Green's Functions via Integral Transforms

A Green's function is the inverse of a differential operator written out as a kernel: solve $L_x G(x,x')=\delta(x-x')$ once, and every inhomogeneous problem $L\,u=f$ is solved by a single integral, $u=\int G\,f$ . The question this article answers is the practical one — how do you actually find $G$ ? — and the answer, whenever the operator has constant coefficients, is the integral-transform method: one machine that produces the Coulomb potential, the heat kernel, the damped-oscillator impulse response, and the propagators of quantum field theory by the same three moves. The trick is that transforms trade calculus for algebra. Under the Fourier transform every $\partial_x$ becomes multiplication by $ik$ , and under the Laplace transform $\partial_t$ becomes $s$ ; a constant-coefficient operator therefore collapses to a polynomial, its symbol $L(ik)$ or $L(s)$ . The defining equation $L\,G=\delta$ transforms to the purely algebraic $L(ik)\,\tilde G(k)=1$ — because the point source $\delta$ has the flat spectrum $1$ — and we read off $\tilde G(k)=1/L(ik)$ by division. The differential equation is inverted without a single derivative surviving. Everything subtle, and everything physical, then lives in the inverse transform. The function $1/L(ik)$ has poles wherever $L(ik)=0$ — exactly the operator's free modes, its resonances — and may carry branch cuts. How the inversion contour is routed around those singularities is the boundary condition: it selects retarded versus advanced, outgoing versus incoming, causal versus not. Made rigorous in the theory of tempered distributions, the method delivers $G$ as the inverse transform of the distribution $1/L(ik)$ — one of the most transferable techniques in all of mathematical physics.

Key takeaways

A Green's function is the response of a linear operator $L$ to a unit point source, $L_x G(x,x')=\delta(x-x')$ ; convolving it against any forcing gives $u=\int G f$ , the solution of $L\,u=f$ .
Integral transforms turn the defining PDE into algebra: differentiation becomes multiplication ( $\partial_x\mapsto ik$ , $\partial_t\mapsto s$ ), so $L\,G=\delta$ becomes $L(ik)\tilde G(k)=1$ , giving $\tilde G(k)=1/L(ik)$ .
The Green's function is recovered by an inverse transform — a contour integral whose pole structure is the operator's spectrum and whose contour routing fixes the boundary and causality conditions.
Poles off the real axis give exponentially decaying kernels (screened, Yukawa); poles on the real axis force an $i\epsilon$ choice that separates retarded, advanced, and Feynman Green's functions.
The same machine yields the Coulomb kernel $1/4\pi r$ , the screened kernel $e^{-\mu|x|}/2\mu$ , the damped-oscillator ringdown (poles $s=-3\pm4i$ , peak $G\approx0.100$ at $t=0.232\ \mathrm{s}$ ), the heat kernel, and the QFT propagator from $1/(p^2-m^2)$ .
Key idea: transform $L\,G=\delta$ so differentiation becomes multiplication, $L(ik)\tilde G(k)=1$ , hence $\tilde G(k)=1/L(ik)$ ; the Green's function is the inverse transform, and the contour chosen around the poles of $1/L(ik)$ fixes the boundary and causality conditions.

The core insight

Transform $L\,G=\delta$ to turn differentiation into multiplication: $L(ik)\tilde G(k)=1$ , so $\tilde G(k)=1/L(ik)$ . The Green's function is the inverse transform, and the contour chosen around the poles of $1/L(ik)$ fixes the boundary and causality conditions.

Where this lives today

The transform construction of Green's functions is the daily arithmetic of theoretical physics. In quantum field theory the Feynman propagator is literally the inverse Fourier transform of $1/(p^2-m^2)$ , and the $i\epsilon$ that defines it ( $m^2\to m^2-i\epsilon$ ) is the same contour choice made here for the Helmholtz operator; every Feynman diagram is a product of such Green's functions. In many-body theory the retarded Green's function $G^R(\omega,\mathbf k)$ , built by exactly this machinery, is what angle-resolved photoemission (ARPES) measures — its poles map electronic quasiparticle dispersions and lifetimes. Numerical relativity and antenna design use retarded wave-operator kernels to propagate radiation causally, and lattice QCD computes propagators as inverses of a discretised operator — the modern descendant of $\tilde G=1/L(ik)$ . The open frontier is operators that are not translation-invariant — disordered, interacting, or strongly time-dependent — where $1/L(ik)$ fails and $G$ must be built from spectral or numerical resolvents instead.

