A PDE as a Machine: the Operator L
In the last guide you sorted equations onto the linear / semilinear / quasilinear / fully nonlinear ladder. The reward for landing on the very first rung is this whole guide. To see why, it helps to repackage a PDE as a single machine. Write L for the bundle of derivative operations that act on the unknown — for the heat equation, L u = u_t - k u_xx; for Laplace's equation, L u = u_xx + u_yy. The equation is then just L u = f, where f is whatever sits on the right.
Calling L linear means exactly two things, the same two you know from matrices and from straight lines through the origin: L(u + v) = L u + L v, and L(c u) = c L u for any constant c. Both hold here because differentiation itself is linear — the derivative of a sum is the sum of derivatives, and constants slide out front. The principal part (the highest-order terms) and the lower-order terms all inherit this, provided no coefficient depends on u and u never appears squared, multiplied by its own derivatives, or inside a function like sin(u).
The Homogeneous Miracle: Solutions Add
Now suppose the right-hand side is zero, so the equation is homogeneous: L u = 0. If u solves it and v solves it, then L(c1 u + c2 v) = c1 L u + c2 L v = c1 · 0 + c2 · 0 = 0. The combination solves it too. That single line is the superposition principle: for a homogeneous linear PDE, any linear combination of solutions is again a solution. The solutions form a vector space, exactly like solutions of a homogeneous linear ODE — but here the space is usually infinite-dimensional.
L(c1 u1 + c2 u2 + ... + cn un) = c1 L(u1) + c2 L(u2) + ... + cn L(un) = c1*0 + c2*0 + ... + cn*0 = 0 (if every uk solves L u = 0)
Here is the payoff in miniature. The heat equation u_t = k u_xx on a rod has, for each whole number n, the simple building-block solution u_n(x, t) = e^(-k n^2 t) sin(n x). Each one decays and wiggles on its own; none of them, alone, can match a complicated starting temperature. But superposition lets you add them with any weights b_n you like, and the sum still solves the equation. Suddenly you have a vast, tunable family — and the only question left is how to choose the weights.
From Finite Sums to Fourier Series
Choosing the weights is where superposition meets data. The building blocks were engineered so that at t = 0 each one is just sin(n x); their combination at t = 0 is the sum of b_n sin(n x), which is precisely a Fourier sine series. So matching the initial condition f(x) = u(x, 0) means expanding f as a Fourier series, and the coefficient formulas hand you each b_n. Pin the data, get the weights, and superposition propagates them forward in time — that is the entire arc of separation of variables in one breath.
Notice the quiet upgrade: we went from a finite sum to an infinite one, and even to an integral when the spatial domain is the whole line and the index n becomes a continuous frequency. This continuous superposition is what the Fourier transform and the fundamental solution are really doing — adding up an uncountable continuum of building blocks, weighted by the data. The bookkeeping changes from sums to integrals, but the engine is identical: linearity lets responses pile up without interfering with one another.
Inhomogeneous Equations: Particular plus Homogeneous
What if a source is present, so the equation is inhomogeneous: L u = f with f not zero? Superposition still helps, in two complementary ways. First, the structure theorem: if u_p is any one particular solution of L u = f, then every solution has the form u = u_p + u_h, where u_h ranges over all solutions of the homogeneous equation L u_h = 0. You reach the general solution by adding the full homogeneous space — exactly the same scaffolding as for linear ODEs and linear systems.
Second, superposition lets you split a hard source into easy pieces. If f = f1 + f2 and you can solve L u1 = f1 and L u2 = f2 separately, then u1 + u2 solves L u = f. Push this to the limit — chop the source into a continuum of point pulses (delta sources), solve for each, and add up the responses — and you arrive at Duhamel's principle and the Green's-function method. Both are superposition wearing a more grown-up coat: build the answer from the responses to elementary inputs.
It is worth saying the structure aloud, because it is the skeleton of every linear solution method: the general solution of L u = f is u = u_p + u_h, one particular solution u_p (any single function that hits the source f) plus u_h ranging over the entire homogeneous solution space (everything with L u_h = 0). The particular solution carries the source; the homogeneous part carries the freedom you later spend on matching boundary and initial data.
Where the Engine Stops
Superposition is a privilege of linearity, and it is genuinely lost the moment nonlinearity enters. For Burgers' equation u_t + u u_x = 0 the term u u_x is the unknown multiplying its own derivative; if u and v each solve it, u + v generically does not, because (u+v)(u+v)_x is not u u_x + v v_x. Two valid travelling waves cannot simply be summed. This is not a technicality — it is why nonlinear waves can steepen and form shocks while linear waves merely pass through each other unchanged.
Two finer cautions even within the linear world. Superposition combines solutions of L u = 0, so it respects only homogeneous linear side conditions: you may add two solutions that both vanish on the boundary, but adding two that both equal the value 5 on the boundary gives one that equals 10 there. And the principle says nothing about which combination is the right one — that is fixed by the data and by whether the problem is well-posed for that data, a question the elliptic / parabolic / hyperbolic type will decide. Superposition supplies the raw material; the data and the type choose the answer.
Keep this engine in view as you climb. Separation of variables, Fourier series, the Fourier transform, eigenfunction expansions, Duhamel's principle, Green's functions — every one of them is the same move: break the data into elementary pieces, solve each piece, superpose. Linearity is what makes that strategy legal, and it is no accident that nearly every clean, closed-form method you will meet lives on the linear rung.