Imaging

Now you can image the data. HORUS and UVMAP both Fourier transform the visibilities and create a dirty image and a dirty beam. Alternatively, you could use MX which images and optionally deconvolves with CLEAN. MX can image up to 16 fields in the primary beam simultaneously (hence its name, because the MX missile makes [or used to] multiple warheads) whereas UVMAP and HORUS can only do one at a time. I recommend that you do not deconvolve (CLEAN or MEM) blindly. Until you become familiar with the source that you are imaging, always make a dirty image first and inspect it for emission. You can tell MX to make a dirty image without CLEANing, and then restart it for the CLEAN step later (see § 16). MX\ is very useful for discrete and widely separated emission, because you can use the multi-field capability to make small images centred on the sources of emission, rather than having to make one huge image. MX\ also has advantages in the CLEAN step (see § 16). Note that there is no MX like program in MIRIAD.

The inputs for HORUS, UVMAP, and MX (for just the imaging stage) contain many of the same adverbs, so I will just show the UVMAP adverbs, and any additional ones for the other tasks that might cause confusion. Note of course, that if you use HORUS on a multi-source file, you should turn on the relevant calibration adverbs. You must apply the calibration just as in SPLIT (§ 12), so I will not discuss the calibration adverbs again here. For single-source files, the calibration adverbs should be turned off (as there are no calibration tables to apply anymore) For MX, I will defer discussion of the CLEAN step adverbs to the deconvolution section (except to explain how to turn the CLEAN off). Finally, note that MX also creates what is called a visibility work file. You will need this in the CLEAN step later on, so don't delete it. Now, on with the nitty gritty.

• Important adverbs in the imaging tasks are imsize and cellsize. The former is the size of the image in pixels and should be a power of 2. The latter is the size of each pixel in arcseconds. Both specify an x and a y dimension. They needn't be the same (e.g. if your beam is elongated significantly in the north-south direction). The product of imsize and cellsize should be large enough to encompass the emission. If you are using UVMAP or HORUS, this emission must fit into the inner quarter of the image for correct deconvolution. If you are using MX then it can fill virtually the entire image and still be succesfully deconvolved. Choose the cellsize by allowing about 3 pixels across the FWHM of the main lobe of the dirty synthesised beam. If you under-sample the beam, you will induce some unpleasant effects in your images, especially if you plan to deconvolve. If you do not plan to deconvolve, then you could probably (but only if necessary) set the pixel size so that there are 2 pixels across the FWHM of the synthesized beam without sacrificing too much.

  Consider also the effect of making an image from too few visibilities. You are not doing anything very meaningful if you make an image with more pixels than you have visibilities to constrain the pixel values. This can easily be the case with ATCA snapshot data and large images.

  Since , the size of the gridded (u,v) plane is  wavelengths. This must be big enough to encompass your longest spacings. If you find the imaging programs tell you that they have not used all of the visibilities (they may say ``Note: using only 1 of 345430 visibilities'' or the like), you may have set the cell size to too large a value. Essentially you have not matched your resolution to the resolution expected from the longest spacing. This means you would throw away some of your data.

• As mentioned earlier, it is always a good idea to make a low resolution image of the entire primary beam of new fields. This is done by weighting (tapering) the short spacing data more than the longer spacing data and imaging with a larger cell size. With the ATCA, unless you are imaging a visibility file with multiple configurations, you do not have a lot of spacings to play with, so you should do this rather carefully. If you taper too heavily, you will weight down too much data and there will be insufficient left to make a decent image. The tapering procedure multiplies the visibilities by a Gaussian, and you specify, via uvtaper, its 30% level in in the u and v directions. See the HELP file for more information.

    

• As well as applying a Gaussian taper, you also have control over how the visibilities are weighted through the uvwtfn adverb. You can select uniform or natural weighting. When gridding, N visibilities contribute to a `summed' (really convolved and resampled) visibility in each grid cell. Uniform weighting means that the gridded visibility is normalized by N; it is weighted down by the local density of points. This means that each gridded cell contributes `uniformly' to the Fourier transform. In natural weighting, the normalization by N is not done, so that each gridded visibility contributes its `natural' weight to the Fourier transform. Uniform weighting gives slightly higher resolution and a better sidelobe response. Natural weighting gives better sensitivity and is generally preferred for detection experiments. The default is uniform weighting.
• uvbox controls the width of the box used in normalizing the visibilities for uniform weighting. The default of zero means that the normalization is determined from the number of ungridded points falling in each grid cell. If uvbox=1 then the total number of ungridded points in a box is found. If uvbox=2 the sum is done in a box. These bigger boxes do what is called super-uniform weighting. They are sometimes useful for sparsely sampled data. Note that if the box was as big as the longest spacing, then you would have natural weighting again, as all the visibilities would be normalized by the same number. Experiment with caution here.
• A somewhat capricious adverb is zerosp. This allows you to specify the total flux in the primary beam (should you know it from single-dish observations) and its weight. The total power is known as the `zero spacing', because this is what the visibility function sampled at u=0 and v=0 corresponds to. It is the source of a great many problems in interferometry because we can't measure it. It causes sources that have spatial frequencies larger than that corresponding to the shortest spacing to sit in negative bowls in the image. Any deconvolution algorithm (CLEAN or MEM say) actually tries to extrapolate the visibility function to the zero spacing value. However, if you provide the correct value, the interpolation through the unsampled part of the visibility function (at spacings less than the shortest baseline) is helped considerably. If the visibility function is very steep for the innermost spacings, the presence of a correct zero spacing value can prove a great advantage. But there are limits. If the correct zero spacing is more than about 4-5 times the amplitude of the visibility on the shortest baseline, you will be asking too much of the zero spacing insertion.