The physical picture

Read $L\,u=f$ as an infinite system of coupled linear equations. Its solution is $u=L^{-1}f$ , and the Green's function is that inverse operator written as a kernel, $u(x)=\int G f\,\mathrm{d}x'$ . Inverting a matrix is trivial in the basis that diagonalises it — and for a translation-invariant operator that basis is the plane waves $e^{ikx}$ . There $L$ is multiplication by the number $L(ik)$ , so $L^{-1}$ is multiplication by $1/L(ik)$ , and transforming back is the inverse Fourier transform. The source $\delta$ becomes $1$ because it is flat in frequency, so $L^{-1}$ weights mode $k$ by exactly $1/L(ik)=\tilde G(k)$ . The whole story hangs on where $1/L(ik)$ blows up. Its poles sit at the operator's free modes — the resonances — and the inversion contour must decide how to pass them. That single decision carries the entire boundary-condition content: route above for outgoing radiation, below for incoming, indent for the Feynman kernel. The first figure shows this geography for the Helmholtz operator, whose poles sit right on the real axis so the $i\epsilon$ nudge is unavoidable.

Figure

Inversion in the complex k-plane for the Helmholtz operator: the poles of G̃(k) = −1 divided by (k² − k₀²) sit on the real axis at ±k₀, and the retarded prescription nudges them to +k₀+iε and −k₀−iε; closing the contour upward for positive x encloses the +k₀ pole and yields the outgoing wave e^{ik₀x} scaled by 1 over 2ik₀

Change the operator and the geography changes with it, and so does the answer. The next figure is the punchline in the time domain: the retarded Green's function of a damped oscillator, by the Laplace route. Its two poles at $s=-\gamma\pm i\omega_d$ sit in the left half-plane, so closing the Bromwich contour to the left for $t>0$ (and getting nothing for $t<0$ ) makes the response strictly causal — it rings and decays, and is exactly zero before the kick.

Figure

The retarded Green's function of a damped oscillator with ω₀ = 5 rad/s and γ = 3 s⁻¹, from Laplace inversion: poles at s = −3 ± 4i give ω_d = 4 rad/s, and G(t) = θ(t) e^{−3t} sin 4t divided by 4 rings under the envelope e^{−3t}/4, peaking at G ≈ 0.100 at t = 0.232 s and vanishing for t less than 0 by causality

The core equations

A Green's function is defined by feeding the operator a unit point source. The delta is that source; $G$ is the response at $x$ when the source sits at $x'$ .

L_x\,G(x,x') = \delta(x-x') \qquad (1)

where: $L_x$ — linear differential operator acting in $x$ ; $G(x,x')$ — Green's function, the response at $x$ to a unit source at $x'$ ; $\delta$ — Dirac delta, the unit point source. Read it as: hit the operator with a spike, record what comes out. Once $G$ is known, every forcing is handled by superposition — write $f$ as a continuous sum of point sources and add the responses.

u(x)=\int G(x,x')\,f(x')\,\mathrm{d}x' \qquad (2)

where: $u$ — the solution of $L\,u=f$ ; $f$ — an arbitrary forcing term; the integral is a convolution when $G$ depends only on $x-x'$ . This is why $G$ is worth computing once: it solves the equation for all right-hand sides. Transforming (1) is where the method earns its keep. For constant-coefficient $L$ the Green's function depends only on $x-x'$ , each $\partial_x^n$ becomes $(ik)^n$ , and $\delta$ becomes $1$ , so the PDE collapses to a one-line algebra problem.

L(ik)\,\tilde G(k)=1 \;\Longrightarrow\; \tilde G(k)=\frac{1}{L(ik)} \qquad (3)

where: $\tilde G(k)$ — Fourier transform of $G(x-x')$ ; $L(ik)$ — the symbol of the operator, each $\partial_x$ replaced by $ik$ ; $k$ — the conjugate wavenumber. The operator is now a polynomial, and its inverse is just division. The Green's function is the inverse transform of that reciprocal — the step where all the analysis lives.

