Are you using the NevenPsf? What method are you using? How do you determine the position at which to realise the PSF?

The current code used to generate the uniform illumination example is:

sci_image_unit=np.ones((20,20))
var_image_unit=np.ones((20,20))

allparameters_proposal=np.array([0.1,0.09,0.02,-0.1,0.0,0.1,0.02,0.2,
0.7,0.1,0.3,0.0,0.07,0.0,
0.17,0.67,1,1,
0.0,-0.0,0.80,
0.01,0.51,1.06,0.28,
50000,2.5,0.027,
0.37,1.85,1.0])

model = LN_PFS_single(sci_image_unit,var_image_unit,dithering=1,save=1,verbosity=1,npix=1536)
res=model(allparameters_proposal,return_Image=True)
print(res[0])

sci_image_fake=res[1]
var_image_fake=0.97*res[1]+40
model_fake = LN_PFS_single(sci_image_fake,var_image_fake,dithering=1,save=1,verbosity=1,npix=1536)
res_fake=model_fake(allparameters_proposal,return_Image=True)

print(res_fake[0])
plt.imshow(res[1]-res_fake[1],cmap='bwr')
plt.colorbar()

To my knowledge this isn't associated with NevenPsf. This is a test to see if LN_PFS_single can reconstruct a PSF that it's created. The position of the PSF is determined by the code in the create_complete_realization function and is the issue to be investigated.

I was worried that an important part of the 2D pipeline wasn't working for you, but apparently not. Sorry to have bothered you!

My apologies, that's my fault for not specifying the issue precisely. Thank you very much for checking.

I have uploaded the new version of the code (internal version 0.34b). I have added a small finishing step that optimizes with Nelder-Mead method (as discussed on technical DRP 2D telecon on October 12). The reason why I did not use it at the start is that one can not put bounds on that method, and sometimes the solution would escape and crash the code. I found it easier to use scipy.shgo that uses bounds. Once when we are very close to the actual solution we can use Nelder-Mead  to finish up.

Your fix gets the center to a very high degree of precision, orders of magnitude better than before! I've uploaded v034b_UniformInitial_PSFshift.png and v034b_RandomInitial_PSFshift.png that both show residuals on the order of 1e-6, which is consistent. This works perfectly for my case, and I believe you've completed this ticket.

Is this code you see being used for the 2D_DRP or does the Nelder-Mead add too much runtime?

No, we probably won't be using this code.  Regular, and fast, centroiding is good enough but we'll have to make sure that it's consistent with the definition of the centroid for Neven's models.   For now, let's use this.

I have closed this ticket. Brent Belland continue with PIPE2D-636 and inform me if there are fresh problems.