Image Subtraction

from: Applications and Algorithms

Thoughts by ACB from an email I have composed but not yet sent…

• Functional form; damping their wings : Currently we are using *the* most general mechanism for generating a kernel. This can (and does) lead to noise in their wings. It has been suggested that we use some sort of functional form in the outer portions of the kernels to damp the power out there. A problem with this is for mis-registrations, where the kernel needs to shift the stars by a significant number of pixels, meaning the power in the kernel is off-center. The damping wings would inhibit this functionality.

• Compensating for mis-registration : Its often quoted that "kernels can compensate for misregistrations". I believe the power in this functionality is perhaps overestimated by some. In the case where you have a functional form for the kernel (e.g. a traditional Alard-Lupton basis, the sum of 3 gaussians each of which is multiplied by a spatial polynomial) most of your power is in the center of the kernel by design. If you need off-center power - the case when the kernel compensates for misregistrations - you start to depend on the high-order modifications to these symmetric basis functions. So if you need to shift things at a 45 degree angle, you need to depend primarily on the (I think) third order of the Gaussian whose width is closest to the amplitude of the shift. This is not ideal, and means many of your basis functions have zero weight. If you need to shift things at a 37.3 degree angle, well you can't so well. The delta function basis functions give you this generality, although the pixelization of the kernel hinders its usefulness somewhat.

• Regularization : TBD

• DiffIm? Quality : We can use the number of "extended" sources as a metric of how well it went. Also number of cosmic rays.
  • NOTE : We should try and propagate the knowledge of any CR discovered during diffim to the deep detection pipeline, e.g. CR on top of galaxies.