Photo-z for strong lensing science?

A science question regarding photo-z came to me by email, which I’m reposting here for further discussion.

Has any thought been given for photometric redshifts for strong lenses? LSST will find 100,000 strong lens systems and this is a top science case, they are still 1 in 1000 or 1 in 10,000 objects and require extra ‘care’ as source and deflector are always very close and/or blended and you want a good photo-z for both.

I can offer a few thoughts in the context of Data Management’s (DM’s) Roadmap to Photo-z for the LSST Object catalog. This particular science case might be a good topic for a Letter of Recommendation. For example, that LoR could outline the minimum qualities (accuracy, types of outputs, flags) for a photo-z data product that would enable strong lensing science. If possible, the LoR could suggest photo-z estimators that have been shown to do well for strong lensing science.

However, this topic thread can also discuss the research and development of photo-z algorithms (i.e., beyond the scope of DM’s PZ Roadmap) to enhance and optimize strong lensing science with LSST.

1 Like

Two thoughts on this

One is to simply deblend the lens and source images, combine the latter and then feed to the regular photo-z code. One might need a second pass deblending for any system flagged as a potential lens.

Next is to do a ‘photogeometric redshift’ (I think that is the term I saw) where you try and model the lens system and the photo-z simultaneously. Both the colours of the sources and their positions provides redshift information. This has a bonus that if you get a good solution you are more sure it is a lens. I expect the SL community would want to do both.

I think that the sample of “only” 100,000 objects (compared to the 3-4 billion object photo-z’s expected in cosmology samples and tens of billions overall) is small enough that we could think of doing a standalone custom photo-z measurement for the SL candidates. A custom run could include some specialized settings tuned to the expected SL sample and possibly more detailed estimates of physical parameters such as stellar mass that might not be appropriate or computationally feasible for the full “generic” photo-z catalog, yet still be fairly low demand in terms of computational resources on the limited SL candidate level. I believe that is what we are thinking within DESC, though this is preliminary and subject to change.