[daip] msclean

[Eric Greisen]

Volker Heesen wrote:
> Dear Eric,
> 
> I tried recently msclean. It seems to work well, but I would like to 
> understand a little better how it works. Particularly I tried to 
> convolve a noise image (line free HI map) to the resolution of the 
> second map. The first map has a resolution of 6x8 arcsec and the second 
> map has 20x23 arcsec as stated by imhead. Contrary, I chose for the 
> scales  0 and 15 arcsec, respectively. Now I wonder why the second map 
> has a resolution not identical to the scale I specified. Also I tried to 
> convolve the first map to the resolution of the second map. The noise 
> level of the convolved map is lower than I expect from the ratio of the 
> beam sizes. Thus, there seems to be a scaling factor between the two. 
> Summarizing my questions are:
> 
> 1) What is the resolution of the 2nd map? Is it the resolution stated in 
> imheader? Or is it the resolution of the 2nd beam? This beam has a FWHM 
> of 19 arcsec. That doesn't fit neither to the 15 arcsec (of IMAGR) nor 
> to the 20x23 arcsec of imhead. I also wondered whether the second beam 
> is the beam one has to convolve with in order to get from the first map 
> to the second map.
> 2) How do you get the resolution of the second map? Do you convolve the 
> first map with the scale that is specified in IMAGR?
> 3) What is the scaling of the 2nd map? If I convolve the first map to 
> the resolution of the second map, I get a higher noise level than the 
> 2nd map (a factor of 2 roughly). For this I have used the ratio of the 
> beam sizes stated in imhead as the scaling factor.

The resolution of the 2nd scale is a result of applying a taper to the 
data and tends to be the modulus of the point resolution and the 
requested resolution.  Your 20x23 seems rather larger than that.  Of 
course, there can be a problem with the shape of the dirty beams not 
being a Gaussian - esp the point source one since the Gaussian taper 
moves the others to be more like Gaussians.   The apparent point spread 
function beam of the 2nd scale was fit with a 20x23 Gaussian and so that 
is the apparent resolution in that scale.  Note that the beam that is 
left over from that scale is not the point spread function for that 
scale but the "Gaussian-spread" function of the model size - about the 
convolution of the point-spread function with a Gaussian.  It is likely 
to have a width in its header that is larger still since it is not easy 
to stop the fitting routine from running and adjusting the header.

The scaling of the 2nd image is produced by finding the peak of the 
point-spread function for that taper and dividing by it.  Thus it should 
be Jy/beam for the point-beam size present in that image.

The 2nd image that is left behind is a residual image - so if any CCs 
were taken at that resolution, then its noise is not correct.  Only the 
first image has the components restored to the residual and they are 
restored scaled to the Jy/beam of the beam of the 1st image.

CONVL normally does not change the noise level for a simple spatial 
smoothing since the noise is reduced per the smaller beam by averaging 
but is then summed over a larger beam more or less exactly compensating.
The improvement in noise in smoothed HI cubes comes only from the 
frequency smoothing.

When I just ran a test I noticed that the fit widths no longer appear in 
the beam headers even for the 0-width case.  The fit sizes come from the 
point-spread function beam (the first of 2 computed for all non-zero 
scales) and are put in the Clean image headers.

Eric Greisen