Why a derivative is the hard part
You arrive at this rung having solved PDEs on bounded intervals by hand: separate the variables, expand the data in a Fourier series, and let each mode evolve on its own. That worked because a finite bar has a discrete ladder of natural frequencies. But what about an infinite bar, or the whole real line, where there is no box to set the modes? The sum becomes an integral, the series becomes a transform, and the same idea scales up. This rung is about that move to the whole line, and every guide in it leans on one mechanical fact you will meet right now.
Here is the obstacle, stated plainly. The equations of this rung — the heat equation u_t = k u_xx, the wave equation u_tt = c^2 u_xx — are easy to write and hard to solve for exactly one reason: they contain derivatives in x. Multiplying, adding, integrating a number against a fixed weight — these we handle without breaking a sweat. Differentiation is the operation that couples a value to its infinitesimal neighbours, and that coupling is what makes a PDE a PDE. If only we could find a viewpoint in which u_xx were just u multiplied by some number, the whole thing would collapse into algebra. The Fourier transform is exactly that viewpoint.
Building the function out of pure waves
The Fourier transform rests on one bold claim: any reasonable function f(x) on the line can be written as a continuous blend of pure waves e^(i k x), one for every frequency k from minus infinity to plus infinity. The amount of each wave you need is a number F(k), the transform of f. So F(k) is a recipe: it tells you, frequency by frequency, how much of each wave to stir in to rebuild f. A spike concentrated at one point needs a little of every frequency in equal measure; a slow gentle bump needs mostly low frequencies. Reading f off from F is the inverse transform, the integral that adds all those waves back up.
forward (analysis): F(k) = integral over x of f(x) e^(-i k x) dx
inverse (synthesis): f(x) = (1/2pi) integral over k of F(k) e^(i k x) dk
f(x) <-- a function of position --
F(k) <-- the same thing, sorted by frequency --The one fact this whole rung runs on
Now the payoff. Take one pure wave, e^(i k x), and differentiate it in x. By the chain rule you get i k times e^(i k x) — the very same wave, merely scaled by the number i k. Differentiation does not change a pure wave's shape at all; it only multiplies it. Since every function is a blend of these waves, differentiating the whole function f just multiplies each wave in the blend by its own i k. Translated into transforms: the transform of f_x is (i k) F(k). This is the differentiation rule, and it is the engine of the entire rung.
transform of u_x = (i k) U(k) transform of u_xx = (i k)^2 U(k) = -k^2 U(k) transform of u_xxx = (i k)^3 U(k) derivative in x <--> multiply by i k in k (the hard operation) (the easy operation)
Sit with what just happened. The hardest operation in the original problem, the x-derivative that coupled every point to its neighbours, has become the easiest operation there is: multiply by a number. Differentiating twice multiplies by (i k)^2 = -k^2. So on the transform side, u_xx is simply -k^2 times U(k) — no derivatives left at all. The PDE, which tied together the values of u at neighbouring points, becomes a separate, uncoupled relation at each frequency k, with no k talking to any other. We have traded one hard problem in x for infinitely many trivial problems in k.
Watch a PDE collapse into algebra
Let us run the heat equation through the machine, just far enough to feel it click — the next guide does the full job. Start with u_t = k u_xx on the whole line, and transform both sides in x. The left side, u_t, is a time-derivative, which the x-transform leaves alone except to make U a function of both k and t: it becomes U_t(k,t). The right side, k u_xx, becomes k times -k^2 U, that is -k k^2 U. (Be careful: the small k here is the heat constant, the big k is frequency — two different k's that the symbol unhappily shares.) The PDE has turned into U_t = -(constant) k^2 U.
- Transform the PDE in x. Using the differentiation rule, u_xx becomes -k^2 U, so u_t = k u_xx turns into U_t = -k k^2 U at each fixed frequency.
- Solve the resulting ODE in t. For each frozen k this is just exponential decay, U(k,t) = U(k,0) e^(-k k^2 t) — an ordinary differential equation you mastered two rungs ago.
- Read U(k,0) from the initial data: it is the transform of u(x,0) = f(x). The whole solution lives, fully assembled, on the frequency side.
- Invert. Apply the inverse transform to turn U(k,t) back into u(x,t). This last integral is where the genuine work hides — and where the heat kernel will appear in the next guide.
Notice the strategy in miniature: transform, solve the easy problem, invert. The first step trades the PDE for an ODE; the middle step is something you can already do in your sleep; only the last step, the inversion, demands new care. And look at the physics that drops out for free. The decay rate is k^2: high frequencies, large k, die away vastly faster than low ones. That is the smoothing of the heat equation seen in its purest form — the wiggliest features vanish first, leaving a gradually smoother profile. The transform did not just solve the equation; it explained, in one line, why heat smooths.
The second rule, and what waves cost
Differentiation is the headline rule, but a partner rides with it. Many solutions arrive in the form of a product on the frequency side — say U(k,t) = F(k) G(k), the initial data multiplied by some response. The convolution theorem tells you what that product means back in x: a product of transforms is the transform of a convolution, a sliding-and-blending integral that smears f against g. This is the structural reason a solution so often looks like the initial data blurred by a fixed shape — that fixed shape is a fundamental solution, and the blurring is convolution. Multiplication in k is convolution in x; the two faces of one fact.
There is also a conservation law hiding in the transform, the Plancherel theorem: the total energy of f, the integral of |f|^2, equals (up to the convention's constant) the total energy of F, the integral of |F|^2. Nothing is lost in passing to the frequency side; the transform merely re-sorts the same energy by frequency. For the wave equation this is precisely conserved energy. For the heat equation it makes the smoothing quantitative: as the high-k part of |F|^2 decays, the total energy genuinely drops, and you can measure exactly how fast.
What the transform reads off at a glance
One last reason this single trick earns five guides: once a PDE is on the frequency side, you can often read its physics straight off without solving anything. Feed a trial wave e^(i (k x - omega t)) into a linear constant-coefficient PDE and the differentiation rules turn the equation into a plain algebraic link between frequency omega and wavenumber k. That link is the dispersion relation, omega = omega(k), and it is a fingerprint of the equation. From it you read the phase and group velocities — how fast a single crest travels versus how fast a packet of waves, and its energy, moves — which is the closing guide of this rung.
Be honest about the boundaries of the method, though. The differentiation rule rests on linearity and constant coefficients: it works because differentiating a wave just multiplies it, and superposing solutions stays a solution — the superposition principle in action. Throw in a coefficient that varies with x, or a nonlinear term like u u_x, and a single frequency no longer evolves on its own; frequencies start feeding each other, and the clean uncoupling breaks. The transform method is a precision tool for the linear, constant-coefficient, whole-line world — which is exactly the world this rung lives in, and an enormous and useful world it is.