G(x-x')=\frac{1}{2\pi}\int_{-\infty}^{\infty}\frac{e^{ik(x-x')}}{L(ik)}\,\mathrm{d}k \qquad (4)

where: $e^{ik(x-x')}$ — plane-wave kernel; $1/(2\pi)$ — normalisation in the convention $\tilde f(k)=\int f\,e^{-ikx}\,\mathrm{d}x$ . The integrand's poles and cuts, and the contour skirting them, decide everything about $G$ . When the symbol has no real zeros the integral is unambiguous. The screened 1D operator $-\partial_x^2+\mu^2$ is the cleanest case: poles off the axis at $k=\pm i\mu$ , an exponentially decaying answer.

G(x)=\frac{1}{2\mu}\,e^{-\mu|x|}\qquad\left(L=-\partial_x^2+\mu^2\right) \qquad (5)

where: $\mu$ — inverse screening length (a "mass"); $|x|$ — distance from the source. The decay length $1/\mu$ is set by how far the pole sits off the axis. When the symbol does have real zeros — the wave and Helmholtz operators — the bare integral is ill-defined and an $i\epsilon$ prescription must route the contour. The retarded (outgoing) choice pushes the poles just off the axis, giving a purely radiating kernel.

\tilde G(k)=\frac{-1}{k^2-k_0^2-i\epsilon}\;\Longrightarrow\;G_{+}(x)=\frac{e^{ik_0|x|}}{2ik_0}=-\frac{i}{2k_0}\,e^{ik_0|x|} \qquad (6)

where: $k_0=\omega/c$ — on-shell wavenumber; $\epsilon\to0^+$ — infinitesimal displacing the poles to enforce outgoing radiation; $G_+$ — retarded (Sommerfeld) kernel. The advanced $G_-$ follows from $\epsilon\to-\epsilon$ , giving incoming $e^{-ik_0|x|}$ . For initial-value problems the Laplace transform is natural, and its inversion is the Bromwich contour, closed to the left for $t>0$ .

\bar G(s)=\frac{1}{L(s)},\qquad G(t)=\frac{1}{2\pi i}\int_{\gamma-i\infty}^{\gamma+i\infty}\frac{e^{st}}{L(s)}\,\mathrm{d}s \qquad (7)

where: $s$ — Laplace variable, $\partial_t\mapsto s$ for homogeneous initial data; $\bar G(s)$ — transform of the causal Green's function; $\gamma$ — abscissa right of all poles of $1/L(s)$ . Closing left for $t>0$ builds causality in automatically, $G(t)\propto\theta(t)$ . Finally, the 3D Poisson operator gives the most famous kernel of all — the Coulomb potential — as the $\mu\to0$ limit of the screened result.

G(\mathbf r)=\frac{1}{4\pi r}\qquad\left(L=-\nabla^2,\ 3\mathrm{D}\right) \qquad (8)

where: $r=|\mathbf r-\mathbf r'|$ — distance from the source; $-\nabla^2$ — the negative Laplacian. The long-range $1/r$ tail is the fingerprint of a massless operator.

Derivation AFrom the defining equation to one over L of ik

The goal is to justify the two claims the whole method rests on: that convolving with $G$ actually solves $L\,u=f$ , and that the Fourier transform converts $L\,G=\delta$ into the algebraic relation $\tilde G=1/L(ik)$ . We prove superposition from the sifting property of the delta, then transform the defining equation term by term, using only that differentiation maps to multiplication and that a point source has a flat spectrum. Nothing assumes a specific operator, so this is the template every later example specialises.

Step 1. Apply $L$ to the candidate $u(x)=\int G(x,x')f(x')\,\mathrm{d}x'$ . Because $L_x$ differentiates in $x$ while the integral runs over $x'$ , it passes straight through: $L_x u(x)=\int \big(L_x G\big)f\,\mathrm{d}x'$ . Why: operator and integral act on independent variables, so their order can be exchanged.

Step 2. Insert $L_x G=\delta(x-x')$ and use sifting, $\int \delta(x-x')f(x')\,\mathrm{d}x'=f(x)$ , giving $L_x u=f(x)$ exactly. Why: this is what calling $G$ an "inverse" means — one integral against it undoes the operator, for every $f$ .

