One pattern behind every transform
Step back and look at the machines you have built across this rung. The Fourier transform multiplies a function f(x) by e^{-i k x} and integrates over all x. The Laplace transform multiplies f(t) by e^{-s t} and integrates from 0 to infinity. Different letters, different limits — but the *shape* is identical: take your function, multiply it by some chosen function of two variables, integrate out the original variable, and what comes back is a new function of the second variable. That two-variable function you multiply by has a name: it is the kernel of the transform.
Write the general pattern once and the family snaps into focus: F(p) = integral over the domain of K(p, x) f(x) dx, where K is the kernel. Choose K(k, x) = e^{-i k x} and you have Fourier; choose K(s, t) = e^{-s t} and you have Laplace. Every member of the clan is just this same integral with a *different kernel chosen to match a different symmetry*. That is the secret the whole guide unpacks: the kernel is the personality, and once you know which symmetry your problem respects, you know which kernel — which transform — to reach for.
Hankel: the transform for round problems
The Fourier transform's building blocks are the straight travelling waves e^{i k x}, and they fit beautifully when a problem is laid out along a line or in a rectangular room. But so much of physics is *round* — a circular drumhead, heat spreading outward from a point in a flat disk, light focused by a circular lens. Try to describe a ripple expanding in a perfect circle using straight sine waves and you fight the geometry the whole way. The right move is to switch to building blocks that already respect circular symmetry, and those building blocks are the Bessel functions J_n(k r) — the radial standing-wave shapes you met as solutions of the Bessel equation back in the special-functions rung.
Swap the Fourier kernel e^{i k x} for the Bessel kernel J_n(k r) and you get the Hankel transform: F(k) = integral from 0 to infinity of f(r) J_n(k r) r dr. Notice the small but crucial extra factor of r — that is the radial slice of area in polar coordinates, the dr-and-r-d-theta you remember from double integrals, and it is exactly what makes the round geometry come out right. The payoff matches Fourier's: the messy radial part of the Laplacian in cylindrical coordinates, that awkward second derivative plus (1/r) first derivative, turns into plain multiplication by -k^2. A partial differential equation on a disk collapses to ordinary algebra in k, just as Fourier collapsed one on a line.
Mellin: the transform tuned to scaling
Fourier and Laplace are built around *sliding*: shift a function along, and the transform answers cleanly, turning a shift into a phase or an exponential factor. But some problems are not about sliding at all — they are about *scaling*, about zooming in and out, about how a function behaves near zero versus near infinity. Power laws, self-similar shapes, the tails of probability distributions: these live in the world of multiplication, x to a x, not addition, x to x plus a. For them you want a transform tuned to stretching the way Fourier is tuned to shifting, and that is the Mellin transform.
Its kernel is the pure power x^{s-1}: F(s) = integral from 0 to infinity of f(x) x^{s-1} dx, where now s is allowed to be complex. If that integral looks familiar, it should — set f(x) = e^{-x} and you are looking at the very definition of the Gamma function, Gamma(s) = integral from 0 to infinity of e^{-x} x^{s-1} dx. The Gamma function is literally the Mellin transform of the decaying exponential, and the same kernel quietly underlies the Riemann zeta function too. There is also a clean bridge to what you already know: substitute x = e^{-t} and the Mellin transform becomes a two-sided Laplace transform. Mellin is not a stranger; it is Laplace seen through the logarithm, the transform that turns multiplicative structure into additive structure.
Hilbert: a transform that stays at home
The Hilbert transform is the odd one out, and the difference is instructive. Fourier, Laplace, Hankel, Mellin all carry a function *into a new domain* — frequency, the s-plane, the k-axis, the complex s-strip — where the variable means something new. The Hilbert transform takes a function of time and hands back another function of time: same kind of object, same axis. It does not change your domain at all. What it changes is *phase*. It rotates every frequency component by exactly ninety degrees — turning every cosine into a sine and every sine into a minus-cosine — while leaving every amplitude untouched.
