3 Data reduction

Data were reduced with the reduction package Miriad (Sault et al. 1995); five- and six-letter acronyms in capitals throughout this paper will indicate the Miriad programs we used. To make the data quality uniform across the whole survey area, only baselines with lengths between 5 k$\lambda$ and 55 k$\lambda$ were used (cf. Table 1). The primary and secondary calibrators were edited first with a custom-made RFI flagger WSRFI (see Sect. 3.1) and further by hand. After calibration (MFCAL), the data from all pointings were edited only with WSRFI.

The searching was performed largely as described in Paper I, with MPFND, a custom-made derivative of INVERT, aimed at searching large amounts of modest-quality data for point sources. This routine Fourier-transforms spectral channels one by one, retaining only the position and flux density of the brightest pixel for each image. Subsequently, the peaks are correlated in the spectral direction, looking for statistically significant (OH/IR-type) spectral features. If one has been found, an appropriate point-source model is subtracted from the visibilities and the same routine is performed at a lower detection level. All details are given in Appendix A of Paper I.

The image size was always 1665², with square cells of 1 $.\!\!^{\prime\prime}$5, searching all but 10 cells along each border. Four passes were performed, the last one well below 3$\times$ the theoretical noise (as given by MPFND).

Figure 3: a) and b) These histograms display the net integration times and residual noises for all 965 fields used in the searching process (see Sect. 2). For each pointing 5 min integration time was scheduled

A few changes were made to accommodate specific properties of VLA data. To save time, with the VLA cubes three times larger than the ATCA cubes due to higher resolution, an extra first pass through the data was performed in the visibility domain. Spectral peaks were identified in the scalar-averaged spectra and then only the peak channel was Fourier transformed to find the source position.

With the modest spectral resolution of the VLA correlator, a large fraction of the sources induced spectral aliassing; the narrowest, mostly brightest, sources creating responses in all 255 channels. Obviously, these reponses may be negative. To account for this, we adapted the subroutine UVSUB to allow for negative point source models. Input models for the new UVSUB were "log'' output files from UVSPEC, containing the real part of the visibilities, offset to the source position, for the whole spectral range (215 channels in calibrated data). Hence, in all passes, once a source position was established, either directly as a real source or after shifting to the real position in the case of detecting a sidelobe (see Paper I), a full spectral point-source model was subtracted from the visibility data. This clearly took out the random noise at the source positions as well, but the data on the whole were not significantly influenced.

The visibility-based point-source subtraction was not as effective for the VLA data as for the ATCA, due to the non-negligible third dimension in the antenna-position coordinates (u, v, w) for the VLA. Sources at higher offsets were occasionally redetected in later passes. This was corrected for in the post-searching cross-identification process.

3.1 Radio-frequency-interference excision

Three kinds of RFI corrupted the data during most of the observing dates. One was the usual broad-band Glonass RFI ("G" in Table 1), often accompanied by a single-channel spike ("S", sub-/superscripts indicating channel of spike). The third was interference from a nearby military base ("W"), depending in strength on u-v direction rather than baseline length, which saturated the correlator to give non-random noise characteristics (see Fig. 2).

The routine UVLIN, used for the ATCA data (Papers I, II), had no positive effect when applied to these data with the required high order of polynomial, probably again due to the non-negligible third dimension of the VLA array. However, with a custom-made visibility-flagging algorithm WSRFI, written for use within Miriad, we managed to excise the worst of all three types of RFI. In Fig. 2, we show the spectrum of one of the primary calibrators before and after running WSRFI on the data.

With the combined losses due to antenna downtime, data flagging to delete interference and retaining only baselines between 5 k$\lambda$ and 55 k$\lambda$, the resulting net "visibility time" (all visibilities used divided by the number of baselines used, times 30 s) is of the order of 1-3 min (Fig. 3).