Ideas for final improvements of lens model

[PrayaagKatta]

I have the model for this system, but I would like some ideas for reducing the residuals in the arcs even further, as introducing shapelets etc did not help. Thank you!

[ajshajib]

It looks to me that the lens model fit is most likely incorrect, given that it is much more flattened than the observed light. Also, this looks like AO imaging; how are you obtaining the PSF? Is this lensed quasar or a galaxy-galaxy lens?

[PrayaagKatta]

I am actually just using a Gaussian PSF of FWHM = 0.05. And I think it is a galaxy galaxy lens.

[ajshajib]

Ok. For accurate modeling, it is necessary to use your best estimate of the PSF. The residual over image B looks symmetric, which most probably means that the PSF needs to be more accurate for your current model setup (I mentioned "current" because it looks to me that you have added a point source in your model setup). If the PSF cannot be known with sufficient accuracy or cannot be simultaneously reconstructed during modeling, one way to account for the "unknownness" of the PSF is to boost the error on the PSF. This can be done for a pixelated PSF using the provided error map in kwargs_psf['psf_error_map']. To me, it looks like a point source is being added on top of the source galaxy, thus the effect of an inaccurate PSF is probably being amplified in the residual. For galaxy-galaxy lenses, I don't add a point source and model the source galaxy's light just using a combination of Sersic profile and shaplets. In that case, the PSF accuracy can be less of a concern (e.g., https://ui.adsabs.harvard.edu/abs/2021MNRAS.503.2380S/abstract).

Another source of spurious residual may be the lens mass profile being very elliptical and then having a large shear on top, which looks odd to me, given the roundness of the lens light. Are you adding a point source to the model with point_source_model_list = ['LENSED_POSITION'] and then providing a non-None value in the kwargs_constraints['SOLVER_TYPE']? That might be causing the lens model to be this highly elliptical in the absence of any restriction against such a solution. One possible restriction on the mass ellipticity, if desired, could be to make the ellipticity parameters of the mass and light profiles have the same values through kwargs_constraints['joint_lens_with_light']. Alternatively, a more fine-tuned restriction can be imposed through a function provided in kwargs_likelihood['custom_logL_addition'].

Now, getting a well-fitted lens model can sometimes be more like art than science, and it's not always possible to give a set of instructions with guaranteed success. Also, given that lens systems come in diverse morphologies and the kind of lens modeling often depends on the science questions, I believe that more finely-tuned modeling choices should be made by the modeler in consultation with their co-authors. A discussion board like this is only suitable for introducing/discussing the features of lenstronomy to enable various modeling choices (as I have done above), and it should be up to the lens modeler and their co-authors to make the appropriate choices out of all the available features.

[PrayaagKatta]

Thank you so much for the detailed answer! I opened the discussion here since I wanted to be sure that I was not missing something that I could use in lenstronomy to help us in fine-tuning the parameters. I will try with a pixelated PSF! I think the ellipticity comes from the way the data was processed, I will have a look at that too.
(I initally started the fit with just Sersic profiles and shapelets but got a similar result, which I thought could be solved by adding a point source too.)