Point source convolution with symmetric PSF yields asymmetric image

[oczoske]

Asymmetric point source images have been reported for METIS with PSF_SCAO_9mag_06seeing.fits, most noticably with the H2O-ice filter. This filter is observed with two FOVs, one using the PSF layer at 2.2um, the other the PSF layer at 3.8um. The figures show the two FOVs, with the effect more pronounced with the narrower 2.2um PSF. The source is at (0, 0) arcsec, i.e. at pixel (1023.5, 1023.5) . Since the PSF is symmetric, the central four pixels should have equal signal.

cmd =  sim.UserCommands(use_instrument="METIS", set_modes=["img_lm"], properties={"!OBS.filter_name": "H2O-ice", "!OBS.psf_file": "PSF_SCAO_9mag_06seeing.fits"})
metis = sim.OpticalTrain(cmd)
star = sim.source.source_templates.star()
metis.observe(star)

The two FOVs are not even aligned (image is fov at 2.2um, contours at 3.8um):

[oczoske]

The culprit is presumably the _rescale_kernel method of FieldConstantPSF. This uses the zoom function and some cutting using integer division, which is not tested to preserve centering of the kernel. The approach in make_psf_cube using RegularGridInterpolator offers more control and appears to be robust. Refactoring is needed to avoid redundancy.

[oczoske]

Replacing the interpolation scheme did not help (but will be implemented anyway). The real problem was convolve() with mode='same' when either the image or the kernel have odd pixel numbers. This causes a shift of a point source by 0.5 pixels in the output even when source and kernel are perfectly centred on input.