Step 3. Transform the defining equation. With constant coefficients $G(x,x')=G(x-x')$ , and each $\partial_x^n$ contributes $(ik)^n$ , so $\widehat{L_x G}(k)=L(ik)\,\tilde G(k)$ . Why: the transform diagonalises translation-invariant operators — in the plane-wave basis $L$ is just the number $L(ik)$ .

Step 4. Transform the right-hand side: $\widehat{\delta}(k)=1$ , the flat spectrum of a point source. Equating gives $L(ik)\tilde G(k)=1$ , so $\tilde G(k)=1/L(ik)$ ; the inverse transform (4) reconstructs $G$ . Why: the delta's flatness is what makes the algebra trivial — every mode filtered equally by $1/L(ik)$ returns $G$ mode by mode.

Go deeper · Derivation B · 1 minThe screened Green's function by contour integration

Derivation BThe screened Green's function by contour integration

The aim is to carry the abstract recipe through to a closed-form answer and see, concretely, how the location of a pole becomes the decay of a kernel. We solve $(-\partial_x^2+\mu^2)G(x)=\delta(x)$ on the whole line with $G\to0$ at infinity, form $\tilde G$ by the previous derivation's algebra, and invert by closing the contour where the plane wave decays. The residue theorem does the work, and a final check on the derivative jump confirms the delta source.

Step 1. The symbol of $-\partial_x^2$ is $k^2$ , so $(k^2+\mu^2)\tilde G=1$ and $\tilde G(k)=1/(k^2+\mu^2)$ . The inversion is $G(x)=\frac{1}{2\pi}\int_{-\infty}^\infty \frac{e^{ikx}}{k^2+\mu^2}\,\mathrm{d}k$ , poles at $k=\pm i\mu$ — both off the real axis, so no prescription is needed. Why: a positive symbol $k^2+\mu^2$ never vanishes for real $k$ — the screened operator is the unambiguous case.

Step 2. For $x>0$ the factor $e^{ikx}$ decays as $\mathrm{Im}\,k\to+\infty$ , so close in the upper half-plane, enclosing $k=i\mu$ ; for $x<0$ close below, enclosing $k=-i\mu$ . Why: by Jordan's lemma the added semicircle contributes nothing only where the exponential is damped, which the sign of $x$ dictates.

Step 3. The residue at $k=i\mu$ is $\dfrac{e^{-\mu x}}{2i\mu}$ , so $G(x)=\frac{1}{2\pi}\cdot 2\pi i\cdot\frac{e^{-\mu x}}{2i\mu}=\frac{e^{-\mu x}}{2\mu}$ for $x>0$ ; the mirror gives $G(x)=\frac{1}{2\mu}e^{-\mu|x|}$ . Why: the pole's imaginary part $\mu$ lands in the exponent as a decay rate — pole position becomes decay length $1/\mu$ .

Step 4. Check the source. With $G'(x)=-\tfrac12\,\mathrm{sgn}(x)\,e^{-\mu|x|}$ the derivative jumps by $G'(0^+)-G'(0^-)=-1$ , so $-\partial_x^2 G$ contributes $\delta(x)$ , and away from the origin $(-\partial_x^2+\mu^2)G=0$ . Why: a Green's function is smooth except at the source, where its derivative jump manufactures the delta — the signature to check on any 1D result.

Go deeper · Derivation C · 1 minThe retarded Green's function and the i-epsilon prescription

Derivation CThe retarded Green's function and the i-epsilon prescription

This derivation isolates the one new idea beyond the screened case: when the symbol has real zeros, the inverse transform does not exist until a contour is chosen, and that choice is the boundary condition. We solve $(\partial_x^2+k_0^2)G(x)=\delta(x)$ , whose symbol $k_0^2-k^2$ vanishes at the real points $k=\pm k_0$ , displace the poles by $i\epsilon$ to enforce outgoing radiation, close the contour, and read off the kernel — then note that the opposite displacement gives the advanced one, the two differing only by a free solution.

Step 1. The symbol of $\partial_x^2+k_0^2$ is $-(k^2-k_0^2)$ , so $\tilde G(k)=-1/(k^2-k_0^2)$ with poles at $k=\pm k_0$ sitting on the integration path — the integral is meaningless until the path is defined. Why: real poles signal undamped free modes (running waves); physics, not algebra, must say how energy enters or leaves.

