A: The short answer is it prevents a systematic bias in the final output. The cause of this bias was traced to an error in floating point representation of the profile image. When a perfect (no noise, straight trace, no bad pixels, ...) image was tested, there was symetric error at machine level precision in the profile image created by looping procvect.pro. However, as this image was scaled to make sure each wavelength summed to 1, the error became off-center. This error was componded in later steps and eventually showed up in the final spectrum. There was no way to scale a floating point array without this occuring, so instead everything is done in double precision.
Q: Help! The program keeps masking all of my data points. What's wrong?
A: The probable answer lies in the variance image creation. The routine uses and iterative sigma-cliping calculation using: Residual = (Data - Background - Spectrum*Profile)^2 / Variance If the variance estimate is wrong, the procvect routine can end up clipping way too many or too few pixels easily. The background fiting and the first pass of the profile fitting use the inital variance estimates that are passed in. However, each subsequent pass in profile fitting, and in the optimal extraction iteration, the variance image used is recalculated based on: Variance = Readnoise^2 + Abs(Spectrum*Profile + Background) / EPADU Therefore, if your estimates for read noise or gain are off, each revisted estimate will be inaccurate. The sigma-clipping routine is sensitve to both rdnoise and gain. If background pixels are wrongly masked, the readnoise is proabably off. If the high-intensity pixels are masked, then it is probably the gain.
Another problem could stem from inaccurate estimation of the fit. If the profile image is changing more rapidly than can be modeled by a third degree polynomial, the degree must be increased, or a different fit used. Similarly, if skylines are curved too much (more than three pixels spell distator for polynomial fitting) the background fitting routine won't be able to model the background.
Q: Why isn't the answer to my question listed here?
A: Because you haven't sent it to us. Email to jh@oobleck.astro.cornell.edu