PSF artifacts

[kbarbary]

There are some SNe where the PSF model is pretty obviously wrong. Of the 30 SNe I test on locally:

LSQ12dbr (R):

SN2005ki (B):

SN2005ki (R):

SN2006ac (B):

SN2011by (R):

SNF20060512-002 (B):

SNF20060512-002 (R):

SNNGC4424 (R):

SNNGC6801 (B):

SNNGC6801 (R):

[kbarbary]

Note that the ringing is caused by FFT shifting the PSF, but this is a separate issue. For example, running SN2005ki (B) with a fully resampled PSF (not FFT shifted) removes the ringing, but not the large scale residuals:

Fully resampled PSF:

FFT-shifted PSF (same as above):

[ycopin]

I don't understand how you can say this is coming from the PSF: which PSF, of which epoch? All epochs seem equally bad. Furthermore, the final model for the "fully resampled" and "FFT-shifted" PSF look rather different, while the PSF is the same.

PS: could you change the scale for the reference images, so that we could see if there's any structure in there?

[kbarbary]

As one concrete example, take this epoch from SNNGC4424 (R) as an example:

It seems to me that the PSF in this epoch is too wide, causing positive residuals at large radius and negative residuals at small radius. I suggested that it has at least something to do with the PSF because the only free parameters in cubefit at this stage are the position of the galaxy, the position of the SN, and the SN amplitude. Those all seem to be optimized pretty well - or at least, there's no way to change those parameters to make this residual go away.

Of course, this could be a problem with how cubefit is treating the PSF model, rather than the PSF model itself.

[ycopin]

But then I'm wondering why all the PSF of a SN sequence (and not only a couple of them) would be wrong at the same time.

[kbarbary]

Note: the scale labels on the residuals in the plots above are wrong. The residuals are actually typically on the scale of 1-5% (of the maximum value in the scene model). I won't repost all the images, but here's LSQ12dbr (R) for example:

(This is with resampled PSF fitting and subpixel sampling of 3x3, which is now the default.)