I've thought about finding a signal processing transformation that is lossless, and yet brings a decent separation into several frequency subbands.

By lossless I mean not just "lossless if quantization uses an infinite number of bits" but really, a bijective transformation from time to subbands, say from 16-bit PCM to 16-bit subband samples without any loss..

Does someone know if such a thing is possible ?

By the way, Dibrom, what would you think about creating a more specific forum in HA ? That would be one that's related to audio research/development in audio coding - such as signal processing - but from a developer's (not user's) perspective as in General.

Of course, the subject would have to be pretty general.. so the name "Tech - General" would apply, as opposed to "Tech - MP3" or "Tech - Vorbis".

For example, if there would be a "Tech - General" I guess I'd have posted my message there (instead of General or say, Tech-MPC).

Cheers

Look for transforms implemented with lifting steps, like the IntMDCT

Do you know that there is a forum on Mp3-tech.org that is mainly technical?

I guess you are looking at transforms which don't give irrational results, so anything involving pi, e and trig functions are out.

That leaves things like Walsh, Hadamard and Haar (max order 2) transforms. Unfortunately these transforms aren't as efficient to encode and do not provide accurate frequency information.

Lifting integer wavelet transforms can be pretty good and they can be implemented entirely lossless. The trick is to get the right coefficients, in the Off-Topic forum there is a thread titled Good documentation (http://www.hydrogenaudio.org/forums/index.php?act=ST&f=17&t=3524), it contains some links to good sites about things like this.