Frequently Asked Questions

Users should also be aware that no attempt was made in NEWSIPS to eliminate cosmic rays from the high-dispersion images, and the attempts made for low dispersion probably err on the conservative side. The reason for this was a concern that a provision for cosmic rays would lead NEWSIPS inadvertently to remove actual emission lines. Users should especially be cautious about the influence of oblique, diffuse cosmic rays, which can subtlely affect the background surface by 150 pixels or more from the centroid of the "hit" region and be responsible for less than optimal fits to the background surface. We suggest that the SIHI or SILO images be first visually screened before the extracted spectra are analyzed.

Since "clean" continuum is not always easy to find, let's consider an empirical procedure which works well even for a spectrum containing features. As long as the flux errors are primarily gaussian- distributed, we can make use of the fact that randomly drawn samples will differ from one another, on average, by exactly one standard deviation. In the computing language IDL, an estimate of the rms may be computed conveniently by a few steps. Defining "f" as a spectral flux array taken from pixels near the blaze maximum (so the noise will be approximately uniform along the spectrum) and containing only zero-valued quality flags, we compute the mean point to point fluctuation of two pixels separated by a short distance n:

The distribution of noise fluctuations among these pixels can be obtained from the computation:

rmsdist = abs( f - shift( f , n) )  .

The median of this distribution is computed by sorting the distribution rmsdist and finding the middle element of the sorted rmsdist array, which by definition is half the number of elements in the distribution. Thus:

ntot = n_elements(rmsdist),     so   ntot2 , = ntot/2  ,

Now we compute the sorted index distribution and take the value of rmsdist we need:

isort = sort(rmsdist)  ,

medrms = rmsdist( isort(ntot2) )  . medrms is the our noise estimate using the median average.

The mean of the distribution can likewise easily be determined from:

meanrms = total( rmsdist )/ntot  .

This value will be biased on the high side if there are outlying fluctuations in the spectrum. In practice the pixel separation distance n should be at least a spectral resolution element (3 or more).