Hi Cheng and Cameron,

As far as detrending is concerned, it is a linear operation. As such, averaging detrended signals is equivalent to detrend the average. For filtering it depends on how exactly it is implemented. NIAK uses (as SPM I believe) a least-square fitting of a basis of discrete cosines. It is therefore linear also. As a side remark, I believe that detrending and high-pass filtering are highly redundant, and I would rather skip detrending altogether, which is what's implemented in NIAK. 

Bottom line is, I don't understand why there are strong linear trends in the average time series, whatever the parcellation may be. There's clearly something I am missing there. I have never performed a time-space frequency analysis on the rois_1000 or rois_3000 average time series, but I'd be really curious to see if that phenomenon can be observed there as well. All I can tell is that the time course estimated through ICA have no power below 0.01 Hz, besides small non-stationarities. I don't know which tool Cheng is using to assess the power spectrum, but if there is any doubt on that front I'd suggest to try niak_visu_wft which is available in NIAK.

I hope this helps,

Pierre