I think the reason for this to work is because 32 floating point decimal means that the values are constantly calculated (that´s why it´s floating) so when he lowers the volume so that it peaks under 0 dB the shape is calculated back to a pure sine . It will be shite if you export to 16 or 24 bit wave file...
Some nerdy techniocal stuff about floating point:
http://en.wikipedia.org/wiki/Floating_point