Green's Function Construction by Matching

Every linear boundary-value problem $Ly=f$ asks the same question in a different disguise: given a distributed source $f$ , what is the response $y$ that also satisfies the boundary conditions? The Green's function is the answer packaged once and for all. It is the response to the sharpest possible source — a unit spike at a single point $\xi$ — and because the operator is linear, the response to any $f$ is just the superposition $y(x)=\int G(x,\xi)f(\xi)\,\mathrm{d}\xi$ . Build $G$ once and you have inverted the operator for every right-hand side at the same time. The matching construction is the most direct way to build it for a regular second-order operator on a finite interval. Its engine is a single observation: away from the source the spike is absent, so $G$ obeys the homogeneous equation $LG=0$ and must be a combination of the two independent homogeneous solutions. The boundary conditions decide which combination on each side — a solution $y_1$ obeying the left condition governs $x<\xi$ , a solution $y_2$ obeying the right governs $x>\xi$ . All that remains is to glue the pieces at the source, and the delta dictates how: $G$ stays continuous, but its slope kinks by a precisely calibrated amount. Those two gluing conditions form a $2\times2$ linear system whose determinant is the Wronskian $W$ . When $W\neq0$ the system inverts to the symmetric kernel $G=y_1(x_<)y_2(x_>)/[p\,W]$ ; when $W=0$ a homogeneous solution has slipped through both boundary conditions, the ordinary Green's function does not exist, and the Fredholm alternative takes over. Knowing when the Wronskian vanishes is as much the content of the method as the formula.

Key takeaways

The Green's function solves $LG=\delta(x-\xi)$ with the same homogeneous boundary conditions as the target problem, so a single integral $y(x)=\int G(x,\xi)f(\xi)\,\mathrm{d}\xi$ inverts $Ly=f$ for every source at once.
Away from the source $G$ is a homogeneous solution; the matching method builds it piecewise from $y_1$ (obeying the left boundary condition) for $x<\xi$ and $y_2$ (obeying the right) for $x>\xi$ .
Two conditions pin the pieces together: $G$ is continuous at $x=\xi$ , and its derivative jumps by $1/p(\xi)$ , where $p$ is the leading coefficient in self-adjoint form.
Solving the resulting $2\times2$ system gives the symmetric closed form $G(x,\xi)=y_1(x_<)\,y_2(x_>)/[p(\xi)W(\xi)]$ , with $W$ the Wronskian and $x_<,x_>$ the lesser and greater of $x,\xi$ .
The construction exists only when $W\neq0$ — no nontrivial homogeneous solution may satisfy both boundary conditions — otherwise the Fredholm alternative and a modified Green's function are required.
Symmetry $G(x,\xi)=G(\xi,x)$ is reciprocity: it follows from the self-adjointness of the operator with these boundary conditions.
Key idea: away from the source the Green's function is an ordinary homogeneous solution; the entire delta lives in a single unit-flux kink of the derivative, $\Delta G'=1/p(\xi)$ , and matching that kink against continuity fixes $G$ completely.

The core insight

Away from the source the Green's function is a homogeneous solution obeying the boundary conditions; the delta lives entirely in a unit-flux jump of the first derivative, $\Delta G'=1/p(\xi)$ . Match the boundary-respecting solutions $y_1,y_2$ across the source with continuity plus that jump and $G$ is fixed uniquely — whenever the Wronskian does not vanish.

Where this lives today

The matching construction is the working core of one-dimensional Green's function calculations wherever a problem reduces to an ordinary differential operator on an interval. In computational electromagnetics and photonics, the transfer-matrix and scattering-matrix methods that design multilayer optical coatings, fibre Bragg gratings, and semiconductor laser mirrors are the matching construction in matrix notation: $y_1,y_2$ become forward- and backward-decaying modes and the Wronskian denominator becomes the scattering denominator whose zeros mark the guided modes. In quantum scattering the same $y_1(x_<)y_2(x_>)/(pW)$ form with energy-dependent solutions is the one-dimensional resolvent, and its Wronskian is the Jost function whose complex zeros locate bound states and resonances — the backbone of the S-matrix. Structural engineers computing beam deflections, geophysicists modelling layered media, and numerical analysts building boundary-integral preconditioners all lean on the same kernel. The failure mode is equally alive: whenever a resonant cavity or a degenerate quantum level pushes the Wronskian toward zero, the Fredholm alternative and the modified Green's function keep the calculation finite — which is why the solvability condition is checked as routinely as the formula.

