Per usual, as I remember, good post Vorpal. I figured I might 'high-school-ify' some of the things if that's alright. =)
VorpalNeko
Understanding the Fourier transform via the Fourier series first is a good approach. That's the f(x)cos(nx), and f(x)sin(nx) integrations referenced to above, since you would need to know enough calculus to do that anyway. It really helps in understanding how Fourier "picks out" the contribution of a particular frequency. It's literally a projection.
Explanation on projecting:
'Projection" is a good name for it. Think of 3 dimension space. The ground is 2 dimensional, and the space above it provides the third dimension. Picture a stick pocking out of the ground at an angle. That's like a vector. Now shine a light from above...the stick casts a shadow on the ground right below it! The shadow is like another vector: the projection of the stick onto the ground. So projecting something onto something else is like finding what part of the shadow of thing A touches thing B.
Quote:
A brief roadmap may be something like this:
...
(4) The above obviously works for any dimensions. In fact, it works in infinitely-many dimensions, except in that case we have to be careful to make sure vectors have finite lengths.
(5) In the vector space of functions over some interval, <f|g> = Int[ f(x)g(x) dx ], the integral being over the domain. The only really important concern is that √<f|f> and √<g|g> must be finite. (Again, for complex-valued functions, conjugate f in the integral.)
It might seem strange or confusing, but one can actually think of functions (rules that take one thing (say x) and associate it with another thing (say f(x))) in the same way as you do those 3 dimensional vectors...but there's a catch: it's an infinite dimensional space. Nonetheless, we should be able to choose those 'basis vectors' that Vorpal talked about and find the 'components' (the projections!) of a function vector along those basis functions.
Quote:
(7a) Therefore, we can project any function over I as <cos nx|f>cos(nx)/Ï€ and <sin nx|fsin(nx)/Ï€. It is nontrivial (yet true) that these sines and cosines are all we need to represent a (sufficiently well-behaved) function fully.
(7b) f(x) = (1/Ï€) Sum[ <cos nx|f>cos(nx) + <sin nx|f>sin(x) ].
A function that's periodic with period 2Ï€. Obviously, not all functions are like that. In fact, most aren't periodic. Seems like we should be able to take the limit as the period goes to infinity though, right? Turns out, funny things happen to the vector space of functions then. The basis you'd want to choose becomes 'uncountable' (you can 'count' natural numbers - i.e. 1,2,3,4... - but can't count, for instance, the real numbers - i.e. 0, 0.000000.......(forever), 0.000000(forever)...(infinite number of things), 0.0001, (infinite number of other things), 1,...). So the sum over sin and cos become integrals. It turns out it's easiest to write these integrals with sin and cos in terms of e^(ix) type functions instead.
This is the long way of saying:
Quote:
One way of "getting to" the Fourier transform is to do a certain limit of the Fourier series with complex-valued functions; the summation above will then be a Riemann sum (i.e., an integral).
Happy trails!