Controlling convolution

[amn3142]

I have a Gaussian light source and a separate pixelized point source. It appears that there may be cropping the edges of the Gaussian during the convolution with the PSF. Is it possible to turn this off?

[sibirrer]

Hi @amn3142, I am not sure I understand what you mean. Can you give an example or be more specific?

[amn3142]

Sorry. Here's what I want to do: I want to see if the point spread function is the same for all four images. There is some evidence that the PSF FWHM may vary with image brightness.

The first way I thought to do this was by modelling the 'true' light source as a Gaussian and then convolving with the PSF. I think there is a numerical issue somewhere however because the relative amount of blurring would be small, order ~a tenth of a pixel. So somewhere in the pipeline it looks like I'm getting some aliasing or some numerical feature that doesn't look like a smooth convolution of the point spread function. Currently the oversampling is set to 21.

Fundamentally I just want to have the ability to test out whether there are PSF variations over the image so if you have a better way to try this out with lenstronomy, I'd be open to it.

[sibirrer]

hmm, so the oversampling looks appropriate. The think is that the PSF convolution is done on the regular grid and when the FWHM is of order the pixel then this part can become inaccurate. If you want to convolve on the supersampled PSF, you have to specify that in the numerics (it's not the default). I worry a bit thought that a 21x oversampled PSF is hard to construct and whether or not this may also cause numerical artifacts.

[amn3142]

I do have 'supersampling_convolution' set to true in kwargs_numerics. Is there somewhere else I should pay attention to this?
The oversampled PSF looks ok, it's coming from some numerical model of the optics so it doesn't have any noticeably weird features that I can see by eye.

[sibirrer]

I see. The key (I think) is that the PSF then used in lenstronomy is the one you inputted (it should have the same oversampling factor as the one used in the convolution, otherwise, artifacts may arise). You can check that by generating a PSF() class and calling kernel_point_source_supersampled(supersampling_factor)

[amn3142]

I verified that those are the same.