What this rung has been building toward
You have now assembled the whole toolkit of this rung. You met the idea of a response to a point source; you wrote down the fundamental solution — the response to a Dirac delta — for the Laplace, heat, and wave operators; you upgraded that to the Green's function of a boundary-value problem by adding a corrector that kills the boundary data; and you learned the method of images to write those correctors explicitly on half-spaces and balls. Each piece answered the same question for a single source. This final guide cashes all of it in: if you know the response to one point source, you know the response to every distribution of sources at once.
The mechanism that does the cashing-in has two faces, one for space and one for time, and they are secretly the same idea. In space, Green's identities turn the fundamental solution into a representation formula that prints the solution value at any interior point as a single integral over the data. In time, Duhamel's principle turns a source-free evolution solver into a solver for any forcing, by integrating the responses to a stream of instantaneous pulses. Both are the superposition principle used to its limit: chop the data into points, respond to each, add up. That is the single thread running through the whole rung.
Green's identities: integration by parts grown up
The engine behind the spatial representation formula is a pair of bookkeeping identities you can derive in one line from the divergence theorem. They are simply multivariable integration by parts, and they let you move a Laplacian off one function and onto another at the cost of a boundary term. Green's first identity moves one derivative; Green's second identity, the symmetric one, moves two and is the one that does the real work here.
Green's 2nd identity (over a region D with boundary S):
integral over D of ( u (Laplacian v) - v (Laplacian u) ) dV
= integral over S of ( u v_n - v u_n ) dS
where _n is the outward normal derivative on the boundary S.
Notice it is SYMMETRIC in u and v: swapping them flips a sign.
That built-in symmetry is the seed of Green's-function reciprocity.Here is the trick that turns this into a formula for u itself. Choose the second function v to be the fundamental solution centred at a fixed interior point — the very response-to-a-delta object you built earlier. Then Laplacian v is a Dirac delta sitting at that point, so the term v(Laplacian u) integrates against the equation Laplacian u = source, while the term u(Laplacian v) integrates against the delta and simply picks out the value of u at the centre. The whole integral collapses into a statement about one number: the value of u at the chosen point. That collapse — a delta reaching in and sampling u — is the entire magic.
The representation formula: solution = data, integrated against a kernel
Carry that collapse out and you get the representation formula: the value of u at an interior point equals a volume integral of the source against the fundamental solution, plus a boundary integral of the data against that same kernel and its normal derivative. Read it slowly, because it is the punchline of everything: the unknown solution at a point is written entirely in terms of known things — the source inside and the data on the boundary — each weighted by how strongly a unit source there influences this point. The kernel is the influence; the formula is just adding up influences.
The raw fundamental solution still leaves an awkward boundary term involving the normal derivative of u, which you usually do not know. This is exactly the gap the Green's function was built to close. Replace the bare fundamental solution by the full Green's function — fundamental solution plus the boundary corrector that makes it vanish on the boundary — and the awkward term switches off. For a Dirichlet problem you are left with a clean formula using only the prescribed boundary values and the normal derivative of the Green's function. On a ball that normal derivative is computable in closed form, and the result is precisely the Poisson integral formula you may have met as a standalone miracle — now revealed as a special case of one general machine.
Symmetry, reciprocity, and the meaning of the kernel
Look back at Green's second identity and notice it was symmetric in u and v. Feed it two Green's functions centred at two different points and that symmetry hands you a beautiful fact for free: the Green's function is symmetric in its two arguments — G(x, y) = G(y, x). Physically this is reciprocity: the influence a unit source at y exerts at x equals the influence a unit source at x exerts at y. Wiggle a charge here and measure the potential there, then swap the roles, and you read the same number. This is not a coincidence of the formula; it is the symmetry of the Laplacian (its self-adjointness) made visible.
Reciprocity is more than elegant — it is a load-bearing sanity check and a computational shortcut. It tells you the kernel really is the inverse of the differential operator: applying the operator and then integrating against the Green's function returns the identity, exactly as multiplying by a matrix and then its inverse does. It also halves your work, since you can compute the response at many points from a single well-chosen source rather than the other way round. And it is the discrete echo of a fact you already trust from finite-dimensional linear algebra, where a symmetric matrix has a symmetric inverse.
Duhamel's principle: the same idea, but in time
Everything so far solved a steady problem — Laplace and Poisson, where there is no time. Evolution equations like the heat equation and the wave equation add time, and with it a new kind of source: a forcing term that keeps injecting at every instant, not just an initial condition. Duhamel's principle is the representation formula's twin for this setting, and it rests on the identical move. Slice the running source into a stream of brief, instantaneous pulses. The pulse fired at time s acts like a fresh initial condition switched on at that moment, and from then on it simply evolves under the source-free equation for the remaining time t - s — using the very solver you already trust, the heat kernel or d'Alembert's formula.
- Freeze the forcing at one instant s. Read off the amount it injects there, f(x, s) ds, and reinterpret that slice as a brand-new initial condition switched on at time s.
- Evolve that frozen pulse forward under the source-free equation for the leftover time t - s, using exactly the homogeneous solver you have already mastered for this operator.
- Superpose: integrate these independently-evolving pulses over every firing time s from 0 to t. Writing S(tau) for the source-free evolution over a time tau, the forced part is u(x, t) = integral from 0 to t of S(t - s) f(., s) ds.
- If the initial data is nonzero too, add on its own separately-evolved response by superposition — forcing handled by Duhamel, the start handled by ordinary evolution, and the two simply sum.
Written with the kernel, the heat response to a source is u(x, t) = integral from 0 to t of [ integral of G(x - y, t - s) f(y, s) dy ] ds. This is the exact PDE analogue of variation of parameters from ODEs, and the inner object G(x - y, t - s) is the causal Green's function in time — sometimes the Duhamel formula is taken as its definition. The word causal carries real content: the kernel lives only where t - s > 0, encoding that the temperature now can depend only on heat added in the past, never the future. Duhamel's principle is, quite literally, Green's-function thinking with time as one of the variables and causality built into where the kernel is allowed to be nonzero.
Honest limits, and the one idea under all of it
Both representation and Duhamel are linearity statements, head to toe — they work only because the responses to separate point sources or separate pulses superpose. The moment the equation goes nonlinear, the clean sum breaks: for a source f that depends on u itself, the Duhamel formula stops being a final answer and becomes an integral equation you must iterate toward a fixed point, and a true representation formula for the full nonlinear solution generally does not exist. This is the same wall that lets smooth solutions of nonlinear conservation laws form shocks in finite time, after which a weak solution is not even unique without an extra entropy condition. Superposition is a privilege of linear problems; spend it knowingly.
Causality and well-posedness also discipline what you may write. The causal Green's function points one way in time precisely because the forward heat and wave problems are well-posed while their backward versions are not — running diffusion backward is ill-posed, and the kernel's t - s > 0 support is that fact made into geometry. And the very shape of the time-kernel changes the physics: for the wave equation in odd space dimensions the response to a pulse is a sharp shell that passes and is gone, which is Huygens' principle — a clean, echo-free signal — whereas in even dimensions the response leaves a lingering tail. The representation formula does not just record the answer; its kernel's support is the propagation law.