An extended re-implementation of Gavazzi et al's RingFinder galaxy-scale strong gravitational lens detection robot. Subtract red light from blue, analyze residuals, classify.
MIT License
1
stars
3
forks
source link
Improve sensitivity by fitting SEDs during subtraction #6
In the basic implementation we subtract a scaled z-band image from each of the g, r, and i band images. Can we make better use of these data to somehow identify an LRG-colored image, that represents the light from the lens galaxy, separate from the rest of the features? If we assume the other features are also all of one SED (or at least dominated by one SED) then we could project out these as well.
Two model images, each associated with an SED, would lead to a linear optimization (maximizing the pixel log likelihood) for the model images, and a search for up to 6 non-linear parameters (the 3 colors in each 4-band SED). One possibility could be to sample this space simply, drawing colors from a prior PDF defined by some large sample of exemplar LRGs and sources, and evaluating the maximum linear log likelihood at each point. (Actually we should compute the evidence for each pair of prior sample SEDs, but the maximum likelihood is probably a reasonable approximation).
In the basic implementation we subtract a scaled z-band image from each of the g, r, and i band images. Can we make better use of these data to somehow identify an LRG-colored image, that represents the light from the lens galaxy, separate from the rest of the features? If we assume the other features are also all of one SED (or at least dominated by one SED) then we could project out these as well.
Two model images, each associated with an SED, would lead to a linear optimization (maximizing the pixel log likelihood) for the model images, and a search for up to 6 non-linear parameters (the 3 colors in each 4-band SED). One possibility could be to sample this space simply, drawing colors from a prior PDF defined by some large sample of exemplar LRGs and sources, and evaluating the maximum linear log likelihood at each point. (Actually we should compute the evidence for each pair of prior sample SEDs, but the maximum likelihood is probably a reasonable approximation).