The physical picture

Picture a taut elastic string pinned at both ends, and press on it at one interior point with a unit downward force. Everywhere except under your finger the string is unloaded, so it hangs straight — the homogeneous solution of $-Ty''=0$ . Under your finger the string develops a sharp corner: the slopes on the two sides differ by exactly what the tension must supply to balance the unit force. That corner is the delta function. The shape the string takes is the Green's function, built from two straight pieces, one anchored at each support, joined at the load with continuous height but kinked slope.

Figure

The point-source response of a stretched string: a unit load at ξ = 0.6 on the interval 0 to 1 produces the tent-shaped Green's function that rises as (1 − ξ)x on the left and falls as ξ(1 − x) on the right, continuous with peak value ξ(1 − ξ) = 0.24 and a slope that kinks by exactly −1 at the load — the signature of the unit delta source

The corner is not a discontinuity in $G$ — it is a discontinuity in $G'$ , and that distinction is the whole regularity story. Integrate $(pG')'=\delta$ once and you get a finite step in $pG'$ ; integrate again and the step smooths into a continuous $G$ with a kink. So $G$ is continuous, $G'$ jumps, and $G''$ carries the spike — the ladder of regularity a point source produces. The second figure separates the three levels so the claim "the delta lives entirely in the jump" becomes visible.

Figure

Regularity ladder of a point source: the response G is continuous with a corner at ξ, its derivative G′ steps down by 1/p = 1 across the source, and applying the operator −G″ returns a pure delta spike of unit area — integrate the spike once for the step, twice for the continuous kink

Because the operator is self-adjoint, one last symmetry falls out for free. The response measured at $x$ from a source at $\xi$ equals the response at $\xi$ from a source at $x$ — press here and read there, or the reverse, and the string sags the same amount. This reciprocity $G(x,\xi)=G(\xi,x)$ is written on the face of the closed form: $y_1(x_<)\,y_2(x_>)$ is symmetric under exchanging the arguments.

The core equations

The whole construction lives inside a self-adjoint (Sturm–Liouville) operator. Writing $L$ in this form is what makes the leading coefficient $p$ appear cleanly in the jump condition, and it is what guarantees the symmetry of $G$ . Any regular second-order operator can be brought to this shape by multiplying through by an integrating factor.

L y \equiv \frac{\mathrm{d}}{\mathrm{d}x}\!\left(p(x)\frac{\mathrm{d}y}{\mathrm{d}x}\right)+q(x)\,y \qquad (1)

where: $L$ — linear second-order differential operator; $p(x)$ — leading coefficient, $p\neq0$ and of one sign on the interval; $q(x)$ — zeroth-order coefficient; $y$ — the function acted upon. Reading it: the operator is a total derivative of $pG'$ plus a smooth multiplier $q$ , and that first term is exactly what an integration across the source will grab. The Green's function is defined as the response to a unit point source, carrying the same homogeneous boundary conditions the physical problem imposes. Those boundary conditions are not decoration — they are half the definition, and changing them changes $G$ .

L_x\,G(x,\xi)=\delta(x-\xi),\qquad a\le x,\xi\le b \qquad (2)

where: $G(x,\xi)$ — Green's function, the response at field point $x$ to a unit source at $\xi$ ; $\delta(x-\xi)$ — Dirac delta point source; $[a,b]$ — the interval on which homogeneous boundary conditions are imposed on $G$ in its first argument. Reading it: $L$ acts on the $x$ dependence, treating $\xi$ as a fixed parameter that merely labels where the spike sits. Away from $x=\xi$ the right side vanishes, so $G$ solves the homogeneous equation and is a linear combination of the two independent homogeneous solutions. The boundary conditions select one solution on each side, leaving only an overall scale to fix.

G(x,\xi)=\begin{cases}A(\xi)\,y_1(x),&a\le x<\xi\\[4pt]B(\xi)\,y_2(x),&\xi<x\le b\end{cases} \qquad (3)

where: $y_1$ — homogeneous solution satisfying the left boundary condition at $a$ ; $y_2$ — homogeneous solution satisfying the right boundary condition at $b$ ; $A,B$ — scale constants depending on $\xi$ , fixed by matching. Reading it: two unknown numbers $A,B$ remain, so we will need exactly two matching conditions. The first matching condition is continuity: $G$ must not jump, or its second derivative would contain a derivative of a delta, a source sharper than the one we posed.

