The variance is discontinuous because the flat-field is discontinuous: there is no flux outside the area of the trace.

What specific problems is this causing?

Ok, I suspected something similar. Thanks for clarifying.

We encountered this problem when he was modeling PSF with PSFx (PIPE2D-355), i.e., when he tried to work with 2D type algorithm on the simulated data and was cutting 40x40 and 20x20 post-stamp images. I never encountered this problem as I never look at the simulated data and only worked with ``real data''.

But what is the problem?
This effect will be present on both simulated and real data.

This effect is not present in the real data. The flux outside traces is not zero in the real data and variance map behaves as expected. If this was present in the real data 2D modeling would be seriously hampered. You might remember, we were even tweaking numbers so that mean of the variance map is the equal to mean of the variance in the science image (PIPE2D-278)? I think we are misunderstanding each other here?

Here's the image (left), mask (center) and variance (right) planes for a part of a calexp produced with the command detrend.py DATA --calib DATA/CALIB --rerun sim2d-119 --id visit=21460 arm=r. I think this shows the same features you're pointing out in the simulated data. If I'm mistaken, could you please be more specific in identifying the problem?

I am very confused. I have never seen this feature when running detrend.py on my side. 

But the variance value of a pixel should not change depending on the fact if the pixel is within a trace or outside a trace. Right? You are showing there is a clear difference in variance value for pixels within trace and outside trace and that has to be unwanted behaviour unless I am misunderstanding something.

All the pixels outside the trace are masked out, and shouldn't be used. We have no variance estimate for them in the flat, but they get set to 1. I could set them to a variance of zero, but the drp_stella code we've inherited wouldn't like that. Alternatively, I could set them to a variance of infinity, but I think rhl has argued against that in the past. What value do you think they should be set to?

Note also that the flat-field outside the traces is set to 1 everywhere, so if you're flat-fielding donut images with the fiber flats, you're probably messing up the images.

How is the computation of variance of pixels outside of trace different than the computation of variance of pixels inside of trace?

Pixels outside of trace are still needed for estimates of scattering etc... 

It is not possible to flat field 2d images for me as we do not have traces (as you correctly pointed out)

I guess why I never noticed this is because I turned off flatfielding when doing detrend ?

That's possible.

I suggest we close this ticket.

price Is the data that you generate above with command detrend.py DATA --calib DATA/CALIB --rerun sim2d-119 --id visit=21460 arm=r avaliable on Tiger somewhere?

My data repo is /scratch/pprice/sim2d-119/DATA.

Could this have been a rerun from the "official" data repo?

I am not aware of an "official" data repo.

We are supposed to have all the data on tiger, so that would be the official one.

rhl, did you decide how you want to handle this? The problem is that pixels outside the trace have zero weight when normalising by the trace.