Step 2. Impose the retarded (outgoing) condition by displacing the poles to $k=+k_0+i\epsilon$ and $k=-k_0-i\epsilon$ , i.e. $\tilde G(k)=-1/\big[(k-k_0-i\epsilon)(k+k_0+i\epsilon)\big]$ . Why: this displacement is the analytic form of the Sommerfeld radiation condition — whichever pole is enclosed gives a wave leaving the source.

Step 3. For $x>0$ close in the upper half-plane, enclosing $k=+k_0+i\epsilon$ ; the residue $-e^{ik_0x}/(2k_0)$ gives $G_+(x)=\frac{1}{2\pi}(2\pi i)\big(-e^{ik_0x}/2k_0\big)=\frac{e^{ik_0x}}{2ik_0}$ . For $x<0$ the pole at $-k_0-i\epsilon$ gives $e^{-ik_0x}$ , so both half-lines combine to $G_+(x)=\dfrac{e^{ik_0|x|}}{2ik_0}$ . Why: the $|x|$ says waves leave the source in both directions — what "outgoing" means on a line.

Step 4. Confirm by direct construction. Writing $G=A\,e^{ik_0|x|}$ and differentiating twice gives $G''+k_0^2G=2Aik_0\,\delta(x)$ , so $A=1/(2ik_0)$ — identical to the contour result, no stray sign. The advanced kernel $G_-=-e^{-ik_0|x|}/(2ik_0)$ comes from $\epsilon\to-\epsilon$ , and $G_+-G_-\propto \sin(k_0x)/k_0$ is a source-free standing wave. Why: two kernels differing by a homogeneous solution both satisfy $L\,G=\delta$ ; the contour-encoded boundary condition selects the physical one. *Note: some sources quote $-e^{ik_0|x|}/2ik_0$ ; the correct outgoing kernel is $+e^{ik_0|x|}/2ik_0$ , as the direct construction $A=1/2ik_0$ fixes unambiguously.*

Limiting cases

The character of $G$ is dictated entirely by the singularity structure of $1/L(ik)$ — where the poles sit, whether they are on the real axis, their order, and whether cuts appear:

Case	Result	Meaning
Simple poles off the real axis	$\tilde G(k)=\dfrac{1}{k^2+\mu^2}$	Poles at $k=\pm i\mu$ give an exponentially decaying $G(x)=\tfrac{1}{2\mu}e^{-\mu\|x\|}$ (screened / Yukawa type); the integral converges absolutely and needs no prescription.
Poles on the real axis	$\tilde G(k)=\dfrac{1}{k^2-k_0^2}$	Real poles at $k=\pm k_0$ make the inversion ambiguous; the $i\epsilon$ prescription $k_0\to k_0+i\epsilon$ selects outgoing (retarded) versus incoming (advanced) Green's functions.
Massless / gapless operator	$\tilde G(\mathbf k)=\dfrac{1}{k^2}$	A pole at $k=0$ (no mass gap) gives a long-range kernel: the Coulomb $1/4\pi r$ in 3D; a screening regulator $\mu$ tames the origin, then $\mu\to0$ .
Branch points	$\tilde G(k)=\dfrac{1}{\sqrt{k^2-k_0^2}}$	A square-root cut signals a continuum of modes (2D/3D propagation); the inverse transform yields Hankel and Bessel functions, not simple exponentials.
Higher-order poles	$\tilde G(k)=\dfrac{1}{(k^2+\mu^2)^2}$	A double pole gives a polynomial-times-exponential profile $\propto(1+\mu\|x\|)e^{-\mu\|x\|}$ — biharmonic or critically damped operators.

Worked examples

1 · Impulse response of a damped oscillator by Laplace inversion (fully worked)

A damped harmonic oscillator obeys $\ddot x+2\gamma\dot x+\omega_0^2 x=f(t)$ with $\omega_0=5\ \mathrm{s^{-1}}$ and $\gamma=3\ \mathrm{s^{-1}}$ , at rest before the forcing. Find its Green's function (impulse response), locate the peak, and evaluate it at $t=0.5\ \mathrm{s}$ .

