Issue with convolution over pixel using astropy.convolution.Box2DKernel? about webbpsf HOT 4 OPEN

I think that's another way of showing the same thing using Eddie's approach in his gist: for odd oversampling, the Scipy and Astropy convolutions give (visually) the same things, whereas even oversampling yields to sharper differences.

from webbpsf.

In my head I was guessing that astropy is replacing the even kernel with some half-pixel shifted odd kernel, and so replacing the top-hat with something where the wings ramp linearly down or something, which sounds sane but for a discontinuous kernel like this one would be problematic. But I didn't push hard enough to actually find the code that was doing that and may be off-base. That would make the effective kernel a bit wider than it should be for even oversampling and have an effect like the one we see here.

from webbpsf.

Right - as @obi-wan76 pointed out, Astropy Boxkernel2D returns an odd-size kernel regardless of the width specified, so that's one difference with the Scipy approach used above.

With that said, I don't measure any difference in the centroids of the images I posted above...

from webbpsf.

Ah, sorry, I'm not sure I can see obi-wan76's message, thanks. Yes, I didn't test centroid differences for my test image, only for the real Roman images, but you can imagine that I was surprised when I learned that I could reproduce the input centroids well for Roman only for odd-sized convolutions! Note that the systematics were small (~0.01 pix), but we obviously shouldn't be limited by this particular systematic. The real Roman PSFs are of course a bit more complicated than these test images.

from webbpsf.