text{ has }x^0\geq a^0 \text{ and }(x-a)^2\geq0 \right\},\\ J^-(A) &= \left\{ x:\text{some }a\in A \text{ has }x^0\leq a^0 \text{ and }(x-a)^2\geq0 \right\}. \end{aligned}$$

The mass term changes dispersion but not the principal symbol. Thus the characteristic set of $P$ is the null cone even when $m>0$. That local classification does not yet choose an inverse.

The Cauchy problem on a constant-time hypersurface $\Sigma_{t_0}$ is

$$\begin{aligned} Pu&=f,\\ u|_{\Sigma_{t_0}}&=u_0,\\ \partial_tu|_{\Sigma_{t_0}}&=u_1. \end{aligned}$$

For smooth compactly supported data, it has a unique smooth solution that depends continuously on $(f,u_0,u_1)$. Its support is contained in $J(K)$, where $K$ is the combined support of the source and initial data. In particular, for evolution to the future of $\Sigma_{t_0}$, the value at $x$ depends only on source and initial data in the backward causal cone of $x$. This is the domain-of-dependence statement, not merely a statement that plane-wave frequencies are real. The corresponding global theorem for normally hyperbolic operators requires a globally hyperbolic spacetime; Bär, Ginoux, and Pfäffle 2007, §3.2, especially Theorems 3.2.11–3.2.12, PDF, give precise hypotheses and topologies.

For a real solution, the local energy calculation makes the mechanism visible. Multiplying $Pu=f$ by $\partial_tu$ gives

$$\partial_t \left[ \frac12(\partial_tu)^2 +\frac12|\nabla u|^2 +\frac12m^2u^2 \right] - \nabla\mathbin{\cdot} \left[ (\partial_tu)\nabla u \right] = f\,\partial_tu.$$

Integrating this identity over a truncated backward cone controls the localized energy on each spacelike cross-section by data and forcing in the cone. If those vanish, the localized energy vanishes on every cross-section, so the solution vanishes through the cone and at its tip. This energy argument supplies uniqueness and finite propagation for the wave and Klein–Gordon equations; it also explains why a support theorem needs estimates, not just a characteristic polynomial.

Stop rule. Do not infer a global causal inverse from the word “hyperbolic” alone. One still needs a well-posed operator realization, admissible initial or boundary data, and a global setting in which the relevant support condition determines a unique solution.

Support conditions select two Green operators

Let $\mathcal D=C_c^\infty(\mathbb R^{1+n})$. The retarded and advanced Green operators are linear maps

$$G_{\mathrm{ret}},G_{\mathrm{adv}}: \mathcal D \longrightarrow C^\infty(\mathbb R^{1+n})$$

such that, for every $f\in\mathcal D$,

$$\begin{aligned} P G_{\mathrm{ret}}f &= f, & G_{\mathrm{ret}}Pf &= f, & \operatorname{supp}(G_{\mathrm{ret}}f) &\subseteq J^+(\operatorname{supp}f),\\ P G_{\mathrm{adv}}f &= f, & G_{\mathrm{adv}}Pf &= f, & \operatorname{supp}(G_{\mathrm{adv}}f) &\subseteq J^-(\operatorname{supp}f). \end{aligned}$$

These equations define the names on this page: retarded means future-supported response, and advanced means past-supported response. This avoids an ambiguity in the literature, where the labels attached to $G_+$ and $G_-$ are not uniform. Existence and uniqueness in the normally-hyperbolic, globally-hyperbolic setting are stated in Bär, Ginoux, and Pfäffle 2007, Definition 3.4.1 and Corollary 3.4.3, PDF; sequential continuity is Proposition 3.4.8.

