Why properties, not just the integral
In the previous guide you met the Fourier transform pair: a function f(t) in time and its transform F(omega) = integral of f(t) e^{-i omega t} dt in frequency, two faces of the same object. You could, in principle, attack every new problem by grinding out that integral from scratch. But almost nobody does, and here is why: the few transforms you compute by hand become a small dictionary, and a handful of properties let you slide, stretch, and combine those dictionary entries to reach almost anything you need — without touching the integral sign again.
Every property below flows from one source: the transform is linear and built on the wave e^{-i omega t}. Linearity alone already says the transform of a sum is the sum of the transforms, and a constant multiple pulls straight through — the same superposition principle that ran through Fourier series and linear differential equations. The richer properties come from asking a sharper question: when you do something simple to f in the time domain, what is the matching move on F in the frequency domain? The answers are strikingly clean, and once memorized they turn the transform into a calculator.
Shifting and modulation: a tug-of-war between the two domains
Start with the simplest move: slide the signal later in time by an amount a, replacing f(t) with f(t - a). What happens to F? The time-shifting property says the magnitude of F(omega) does not change at all — every frequency is still present in the same amount — but each frequency picks up a phase twist e^{-i omega a}. That is exactly right intuitively: delaying a song does not change which notes it contains, only when each wave crests. A pure delay is invisible to the energy spectrum and lives entirely in the phase.
Now flip it around. Multiply the signal by a wave instead of shifting it: replace f(t) by e^{i omega_0 t} f(t). This is the modulation property, and it does in frequency exactly what shifting did in time — it slides the whole spectrum, sending F(omega) to F(omega - omega_0). Multiplying by a wave in one domain is shifting in the other; the two operations are mirror images across the transform. This is precisely how AM radio works: multiplying your voice by a high-frequency carrier lifts its spectrum up to the broadcast band, and the receiver multiplies again to slide it back down.
Scaling and duality: narrow here means wide there
Squeeze the signal in time — replace f(t) by f(a t) with a > 1, compressing it into a shorter burst. The scaling property says its transform becomes (1/|a|) F(omega/a): the spectrum stretches WIDER as the signal gets narrower, and dims a little to conserve area. This is not a quirk; it is a law. A sharp, brief click necessarily contains a broad band of frequencies, while a long, lazy tone is nearly a single pure frequency. You cannot have both a signal that is tightly localized in time AND tightly localized in frequency — squeeze one and the other spreads. That trade-off is the mathematical bone beneath the Heisenberg uncertainty principle in quantum mechanics, which is, at heart, this scaling property applied to a wave packet.
Behind the symmetry of all these property-pairs sits one deeper fact: the duality of the transform. The forward transform (time to frequency) and the inverse transform (frequency to time) are nearly identical integrals — they differ only by the sign in the exponent and a factor of 2 pi. So any transform pair, read backwards, gives you a second pair for free. The classic example is the Gaussian, the bell curve from the Gaussian integral: its Fourier transform is again a Gaussian. It is its own echo across the transform — which is exactly why the bell curve is the boundary case of the width trade-off, the one shape that is equally compact in both domains.
Convolution: the smearing of one signal by another
Now the centerpiece. Many of the most useful operations in the world are not pointwise — they mix a value with its neighbors. Blurring a photo replaces each pixel by a weighted average of those around it. A reverberant room replaces each instant of sound by a fading echo of the instants before it. The output of any linear, time-invariant system is its input smeared by the system's impulse response. That smearing operation has a name: convolution, written (f * g)(t) = integral of f(tau) g(t - tau) d tau. Read it slowly: for each output time t, you flip g, slide it to position t, multiply it against f everywhere, and add up the overlap. It is a sweep of one function dragged across the other.
Computed directly, convolution is heavy: for every one of the output points you must do a whole integral, an overlap sum that drags across the entire signal. For a long recording or a megapixel image, that is an enormous amount of arithmetic. This is the wall — and the convolution theorem is the door through it. You already met its sibling in the Laplace world, the Laplace convolution theorem; the Fourier version is the same beautiful idea, the engine behind the whole frequency-domain toolkit.
The convolution theorem: smearing becomes multiplication
Here is the statement that earns this guide its title. The Fourier convolution theorem says: the transform of a convolution is the ordinary product of the transforms. In symbols, if f * g is the convolution, then its transform is simply F(omega) times G(omega), pointwise. The clumsy sweeping integral in the time domain becomes a plain multiplication, frequency by frequency, in the transform domain. (The duality cuts both ways: multiplying two signals in time convolves their spectra — the partner statement.) Why should this be true? Because convolution is built precisely so that each pure wave passes through untangled, and the transform sorts a signal into exactly those pure waves.
The intuition is worth dwelling on. Feed a single pure wave e^{i omega t} into a smearing operation and it comes out as the same wave, only scaled and phase-shifted — its frequency is untouched. So in the basis of pure waves, smearing is diagonal: each frequency is just multiplied by its own number, and that number is G(omega), the transform of the smearer. Sorting f into its frequencies, multiplying each by the right factor, and reassembling is precisely transform, multiply, invert. The convolution theorem is the statement that the Fourier transform diagonalizes convolution — the same reason eigenvalues made matrices easy back in linear algebra.
- Transform both signals: send the input f and the smearing kernel g (a blur shape, a filter, an impulse response) to their spectra F(omega) and G(omega).
- Multiply pointwise: at each frequency, form the product F(omega) G(omega). This single multiplication replaces the entire sweeping convolution integral.
- Invert back to time: take the inverse Fourier transform of the product, and you have f * g, the smeared output — no overlap integral ever computed.
Filtering, and an honest look at the dictionary
This theorem is why filtering is so simple in frequency space. To remove a 60 Hz hum from a recording, you do not design some intricate averaging in time — you transform, set the spectrum to zero near 60 Hz, and transform back. To sharpen or blur an image, you reshape its spectrum and invert. Every equalizer, noise gate, and edge detector is, underneath, a multiplication G(omega) chosen to pass the frequencies you want and suppress the rest. And because there is a fast algorithm to compute the transform of sampled data, the round trip 'transform, multiply, invert' is dramatically cheaper than the direct convolution it replaces — the reason this is not just elegant theory but the workhorse of modern signal and image processing.
One more property closes the circle and connects to energy. The Parseval-Plancherel theorem says the total energy of a signal — the integral of |f(t)|^2 — equals the total energy of its spectrum, the integral of |F(omega)|^2, up to a constant. Nothing is created or lost in crossing to the frequency side; the transform only redistributes the same energy among the frequencies. It is the continuous heir of the Parseval identity you met for Fourier series. Together with the property-pairs and the convolution theorem, it completes your working grammar of the transform: shift, scale, modulate, convolve, and account for energy — all without ever returning to the defining integral. Next in this rung, that grammar gets turned loose on the differential equations of physics.