Actually a common way to compress a signal is to transform it from the time domain to the frequency domain, by a Fourier transform. Then you can throw away half of the signal because the Fourier transform produces both positive and negative frequencies, for a real signal these are mirror images of each other, in other words for a real signal S(t), which has a Fourier transform F, F(f) = F(-f) (note the change of variables from time t to frequency f). Something like this is used in quadrature amplitude modulation (QAM)
Doesn't the Fourier transform of the real signal contain both real and imaginary components (amplitude and phase) though? Then half of the (complex) transform is still the same size as the (real) input signal, so it's not compression per se as far as I can tell. But yes there are schemes that compress in the frequency domain since when it comes to many types of data, in particular, sound and images, the frequency domain is much more predictable than the spatial domain and is more readily compressed. It's also way easier to quantize a signal without losing too much perceived accuracy (audio, video, ..) by working in the frequency domain rather than in the spatial domain. JPEG does this for instance, as well as MP3, though I don't know all the details.
yes.
typically these are not actually using the actual Fourier transform, but other frequency based transforms:
for example, JPEG (and the MPEG family) uses the DCT (Discrete Cosine Transform);
MP3, Vorbis, and AAC use MDCT(Modified Discrete Cosine Transform).
another option (used, for example, in HDPhoto / JPEG-XR) is the WHT (Walsh Hadamard Transform).
the WHT has an advantage in terms of being faster and can more easily be made lossless (while the DCT can be made fully-reversible / lossless, doing so is computationally expensive). however, the WHT has a drawback in terms of not compressing as well as DCT.
...
the advantage of these transforms is, as noted, that they compress well and are fairly easy to quantize.
they also offer a reasonable tradeoff between conceptual elegance and computational performance.
another possible strategy is Vector Quantization, where the values are instead approximated via either replacing them with indices into lookup tables (similar in concept to a font or color-palette), or via using very coarse numerical approximations (such as selecting between a palette of interpolated values, ...).
the advantage of VQ is that decoders can be made to be very fast (compared to what is generally possible with DCT or WHT).
the drawback is generally a tradeoff that either the format has poor quality, poor compression, or both, or alternatively needs to be overly complex and requires a computationally expensive encoder to get good compression (pretty much the entire codec may turn into an ugly mass of special cases and hair...). so, it may become a lot of tables to build tables to index other tables, with lots of special cases for how to represent the table indices, and lots of dedicated special-case code-branches, ...
going the other direction, you find things like the DWT (Discrete Wavelet Transform), which tend to offer some interesting properties (and was used, for example, in JPEG-2000 and in some video codecs, such as Dirac), and the advantage of being conceptually elegant, but the drawback of generally (in practice) being a bit slower than what is possible with either the DCT or WHT transforms (while not really offering much in terms of a compression advantage).
the result then is that thus far DCT based formats have generally held the dominant position in the image/audio/video space.