Top panel: the distribution of S20/S3.6 in the general source population, in the sample of HzRG by Seymour et #160;al. (2007), and in our IFRS sample. The IFRS clearly occupy a different regime than the general population, and tend to overlap more with the HzRG. Bottom panel: the histogram of the IFRS S20/S3.6 ratios as in the upper panel, shifted to the right by log(5) #160;= #160;0.7. This takes into account that Norris et #160;al. (2010) found no IR counterparts for IFRS in a stacking analysis with a 5 times higher sensitivity. On average the IFRS then have a S20/S3.6 which is about two times higher than that of the HzRG.

Slices through three-dimensional ensemble mean (left panels) and variance fields (right panels). The top panels show results obtained with foreground correction, while the bottom panels show results without any foreground correction. As for the power spectrum, we find an excessive large-scale power when foreground corrections are not applied. When the foreground is computed self-consistently, the result is a non-contaminated reconstruction. The variance fields are also affected, as is shown by the notably darker bottom on average compared to the top slice, which indicates higher variance.