\lim_{x\to\xi^+}G-\lim_{x\to\xi^-}G=0 \qquad (4)

where: $x\to\xi^\pm$ — the one-sided limits approaching the source from above and below. Reading it: the two pieces must meet at the same height at $x=\xi$ . The second condition is the jump in the derivative, obtained by integrating the defining equation across an infinitesimal window straddling the source. This is the single place the delta leaves its mark.

\left.\frac{\partial G}{\partial x}\right|_{\xi^+}-\left.\frac{\partial G}{\partial x}\right|_{\xi^-}=\frac{1}{p(\xi)} \qquad (5)

where: $\partial G/\partial x$ — derivative with respect to the field point $x$ ; $p(\xi)$ — leading coefficient at the source. Reading it: the slope kink is not $1$ but $1/p(\xi)$ — forgetting $p$ is the classic error, and Derivation 1 shows exactly where it comes from. Solving the two matching conditions collapses everything to a single symmetric formula. This is the payoff of the method — one line that inverts the operator.

G(x,\xi)=\frac{y_1(x_<)\,y_2(x_>)}{p(\xi)\,W(\xi)},\qquad W=y_1y_2'-y_1'y_2 \qquad (6)

where: $x_<=\min(x,\xi)$ ; $x_>=\max(x,\xi)$ ; $W(\xi)$ — Wronskian of $y_1,y_2$ at $\xi$ . Reading it: the boundary-respecting solution always takes the smaller argument, so $y_1$ governs whichever of $x,\xi$ is nearer the left end. The denominator $p\,W$ is a single constant by Abel's identity. Once $G$ is known the target problem is solved by one integral — the reason the whole exercise is worth doing.

y(x)=\int_a^b G(x,\xi)\,f(\xi)\,\mathrm{d}\xi \qquad (7)

where: $f(\xi)$ — the inhomogeneous source in $Ly=f$ ; $y(x)$ — the solution, automatically satisfying the homogeneous boundary conditions. Reading it: acting with $L_x$ under the integral turns $G$ back into $\delta$ , which sifts out $f(x)$ , so $Ly=f$ holds identically. Finally the reciprocity relation, a direct fingerprint of self-adjointness and visible in the symmetric numerator of (6).

G(x,\xi)=G(\xi,x) \qquad (8)

where: $G(\xi,x)$ — the Green's function with source and field points exchanged. Reading it: source and receiver are interchangeable, the mathematical content of physical reciprocity.

Derivation AThe unit jump from integrating across the source

The goal here is to derive the one condition that encodes the delta function — the jump in the derivative of $G$ — starting only from the defining equation $L_xG=\delta(x-\xi)$ in self-adjoint form, and to see precisely why the leading coefficient $p(\xi)$ appears in the jump rather than a bare $1$ . The strategy is to integrate over a vanishingly thin window centred on the source, track which term survives, and read off the boundary term the fundamental theorem of calculus leaves behind.

Step 1. Integrate both sides of $(pG')'+qG=\delta(x-\xi)$ over the window $[\xi-\epsilon,\xi+\epsilon]$ . The right side collects the entire unit weight of the delta, giving $\int_{\xi-\epsilon}^{\xi+\epsilon}\delta(x-\xi)\,\mathrm{d}x=1$ for any $\epsilon>0$ . Why: the delta is normalized to unit area, so no matter how narrow the window, integrating it returns exactly one.

Step 2. On the left, the derivative term integrates exactly by the fundamental theorem: $\int_{\xi-\epsilon}^{\xi+\epsilon}(pG')'\,\mathrm{d}x=\big[p(x)G'(x,\xi)\big]_{\xi-\epsilon}^{\xi+\epsilon}$ . The remaining piece $\int_{\xi-\epsilon}^{\xi+\epsilon}qG\,\mathrm{d}x$ shrinks to zero as $\epsilon\to0$ because $q$ and $G$ are bounded and the interval length vanishes. Why: only the total-derivative term can survive an infinitesimal window — a bounded integrand over a zero-width interval contributes nothing.