Define the causal propagator with the local sign convention

$$E \equiv G_{\mathrm{ret}}-G_{\mathrm{adv}}.$$

Then

$$PEf=0, \qquad EPf=0 \qquad (f\in\mathcal D),$$

because the two inverse identities cancel. Moreover,

$$\operatorname{supp}(Ef) \subseteq J(\operatorname{supp}f) \equiv J^+(\operatorname{supp}f) \cup J^-(\operatorname{supp}f).$$

The four objects needed in this chapter can therefore be separated without relying on the overloaded word “propagator.”

Object	Defining equation	Selection data	Cone support?
$G_{\mathrm{ret}}$	$PG_{\mathrm{ret}}=I$	no response before a compact source	future-supported
$G_{\mathrm{adv}}$	$PG_{\mathrm{adv}}=I$	no response after a compact source	past-supported
$E=G_{\mathrm{ret}}-G_{\mathrm{adv}}$	$PE=0$	difference of the two support-selected inverses	causal, but both time directions
$G_F$	$PG_F=I$	Feynman frequency and pole boundary values	not confined to a causal cone

Thus “causal propagator” does not mean “retarded inverse.” It denotes the homogeneous difference $E$ on this page. The sign is declared because some later identities, including the commutator crosswalk, reverse sign if the subtraction order is reversed.

Constructing the kernels one spatial mode at a time

The flat Klein–Gordon construction reduces the PDE to a family of forced oscillators. Use the spatial transform compatible with the site’s spacetime Fourier convention,

$$\begin{aligned} \widehat f(t,\mathbf k) &= \int\mathrm d^n x\, e^{-i\mathbf k\cdot\mathbf x} f(t,\mathbf x),\\ f(t,\mathbf x) &= \int\frac{\mathrm d^n k}{(2\pi)^n}\, e^{+i\mathbf k\cdot\mathbf x} \widehat f(t,\mathbf k), \end{aligned}$$

and set

$$\omega_{\mathbf k} = \sqrt{\mathbf k^2+m^2}.$$

Each mode obeys

$$\left( \partial_t^2+\omega_{\mathbf k}^2 \right) \widehat u(t,\mathbf k) = \widehat f(t,\mathbf k).$$

The two delta-normalized oscillator kernels are

$$\begin{aligned} g_{\mathrm{ret}}(t;\mathbf k) &= \theta(t) \frac{\sin(\omega_{\mathbf k}t)} {\omega_{\mathbf k}},\\ g_{\mathrm{adv}}(t;\mathbf k) &= -\theta(-t) \frac{\sin(\omega_{\mathbf k}t)} {\omega_{\mathbf k}}. \end{aligned}$$

At $\omega_{\mathbf k}=0$, the quotient is understood by its limit $t$. Away from $t=0$, each expression solves the homogeneous oscillator equation. At $t=0$, the function is continuous and its first derivative has jump $+1$, so

$$\left( \partial_t^2+\omega_{\mathbf k}^2 \right) g_{\mathrm{ret/adv}} = \delta(t)$$

as a distribution. This jump is the normalization check that fixes every later sign.

Transforming back gives the distributional kernels

$$\begin{aligned} G_{\mathrm{ret}}(t,\mathbf x) &= \theta(t) \int\frac{\mathrm d^n k}{(2\pi)^n}\, e^{+i\mathbf k\cdot\mathbf x} \frac{\sin(\omega_{\mathbf k}t)} {\omega_{\mathbf k}},\\ G_{\mathrm{adv}}(t,\mathbf x) &= -\theta(-t) \int\frac{\mathrm d^n k}{(2\pi)^n}\, e^{+i\mathbf k\cdot\mathbf x} \frac{\sin(\omega_{\mathbf k}t)} {\omega_{\mathbf k}}. \end{aligned}$$

For $f\in\mathcal D$, convolution with either kernel is well defined. The retarded solution is explicitly

$$\begin{aligned} (G_{\mathrm{ret}}f)(t,\mathbf x) = \int_{-\infty}^{t}\mathrm dt' \int\mathrm d^n y \int\frac{\mathrm d^n k}{(2\pi)^n}\, e^{+i\mathbf k\cdot(\mathbf x-\mathbf y)} \frac{ \sin[\omega_{\mathbf k}(t-t')] }{ \omega_{\mathbf k} } f(t',\mathbf y). \end{aligned}$$

The upper limit implements “no response before the source” in the chosen time orientation. Finite propagation strengthens this time-order statement: only source points in $J^-(t,\mathbf x)$ can contribute.

