Well some of the lower bits are noise. There are plenty of ways to add non-noise to those bits with an algorithm that would later subtract the noise leaving the original added bit. Which then gets added into the output. So if you can do this you do reduce file size.
What I understood is that the original recording is analyzed and that the input characteristics of the used recording equipment is taken into account and corrected for. Not all tape equipment is really linear.
The original signal is recorded at 192/24 and analyzed (for bias tones etc which can also be used to 'correct' for tape speed speed variances when analysed. This can be done using a not steep filter as there isn't much that peeps out above say 90kHz in any recording anyway so no chance of aliasing.
Then that original recording is 'EQ'ed or phase corrected or other aspects that needed correction is done.
That would be the actual brilliance of the encoding side (if this is actually true).
That 'corrected' file is then cut into 3 bands.
The upper 2 bands are subsequently compressed using a lossy encoder. It can be lossy as no one can actually hear ultrasonics anyway and as Blumlein mentioned can be encoded with just a few bits.
Those smaller in amplitude and > 20kHz lossy files are then 'encoded' with an algorithm that 'sounds like white noise' but when decoded retrieves the 2 upper bands.
Now at the same time the lowest bits of the 'improved' master are removed.
This could be either truncated or, what I would have done if I designed the codec, dithered so to get a larger dynamic range beyond the amount of remaining bits. This is part not lossy, execpt for the obvious bit reduction but audibly not so much because of dithering.
This result is encoded in the upper bits. As it is digital there is no noise... just bits so it will only work in the digital signal and won't work anymore when one would decode it with a non MQA DAC and then record the analog signal again as noise would have f'ed up the > 20kHz unfold.
Anyway the lowest bits get the 'disguised as noise by an algorithm' and this is encoded in the lower, removed bits.
So not added within noise but granted in most recordings those bits would contain mostly noise anyway.
This is the 'clever' part of restricting the bandwidth to 44.1 or 48kHz while keeping the same amount of bits.
Bob could easily not remove the lowest bits of the original and simply do 44.1/20 bits or 48/30 bits as the digital signal but most DACs would not recognize the signal or not truncate correctly so it would not play which would be an obvious DRM.
So they had to remove the less important bits < 20kHz in order to let MQA play on normal DACs.
Now if the processing at the encoding side indeed takes place then the not unfolded recording may well sound 'better' because of the processing but has a slightly higher noise floor.
The first 'unfold' is done with software where they extract and decode the unfolded frequencies and then simply 'paste' the first folded band to the lower bit < 20kHz but (dithered ?) signal and so one gets 'improved (in the encoding process), not steep filtered' 0-40kHz signal in return. The highest band is not used unless one has a hardware decoder as well which 'pastes' the highest band as well to the first unfolded signal.
If this is really how MQA works I can easily see the technical brilliance of it and may actually have better 'master' because of the processing at the encoding stage.
Now the part where it becomes 'tricky'. That first 'sampling and correcting' the 24/192 recording could easily be distributed and simply sold as improved master quality recording lets call this IMQ. But that could be illegally copied/ripped etc. They could only ask the studios money for improving the master.
So the encoding bit with last unfold NOT being free but paid for with a license and the 'correct' filters being used which are similar to the used anti-aliasing at the encoder side is what lets Bob get paid for each DAC/license.
Now to the point Blumlein made about the last unfold. The 192 kHz will appeal to those believing such is needed because they have been told that they can secrectly hear ultrasonics above 40kHz when it is harmonically related. These folks will buy the MQA DAC to ensure they are not missing out on the inaudible goodness.
I would say the unfold is already not necessary for the sake of >20kHz but 88.2/96kHz is good enough and at least the 'noise encoded' higher frequencies are not 'noise' anymore but ultrasonics.
So.. I would say just use a normal DAC unfold it for free to 96kHz and if you want to see the 192kHz light on your DAC and use a slow filter then by all means upsample to 192kHz or even higher !
It won't sound worse than the first unfold anyway as there is no useful info anyway.
But I could be wrong in this and is just a scam. Personally I would like to think that the technique used is rather clever... the audible benefits would ONLY be because of that first stage (if that actually is done) but this would benefit even non un-folders and certainly software first unfolders for free.