Computation

           

1. Let us proceed with the actual use of CLEAN and start with APCLN. The inputs for HBCLN, SDCLN, and the CLEAN stage of MX are very similar to those of APCLN, and I won't show them in any detail. You need to provide APCLN with the dirty image and beam, and it will provide you with a CLEANed image, or if you so desire, a residual image. This residual is often useful to examine, when trying to work out if you need to CLEAN some more; if there is source left in the residual image, you probably aren't finished yet.
   • An important parameter in CLEANing is the CLEAN window. This is the region of the image in which CLEAN looks for CLEAN components. It is very helpful to CLEAN if this window snugly encompasses the real emission in the image. At first, you may not know where this is. However, after some quick and dirty (sic) CLEANing, you could redo it properly with a better idea of the CLEAN window. Essentially, you are giving CLEAN a priori information about the source, which better constrains the deconvolution problem. You can specify up to 10 CLEAN windows with the adverbs nbox and box. Make sure they don't overlap and put them tightly around the real emission. You can set the CLEAN window(s) interactively with the TV and the command TVBOX.
   • A good number for the loop gain is 0.1. If CLEAN stripes occur, then try lowering the loop gain by a factor of about 10, or give SDCLN\ a go.
   • You can tell CLEAN when to stop (some say you should never start) in two ways. You can tell it to quit once the brightest pixel in the residual image reaches some multiple of the r.m.s. noise level (3- or so). Specify this level with the adverb flux. Otherwise, CLEANing will proceed until niter CLEAN iterations have been performed. For small and simple sources, a few hundred iterations are usually sufficient. For complicated and large sources, you can CLEAN forever.
   • There is a knob that protrudes from all Clark CLEAN algorithms, usually with the inspiring name of factor. This controls how often major cycles are performed. The more major cycles, the more thorough is the CLEAN, and the longer it takes. For point sources, this should be about 1. For extended sources, something in the range -1 to -0.5 is good. Use 0 for in-between sources.
   • The parameter minpatch is the minimum half-width of the beam patch used in the minor cycle; the default is the maximum value of 127 (this number is for historic reasons). If your beam has a good sidelobe response, you can set minpatch to something smaller; 51 is the usual number. It is a good idea to examine your beam first before using the Clark CLEAN, so that you can assess whether the small beam-patch algorithm will be likely to cause you trouble or not. If it does, use HBCLN.
   • If bmaj and bmin are zero, the main lobe of the dirty beam will be fitted by a Gaussian. This is the CLEAN (or restoring) beam used to convolve the CLEAN component image before it is added to the residual image. The fitting algorithm is simple and often fails for messy or elongated beams. If so, make a contour plot (with CNTR) and measure the FWHM of the beam, and fill in bmaj, bmin, and bpa explicitly. This is a reasonable way to make your beam circular if you so desire. Don't make the CLEAN beam much smaller than the real FWHM of the main lobe of the dirty beam. Super-resolving is a nasty business, and best left to MEM. Similarly, don't make the restoring beam much bigger than the natural resolution of the image, because the residuals are not convolved by the restoring beam.

     If you want the final image to be the residual image rather than the restored image, make bmaj negative.

     If you really want to see exactly where the CLEAN components come from in the image, make bmaj and bmin very small ( say). The restored image will then be the residuals plus delta functions where the CLEAN components were found. If you are truly a masochist, and want to see an image of the CLEAN components (the true CLEAN model), then subtract the residual image from this image with COMB.

     For beams with a high sidelobe response, the simple algorithm that fits the beam is prone to failing. Sometimes it knows it fails, and tells you, other time the program crashes. For poor beams, you should measure the FWHM of the beam yourself (make a contour plot) and enter the beam parameters into bmaj, bmin and bpa. When the beam fitting algorithm fails, it gives an arbitrary value for the FWHM. The usual signature is that the major and minor axis FWHM are identical.

   • Note that APCLN can only CLEAN one plane of a cube at a time, but that the output image will have the same dimensions as the input cube, no matter which plane you happen to be processing. You select the desired plane with blc(3). You might write an AIPS procedure to loop over all the planes. Make sure that you specify the output file completely in the INPUTS so that APCLN does not generate as many output files as there are planes when CLEANing planes after the first one.
   • The CLEAN components are stored in the CC extension file. For each run of APCLN on the same cube, you should generate a new CC file.
   • If you CLEAN an image, and find that you need to CLEAN some more, APCLN can be restarted from where you left off. Do this by specifying the output file completely (use GETONAME), and setting biter at the number of compnents already CLEANed.

   inname,inclass Dirty image inseq,indisk

   in2name,in2class Dirty beam in2seq,in2disk

   outname,outclass Fully specify when

   outseq,outdisk restarting CLEAN

   blc(3)=10 Clean plane 10, say, of input image

   invers=0 Make CC file next highest version

   gain =0.1 Loop gain

   phat=0 No prussian helmet

   flux=0 Terminate CLEAN at this residual level or

   niter=500 specify total number of CLEAN components

   biter=0 Number already CLEANed if restarting

   bmaj=0 CLEAN beam. If 0, program tries

   bmin=0 to fit main lobe of dirty beam to

   bpa=0 work it out. If bmaj negative the output image is the residual image

   nbox=0 CLEAN window defaults

   box=0 to inner quarter of image

   factor=0 Speedup factor; -1 for extended sources, +1 for point sources

   minpatch=127 Minimum beam size for minor cycles

   dotv=1 Display residual images on TV

   The total CLEANed flux density (i.e., the cumulative sum of the CLEAN components) should eventually settle down to a roughly constant number (you can look at this with the task PRTCC). This indicates that you are just picking up noise, and that there are no side lobes left to remove. If the total CLEANed flux starts to decrease again, this usually indicates that you have been a bit heavy handed, and CLEANed too much. You might also look at the result and see if you can see any side-lobes left over.

   Some discussion of the zero spacing was given in the imaging section (§ 15). It is the deconvolution step that really tells you whether you managed to get the combination of the zero spacing and its weight correct. If you did, the total CLEANed flux should be equal to the zero spacing. If you got it wrong, this will not be the case, and one often sees the CLEAN window region badly offset from the rest of the image. You will know when you need to try again.

2. Note that for MX, you can specify a restart CLEAN component number for each field through bcomp instead of biter. When restarting the MX CLEAN, you must also fill in the correct work file (use 2name\ to get it). This contains the current residual visibility data base (i.e., the original ungridded visibility data with the Fourier transform of the CLEAN components subtracted). Also in MX, you have the choice of a DFT or gridded FFT when Fourier transforming the CLEAN components. This is controlled by the adverb cmethod. The DFT is more accurate but very slow for large CC lists. Leave cmethod blank, and MX will make an informed guess on which is the better choice. Note that you cannot restart HBCLN, you always have to CLEAN from the beginning.
3. If you use SDCLN, there are three additional controls which you will probably need to experiment with. See the EXPLAIN file for details on these. First, all pixels brighter than cutoff times the current residual image peak are considered CLEAN components. Second, stfactor is the loop gain in the SDI phase of the CLEAN (SDCLN switches from a normal Clark CLEAN to an SDI CLEAN). Third, a normal Clark CLEAN is used until ncount pixels are included in an SDI iteration. That is, the residual image must reach some degree of flatness before the SDI CLEAN cuts in. See the EXPLAIN file.