Step 3. Take $\epsilon\to0$ . Since $p$ is continuous it approaches the common value $p(\xi)$ from both sides, so the boundary term factorizes to $p(\xi)\big[G'(\xi^+)-G'(\xi^-)\big]=1$ , i.e. $\Delta G'=1/p(\xi)$ . Why: this is the calibrated kink — the derivative must jump by exactly $1/p(\xi)$ , and if you had left $L$ in the form $a_2G''+\ldots$ the jump would be $1/a_2(\xi)$ , never a bare $1$ unless $a_2\equiv1$ .

Go deeper · Derivation B · 1 minMatching to the symmetric closed form

Derivation BMatching to the symmetric closed form

The goal is to turn the two matching conditions — continuity and the unit jump — into a determinate pair of equations for the scale constants $A(\xi),B(\xi)$ , solve that $2\times2$ system explicitly, and show it collapses to the compact symmetric kernel $G=y_1(x_<)y_2(x_>)/[p\,W]$ . Along the way the Wronskian appears as the system determinant, which is what makes its non-vanishing the exact condition for success. The strategy is Cramer's rule followed by a reassembly that exposes the $x\leftrightarrow\xi$ symmetry.

Step 1. Impose continuity (4) on the ansatz (3): the two pieces must meet at the source, so $A(\xi)\,y_1(\xi)=B(\xi)\,y_2(\xi)$ . Why: without this the second derivative would contain $\delta'$ , a source one order sharper than the one posed, so continuity is forced by the physics of the problem.

Step 2. Impose the jump (5). The upper piece contributes $B\,y_2'(\xi)$ and the lower $A\,y_1'(\xi)$ to $G'$ at the source, so $B(\xi)\,y_2'(\xi)-A(\xi)\,y_1'(\xi)=1/p(\xi)$ . Why: the derivative must gain exactly $1/p(\xi)$ crossing the source from left to right, the flux condition derived above.

Step 3. Write the two conditions as a linear system for $(A,B)$ and read its determinant: it is $y_1(\xi)y_2'(\xi)-y_1'(\xi)y_2(\xi)=W(\xi)$ , the Wronskian of the two homogeneous solutions. Cramer's rule then gives $A(\xi)=y_2(\xi)/[p(\xi)W(\xi)]$ and $B(\xi)=y_1(\xi)/[p(\xi)W(\xi)]$ . Why: the system inverts only when $W\neq0$ — a vanishing Wronskian means $y_1$ and $y_2$ are proportional and the two matching conditions become inconsistent, which is the failure mode analysed next.

Step 4. Substitute back: for $x<\xi$ , $G=A\,y_1(x)=y_1(x)y_2(\xi)/[pW]$ ; for $x>\xi$ , $G=B\,y_2(x)=y_1(\xi)y_2(x)/[pW]$ . Both branches are the single expression $y_1(x_<)y_2(x_>)/[p(\xi)W(\xi)]$ , since the boundary-respecting $y_1$ always carries the smaller argument. Why: writing it with $x_<,x_>$ makes the reciprocity $G(x,\xi)=G(\xi,x)$ manifest — exchanging $x$ and $\xi$ leaves $\min$ and $\max$ untouched.

Step 5. Confirm consistency with Abel's identity: for a self-adjoint operator $p\,W=p(y_1y_2'-y_1'y_2)$ is independent of $x$ , so the denominator is one constant, evaluated once. Why: compute $pW$ once and forget its $\xi$ dependence entirely.

Go deeper · Derivation C · 1 minWhen the construction fails and what replaces it

Derivation CWhen the construction fails and what replaces it

The goal is to characterize exactly when the matching method breaks — the case $W(\xi)=0$ — connect that breakdown to a zero mode of the operator, and state the replacement: the Fredholm solvability condition plus the modified Green's function. This matters because a vanishing Wronskian is not a computational accident but a statement that the operator has a zero eigenvalue compatible with the boundary conditions, exactly where the naive inverse cannot exist. The strategy is to trace what a zero mode does to the two boundary-respecting solutions.

Step 1. Suppose a nontrivial $\phi$ satisfies $L\phi=0$ and both boundary conditions. Then $\phi$ qualifies simultaneously as $y_1$ (it obeys the left condition) and as $y_2$ (it obeys the right), forcing $y_1\propto y_2$ and hence $W\equiv0$ by Abel. Why: the whole method rested on $y_1,y_2$ being independent; a zero mode collapses them onto the same line, and the $2\times2$ system loses rank.

