We have given the results of a survey of the region and in the OH 1612.231 MHz maser line. The survey is complete for sources brighter than 500 mJy and 80% complete for sources brighter than 300 mJy. The absolute flux density limit is 160 mJy. We have found 307 compact OH-maser sources, 145 of which are new detections. The sources are mainly OH/IR stars, with a few related sources, like PPNe and supergiants. The sources have positions accurate to , velocities accurate to 1 and flux densities accurate to 5%. For 201 sources, an associated IRAS point source is found. A special CLEANing method was developed to search a very large data set for spectral-line point sources.
The authors thank ATNF staff for valuable discussions about observations and reduction, especially Ron Ekers, Wim Brouw and Jim Caswell. MS thanks the NFRA for financial support, the ATNF for infinite hospitality, Richard Arnold for all his useful ideas and Laurens Smulders for many a statistical eye-opener. ML is supported by an ESA external fellowship. This research has made use of the Simbad database, operated at CDS, Strasbourg, France. All FFT computing in the reduction proces was done on the Cray-C98 of the National Facility for Supercomputing in Amsterdam.
Appendix A: Reduction method
The reduction of the acquired data calls for a special strategy for a number of reasons. What we describe here is the process of searching the data for sources after the calibration described in Sect. 3.
truecm In creating this strategy the following facts have to be dealt with :
Corresponding requirements were made for a possible reduction strategy:
There are three major assumptions we make to justify our strategy.
Firstly, we assume the sources we are looking for are point sources. Secondly, we assume the whole region of the survey is covered entirely within the width of the primary beam out to the offset where the primary-beam response in the main lobe equals the response of the first primary-beam sidelobe (see the mark max in Fig. 2 (click here)a). It then follows that if we see the same spectrum (save a factor in flux density) at different positions in different fields, the brightest represents the real star. Thirdly, we assume the flux density of most of the stars we expect to find is too low for detection in a vector-averaged visibility spectrum with zero phase offset, so we do have to Fourier transform the visibility data and search in the image domain.
These three assumptions are all justified for the data we are presently discussing; there is only one exception to the first assumption in that source #153 (Table 1) is probably a nearby object and slightly resolved.
The stategy developed is as follows:
We Fourier transform one channel at a time, keeping it in memory. This image is then searched for its one highest peak value. If the shape of the peak is approximately gaussian (i.e. a point source) and above a certain detection level, we write its velocity and fitted spatial position to an output file, otherwise we discard it. The image is then discarded. This process is repeated for all channels and for a number of neighbouring fields. Each field has a separate output file.
2) Model correlation
The output files of different fields are then individually searched for detections at the same position in a number of neighbouring channels. These detections are then marked as models. If there are models in two (or more) different fields at the same velocity within a certain distance of each other, the brightest is assigned real and the others (assumed to be sidelobes) are shifted in position to the real position.
The consolidated point source models for each field are subtracted from the visibility data and the process starts again at step (1), now with a lower detection level and the new visibility data.
The whole cycle from step (1) to (3) is called a pass. This way we build the stellar spectra in subsequent levels and find fainter sources in channels where bright sources were found in the first passes.
Fig. A1. Schematic representation of the reduction method. Every main cycle through MPFND, the model correlation and UVSUB is called one pass. At the end of the first pass (P2) UVLIN is applied. The input parameters at (1) and (2) can be adapted according to spectral features of the sources looked for and several features of the visibility data (see Sect. A4)
The main routine of the method is MPFND which does the imaging and the searching. It is a derivative of the existing MIRIAD imaging routine INVERT (version 18 nov. 94). This is the most time-consuming part of the reduction and this routine as well as UVSUB and UVLIN was run on a Cray-C98. The less time-consuming part of correlating the model files was done on a local workstation since it requires more interaction and does not act directly on the visibility data.
When imaging we used a natural weighting scheme of the visibilities, without any tapering, to get the maximum SNR. This does increase the sidelobe levels, but the method deals with that. The visibility data from the shortest baseline were always excluded from the imaging because of RFI.
The input at label (1) in Fig. A1 (click here) consists of the cell size, image size, detection level and the width of the boundary region of each image to exclude from the searching.
