The PLL removes the low frequency components of the phase noise in the output oscillator that is locked to a reference, and passes through the low frequency components and removes the high frequency components of the reference it is locked to. Thus the phase noise of the output oscillator sees a high pass filter with the use of the PLL, and the phase noise of the reference oscillator see a complementary low pass filter. The filtering operation has a slope consistent with the order of the loop.
To evaluate noise in the PLL due to phase noise. It is helpful to first understand phase noise in terms of its frequency performance, and then with that understand how the phase lock loop filters this phase noise. Below is a graphic demonstrating the concept of phase noise components and what the power spectral density of phase noise is showing us.
Here we see the power spectral density (PSD) as a typical Phase Noise plot, showing us the power due to phase fluctuations in the signal at each frequency component of those fluctuations. I have taken a slice at a low and high frequency component, to show if we were able to filter the phase noise at those frequency offsets, what the phase fluctuations vs time would be and how they each have a frequency consistent with the frequency offset selected, and a relative magnitude consistent with relative power at that frequency offset.
If we were to lock the oscillator with the phase PSD shown above to a much cleaner reference oscillator (as typically done), using a PLL loop bandwidth of 200 KHz (the plots show “CR Loop BW” as these are graphics I have explaining a carrier recovery loop which similarly tracks out the low frequency phase noise components), the low frequency phase noise components below this bandwidth would be reduced by a high pass filter with a slope consistent with the order of the loop. The resulting slices of phase noise would appear as below.
To determine the total noise, the resulting phase noise would be integrated out to the measurement bandwidth of the system (indicated as "Channel BW" in the plots above), consistent with the shaded area in the plot below. Since the dominant noise is in the low frequencies, there will always be a significant improvement in the reduction of total phase noise after the PLL, assuming we are locked to a lower noise reference.