Subtracting the two kernels removes the step functions:

$$E(t,\mathbf x) = \int\frac{\mathrm d^n k}{(2\pi)^n}\, e^{+i\mathbf k\cdot\mathbf x} \frac{\sin(\omega_{\mathbf k}t)} {\omega_{\mathbf k}}.$$

It has the distributional Cauchy data

$$E(0,\mathbf x)=0, \qquad \partial_tE(0,\mathbf x) = \delta^{(n)}(\mathbf x).$$

Together with the homogeneous mode equation, these data verify $PE=0$ and independently fix the sign in $E=G_{\mathrm{ret}}-G_{\mathrm{adv}}$.

Pole positions encode the same support choice

For independent distributional constructions of the one- and three-dimensional future fundamental solutions and the finite-propagation estimate, see Dyatlov 2022, Proposition 9.11, Theorem 10.14, and Eqs. (10.31)–(10.35), PDF.

With the full Fourier pair

$$\widetilde f(p) = \int\mathrm d^{1+n}x\, e^{+ip\cdot x}f(x), \qquad f(x) = \int\frac{\mathrm d^{1+n}p}{(2\pi)^{1+n}}\, e^{-ip\cdot x}\widetilde f(p),$$

one has $\partial_\mu\mapsto-ip_\mu$ and

$$P \longmapsto m^2-p^2.$$

For fixed $\mathbf p$, let $\omega_{\mathbf p}=\sqrt{\mathbf p^2+m^2}$. The delta-normalized boundary values are

$$\begin{aligned} \widetilde G_{\mathrm{ret}}(p) &= -\frac{1}{ (p^0+i0)^2-\omega_{\mathbf p}^2 },\\ \widetilde G_{\mathrm{adv}}(p) &= -\frac{1}{ (p^0-i0)^2-\omega_{\mathbf p}^2 },\\ \widetilde G_F(p) &= -\frac{1}{ p^2-m^2+i0 }. \end{aligned}$$

Here every $i0$ is a distributional boundary value. The first denominator places both energy poles below the real axis, the second places both above, and the Feynman denominator places the positive-energy pole below and the negative-energy pole above.

Because the inverse transform contains $e^{-ip^0t}$, the contour closes below for $t>0$ and above for $t<0$. The retarded integral therefore vanishes for $t<0$, while the advanced integral vanishes for $t>0$. This reproduces the time-domain oscillator construction by an independent route. The recommended contour page gives the deformation and large-arc conditions that this shorthand suppresses.

The boundary-value identity

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

then yields

$$\widetilde E(p) = \widetilde G_{\mathrm{ret}}(p) - \widetilde G_{\mathrm{adv}}(p) = 2\pi i\, \operatorname{sgn}(p^0) \delta(p^2-m^2).$$

Multiplication by $m^2-p^2$ gives zero. The causal propagator lives on the mass shell in momentum space even though the characteristic set of the differential operator is the mass-independent null cone. Conflating those two sets would erase the distinction between propagation of singularities and the dispersion of solutions.

The Feynman kernel is also an inverse, but it is not a future- or past-supported response. In the site’s free-field normalization,

$$\widetilde\Delta_F(p) = \frac{i}{p^2-m^2+i0}, \qquad P\Delta_F = -i\delta^{(1+n)},$$

so the delta-normalized mathematical kernel in the table is

$$G_F=i\Delta_F.$$

Its opposite pole placements propagate positive and negative frequencies in different time directions. They do not force the position-space kernel to vanish at spacelike separation. State choice, time ordering, and the physical interpretation of $\Delta_F$ belong to Foundations.