The detection levels in subsequent passes were decreased by starting at ( is the noise level as theoretically calculated from the visibilities). In the last pass, for homogeneity, the detection level was fixed to 120 mJy for all fields. This level equals 3 to 4 for 90% of the fields (see Sect. 5.1). The levels were chosen to ensure firstly that all sources bright enough to influence polynomial fitting (see Sect. A5) are subtracted in the first and second passes and secondly that the flux densities of the sources found in a pass are comparable to the detection level of that pass. It is important to find stars in as early a pass as possible, because at lower levels there are more and more noise detections and the process of correlation (see below) becomes more and more time-consuming. This is the reason for not making six passes all with a detection level of 120 mJy, which in principle would yield the same result.
To increase the speed of the transforms in the first two passes the images were made with a cell size of 5 10. This causes some loss in SNR, but, since we only search for very bright sources in those passes, the SNR is still higher than 20. In the following passes cell sizes of 5 were used. The size of the images was always 42 to ensure ample overlap between them. (This corresponds to an image size of and cells respectively.) The cell size was not optimized for all individual pointings but we chose to make the procedure as uniform as possible. To avoid the detection of Fourier transform errors, that are strongest at the boundaries of the images, the outer five cells were excluded from searching in the first two passes and in later passes the outer 10 cells, because at lower detection levels the imaging errors become relatively more important. After finding a peak in an image an area of pixels around it was fitted with a two-dimensional parabola (see Sect. 5.2). If this fit indicates a position that is more than one cell size away from the peak pixel this indicates a highly non-gaussian shape of the peak under consideration, since the peak of a gaussian is roughly parabolic. This can be safely used as a criterion for interference or other unwanted detections and such peaks were discarded.
2) Model correlation
The inputs at (2) in Fig. A1 (click here) are relatively complicated to determine. In all passes, peaks at different velocities are identified as coming from the same source when positional coincidence is smaller than 0.35 cell size. This value was found empirically to be smaller than the typical difference in position between noise or interference peaks, correlated in neighbouring channels and bigger than the scatter expected in the positions of a source found in different channels (Sect. 5.2). Depending on the detection level and the strength of the highest peak found in the pass the distance out to which sidelobes can be expected has to be set. For the first pass this is as much as in declination. After subtracting the brightest sources the distance quickly decreases to about , so that sidelobes are found only in directly neighbouring fields. All 539 fields of the Bulge region were processed through each pass simultaneously, so that all models could be correlated optimally. Only detections in the outermost fields surrounding the whole region of 539 fields had to be checked for sidelobes from unobserved regions, since there the sidelobe detection assumption of complete coverage broke down. This was done by verifying the presence of a point source pattern by eye in the image plane at the detected channels.
Neighbouring channels are correlated by 16% after Hanning smoothing, which is enough to smear out a large number of noise peaks over two channels. Therefore, we required that detections be present at the same position in at least three channels (either directly neighbouring or within a velocity range reasonably expected from outflow velocities) for a detection to be trusted. At lower detection levels statistical considerations put a lower limit on the number of channels for a detection. There can be false detections, that is a number of detections at the same position that meet the requirements mentioned above but are purely a statistical coincidence. At the 5 level the expectation value of the total number of false detections in the whole data set is far less than one when the signal is demanded to be detected in three neighbouring channels at the same position. For the 4 level the expected number is close to one. If we allow the third detected channel to be within a velocity range of 100 from the other two neighbouring channels, the expected number is as much as one hundred. Such false detections could be identified as a double-peaked source. Therefore, at detection levels lower than 5, at least four detections, either in one peak over four neighbouring channels or in two separate peaks, are demanded at the same position. This gives an expected number of false detections of for detection levels 4 up to a few for the lowest detection level of 3.
The flux densities of all the point source models were determined with UVFLUX from the actual data at the velocities and the fitted (and for sidelobes shifted) positions found in the model correlation. They were then subtracted from the visibilities with UVSUB, via a direct Fourier transform.
