[daip] Re: spectral clean?

[Eric Greisen]

Lawrence Rudnick writes:
 > Eric - interesting question for you...  I have Spitzer spectra where I'm 
 > interested in measuring doppler shifts of two lines that are closely 
 > spaced, and each can show either one or two doppler components (or more, 
 > I'd never be able to tell - see attached jpg).   I'm trying to figure 
 > out an easy and efficient way to find between 1-4 doppler velocities (at 
 > each pixel!).  The spectral response is pretty much like a hanning 
 > function.   
 > 
 > Thinking about this manually, I would simply do s slice and fit a 
 > gaussian to a single (unresolved) peak.  Where two well-separated peaks 
 > present, two gaussians, etc.   However, can't do this for 10^4 pixels. 

        XGAUS was intended to do this but it is not really the
program I would like it to be.  You might give it a try.  There is
also NNLSQ but I have no idea how that works.  It is meant to be more
like a Cleaning.  XGAUS lacks the ability to assign the components as
one would like them to be and requires too much interaction in all
likelihood - but it does try to make images of the results.  SLFIT
would be a nightmare to put back into images.

 > 
 > So then I thought of doing something analogous to cleaning.    I would 
 > like the end product at each position to be no more than four spectral 
 > channels with flux in them (plus noise, of course),  which I would then 
 > consider the four velocity components  (it's actually 2 components for 
 > each of two nearby lines).  So if a velocity fell midway between two 
 > channels, I would not want half the flux in one, and half in the
other.  

      Trouble is - that that is the correct fit and having more pixels
(with therefore a wider "beam" measured in pixels) would still suffer
from this problem.  Gaussian separation of 2 blended components does
not work all that well - I was comparing my A-array P and 4 band data
to NVSS and frequently found that their careful fits were not all that
good at separating components which were blended at their resolution.

What you really need is a properly written XGAUS which would let you
fits comopnents 1,3,4 at a pixel (for example) leaving 2 at zero and
then at another pixel do some other selection.  Also some editing and
re-arranging capability in a table of fit results before actually
making the images.  Unfortunately, AIPS' future has not been
guaranteed at the level that would encourage that level of work on a
task that does not seem to be in much demand.

 > A way around this would be to interpolate the spectra onto a 3-5x larger 
 > grid and hope the cleaning would put most of the flux into the right pixel.
 > 
 > Do you have any other ideas how to proceed, or comments on the above 
 > strange procedure?

Good luck with this.  Let me know if you come up with a good
solution.

Eric Greisen