Hi Roger,
Here is a closer look at the fourier amplitudes to the device and network variances.
I look at the "long-wavelength" amplitudes, to use physics jargon; that is, amplitudes which correspond to periods of a day or greater. The basic question is whether there is a day-periodicity ( or, eventually, one-week).
The last plots showed that there is nothing obvious from this analysis.
To look a bit in the dirt, I did the analysis for randomly permuted data. That is, I take the time-series of 2,980,800 variances (for both device and network cases) and scramble the order and redo the Fourier plots. The idea is to see if the ordered and ramdomized data sets give plots that look different in any way.
Again, nothing obvious, but you can peer at them to see what you think.
All of the plots smooth the data with a moving average, as a way of looking for peaks in the dirt. I plotted for 20- and 60- bin moving averages.
One interesting point: if I simulate random data with a weak periodicity, the randomness can shift the Fourier peak slightly off its nominal position. What I have done so far sees this shifting to lower frequency (longer period). So for example, we might see a peak corresponding to a 24 hr period that is centered on 25 hours.
I won't say more for now, but leave you to take a look at the plots.
One plot shows the cumdev of the device variance fourier amplitudes. In a cumdev, a positive peak in the data appears as a positive step.