From probing to redefining
The previous guide left us with a complaint and a clue. The complaint: ordinary functions cannot model a point source, cannot be safely differentiated when they have a corner, and break apart in the kind of limits PDEs constantly take. The clue was a change of habit — instead of asking 'what is the value of u at the point x?', we agreed to only ever ask 'what is the weighted average of u against a smooth, well-behaved weight?'. This guide turns that habit into a precise machine. The smooth weights get a name, test functions, and the objects we probe with them get a name, distributions.
The mental image to carry through everything below is a measuring probe. A test function is an idealized probe you press against an unknown object; the object answers with a single number, its average against that probe. If you have enough probes — enough shapes, enough positions, enough sharpness — the full list of answers pins the object down completely. We never look at the object directly; we only ever read it through probes. That self-denial is exactly what buys us the freedom to differentiate corners and to make sense of a spike.
What makes a good test function
A test function is a real-valued function phi that obeys two strict demands. First, it is infinitely smooth: you can differentiate it as many times as you like and never hit a corner or a jump — it is what analysts call C-infinity. Second, it has compact support: phi is exactly zero outside some bounded interval, so it is genuinely zero (not merely small) everywhere far away. The two demands pull against each other, which is why the existence of even one such function feels surprising the first time you meet it.
Why both? Smoothness is the bribe: because phi has all its derivatives, we will later be allowed to throw every derivative we need from the rough object onto the safe shoulders of phi — that is the trick that gives corners a derivative. Compact support is the safety net: when we integrate by parts to move a derivative across, the leftover boundary terms sit out at the edges of the domain, and a probe that is flatly zero out there makes those terms vanish without a fuss. A probe that is smooth but only decays slowly would leave boundary debris; compact support sweeps it away.
A distribution is a rule, not a graph
Now the bold move. A distribution is defined to be a continuous linear functional on the test functions: a rule T that takes any probe phi and returns a number, written T(phi) or, by long tradition, with angle brackets as < T, phi >. 'Linear' means it respects sums and scalars, < T, a phi + b psi > = a < T, phi > + b < T, psi >, so probing a combination is the same as combining the probe-answers. 'Continuous' means that if a sequence of probes settles down nicely, their answers settle down too — no wild jumps. That is the entire definition. A distribution is not a graph of values; it is the act of measuring against every probe.
The reassuring part is that this redefinition loses nothing — every ordinary function is already a distribution in disguise. Take any reasonable function f (anything you can integrate locally). It induces a rule: feed it a probe phi and let it answer with the integral < f, phi > = integral of f(x) phi(x) dx. That is just the weighted average from the opening section. This rule is linear and continuous, so f is a perfectly good distribution. Different f's give different rules, so no information is thrown away. The new world is strictly bigger than the old one, never smaller — exactly the property a generalized function needs.
And the new world holds objects no ordinary function could be. The headline example is the Dirac delta: define delta by the simplest rule imaginable, < delta, phi > = phi(0). It reads off the value of the probe at the origin and reports nothing else. There is no function whose integral against every phi equals phi(0) — yet as a rule it is flawlessly linear and continuous, so it is a legitimate distribution. The point source that classical functions could not hold is, in this language, just the simplest evaluation rule there is. The next guide takes the delta apart in full; here it stands as proof that the new world is genuinely larger.
What you can read off a rule
Even before differentiating anything, the functional viewpoint quietly hands you the operations you will need. To shift a distribution to the right by a, you shift the probe to the left instead: < T(. - a), phi > = < T, phi(. + a) >. To scale it, you scale the probe. The pattern is always the same — you cannot reach inside T, because there is no inside, so you act on the probe and let the definition carry the operation across. This 'move it onto the test function' reflex is the single most important habit in the whole theory, and it is exactly what will let us define the derivative next guide.
You can also still ask where a distribution 'lives', even though it has no graph. Its support is, roughly, the smallest closed set outside of which T returns zero for every probe that vanishes near it — the region the rule actually feels. For the delta the support is the single point {0}: probe it with any phi that is zero near the origin and you get phi(0) = 0. For an honest function it is just the closure of where the function is nonzero. So 'concentrated at a point' becomes a clean, provable statement, not a hand-wave.
test function phi : smooth (C-infinity) AND zero outside a bounded set bump example : phi(x) = e^(-1/(1 - x^2)) on (-1,1), else 0 distribution T : a continuous LINEAR rule phi |-> < T, phi > (a number) ordinary f as T : < f, phi > = integral of f(x) phi(x) dx Dirac delta : < delta, phi > = phi(0) (no such function f) shift by a : < T(.-a), phi > = < T, phi(.+a) > (act on the probe)
Honest edges of the theory
Three caveats keep you from over-claiming. First, you cannot in general multiply two distributions. Squaring the delta, 'delta^2', has no meaning, and trying to define products of distributions honestly is where many naive calculations go wrong — this is a real limitation, not a temporary gap, and it is the deep reason nonlinear PDEs are so much harder than linear ones. You can safely multiply a distribution by a smooth function, but not by another distribution.
Second, not every distribution is built to ride the Fourier transform. The bump-function probes in D allow wild growth at infinity, and the price is that the Fourier transform may not stay inside the same world. To get a Fourier theory we trade compact support for fast decay and work with the tempered distributions, probed by Schwartz functions instead — that is the whole subject of guide 4 in this rung. So the space of distributions you pick is not one-size-fits-all; you choose the probe space to fit the job.
Step back and the payoff is already visible. We have not solved a single PDE, yet we have rebuilt 'function' on a footing that effortlessly contains a point source, comes with shift, scale, and support for free, and — crucially — is staffed entirely by smooth probes onto which we can dump any derivative we like. That last fact is the spring loaded for the next guide, where moving a derivative onto the probe turns even a non-classical, cornered, or singular object into something you can differentiate as often as you please.