It's not compression. Look it up, it's really cool. Random sampling in a sparse basis. Using the prior. Single pixel camera can reconstruct a 10MP image, etc.
A more approachable example is not needing Nyquist limit of 44.1kHz to reconstruct any frequency up to 22kHz, if the sound is sparse in some basis and samples are random.
For example, if you know a signal is only a solo trumpet, you can probably reconstruct a perceptually perfect recording from just a few kHz of random samples. But, a solo piano would need many more random samples, as there are more harmonics and the hammer strikes, greater complexity and broadband hammer click isn't sparse in frequency.