Step 1. With homogeneous initial data the operator's Laplace transform is its symbol in $s$ : $\bar G(s)=1/(s^2+2\gamma s+\omega_0^2)$ , with poles $s=-\gamma\pm\sqrt{\gamma^2-\omega_0^2}$ . Here $\gamma^2-\omega_0^2=9-25=-16<0$ , so the motion is underdamped with $s=-3\pm 4i$ : ring frequency $\omega_d=\sqrt{\omega_0^2-\gamma^2}=\mathbf{4\ rad/s}$ , decay rate $\gamma=3\ \mathrm{s^{-1}}$ .

Step 2. Invert along the Bromwich contour: the abscissa must lie right of both poles (real part $-3$ ); closing left for $t>0$ collects both residues, closing right for $t<0$ encloses nothing. The two-pole residue sum gives the causal result $G(t)=\theta(t)\,\dfrac{e^{-\gamma t}\sin(\omega_d t)}{\omega_d}=\theta(t)\,\dfrac{e^{-3t}\sin 4t}{4}$ .

Step 3. Locate the first peak: $G'(t)=0$ when $\tan(\omega_d t)=\omega_d/\gamma=4/3$ , so $\omega_d t=\arctan(4/3)=0.927$ and $t=0.927/4=\mathbf{0.232\ s}$ . Evaluate there: $e^{-3(0.232)}=e^{-0.696}=0.499$ and $\sin(0.927)=0.8$ , so $G_{\max}=0.499\times0.8/4=\mathbf{0.0998\approx 0.100}$ .

Step 4. Evaluate at $t=0.5\ \mathrm{s}$ : $e^{-1.5}=0.2231$ , $\sin(2.0)=0.9093$ , so $G(0.5)=0.2231\times0.9093/4=\mathbf{0.0507}$ — past the peak, ringing down under the envelope $e^{-3t}/4$ and first crossing zero at $t=\pi/4=0.785\ \mathrm{s}$ , exactly the causal ringdown in the second figure. Poles $s=-3\pm4i$ ; peak $G\approx0.100$ at $t=0.232\ \mathrm{s}$ ; $G(0.5)=0.0507$ .

2 · The heat kernel and how far heat spreads

A unit of heat is deposited at the origin of a long rod at $t=0$ . The temperature obeys $(\partial_t-D\partial_x^2)G=\delta(x)\delta(t)$ with diffusivity $D=1.0\times10^{-4}\ \mathrm{m^2/s}$ . Find the kernel, and its width and peak height after $t=100\ \mathrm{s}$ .

Step 1. Fourier-transform in space ( $\partial_x^2\mapsto -k^2$ ) and Laplace-transform in time ( $\partial_t\mapsto s$ ), so the symbol is $s+Dk^2$ and $\tilde{\bar G}(k,s)=1/(s+Dk^2)$ — a single simple pole in $s$ at $s=-Dk^2$ , one decaying mode per wavenumber.

Step 2. Invert in $s$ : the residue at $s=-Dk^2$ gives $\tilde G(k,t)=\theta(t)\,e^{-Dk^2t}$ , a Gaussian in $k$ . Inverting the Fourier transform of a Gaussian (completing the square) gives the fundamental solution $G(x,t)=\theta(t)\,\dfrac{1}{\sqrt{4\pi D t}}\,e^{-x^2/4Dt}$ — a spreading bell curve of unit area.

Step 3. Read off the width: comparing to $e^{-x^2/2\sigma^2}$ gives $\sigma^2=2Dt$ , so $\sigma=\sqrt{2Dt}=\sqrt{2(10^{-4})(100)}=\mathbf{0.141\ m}$ — heat spreads about $14\ \mathrm{cm}$ in $100\ \mathrm{s}$ . The width grows as $\sqrt{t}$ , the signature of diffusion (distance $\propto\sqrt{t}$ , not $\propto t$ ).

Step 4. Peak height at $x=0$ : $G(0,t)=(4\pi Dt)^{-1/2}=(4\pi\times10^{-4}\times100)^{-1/2}=(0.1257)^{-1/2}=\mathbf{2.82\ m^{-1}}$ . As the Gaussian widens its peak falls to conserve total heat, the $t^{-1/2}$ decline the prefactor demands. $\sigma=14.1\ \mathrm{cm}$ , peak $=2.82\ \mathrm{m^{-1}}$ at $t=100\ \mathrm{s}$ ; kernel $(4\pi Dt)^{-1/2}e^{-x^2/4Dt}$ .

