The ATESP radio survey

This paper is the fourth in a series reporting the results of the ATESP radio survey, which was made at 1.4 GHz with the Australia Telescope Compact Array. The survey consists of 16 radio mosaics with ~$8\arcsec \times 14\arcsec$ resolution and uniform sensitivity ($1 \sigma$ noise level ~$79 \mu$Jy) over the region covered by the ESO Slice Project redshift survey (~26 sq deg at $\delta \sim -40\degr$). The ATESP survey has produced a catalogue of 2967 radio sources down to a flux limit of ~0.5 mJy ($6 \sigma$). In this paper we present the optical identifications over a 3 sq deg region coinciding with the Patch A of the public ESO Imaging Survey (EIS). In this region deep photometry and 95% complete object catalogues in the I band are available down to $I\sim 22.5$. These data allowed us to identify 219 of the 386 ATESP sources present in the region. This corresponds to an identification rate of ~$57\%$. For a magnitude-limited sample of 70 optically identified sources with $I< 19.0$ we have obtained complete and good quality spectroscopic data at the ESO 3.6 m telescope at La Silla. This data allowed us to determine redshift measurements and reliable spectroscopic classification for all sources (except one). From the analysis of the spectroscopic sample we find that the composition of the faint radio source population abruptly changes going from mJy to sub-mJy fluxes: the early-type galaxies largely dominate the mJy population ($60\%$), while star forming processes become important in the sub-mJy regime. Starburst and post-starburst galaxies go from 13% at $S\geq 1$ mJy to 39% at $S< 1$ mJy. Nevertheless, at sub-mJy fluxes, early-ype galaxies still constitute a significant fraction (25% ) of the whole population. Furthermore we show that, due to the distribution of radio-to-optical ratios, sub-mJy samples with fainter spectroscopic follow-ups should be increasingly sensitive to the population of early-type galaxies, while a larger fraction of star-forming galaxies would be expected in μJy samples. We compare our results with others obtained from studies of sub-mJy samples and we show how the existing discrepancies can be explained in terms of selection effects