EPSFbuilder doesn't seem to be re-centering sources properly resulting in "dotted" PSF centers

[mfloyd21]

We're attempting to use the epsfbuilder function to extract, re-sample, re-center, and combine psf stars in our galaxy pointings to output an empirically determined psf.

• When the oversampling is set to 2, the psfs do have a believable shape and distribution (attached)
  

• When the oversampling is set any higher, the re-centering of the sources seems to fail. In the example I'm attaching, we have set the oversampling to 4, and we get a psf with 4 centers.
  

  • the stars objects before and after the re-centering seem unchanged, I will be attaching the stars before and after being re-centered, as well as their subtraction. The subtraction results in a grid of zeros, so nothing seems to be shifting

So our questions are:

• what is the reason the re-centering doesn't seem to be working properly (centroid measurement failing? Centroids measured but not used?)
• Since we know the stars sub-pixel positions, is there a way to force those values through instead of epsfbuilder estimating the centroids itself

This is my first time coming to a forum, so please let me know if I left anything vital off of my explanation and I would be happy to include it.

[larrybradley]

The ePSF builder algorithm generally requires a (very) large number of stars to build a PSF. And more stars are needed as you increase the oversampling factor. For example, with an oversampling factor of 4, in a perfect scenario you would need at least 16 stars that happen to have centers on each 4x4 subpixel (but even then the resulting PSF would be quite noisy). That's unlikely for a real sampling of stars. Typically you would need hundreds of stars for this to work. Without a good sampling of stars over the oversampled pixel grid, you'll end up with many "holes" in the grid, which will throw off the algorithm.

[emirkmo]

Hi @larrybradley @mfloyd21, please see our Wrapped implementation in the FLOWS photometry pipeline FlowsEPSFBuilder, where these requirements are not met yet we achieve a good fit. It has been tested on thousands of images. Obviously, EPSF is overkill when the stars are few but in a pipeline we had to be robust to input stellar density.

FlowsEPSFBuilder:

https://github.com/SNflows/flows/blob/c7e98d25e6b5c84248c8cb1ae9dad7013837178a/flows/epsfbuilder/epsfbuilder.py#L17-L56
The function to make the epsf with defaults:
https://github.com/SNflows/flows/blob/c7e98d25e6b5c84248c8cb1ae9dad7013837178a/flows/photometry.py#L105-L123

Be aware that currently the pipline is semver locked to photutils == 1.1.0. We only update when we're going to re-run the entire database for consistent results and I haven't had the chance to test the dependency upgrade path.

Details

We use oversampling = 1 by default as most fields are not crowded and most images are not oversampled. We also keep a small fitting & recentering boxsize since that works well with the default weights in the init Gaussian shape. We have maxiters high to ensure convergence and to test for a now fixed bug where sometimes iterations were diverging.