Physics of the portrait

The Green's function is named for a man who had barely a year of formal schooling when he wrote it down — a Nottingham miller who taught himself the newest continental mathematics from library books and, in a single self-published essay, invented both the function and the theorem that carry his name.

Origins · 1 minGeorge Green and how this was discovered

George Green

1793–1841 · self-taught Nottingham miller who founded potential theory and named the potential

George Green ran his father's windmill and had almost no formal education, yet in 1828 he privately printed An Essay on the Mathematical Analysis of the Theories of Electricity and Magnetism — the work that introduced the "potential" by name, proved what we now call Green's theorem, and defined the Green's function as a region's response to a unit point source with prescribed boundary behaviour. Only 51 subscribers bought it, and it sank almost without trace. Green entered Gonville and Caius College, Cambridge, in 1833 at the age of forty, graduated in 1837, and died in 1841 with his essay still unknown. In 1845 the young William Thomson — later Lord Kelvin — stumbled on a copy and had it republished, launching the ideas into mainstream physics just as Maxwell's generation needed them. The integral-transform construction is the modern, computational face of Green's boundary-value idea: where Green built his function from a domain's geometry, the transform method builds it from the algebra of the operator's symbol, $\tilde G=1/L(ik)$ — the two meeting in the distribution theory that finally made $1/L(ik)$ rigorous.

History

1822

— Fourier's Théorie analytique de la chaleur establishes the transform that turns diffusion into algebra

1828

— Green's privately printed Essay introduces the potential and the function that bears his name

1845

— Thomson (later Lord Kelvin) rediscovers Green's Essay and has it republished

1916

— Bromwich puts the Laplace-inversion contour on rigorous footing, taming Heaviside's operational calculus

1950

— Schwartz's theory of distributions makes the object 1/L(ik) a rigorous inverse Fourier transform

Going further

Texts. Stone & Goldbart, Mathematics for Physics, ch. 6 (Green Functions) · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., chs. 10 and 20 · Hassani, Mathematical Physics, 2nd ed., chs. 21–22 (Green's functions in one and several dimensions) · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., ch. 21 · Morse & Feshbach, Methods of Theoretical Physics, ch. 7 · Byron & Fuller, Mathematics of Classical and Quantum Physics, ch. 7. Key results to keep in mind: the residue inversion of $1/(k^2+\mu^2)$ gives $e^{-\mu|x|}/2\mu$ with a unit derivative jump; the outgoing Helmholtz kernel is $G_+=e^{ik_0|x|}/2ik_0$ , with the sign fixed by the direct construction $A=1/2ik_0$ ; the integral $\int_0^\infty k\sin(kr)/(k^2+\mu^2)\,\mathrm{d}k=\tfrac{\pi}{2}e^{-\mu r}$ , from which $e^{-\mu r}/4\pi r\to 1/4\pi r$ as $\mu\to0$ ; the damped-oscillator poles are $s=-3\pm4i$ , with peak $G\approx0.100$ at $t=0.232\ \mathrm{s}$ , and the heat-kernel width is $\sigma=\sqrt{2Dt}=0.141\ \mathrm{m}$ with peak $2.82\ \mathrm{m^{-1}}$ . Convention throughout: $\tilde f(k)=\int f\,e^{-ikx}\,\mathrm{d}x$ .

Prerequisites. The Fourier Transform · The Laplace Transform · The Dirac Delta Function · The Convolution Theorem · Laplace Transform Inversion. Leads to. Solving Differential Equations by Transform Methods · Tempered Distributions and Fourier Transforms · Theory of Distributions.

Ask the tutor: Want to see how the very same $i\epsilon$ that made our Helmholtz kernel outgoing becomes the Feynman prescription $m^2\to m^2-i\epsilon$ for the QFT propagator?

What is a Green's function, really, in one sentence?

