What Is the Fourier Transform?

I have a pet theory that the reason why the FT, and other transforms (generating functions, Mellin/Laplace/Legendre/Haar), are so useful is because many real world functions are sparse and lend themselves to compressed sensing.

The FT, as are many other transforms, are 1-1, so, in theory, there's no information lost or gained. In many real world conditions, looking at a function in frequency space greatly reduces the problem. Why? Pet theory: because many functions that look complex are actually composed of simpler building in the transformed space.

Take the sound wave of a fly and it looks horribly complex. Pump it through the FT and you find a main driver of the wings beating at a single frequency. Take the sum of two sine waves and it looks a mess. Take the FT and you see the signal neatly broken into two peaks. Etc.

The use of the FT (or DCT or whatever) for JPEG, MP3 or the like, is basically exploiting this fact by noticing the signal response for human hearing and seeing it's not uniform, and so can be "compressed" by throwing away frequencies we don't care about.

The "magic" of the FT, and other transforms, isn't so much that it transforms the signal into a set of orthogonal basis but that many signals we care about are actually formed from a small set of these signals, allowing the FT and cousins to notice and separate them out more easily.

As mentioned by other commenters, a reason for the FT's dominance in particular is because sine, cosine, and complex exponentials are the eigenfunctions of the derivative operator. Since so many real-world systems are governed by differential equations, the Fourier Transform becomes a natural lens to analyze these systems. Sound waves are one (of many) examples.

And there's another good reason why so many real-world signals are sparse (as you say) in the FT domain in particular: because so many real-world systems involve periodic motion (rotating motors, fly's wings as you noted, etc). When the system is periodic, the FT will compress the signals very effectively because every signal has to be harmonic of the fundamental frequency.

The question is why "so many real-world systems are governed by differential equations" and "so many real-world systems involve periodic motion".

Well, stable systems are can either be stationary or oscillatory. If the world didn't contain so many stable systems, or equivalently if the laws of physics didn't allow so, then likely life would not have existed. All life is complex chemical structures, and they require stability to function. Ergo, by this anthropic argument there must be many oscillatory systems.

Differential equations aren't limited to describing stable systems, though, and there are chaotic systems that are also in some sense stable.

Ordinary differential equations can describe any system with a finite number of state variables that change continuously (as opposed to instantaneously jumping from one state to another without going through states in between) and as a function of the system's current state (as opposed to nondeterministically or under the influence of the past or future or some kind of supernatural entity).

Partial differential equations extend this to systems with infinite numbers of variables as long as the variables are organized in the form of continuous "fields" whose behavior is locally determined in a certain sense—things like the temperature that Fourier was investigating, which has an infinite number of different values along the length of an iron rod, or density, or pressure, or voltage.

It turns out that a pretty large fraction of the phenomena we experience do behave this way. It might be tempting to claim that it's obvious that the universe works this way, but that's only because you've grown up with the idea and never seriously questioned it. Consider that it isn't obvious to anyone who believes in an afterlife, or to Stephen Wolfram (who thinks continuity may be an illusion), or to anyone who bets on the lottery or believes in astrology.

But it is at least an excellent approximation that covers all phenomena that can be predicted by classical physics and most of quantum mechanics as well.

As a result, the Fourier and Laplace transforms are extremely broadly applicable, at least with respect to the physical world. In an engineering curriculum, the class that focuses most intensively on these applications is usually given the grandiose title "Signals and Systems".

One amazing application of spectral theory I always harp on when this topic comes up is Chebfun[1]. Trefethen's Spectral Methods in Matlab is also wonderful.

[1] http://www.chebfun.org/

I haven't read it! Thanks for the recommendation!