Step 2. With $W=0$ the closed form (6) has $G\sim1/(pW)\to\infty$ : no ordinary Green's function exists. Why: this mirrors the spectral picture $G=\sum_n\phi_n(x)\phi_n(\xi)/(\lambda-\lambda_n)$ , which diverges precisely when the operator's parameter $\lambda$ lands on an eigenvalue $\lambda_n$ — here $\lambda=0$ is such an eigenvalue.

Step 3. The problem $Ly=f$ is then solvable only if $f$ is orthogonal to the zero mode, $\int_a^b\phi(\xi)f(\xi)\,\mathrm{d}\xi=0$ — the Fredholm alternative. Why: projecting $Ly=f$ onto $\phi$ and integrating by parts twice sends $L$ onto $\phi$ , giving zero on the left, so the right side must vanish too or there is no solution at all.

Step 4. When the condition holds, one inverts $L$ on the complement of the zero mode using the modified Green's function $L_xG_M=\delta(x-\xi)-\phi(x)\phi(\xi)/\langle\phi,\phi\rangle$ . Why: subtracting the zero-mode projection removes the resonant component the operator cannot invert, and $G_M$ is then determined up to arbitrary additive multiples of $\phi$ , which drop out against an orthogonal source.

Limiting cases

The matching formula either simplifies or breaks in four informative regimes; each is a check on whether you have understood what $p$ , $W$ , and the boundary conditions are doing.

Case	Result	Meaning
Wronskian vanishes	$W(\xi)=0$	A nontrivial homogeneous solution satisfies both boundary conditions; $\lambda=0$ is an eigenvalue. The ordinary $G$ does not exist — use the Fredholm alternative and a modified Green's function.
Constant-coefficient string	$p\equiv1,\ q\equiv0$	For $-y''=f$ on $[0,1]$ with $y(0)=y(1)=0$ the kernel is the tent function $G=x_<(1-x_>)$ , peaked at the source.
Abel's identity	$p(\xi)W(\xi)=\text{const}$	In self-adjoint form $pW$ has no $\xi$ dependence, so the denominator is evaluated once at any convenient point.
Initial-value (causal) limit	$G=0$ for $x<\xi$	With both conditions at the left end the right solution is unconstrained; one recovers the retarded Green's function of a driven oscillator.

Common trapThe derivative always jumps by exactly 1

It does not. The jump is $\Delta G'=1/p(\xi)$ , and it equals $1$ only when the operator is already normalized so that the coefficient of $G''$ is unity. If you start from a raw operator $a_2(x)G''+a_1(x)G'+a_0(x)G=\delta$ , the jump is $1/a_2(\xi)$ — and if you first divide through by $a_2$ to normalize, you have also rescaled the delta, so the same $1/a_2(\xi)$ reappears. The safe recipe is to write $L$ in self-adjoint form $(pG')'+qG$ first, whereupon the jump reads $1/p(\xi)$ directly. A second, related trap: the jump lives in $\partial G/\partial x$ , the derivative with respect to the field point; differentiating with respect to $\xi$ picks up an extra sign.

Worked examples

1 · The tent function, from scratch (confidence builder)

Solve $-y''=\delta(x-\xi)$ on $[0,1]$ with $y(0)=y(1)=0$ by matching, take the source at $\xi=0.6$ , and check every condition numerically.

Step 1. Homogeneous solutions of $-y''=0$ are linear. The one obeying the left condition $y(0)=0$ is $y_1(x)=x$ ; the one obeying the right condition $y(1)=0$ is $y_2(x)=1-x$ . These are the two straight pieces the string will take on either side of the load.

Step 2. Put the operator in self-adjoint form. Here $L=-\mathrm{d}^2/\mathrm{d}x^2=(pG')'$ with $p=-1$ , $q=0$ . The Wronskian is $W=y_1y_2'-y_1'y_2=x(-1)-(1)(1-x)=-1$ , a constant as Abel guarantees, so the denominator is $p\,W=(-1)(-1)=\mathbf{+1}$ .

Step 3. Apply the closed form (6): $G=y_1(x_<)y_2(x_>)/(pW)=x_<(1-x_>)$ . Explicitly, $G=(1-\xi)x$ for $x\le\xi$ and $G=\xi(1-x)$ for $x\ge\xi$ . At the source $\xi=0.6$ the peak is $G(\xi,\xi)=\xi(1-\xi)=0.6\times0.4=\mathbf{0.24}$ .

