Again, I would like to stress that downsampling data is not compression.
The concept is fundamentally different from data compression, which uses fewer bits to store information than the full data set and requires decompression when reading the data. That is, the data compression process finds repeating blocks of bits or other patterns in the data, and is able to call these out in a way that each iteration of those patterns doesn't have to stored, thereby saving space. Those repeating patterns are then reconstructed (i.e. decompressed) when reading the file. Lossy compression also uses intelligent ways of eliminating information that is unlikely to be discernible by the end user when encoding the file - but this is not integral to the concept of compression; only to lossy compression. This is not downsampling, because, for example, a LAME V0 mp3 file may still retain a 20 kHz signal if it is insanely loud - yet most of the time it throws away all the information about 16 kHz or so because we can't tell the difference. The sample rate isn't decreased, nor is the bit depth, so it is not downsampling.
Yes, downsampling does throw away information. It is lossy. But it is not compression. The data is not put in a container that is smaller than the full data set. Yes, it is smaller than the original data set, but the container the downsampled data is in is not smaller than the downsampled data set itself. There can be no decompression of an uncompressed but downsampled file. Yes, you can upsample the data back to the original size, but that is a pure fabrication of data based solely on the content of the downsampled file, not a reconstruction of the original, non-decompressed file as occurs with the decompression of a compressed file.
So, I repeat, 16 bit, 44.1 kHz LPCM audio is not data compressed from a 24 bit, 96 kHz (or whatever) master. It is downsampled and in most cases dithered (i.e. noise added to decrease distortion from downsampling - sort of like anti-aliasing).