iraf.net

>From P.Werner@bristol.ac.uk Mon Jun 21 05:35:47 1999
Date: Mon, 21 Jun 1999 13:35:37 +0100 (BST)
From: P N Werner <P.Werner@bristol.ac.uk>
To: Frank Valdes <valdes@noao.edu>
Subject: re: "onedspec" questionDear Dr Valdes
Thank you very much for your email regarding sky subtraction.I wonder if I could ask you another question that has been bugging me,
regarding the 'd' deblending function in SPLOT. I would like to get errors on the fitted central wavelengths, so that I
can combine the velocities of the various emission lines in my spectra by
calculating a weighted mean based on the errors on each individual line. I only get the RMS of the fit, and the gaussian FWHM of each line, neither
of which really correspond to the error on the wavelength, and using them
gives me enormous errors on the final velocity of the combined emission
lines. I would appreciate it if you could let me know if there is a way of
getting SLOT to work out an estimated error on the fitted gaussian
positions.
Regards,
Pierre Werner
-------------------------------
Pierre Nicolas Werner
Astrophysics Group
Department of Physics
University of Bristol
Tyndall Avenue, Bristol BS8 1TLTel: +44 (0)117 928 7561
Fax: +44 (0)117 925 5624p.werner@bristol.ac.uk
-------------------------------
>From valdes Mon Jun 21 08:45:44 1999
Date: Mon, 21 Jun 99 08:45:43 MST
From: valdes (Frank Valdes)
To: P.Werner@bristol.ac.uk
Subject: re: "onedspec" questionHi Pierre,SPLOT, in the recent releases, does allow calculation of errors if
you can give some numbers that "characterize" the noise properties
of the spectrum. You can give a constant sigma value in the "sigma0"
parameter, a Poisson gain scaling (sigma = sqrt(gain * pixelvalue)) in
the "invgain" parameter, or some combination of the two.The way the error calculation works is that the actual spectrum is
taken as a "truth" spectrum. Then a number of realizations of the
spectrum with random noise added using the noise model are computed and
the measurements done with exactly the same algorithm. The "nerrsample"
parameter sets the number of trials and should be 50-100 for reasonable
accuracy. The scatter in each measured value, such as the wavelength,
then is a very good indicator of the uncertainty even in such complex
non-linear algorithms as the deblending. Note the data must be positive
everywhere for the error estimation to work.This provides very good uncertainties for the derived parameters for
each line. There is also the errors between the velocities from various
lines. This is sometimes greater than the error in a measurement of
an individual line center. This is caused by the line center being
affected by other lines. You might want to look at RVIDLINES in the
RV package for another tool for measuring velocities from multiple lines.
This computes uncertainties in the velocity from the scatter ni the
velocities from individual lines but it does not include the uncertainties
in the centroid measurements of individual lines.Let me know if you have questions.Cheers,
Frank ValdesDate: Tue, 22 Jun 1999 14:21:16 +0100 (BST)
From: P N Werner <P.Werner@bristol.ac.uk>
To: Frank Valdes <valdes@noao.edu>
Cc: P N Werner <p.werner@bristol.ac.uk>
Subject: re: "onedspec" questionDear Dr ValdesMany thanks indeed for your reply. I have tried the SPLOT error calculation,
and it gives very reasonable answers. The only problem I now have is
justifying the values of 'sigma0' and 'invgain' which I use. I tried
using some of the IRAF tasks for determining noise characteristics of
spectra (like 'noisemodel' in the 'stsdas.hst_calib.wfpc' package). I
don't appear to be able to successfully use these to find out what values
of 'sigma0' and 'invgain' I should use.For example, I have found that by using 'sigma0=3' and 'invgain=1', I get
very reasonable errors (of the order of 1 Angstrom or 40 km/s, but I
would have trouble justifying these values, and the errors I get seem a
bit arbitrary, as I can edit the parameters in such a way that I get the
errors I would expect.I would appreciate any further ideas you could give me.King regards,
Pierre Werner
>From valdes Fri Jun 25 22:07:46 1999
Date: Fri, 25 Jun 99 22:07:44 MST
From: valdes (Frank Valdes)
To: P.Werner@bristol.ac.uk
Subject: re: "onedspec" questionHi Pierre,Yes you have hit on a problem. Ideally the noise estimates for the
data, on a pixel-by-pixel basis, would be propagated down through all
the CCD reduction steps and the spectral extraction and calibration.
This is not currently down so you would need to do this by hand, a
really difficult problem and signal level dependent.Short of that you have to determine it from the data. One way is using
the task you noted. However that task requires alot of data with
regions of nearly constant values. Thus it is primarily meant for
direct imaging data. The only way I could think off is if you actually
took multiple observations. It think this is what you did. If you
scale them all to the same continuum level and subtract one from the
rest and use a region of continuum where there are no lines then the
RMS of the data would be sigma0. You could ignore the gain dependence
in photon counts. If the data is flux calibrated this is all you can
do since the relation to photon statistics is largely gone. I could
ramble on but I can't really give you any real answer. It is up to you
to justify what you think is the noise of a pixel in the spectrum
possibly with a dependence on signal strength (which is what the
invgain represents).Sorry I can't be of more help,
Frank Valdes