Here is a genuinely clever trick for taming many of the classical linear PDEs. A PDE mixes several variables together — space and time, both changing at once — which is exactly what makes it harder than an ordinary differential equation. So make a bold guess: suppose the solution actually splits apart into a product of separate, single-variable functions,

one factor that only cares about x and another that only cares about t. It sounds almost too optimistic to work — most functions of two variables can't be split apart like that at all. But plug this guess into a linear PDE like the heat equation, and something remarkable happens: the single, coupled PDE falls apart into two separate, much easier ordinary differential equations, one for X alone and one for T alone. That collapse from one hard problem into two easy ones is the entire engine behind separation of variables.

The method, as a recipe, has four moves:

1. Assume a product. Look for solutions of the form u(x, t) = X(x)\,T(t) — one factor per variable.
2. Substitute and separate. Plug in and rearrange algebraically until everything depending on x sits on one side and everything depending on t sits on the other.
3. Introduce a separation constant. A function of x alone can equal a function of t alone, for every x and every t, only if both sides equal the same constant — call it -\lambda.
4. Solve the resulting ODEs and superpose the pieces to match the given data.

Watching the split happen

It's worth seeing the algebra once, in full, so the trick stops feeling like magic. Take the one-dimensional heat equation u_t = \alpha\,u_{xx} and substitute the product guess u = X(x)T(t). Because X doesn't depend on t and T doesn't depend on x, the partial derivatives simplify to ordinary derivatives of the single-variable factors:

Now divide both sides by \alpha\,X(x)\,T(t) (wherever it's nonzero) to push every t to the left and every x to the right:

Look closely at what this equation is claiming: the left-hand side is a function of t only, and the right-hand side is a function of x only — yet they're equal for every choice of x and t, independently. Wiggle x while holding t fixed and the left side can't change, so the right side can't either. The only way a purely-t expression can equal a purely-x expression for all inputs is if both are secretly the same constant. Call that constant -\lambda, and the single PDE has split cleanly into two ordinary differential equations:

The T equation is immediately solvable — it's exponential decay (or growth), T(t) = e^{-\alpha\lambda t}. All the remaining work goes into the X equation, and that's where the boundary conditions finally get to speak.

The eigenvalue problem at its heart

Suppose the rod has length L and is held at zero temperature at both ends: u(0, t) = u(L, t) = 0 for all t. Since u = X(x)T(t) and T(t) is (generically) never zero, those boundary conditions force X(0) = X(L) = 0 too. So the spatial factor must solve the boundary-value problem

This is an eigenvalue problem — the continuous cousin of eigenvectors and eigenvalues. Try \lambda \le 0 and the only solution satisfying both endpoint conditions is the useless zero function. Try \lambda > 0 instead, and the general solution X(x) = A\sin(\sqrt{\lambda}\,x) + B\cos(\sqrt{\lambda}\,x) meets X(0) = 0 only if B = 0, and then meets X(L) = 0 only if \sqrt{\lambda}\,L is a whole multiple of \pi. That single geometric requirement — fit a whole number of half sine-waves between the two fixed ends — sieves out a discrete list of allowed values:

with matching eigenfunctions X_n(x) = \sin(n\pi x / L). Every other value of \lambda gives nothing but the trivial zero solution — the boundary conditions act as a strict sieve, and only these countably many "modes" survive it. Multiplying each X_n by its matching time factor T_n(t) = e^{-\alpha\lambda_n t} gives one basic building-block solution u_n(x, t) = \sin(n\pi x/L)\,e^{-\alpha (n\pi/L)^2 t} for every n. The full worked solution of this exact problem, matched to a specific starting temperature profile, is carried out in heat equation on an interval.

The payoff: infinitely many modes, one general solution

Each allowed n hands you one valid "mode" solution u_n(x,t) — but a single mode is a very special-looking wave: it starts as a pure sine shape and just fades in place. Almost no real initial temperature profile looks like a single clean sine wave. The move that rescues the method is the superposition principle: because the heat equation is linear, any sum (even an infinite sum) of solutions is itself a solution. So build the general solution as a whole stack of modes, each with its own weight b_n:

At t = 0 this collapses to u(x, 0) = \sum_n b_n \sin(n\pi x/L) — exactly a Fourier sine series. Choosing the b_n to match any given initial temperature profile is precisely the Fourier-coefficient calculation, and it is no coincidence at all: the boundary conditions selected sines as the allowed spatial shapes, and sines are exactly the basis Fourier series are built from. Separation of variables and Fourier series are, underneath, the same idea wearing two different hats.

The allowed shapes

On [0, \pi] the eigenfunctions are \sin(nx). Step n: each is zero at both ends (satisfying the Dirichlet conditions) and packs in one more half-wave. These are the only spatial shapes separation allows — the building blocks every solution is assembled from.