jiwoncpark
commented
3 years ago @mueland this might help with your literature review!
Open jiwoncpark opened 4 years ago
jiwoncpark
commented
3 years ago @mueland this might help with your literature review!
Strong lensing happens when a massive foreground lens galaxy is aligned with a background source. But galaxies and groups in the environment (ENV) of the lens and projected along our line of sight (LOS) produce a second-order, weak lensing effect. This effect is described by a quantity called "external convergence" (kappa_ext, kappa). See McCully et al 2017 and Birrer et al 2017 for some theory.
Previous studies (Greene et al 2013, Rusu et al 2017, Birrer et al 2019) have shown that information available in astronomical survey catalogs for a sky region -- such as the observed objects' positions, brightnesses in each color filter, sizes, shapes, and approximate redshifts (~distances from us) -- can constrain kappa in that region reasonably well. So we know that there's information in the catalogs that tells us about kappa.
Where we can do better is in the modeling. Most studies so far have relied on manually calculating some summary statistics of the catalog, such as the number of galaxies in a region, and mapping it to kappa. As you can probably guess, this is a simplified model that potentially discards valuable information in the catalogs. Machine learning models can process all of the information.
The ultimate goal of this project is model the probability distribution over kappa, given the astronomical catalog entries of objects within ~2 arcminutes of a sky position. Models I'm thinking of are some various conditional density estimators (Dalmasso et al 2019), deep sets (Zaheer et al 2017), variational Bayes (Gal et al 2016 and lots of other papers), and graph nets (Battaglia et al 2018). We can start with getting point estimates of kappa rather than a full distribution, to see which features our model is picking up.
One important implementation detail is that the number of objects in our input X is variable, i.e. we are taking variable numbers of rows with a fixed set of columns from a catalog. Our output Y, for now, is one scalar, kappa.
You now have enough information to go do some literature review and write the introduction in your notes! I made an Overleaf template for your notes (view link here; log in with your Stanford email to edit).