A.5. Further implementations
To remove RFI from the visibilities, an 11-order polynomial fit (UVLIN, Sault 1994) to the real and imaginary parts of the visibilities was subtracted from the data between the second and third pass. By doing it at this stage, fitting bright OH sources was avoided, because all sources brighter than 1 Jy had been subtracted. Fitting spectra of brighter sources could cause serious problems over the whole spectral band and especially at the edges, because too much intensity would be put in the higher order terms of the polynomial.
It should be noted that it is possible for a source to be wrongly identified as a sidelobe of another star. This is not a problem, since the model that is subtracted from the visibility data is correct if the correct flux density level is determined at the position of that other star. If a detection reoccurs after identifying it as a sidelobe and subtracting a corresponding point source model in the previous pass, then obviously the identification as a sidelobe was wrong and it should now be modelled as a real star. Therefore, we never subtracted the same model from the data twice.
Finally, we checked the spectra by eye, extracting them from the original data with UVSPEC (shown in Fig. A2 (click here)). This is not necessary in principle with data that contain only sources and random noise. However, in our data, RFI introduces too much correlated signal that masquerades as sources. After the visual inspection, the limiting flux density, corrected for primary beam attenuation, is found to be 160 mJy (see Figs. 4 (click here)b, f) or, in other words, 4.
Essentially, our method is a modification of the traditional method of CLEAN (Högbom 1974), in the sense that models are subtracted from the data at successive flux density levels. It is, however, stripped of every performance that is superfluous for processing the present data. Since in the present observations the sky does not feature bright extended emission and we want to find only point sources, in principle one cycle is sufficient to find the model. This model is then subtracted with loop gain 1 from the ungridded visibilities (similar to the Cotton-Schwab CLEAN algorithm, Schwab 1984). Therefore RFI and boundary imaging errors do not influence MPFND as much as they do standard implementations of CLEANing. MPFND is a fast method because it does not perform any convolution in the image plane, which is not necessary when one assumes that one iteration and loop gain 1 are best to find the models. Models are subtracted directly from the complex visibilities. Together with the fact that no disk I/O is needed to store cubes, this reduces the time needed for finding and subtracting models by a factor of four. In various fields we tested that the results of MPFND are exactly the same as those of other CLEAN methods with appropriate input values.
Table 1. Compact OH-maser sources in the galactic Bulge region.
The columns of Table 1 contain the following information:
(The table is available in electronic form via anonymous ftp (ftp 126.96.36.199) or through the World Wide Web at http://cdsweb.u-strasbg.fr/Abstract.html).
Table 2. References for previous OH maser detections (Table 1, Col. 15)
01 te Lintel Hekkert et al. 1991
02 te Lintel Hekkert et al. 1989
03 Becker et al. 1992
04 Lindqvist et al. 1992
05 Blommaert et al. 1994
06 Bowers & Knapp 1989
07 David et al. 1993
08 van Langevelde et al. 1992
09 Braz & Epchtein 1983
Fig.A2. On the following pages the spectra for all the sources in Table 1 are shown. They are displayed with 50 (except #008, #200 that have outflow velocities higher than 50 ) on either side of the stellar velocity (if possible; for sources at the edge of the observed velocity range the unobserved part of the spectra is left blank, e.g. #179). The channel width used in the spectra is 1.45 \ which is equal to the velocity resolution. Along the upper border of each spectrum the entry number of the source in Table 1 is given, along with its usual OH name, its identification as a D, S or I source (see Table 1, Col. 3) and the number of the reference in the case of previously known sources. The spectra are extracted from the original visibilities by the MIRIAD routine UVSPEC and the baselines are fitted with polynomial of order up to three. This means that interference still shows in the spectra. Also sidelobes from neighbouring stars are present in some spectra. If there is possible ambiguity in the interpretation of the spectra the detected peaks are marked with an asterisk; other peaks in such spectra are not necessarily connected to the same source. The spectra were left relatively unprocessed in order to give a fair view of the data quality. (This figure is included in the electronic version of the paper, available at . The individual spectra are available upon request in more workable ascii-format (e-mail: email@example.com)).
Fig. A2. The spectra for all the sources in Table 1.
Up: The ATCA/VLA OH