It is the inverse of a differential operator written out as a kernel. Solving $L\,u=f$ means undoing $L$ , and $G$ does the undoing: hit $L$ once with a unit point source ( $L\,G=\delta$ ), record the response, and any forcing is handled by superposition, $u=\int G\,f$ . Physically $G(x,x')$ is the field at $x$ from a unit spike at $x'$ — a point charge's potential, a hammer-blow's ringing, a match's spreading heat — and because the equation is linear, summing point-source responses reproduces the response to any source at all. Finding $G$ is a one-time investment that then solves the equation for every right-hand side.

Why does a transform turn a differential equation into simple algebra?

Because a constant-coefficient operator is diagonal in the plane-wave basis. Each $\partial_x$ acting on $e^{ikx}$ just multiplies it by $ik$ , so $L$ becomes multiplication by the number $L(ik)$ — its symbol — and inverting it is dividing by a number, which is why $\tilde G=1/L(ik)$ . The Fourier transform is precisely the change of basis into plane waves; the inverse transform changes back. The delta source becomes $1$ because it is flat in frequency, so the algebra reduces to weighting each mode by $1/L(ik)$ . This works only for constant coefficients; a spatially varying operator is not diagonal in plane waves and the trick fails.

Retarded, advanced, Feynman — what actually distinguishes them?

Only the contour. All three are inverse transforms of the same $1/L(ik)$ , differing solely in how the path is routed around the poles — and hence differing by homogeneous (source-free) solutions. The retarded kernel routes so the response follows the source (cause before effect, outgoing radiation), and is the physical one for a system driven from rest; the advanced one runs the other way (incoming radiation), a formal device; the Feynman propagator of QFT indents so positive-frequency parts propagate forward and negative-frequency parts backward, giving time-ordered correlators. The lesson: $L\,G=\delta$ has no unique solution when the operator has free modes, and the boundary or causality condition — carried entirely by the $i\epsilon$ — is what picks one.

Why does the 3D Coulomb problem need a screening regulator but the 1D one does not?

Because $-\nabla^2$ has a zero mode. In 3D the symbol is $k^2$ , which vanishes at $k=0$ , so $1/k^2$ is singular at the origin of $k$ -space — the operator is massless, with no gap. A screening term $\mu^2$ shifts the pole to $k=\pm i\mu$ , makes the integral convergent, and yields the Yukawa kernel $e^{-\mu r}/4\pi r$ ; then $\mu\to0$ recovers the long-range Coulomb $1/4\pi r$ . The 1D screened operator $-\partial_x^2+\mu^2$ already carries a mass, so its symbol $k^2+\mu^2$ never vanishes and needs no regulator. The moral: a gapless operator gives a long-range kernel and often needs a regulator; a massive one gives an exponentially localised kernel outright.

Fourier or Laplace — how do I choose?

Match the transform to the geometry. Fourier suits the whole line or all of space with translation invariance and decay at infinity — it treats $\pm\infty$ symmetrically, right for scattering and equilibrium fields. Laplace suits initial-value problems on $t\ge0$ : it absorbs initial data through the boundary terms $s^n\bar G-(\text{initial terms})$ , and its Bromwich inversion, closed to the left, builds causality in automatically so $G(t)\propto\theta(t)$ . A space-and-time problem like diffusion uses both at once — Fourier in space, Laplace in time — as in the heat-kernel example. In short: two-sided steady problem, Fourier; "everything starts at $t=0$ ", Laplace.

When can I not use the integral-transform method at all?

When the operator is not diagonalised by the transform. The whole method rests on constant coefficients (or, more generally, on some transform that turns $L$ into multiplication). A spatially varying operator such as $-\partial_x^2+V(x)$ with non-constant $V$ is not diagonal in plane waves, so $\tilde G\ne 1/L(ik)$ and the algebra step collapses. Then you fall back on eigenfunction expansions $G=\sum_n \psi_n(x)\psi_n^*(x')/\lambda_n$ over the operator's own modes, Sturm–Liouville patching with a Wronskian jump in 1D, or numerical inversion of a discretised operator. Bounded domains with walls also lie outside the plain Fourier method. The transform method owns the translation-invariant, whole-space, constant-coefficient case — which is exactly where it is unbeatable, and where it also equals the free resolvent $\langle x|(H-z)^{-1}|x'\rangle$ , whose poles are the spectrum.