  A more difficult problem is that you must also provide a weight for the zero spacing, and the correct weight has proved rather elusive in the past. One recommended recipe (Jacqueline van Gorkom) is that the weight should be the number of gridded cells in the unsampled region of the (u,v) plane interior to the shortest baseline in the data. Trial and error is usually the way to do it. The effect on the dirty map is simply to add an offset (as the Fourier transform of a delta function is a constant). This will almost be indiscernible. The benefit comes in the deconvolution stage, so the loop is rather tedious (see § 16) if you get it wrong a few times. For your first pass, leave zerosp out, but remember it if need be. It's potentially very useful.

• UVMAP, HORUS, and MX can all make a set of channel images for spectral-line work, although the channel selection adverbs are all rather inconsistently used and confusingly defined. However, since I am only discussing imaging of one channel here, there is some hope of getting it right. I have treated the channel adverbs, in particular, in separate adverb boxes.
• If you want to move the centre of the image from the observing phase centre, then set shift for the shift in x and y in arcseconds on the sky. Positive values shift the image centre to the North and East.
• The adverb uvrange allows you to exclude baselines outside of a certain uv-distance (). For example, you may have a short ATCA configuration such as the 750 m configuration as well as the 6 km antenna. This leaves a very large gap between the spacings involving the 6 km antenna and all the rest. To image just the short spacings you could exclude antenna 6 with this adverb. For example, at 6 cm, 750 m corresponds to 12.5 K, so setting uvrange=0,20 would include the short spacings and exlcude all baselines involving antenna 6.
• Make sure you turn on the grid correction with dogridcr.
• An amusing diversion is to view the gridded tracks on the TV. If you would like to see this, set dotv=1.
• xtype and ytype specify the type of convolving function used in the gridding process. The recommended function is spheroidal, for which these adverbs should both be set at 5. This ensures maximum rejection of aliasing.

inname,inclass Select the XY sorted

inseq,indisk visibility data

channel=0 For channel 0 data

channel=18 or select channel

nmaps=1 One image only

outname,outdisk Output image name outseq, outdisk

stokes='I' Image Stokes I

imsize=512,512 Image size (pixels) in x and y (power of 2)

cellsize=1,1 Cell size in arcseconds in x and y

shift=0 Don't shift image centre from phase centre

uvtaper=0 No taper

uvrange=0 Include all baselines

uvwtfn=' ' Uniform or

uvwtfn='na' natural weighting

uvbox=0 Mimimum width for convolution function

dogridcr=1 Must make grid correction Optionally put gridded data image on TV

zerosp=0 Include total flux density of all sources in the field if you know it

xtype=5 Convolution function type for

ytype=5 gridding, use spheroidal function

xparm=0 Additional convolution

yparm=0 parameters

baddisk=0 AIPS disks to not put scratch files on

The channel adverbs are slightly different for HORUS. In addition, you have the potential to average IFs together. For example, you may have set the two IFs to observe closely spaced frequencies in the same band and you would like to grid and image them together. Just set optype='sum' and select the desired IFs with bif and eif. If you only select one IF, then optype is ignored.

bchan=0 Image only one channel. 0

echan=bchan for channel 0 data, or select

chinc=1 channel explicitly

bif=1 Select IFs to grid together

eif=2 and set the optype

optype='sum' to grid IFs together

Because MX can make up to 16 images at once, some of the adverbs are used slightly differently. Note that shift is replaced by rashift and decshift, which allow you to specify a shift for each field, the number of which is given by nfield. You might want to use this option if it is not possible to fit all the emission present in the primary beam into one reasonably sized image (1024 pixels or less), and the emission is reasonably discrete. The channel adverbs are different again too. Once again you can grid IFs together, but this time, you just need to specify them with bif and eif, there is no equivalent to the optype adverb of HORUS.

in2name,in2class Name of visibility work file; leave blank

in2seq,in2disk when imaging only

bif=2 Select desired IF eif=2

bchan=0 Image only one channel. 0

echan=bchan for channel 0 data, or select

chinc=1 channel explicitly

npoints=1 One image only

channel=0 No CLEAN restart

nfield=1 Just one field for beginners

fldsize=0 For CLEAN step only

rashift=0 Don't shift image centre

decshift=0 from phase centre

niter=0 Don't CLEAN yet