Step 4. Check the jump. The left slope is $\mathrm{d}[(1-\xi)x]/\mathrm{d}x=1-\xi=+0.4$ ; the right slope is $\mathrm{d}[\xi(1-x)]/\mathrm{d}x=-\xi=-0.6$ . The jump is $(-0.6)-(+0.4)=\mathbf{-1}$ , which equals $1/p=1/(-1)=-1$ exactly. Continuity also holds: both pieces give $\xi(1-\xi)$ at $x=\xi$ .

Step 5. Check it inverts a source. For a uniform load $f\equiv1$ , $y(x)=\int_0^1 G(x,\xi)\,\mathrm{d}\xi=\tfrac12 x(1-x)$ , which indeed satisfies $-y''=1$ with $y(0)=y(1)=0$ . The kernel is $G=x_<(1-x_>)$ , peak $0.24$ at $\xi=0.6$ , slope jump exactly $-1$ .

2 · A screened string with a real number

Solve $-y''+\mu^2 y=\delta(x-\xi)$ on $[0,1]$ with $y(0)=y(1)=0$ for $\mu=1$ , source at $\xi=0.5$ , and compare the peak against the unscreened tent.

Step 1. In self-adjoint form $L=(pG')'+qG$ with $p=-1$ , $q=\mu^2$ , so the homogeneous equation $-y''+\mu^2y=0$ has hyperbolic solutions. The one vanishing at $x=0$ is $y_1=\sinh(\mu x)$ ; the one vanishing at $x=1$ is $y_2=\sinh(\mu(1-x))$ .

Step 2. Compute the Wronskian. With $y_1'=\mu\cosh(\mu x)$ and $y_2'=-\mu\cosh(\mu(1-x))$ , the addition formula collapses it to $W=-\mu\big[\sinh(\mu x)\cosh(\mu(1-x))+\cosh(\mu x)\sinh(\mu(1-x))\big]=-\mu\sinh\mu$ . Hence $p\,W=(-1)(-\mu\sinh\mu)=\mathbf{\mu\sinh\mu}$ , a constant.

Step 3. The closed form is $G(x,\xi)=\sinh(\mu x_<)\,\sinh(\mu(1-x_>))/(\mu\sinh\mu)$ . For $\mu=1$ , $\sinh 1=1.175201$ .

Step 4. Evaluate the peak at $\xi=0.5$ : $G(\tfrac12,\tfrac12)=\sinh^2(0.5)/\sinh(1)$ . With $\sinh(0.5)=0.521095$ , this is $0.271540/1.175201=\mathbf{0.2311}$ .

Step 5. Interpret. The unscreened tent peaks at $\xi(1-\xi)=0.25$ ; the screening term $\mu^2$ pulls the peak down to $0.2311$ , a $7.6\%$ reduction, because the restoring $+\mu^2 y$ term resists deflection. The jump check still gives $\Delta G'=-1=1/p$ . $G_{\max}=0.2311$ at the source, versus $0.25$ unscreened — screening softens the response while the jump stays $1/p$ .

3 · The causal limit: a driven oscillator

Build the retarded Green's function for $\ddot x+\omega^2 x=f(t)$ with both conditions imposed at the initial time.

Step 1. Here the "boundary conditions" are both at the left end ( $x=\dot x=0$ before the source), so the left-respecting solution is forced to vanish for $t<t'$ : $G=0$ there. This is the causal choice.

Step 2. For $t>t'$ the homogeneous solutions are $\cos\omega t,\sin\omega t$ ; continuity at $t=t'$ forces $G(t',t')=0$ and the unit jump ( $p=1$ ) forces $\dot G(t'^+)=1$ . The combination meeting both is $G(t,t')=\theta(t-t')\,\sin\omega(t-t')/\omega$ .

Step 3. Verify: $\dot G(t'^+)=\cos(0)=1$ matches the jump, and convolving with $f$ gives Duhamel's formula $x(t)=\int_{-\infty}^{t}\sin\omega(t-t')\,f(t')/\omega\,\mathrm{d}t'$ . The impulse response is $\theta(t-t')\sin\omega(t-t')/\omega$ , the matching method's causal branch.

