A scheme that optimizes when the various attached clients come on the medium to collect there traffic could well have the power gains described.
But, if you're not always listening to the medium, you're bound to miss some opportunity to receive data. The paper seems to say with some methods, the AP wakes the client, and then transfers to it, possibly queuing multiple packets before waking up the client. I'd expect there to be an increase in latency with a scheme like that, and indeed the article seems to mention that. (But, for what one is doing on a mobile device, it probably won't matter.)
Therefore, to amortize the wake-up cost, PSM clients are made to wake-up less frequently, permitting multiple packets to queue up at the AP. Of course, such queuing introduces latency in PSM packet delivery.
The article then improves upon this method, but I didn't see if they actually did anything to remove the latency. I also didn't read the whole 13 pages of research.