Nico,
The power spectral density (the spectrum you would see on a spectrum analyzer) is extremely close to the carrier -- that is no spurs.
The reason for this is that if you make small adjustments to the delay lines every so many clocks, these small steps (~25ps steps) represent a tiny increase or decrease in the period of the clock, which means that there is a correspondingly tiny step of frequency that lasts for one period, and then is gone.
The result is gaussian noise bump in spectra around the carrier.
There are no other frequency components (unless they are already present in the input clock, in which case they come directly through -- a DCM does not attenuate jitter).
In the time domain, if you do a histogram of periods, you will get two overlapping (or more) gaussian distributions of period length: period - tap, period + tap, possible period + two taps, possible period - 2 taps. The time domain is the superposition of period lengths with and without the taps inserted. Since the change in tap is completely random (like flipping a coin each time), there is no determinism, and hence no rate at which changes occur, and no sub-harmonics, or sidebands.
Austin