Its kernel is 1/(pi (x - t)), so the transform is a convolution with 1/x: H(t) = (1/pi) integral of f(x)/(t - x) dx. But that kernel blows up exactly at x = t, right in the middle of the integration, so the ordinary integral does not converge. This is not a flaw to patch over — it is the honest reason the Hilbert transform needs the Cauchy principal value, the careful symmetric limit that approaches the singular point from both sides at once and lets the equal-and-opposite infinities cancel. The principal value is not a fudge; it is the precise prescription that makes a singular kernel define a genuine, well-defined operation.
Why would you ever want to rotate phase by ninety degrees? Because it lets you build the *analytic signal*: take a real signal f(t) and form f(t) + i times its Hilbert transform, and you get a complex signal whose magnitude is the instantaneous amplitude (the envelope) and whose angle is the instantaneous phase. That is precisely how a radio demodulates an AM station, how an engineer extracts the slowly-varying envelope from a fast carrier, and how the famous Kramers-Kronig relations tie the absorption of a material to its refraction — because causality forces a signal's real and imaginary spectra to be Hilbert transforms of each other. A ninety-degree twist, it turns out, is the mathematical heart of envelopes, phase, and causality.
Reading the family tree
Lay the four kernels side by side and the logic of the whole family becomes visible at a glance. Each transform is the same integral pattern, F(p) = integral of K(p, x) f(x) dx; only the kernel and the domain change, and each kernel is the eigen-shape of one symmetry. Pick the transform whose kernel matches the symmetry of your problem, and the operation that was making the problem hard melts into multiplication. That single sentence is the entire navigational map of this rung.
TRANSFORM KERNEL K(p,x) DOMAIN of x BUILT FOR (symmetry it diagonalizes)
--------- --------------- ----------- -------------------------------------
Fourier e^{-i k x} -inf .. +inf translation on a line -> d/dx becomes i k
Laplace e^{-s t} 0 .. +inf decay / one-sided time -> d/dt becomes s
Hankel J_n(k r) * r 0 .. +inf circular symmetry -> radial Laplacian -> -k^2
Mellin x^{s-1} 0 .. +inf scaling x -> a x -> x d/dx becomes -s
Hilbert 1/(pi (x - t)) -inf .. +inf phase (stays in time, rotates by 90 deg)
General pattern, every row: F(p) = integral over domain of K(p, x) f(x) dx
Choose the kernel that fits your symmetry; the hard operation becomes multiplication.How to choose a transform
Faced with a hard integral or differential equation, the family gives you a short diagnostic. You are not guessing; you are matching the geometry of the problem to the symmetry of a kernel. Walk down this checklist and the right tool usually announces itself before you have written a single integral.
- Is the domain an infinite line with no preferred origin, and the key operation a derivative in x? Reach for the Fourier transform — derivatives become i k.
- Is it a one-sided problem in time with initial conditions — a system switched on at t = 0? Reach for the Laplace transform; the initial data rides along inside the derivative rule.
- Does the problem live on a disk or has circular symmetry, with a radial Laplacian? Reach for the Hankel transform with the matching Bessel order n.
- Is the structure multiplicative — power laws, scale invariance, behavior as x to 0 versus x to infinity? Reach for the Mellin transform and read the asymptotics off its poles.
- Do you need an envelope, an instantaneous phase, or a causality relation, staying in the time domain? Reach for the Hilbert transform and build the analytic signal.
And that is the whole arc of this rung. You started by letting a Fourier series stretch its period to infinity until the sum became the Fourier transform integral; then you met its cousin the Laplace transform for one-sided time. Now you can see them not as two isolated tricks but as two seats at a large table, with Hankel, Mellin, and Hilbert in the other chairs — every one of them the same integral, F(p) = integral of K(p, x) f(x) dx, wearing a kernel cut to fit a different symmetry. Learn to recognize the symmetry, and the right transform is already chosen for you.