Cashing in the whole rung
This is the guide the previous four were built for. You now have a test function, a distribution as a linear rule, the Dirac delta as the cleanest such rule, the distributional derivative that differentiates anything, and the Fourier transform extended to tempered distributions. Each tool was introduced with a promise that it would pay off here. The payoff is a single sentence we could not have written honestly five guides ago: a fundamental solution of a linear operator L is a distribution E satisfying L E = delta. We spend this guide unpacking why that sentence is exactly right, and what it buys.
Take the words apart slowly. L is a constant-coefficient differential operator — the heat operator d/dt - k d^2/dx^2, the wave operator d^2/dt^2 - c^2 d^2/dx^2, or the Laplacian d^2/dx^2 + d^2/dy^2. The right-hand side delta is a point source: an idealized unit input switched on at one instant and one place. The unknown E is the fundamental solution, the system's response to that single, infinitely concentrated poke. In the old language none of this parsed — delta was not a function, so L E = delta was not an equation. The whole rung exists to make every symbol on that line mean something.
What 'L E = delta' really asks of E
Here is the move that makes the equation rigorous, and it is the same reflex from every earlier guide: never touch E directly, always test it against a probe. To say L E = delta as distributions means that for every test function phi, pairing L E with phi gives the same number as pairing delta with phi. But L is built from derivatives, and we already know how a distributional derivative works — it hands each derivative over to the smooth probe by integration by parts. So < L E, phi > is rewritten as < E, L* phi >, where L* is the formal adjoint (the same operator with the signs from integration by parts), and the equation becomes < E, L* phi > = phi(0) for every phi.
Read that condition out loud and it stops being scary. It says: whatever E is, when you push the operator's derivatives over onto a smooth probe and then average E against the result, you must always recover the probe's value at the origin. That is a clean, testable demand — no infinities, no division by zero, just a number on each side. A distributional solution is precisely an E that passes this test for every probe, and a fundamental solution is the special case where the source is delta. The previous guides did not just decorate this idea; they are the only reason the equation has a meaning to test at all.
Want (as distributions): L E = delta Test against every probe: < L E, phi > = < delta, phi > Move derivatives onto phi: < E, L* phi > = phi(0) for all phi (L* = formal adjoint: integration by parts flips the signs) So E is a fundamental solution <=> < E, L* phi > = phi(0) for every test phi
One response, every source: the convolution trick
Why pour all this care into the response to a single point source? Because of one fact that makes the effort pay enormous dividends: for a constant-coefficient L, knowing the response to one delta hands you the response to every source. The reason is that a general source f is just a continuous pile-up of point sources — f is the sum over all locations y of f(y) times a delta sitting at y. Because L is linear and does not care where you are, the response to that pile-up is the same pile-up of shifted copies of E. That superposition of shifted fundamental solutions is exactly what convolution computes.
So the recipe is breathtakingly short: to solve L u = f, set u = E * f, the convolution of the fundamental solution with the source. Apply L and the derivatives slide onto E (convolution commutes with the operator), giving L u = (L E) * f = delta * f = f — because convolving with delta gives back exactly what you started with. The delta is the identity element for convolution, the multiplicative '1' of this world, which is why a fundamental solution deserves its name: it is the building block from which every solution is assembled.
The three model operators, seen rigorously
Watch the machinery light up the equations you already know. For the heat operator, the fundamental solution is the heat kernel — that spreading Gaussian, zero before t = 0, which integrates to 1 at every later time. Read distributionally, the statement 'the heat kernel solves u_t = k u_xx with a point of heat at the origin' becomes the exact equation (d/dt - k d^2/dx^2) E = delta. Its infinitely smooth shape for any t > 0 is the rigorous face of the heat equation's instant smoothing: poke it with a delta and the response is already analytic a moment later — and, honestly, that also means heat has infinite propagation speed, since the Gaussian is nonzero everywhere immediately.
The Laplacian tells a different, sharper story. In the plane the fundamental solution of Laplacian E = delta is the logarithmic potential (1/2pi) log|x|; in three dimensions it is the Newtonian potential -1/(4pi|x|), the very 1/r law of gravity and electrostatics. So Coulomb's law is not a separate piece of physics bolted on — it is literally the fundamental solution of an elliptic operator, and convolving it with a charge density f reconstructs the whole potential via u = E * f. Notice the contrast with heat: this E is smooth everywhere except at the source point, where it blows up. That single bad point is its singular support, and away from it the response is as smooth as can be — the fingerprint of an elliptic operator.
The wave operator completes the trilogy and refuses to behave like either. Its fundamental solution does not smooth at all — in three dimensions the response to a flash at the origin lives precisely on the expanding light cone |x| = c t and is zero both inside and outside it, the rigorous statement of finite speed and of Huygens' principle (you hear a clap once, sharply, not as a lingering smear). The fundamental solution is a distribution supported on a surface, no function at all. Three operators, three utterly different E's — and each E's singularities encode the physics: heat smooths and spreads instantly, the Laplacian is singular at one point, the wave keeps its edges sharp forever.
Finding E, and the honest small print
How do you actually compute one of these E's? The clean route uses the previous guide. Take the Fourier transform of L E = delta; because the transform turns derivatives into multiplication and sends delta to the constant 1, the differential equation collapses to an algebraic one, P(k) * E-hat(k) = 1, where P is the operator's symbol. Just divide: E-hat = 1/P(k). Then invert the transform to get E back. The whole problem of solving a constant-coefficient PDE has been reduced to dividing by a polynomial and inverting a tempered Fourier transform — the cleanest possible illustration of why guide 4 was worth the effort.
- Write the equation you want E to satisfy: L E = delta, with L a constant-coefficient operator.
- Fourier-transform both sides; derivatives become multiplication by the symbol P(k), and delta becomes the constant 1, giving P(k) * E-hat(k) = 1.
- Solve the algebra: E-hat(k) = 1/P(k), choosing how to handle any zeros of P (this is where you select a specific E).
- Invert the transform to recover E, then build the solution of L u = f as the convolution u = E * f.
Now the honest limits, because this beautiful machine has a fenced-off yard. First, the existence of E in the first place is a deep theorem, not a triviality: the Malgrange-Ehrenpreis theorem guarantees that every nonzero constant-coefficient operator has a fundamental solution — a genuinely hard result, and one that quietly fails the moment coefficients vary or the operator goes nonlinear. Second, the whole convolution miracle leans on constant coefficients and linearity; with variable coefficients you climb back down to a Green's function tied to the specific geometry, and convolution with a single E no longer suffices.
Third, and most important to remember from the whole rung: this framework is fundamentally linear. The reason E * f works is superposition, and superposition is exactly what nonlinear PDEs destroy. You cannot build a fundamental solution for the equations where a smooth start can erupt into a shock in finite time; there a weak solution need not be unique, and you must reach outside distribution theory for a side condition — the entropy condition — to pick the physical answer. So treasure what you have built here, but hold it at its true scope: distributions made the linear, constant-coefficient world completely rigorous and astonishingly clean, and that is a genuine, bounded triumph — not a universal solvent.