A checked example in three spatial dimensions

For $m=0$ and $n=3$, the mode integral becomes

$$G_{\mathrm{ret}}(t,\mathbf x) = \theta(t) \int\frac{\mathrm d^3 k}{(2\pi)^3}\, e^{+i\mathbf k\cdot\mathbf x} \frac{\sin(|\mathbf k|t)} {|\mathbf k|}.$$

Writing $r=|\mathbf x|$, angular integration and

$$\int_0^\infty\mathrm dk\, \sin(kr)\sin(kt) = \frac{\pi}{2} \left[ \delta(t-r)-\delta(t+r) \right]$$

give

$$\boxed{ G_{\mathrm{ret}}(t,\mathbf x) = \frac{\theta(t)}{2\pi} \delta(t^2-r^2) = \frac{\delta(t-r)}{4\pi r} }.$$

The future-support condition is visible, the dimension is $[G_{\mathrm{ret}}]=2$ in mass units, and integrating the mode equation across $t=0$ gives the unit delta normalization. The advanced and causal kernels are

$$\begin{aligned} G_{\mathrm{adv}}(t,\mathbf x) &= \frac{\delta(t+r)}{4\pi r},\\ E(t,\mathbf x) &= \frac{ \delta(t-r)-\delta(t+r) }{ 4\pi r } = \frac{\operatorname{sgn}(t)}{2\pi} \delta(t^2-r^2). \end{aligned}$$

This exact light-cone support is the sharp Huygens property in $3+1$ massless flat spacetime; it is not the definition of causal support and is not universal. For example, the $1+1$ massless retarded kernel is $\tfrac12\theta(t-|x|)$, which fills the cone. Jaffe 2016, MIT 18.155 Lecture 21, §§5–6, PDF derives the Cauchy representation, finite propagation, and the dimension-dependent Huygens statement.

Mass also produces an interior tail. In $3+1$ dimensions, with $x^2=t^2-r^2$,

$$\begin{aligned} G_{\mathrm{ret}}^{(m)}(x) = \theta(t) \left[ \frac{\delta(x^2)}{2\pi} - \theta(x^2) \frac{ mJ_1(m\sqrt{x^2}) }{ 4\pi\sqrt{x^2} } \right]. \end{aligned}$$

The Bessel tail is nonzero in the timelike interior; its distributional support is the closed future cone because the null boundary is a limit of that interior. It vanishes in the $m\to0$ limit because $J_1(z)\sim z/2$, leaving the massless cone term. This is a useful check that finite propagation means support inside or on the causal cone, not necessarily only on it. A derivation and direct Green equation check appear in Lienert 2018, §3.3, especially equations (3.45)–(3.54), PDF, after translating the source’s source-point and observation-point notation.

For a compact source in the massless $3+1$ case, convolution collapses the time integral:

$$(G_{\mathrm{ret}}f)(t,\mathbf x) = \int_{\mathbb R^3} \frac{\mathrm d^3 y}{ 4\pi|\mathbf x-\mathbf y| } f\!\left( t-|\mathbf x-\mathbf y|, \mathbf y \right).$$

The field at $(t,\mathbf x)$ samples each source point at its retarded time. Applying $\Box$ recovers $f$, while replacing the retarded time by an advanced time produces the other inverse rather than the same solution.

First QFT application: causal scalar response

Consider the classical sourced free scalar equation

$$(\Box+m^2)\phi=J, \qquad J\in C_c^\infty(\mathbb R^{1+n}).$$

The solution whose source-dependent part has no support outside $J^+(\operatorname{supp}J)$ is

$$\phi_{\mathrm{ret}} = \phi_{\mathrm{hom}} + G_{\mathrm{ret}}J,$$

where the homogeneous term carries independently specified incoming Cauchy data and may itself be nonzero before the source. If the total field is required to vanish before the source, the incoming data must be chosen compatibly; in the simplest case, $\phi_{\mathrm{hom}}=0$.

