Bright sources artefacts from diff kernel error

[herjy]

Tests on galsim simulations revealed ring-shaped artefacts in the residuals for bright sources.
The rings are not caused by monotonicity or any of the constraints, neither are they remnants from initialisation.
The suspected cause is an error introduced by the diff_kernel. The diff kernel indeed introduces very small errors to the PSF modeling ~10^-8%, which is small compared to the amplitude of the observed residuals:

[herjy]

For reference, when comparing the data's psf and the psf constructed by convolving the model psf by the diff kernel, errors in log scale look like this:

Renormalisation of the diff kernel to null the flux or amplitude difference between the data PSF and the reconstructed one did not work as the shape of the two do not match.

[herjy]

Double precision or rescaling of the central region did not solve the problem. I am convinced the problem comes from a wrapping or a small contraction/dilation of the Fourier plane.

[pmelchior]

Where would these dilations come from?

[herjy]

fft transform shapes/gridding. I haven't found anything yet playing with the fft code, so that might be wrong.

[pmelchior]

In the example above, the source is clearly truncated at the box. This can easily cause ringing as observed and would be independent of the correctness of the PSF diff kernel. Can you manually increase the box size to see if this has any effect?

There's one more test I like to make when everything becomes unclear. Initialize the source by hand with the correct deconvolved image as morphology. That means the solution is already at the maximum likelihood points. Check if the fitter computes non-zero gradients, i.e. it wants to walk away from it. If so, at least the backward path is biased.

[herjy]

That's not it, I checked before.

The residuals in the model look a lot like a sinc, but the difference between the psfs looks like a gaussian (all positive).

[pmelchior]

Is the original galaxy image from Galsim band-limited before it gets convolved?

[herjy]

Yes, because I convolve it with a small gaussian kernel to make sure it is.

[herjy]

I expanded this kernel from 1 pixel (which I thought was enough) to 1.5 pixels and it seems to do the trick-not*.

[herjy]

I'm still suspicious though.

[pmelchior]

This seems to suggest that undersampling of the model is responsible. If so, we need to increase the width of the model PSF to ensure proper sampling, or (probably better) ensure that no gradient update forces the model to exceed the band-limit set by the model PSF.

[pmelchior]

closed in favor #222