Physics of the portrait

The Green's function carries the name of a self-taught English miller who invented it in near-total isolation, decades before the mathematics existed to justify it. Its modern form — piecewise homogeneous solutions matched at a source — is the distilled residue of a century of work on potential theory and boundary-value problems.

Origins · 1 minGeorge Green and how this was discovered

George Green

1793–1841 · self-taught miller who founded potential theory and the theory of the Green's function

George Green ran his father's windmill in Nottingham and had barely a year of formal schooling, yet in 1828 he privately printed An Essay on the Application of Mathematical Analysis to the Theories of Electricity and Magnetism — the work that introduced both the function now bearing his name and the identities (Green's first and second) that make its symmetry provable. His central idea is exactly the one this article develops: to solve a boundary-value problem, find the potential of a single point source consistent with the boundaries, then superpose. The Essay drew only fifty-one subscribers and sank almost without trace. Green entered Gonville and Caius College, Cambridge, as an undergraduate in 1833 at the age of forty, and died in 1841 still obscure. Four years later a young William Thomson (later Lord Kelvin) stumbled on a copy just after taking his degree, grasped its importance at once, and had it republished — putting Green's function into the hands of Maxwell, Stokes, and the whole Victorian school. The rigorous scaffolding came later: Sturm and Liouville's theory of self-adjoint eigenvalue problems supplied the guarantee of symmetry and completeness, and Fredholm's theory of integral equations settled precisely when the ordinary Green's function exists. What began as a miller's device for electrostatics is now the standard inverse of any linear differential operator.

History

1828

— Green privately prints his Essay, introducing the Green's function and Green's identities

1833

— Green enters Cambridge as an undergraduate at age forty

1837

— Sturm and Liouville establish the theory of self-adjoint second-order eigenvalue problems

1845

— William Thomson rediscovers Green's Essay and later republishes it, spreading it across Europe

1903

— Fredholm's theory of integral equations fixes when the ordinary Green's function exists

Going further

Texts. Stone & Goldbart, Mathematics for Physics, Ch. 5 (Green Functions) — the cleanest modern treatment of the matching construction and the Fredholm alternative · Arfken, Weber & Harris, Mathematical Methods for Physicists, 7th ed., Ch. 10 · Hassani, Mathematical Physics, 2nd ed., Ch. 22 (Green's functions in one dimension) · Byron & Fuller, Mathematics of Classical and Quantum Physics, Ch. 7 · Riley, Hobson & Bence, Mathematical Methods for Physics and Engineering, 3rd ed., §15.2 and Ch. 21 · Courant & Hilbert, Methods of Mathematical Physics, Vol. I, for the classical foundations · Nakahara, Geometry, Topology and Physics, for the operator-theoretic and index-theory perspective. The self-adjoint form and the jump condition follow from integrating $(pG')'+qG=\delta$ across the source. The $2\times2$ matching system has determinant $W$ ; $pW$ is constant by Abel's identity in both worked examples. Key anchor values: the tent function $x_<(1-x_>)$ has peak $0.24$ at $\xi=0.6$ and slope jump $-1$ ; the screened-string kernel $\sinh(\mu x_<)\sinh(\mu(1-x_>))/(\mu\sinh\mu)$ has peak $0.2311$ at $\xi=0.5$ , $\mu=1$ (using $\sinh 0.5=0.521095$ , $\sinh 1=1.175201$ ). Symmetry follows from Green's second identity with vanishing boundary terms.

Prerequisites. Linear Second-Order ODEs · The Wronskian and Linear Independence · Variation of Parameters · Two-Point Boundary-Value Problems · Existence and Uniqueness Theorems for ODEs. Leads to. Sturm–Liouville Theory and Eigenfunction Expansions · The Fredholm Alternative and Modified Green's Functions · Partial Differential Equation Green's Functions.

Ask the tutor: Want to see the matching construction run live on your own operator and boundary conditions — or watch the Wronskian go to zero and the Fredholm alternative take over? Ask and we will build the kernel step by step.

Why does the Green's function satisfy the homogeneous equation away from the source?

Because the delta function is exactly zero everywhere except at the single point $x=\xi$ . For any $x\neq\xi$ the defining equation $L_xG=\delta(x-\xi)$ reads $L_xG=0$ , which is the homogeneous equation. So on each side of the source $G$ must be a linear combination of the two independent homogeneous solutions — there is nowhere else for it to live. All of the source's influence is squeezed into the single point $\xi$ , where it appears not as a value of $G$ but as the calibrated kink in $G'$ . This is why the method works at all: it reduces an inhomogeneous problem to gluing together two homogeneous solutions.