Convolution with $G_{\mathrm{ret}}$ extends from test sources to compactly supported distributions. For a massless point impulse $J=q\,\delta(t)\delta^{(3)}(\mathbf x)$, the sourced part is

$$\phi_{\mathrm{src,ret}}(t,\mathbf x) = \frac{q\,\delta(t-r)}{4\pi r}.$$

This field is distributional. It reaches radius $r$ at $t=r$ and nowhere earlier. Replacing $G_{\mathrm{ret}}$ by $G_F$ would still solve the distributional equation, but would no longer implement this response condition.

The same causal kernel controls the free-field commutator, but only after the normalization is translated. With the causal-propagator sign fixed above and the site’s canonical scalar normalization,

$$\left[ \widehat\phi(x), \widehat\phi(y) \right] = -iE(x-y)\,\mathbf 1.$$

At equal times,

$$\begin{aligned} \left[ \widehat\phi(t,\mathbf x), \widehat\pi(t,\mathbf y) \right] &= \partial_{y^0} \left[ -iE(x-y) \right]_{x^0=y^0}\\ &= i\delta^{(n)}(\mathbf x-\mathbf y)\,\mathbf 1, \end{aligned}$$

because $\partial_tE(0,\mathbf x)=\delta^{(n)}(\mathbf x)$. This recovers the canonical commutator and independently fixes the factor $-i$. Since $\operatorname{supp}E$ is causal, the commutator vanishes at spacelike separation in the free scalar theory.

This calculation is the promised controlled QFT-facing example, not a complete taxonomy of quantum correlators. Vacuum spectral functions, linear-response conventions, state dependence, and the distinction among commutator, Wightman, retarded, advanced, and time-ordered correlators belong to Retarded, Advanced, and Spectral Correlators. Tong 2006, §§2.6.1–2.7.2 supplies the free-scalar commutator, Feynman prescription, and classical retarded-response comparison in the physical normalization.

Where the method stops

The procedure on this page has four inputs: a hyperbolic operator with a well-posed realization, a time orientation, a declared test-function or data space, and a support or pole condition. It outputs a particular inverse or, after subtraction, a homogeneous causal solution operator. The checks are the delta equation, support inclusion, oscillator jump, pole placement, equal-time Cauchy data, and agreement between time- and momentum-space constructions.

It does not establish a causal inverse in each of the following situations:

• a boundary makes additional data necessary but none are supplied;
• the spacetime lacks a suitable global Cauchy surface;
• a constrained or gauge system has not been reduced or gauge-fixed into a well-posed system;
• the formal Fourier quotient is used without a declared distributional boundary value.

Developed curved-spacetime causal geometry and state theory continue in Green Operators, Causal Propagators, and State-Dependent Two-Point Functions. Elliptic boundary conditions and heat evolution instead have a different analytic structure and continue in Elliptic Boundary Problems and Heat Kernels.

Common pitfalls

Calling the causal propagator an inverse. Both $G_{\mathrm{ret}}$ and $G_{\mathrm{adv}}$ invert $P$ on test sources. Their difference $E$ satisfies $PE=0$ because the two delta sources cancel.

Using “causal” to mean Feynman. The Feynman prescription is essential to time-ordered vacuum amplitudes, but its kernel is not confined to the future or past causal cone. Pole boundary data and support data answer different questions.

Reading the mass shell as the characteristic cone. The mass shell $p^2=m^2$ describes on-shell Klein–Gordon modes. The characteristic cone is obtained from the principal symbol and remains $p^2=0$ when a mass term is added.

Assuming every wave travels only on the cone. Sharp Huygens propagation holds for the massless wave equation only in particular dimensions and geometries. Causal support permits timelike interior tails.

Inferring a prescription from a bare denominator. The expression $1/(p^2-m^2)$ is undefined as an ordinary inverse on its real zero set. Retarded, advanced, Feynman, and principal-value boundary values must be distinguished distributionally.