How do I know which homogeneous solution to use on each side?

The boundary conditions decide. On the left interval $[a,\xi)$ the Green's function must satisfy the left boundary condition at $x=a$ , so you pick the homogeneous solution $y_1$ that already obeys it. On the right interval $(\xi,b]$ it must satisfy the right condition at $x=b$ , so you pick the solution $y_2$ that obeys that one. Each side is thereby locked to the correct boundary, and the only freedom left is an overall scale on each piece — the two constants $A$ and $B$ that the matching conditions fix. If a single solution happened to satisfy both boundary conditions, you would be in the degenerate $W=0$ case and the ordinary construction would fail.

Why must the jump be 1/p and not simply 1?

Because the operator in self-adjoint form is $(pG')'+qG$ , and integrating it across the source lands on the boundary term $[pG']$ , not $[G']$ . Setting that equal to the unit weight of the delta gives $p(\xi)\Delta G'=1$ , so $\Delta G'=1/p(\xi)$ . Only when $p\equiv1$ — a string of unit tension, an operator already normalized — does the jump reduce to $1$ . If you instead work with a raw operator whose top coefficient is $a_2(x)$ , the jump is $1/a_2(\xi)$ ; normalizing by dividing through by $a_2$ rescales the delta and reproduces the same factor. The safest habit is to convert to self-adjoint form before reading off the jump.

What is the connection to variation of parameters?

They are two views of the same kernel. Variation of parameters writes a particular solution of $Ly=f$ as $\int[y_1(\xi)y_2(x)-y_1(x)y_2(\xi)]f(\xi)/(pW)\,\mathrm{d}\xi$ , with the integration range split at $x$ so that different combinations of $y_1,y_2$ appear for $\xi<x$ and $\xi>x$ . Collecting the terms by which of $x,\xi$ is larger reproduces exactly $G=y_1(x_<)y_2(x_>)/(pW)$ . Variation of parameters builds the particular solution first and reads off $G$ ; matching builds $G$ first and integrates. Both rest on the same two homogeneous solutions and the same Wronskian denominator, which is why they agree.

What exactly goes wrong when the Wronskian vanishes?

A vanishing Wronskian means $y_1$ and $y_2$ are proportional — the same function, up to scale, satisfies both boundary conditions. That function is a zero mode: a nontrivial solution of $L\phi=0$ obeying all the boundary conditions, equivalently an eigenfunction with eigenvalue zero. The $2\times2$ matching system then has a zero determinant and no solution, and the closed form blows up as $1/W$ . Physically the operator is not invertible — it has a nontrivial null space — so $Ly=f$ has either no solution or infinitely many. The Fredholm alternative says a solution exists only when $f$ is orthogonal to the zero mode, and the modified Green's function subtracts the offending projection to invert $L$ on the rest.

Does the same matching idea work for partial differential equations?

The core idea carries over, but the geometry changes. For a PDE like Poisson's equation the source is still a delta and $G$ still solves the homogeneous equation away from it, but "away from the source" is now a region, and the jump becomes a condition on the normal derivative across a surface (a flux through a small sphere). The clean $y_1(x_<)y_2(x_>)/(pW)$ form is special to one dimension, where a point has exactly two sides and there are exactly two homogeneous solutions. In higher dimensions one expands in eigenfunctions or uses images, though separation of variables often reduces a PDE Green's function back to a one-dimensional matching problem in the radial or longitudinal coordinate.

Why is the Green's function symmetric, and is it always?

The symmetry $G(x,\xi)=G(\xi,x)$ is reciprocity, and it holds whenever the operator is self-adjoint with the given homogeneous boundary conditions. Applying Green's second identity to $G(x,\xi)$ and $G(x,\eta)$ , the boundary terms vanish because both obey the same homogeneous conditions, leaving $G(\eta,\xi)=G(\xi,\eta)$ . The closed form shows it directly: $y_1(x_<)y_2(x_>)$ is invariant under swapping $x$ and $\xi$ . It is not automatic for non-self-adjoint operators, where $G_L$ pairs instead with the adjoint's Green's function, $G_L(x,\xi)=G_{L^\dagger}(\xi,x)$ . Self-adjointness is what makes the two coincide.