Exercises

1. Verify the delta normalization of the retarded oscillator kernel.

   Solution

   Let

   $$g(t) = \theta(t)\frac{\sin(\omega t)}{\omega}.$$

   It is continuous at $t=0$, with $g'(0^-)=0$ and $g'(0^+)=1$. A continuous piecewise smooth function whose first derivative jumps by one has a delta of coefficient one in its second derivative. Away from zero, $g''+\omega^2g=0$, hence

   $$(g''+\omega^2g)=\delta(t).$$

2. Show directly that the causal propagator is not a Green inverse.

   Solution

   By definition,

   $$E=G_{\mathrm{ret}}-G_{\mathrm{adv}}.$$

   Therefore, on a test source $f$,

   $$PEf = PG_{\mathrm{ret}}f - PG_{\mathrm{adv}}f = f-f = 0.$$

   The same subtraction gives $EPf=0$. Its nontrivial information is its Cauchy data and causal support, not a delta source.

3. Derive the momentum-space causal kernel from the pole boundary values.

   Solution

   Writing $x=p^2-m^2$, the retarded and advanced denominators differ by the sign of the boundary value, weighted by the sign of $p^0$. Using

   $$\frac1{x+i0} = \operatorname{PV}\frac1x - i\pi\delta(x),$$

   one obtains

   $$\widetilde E(p) = 2\pi i\, \operatorname{sgn}(p^0) \delta(p^2-m^2).$$

   The principal-value terms cancel. Multiplication by $m^2-p^2$ vanishes, confirming the homogeneous equation.

4. Check the sign relating $E$ to the canonical commutator.

   Solution

   Suppose $[\widehat\phi(x),\widehat\phi(y)]=cE(x-y)\mathbf1$. Differentiating with respect to $y^0$ at equal times gives

   $$\left[ \widehat\phi(t,\mathbf x), \widehat\pi(t,\mathbf y) \right] = -c\, \partial_tE(0,\mathbf x-\mathbf y)\mathbf1 = -c\, \delta^{(n)}(\mathbf x-\mathbf y)\mathbf1.$$

   Matching the canonical value $i\delta^{(n)}\mathbf1$ requires $c=-i$.

References

• Christian Bär, Nicolas Ginoux, and Frank Pfäffle (2007), Wave Equations on Lorentzian Manifolds and Quantization, PDF, §§3.2–3.4, especially Theorems 3.2.11–3.2.12, Definition 3.4.1, Corollary 3.4.3, and Theorem 3.4.7, together with Proposition 3.4.8 for sequential continuity. These results establish Cauchy well-posedness, finite propagation, support-selected Green operators, and the causal difference. It uses signature $(-,+,\ldots,+)$ and calls its future-supported $G_+$ “advanced”; this page translates the signature and defines the physics names directly by support.
• Semyon Dyatlov (2022), Lecture Notes for 18.155: Distributions, Elliptic Regularity, and Applications to PDEs, PDF, Proposition 9.11, Theorem 10.14, and equations (10.31)–(10.35), develops the one- and three-dimensional future fundamental solutions, Cauchy construction, and finite-propagation support estimate. Dyatlov calls the future-supported solution “advanced, or future”; this page translates by support to the physics name “retarded.”
• Ethan Jaffe (2016), MIT 18.155 Lecture 21: Forward Fundamental Solution of the Wave Operator, PDF, §§5–6. This supplies an independent derivation of finite propagation and the dimension-sensitive sharp Huygens result.
• Matthias Lienert (2018), Wave Equations of Relativistic Quantum Mechanics, PDF, §3.3, especially equations (3.45)–(3.54). This supports the explicit massive Klein–Gordon retarded kernel and its interior Bessel tail after a convention translation.
• David Tong (2006), Quantum Field Theory, §§2.6.1–2.7.2, especially equations (2.162)–(2.177). These sections derive the free-field commutator, Feynman prescription, and retarded response.