We have used the ’empirical forward model’ method to reduce 384 images from night 2455864.
Using the DMI CRAY we stepped through 1000 values of alfa from 1 to 2 for each image and required the best possible fit on the sky part of each image.
We plot the best value of alfa found for each filter againts the time of observation:
We see that the values for alfa are narrowly grouped for most filters, except for the VE1 filter which jumps between two values. Unexpectedly we see that the blue filters have high alfas (i.e. narrow halos) while the reddest filter (VE2) has the smallest value of alfa (i.e. the broadest halo). This is opposite of what we might expect from physics, and the code plotting these data must be reviewed.
The scatter of alfa for e.g. the IRCUT filter is 0.0036 which is 0.2 %. Since we had 98 IRCUT values the standard deviation of the mean alfa is about 0.00036 which is about 3 times larger than the required accuarcy estimated on this page below, to achieve a DS accuracy of 0.1%.
Visual inspection of the fits show that improvements are possible, so more experiments with parameter settings is called for. The above were results from fits of the absolute residuals. Next we try fitting the relative residuals, i.e.
(obs – model)/obs
instead of
(obs – model).
We have now tested that – and it was not a good idea: fits got worse.
That the PSFs are broader at the redder wavelengths ius consistent with how wide the PSF shoul dbe given the science behind abberation. However, it is not the only possible cause – a detailed calculation of the PSFs as a function of wavelength, given the details of our optics must be performed. ZEEMAX can do the job.
Inspection of the actual fits reveals that the ‘jump’ we see in VE1 may be due to bad fitting. However, how to fit better then? The ‘bad fit’ is evident only on the disc – which we do not fit … we only fit the sky part … this is because the model only predicts the scattered light, not ES+scattered light.
This may be an indication that the full forward method from synthetic images WITH ES may be the best way, but we have yet to test that. Cray running red hot on present tests!
For each filter the frames are sorted in time, so – yes – it looks like each filter (except IRCUT) changed slightly in time. The filters are actually interspersed so that makes sense (except for IRCUT). I will get times on these data points and plot them on a joint time axis! Whoho!
No idea what VE1 was up to. The hardware reports what it reports. I wonder if looking at absolute flux levels would help us, since the flux through each filter is unique for the filter?
Very interesting indeed!
it looks like there might be a small systematic changes in alpha with time, going from frame to frame at least. Could this be a residual signature of the bias level changing on its 20 minute or so cycle? how long did the sequences take to acquire — e.g. from the first V to the last? The behaviour of VE1 is really amazing. Can’t imagine why one would get a flip of value like that, to the same value that the VE2 filter has. That seems very surprising indeed. Could the VE2 filter simply have been swapped into the beam (and VE1 out) for a while without us being informed by the system? I am still reviewing what accuracy we’ll need on alpha to get ES according to my model fitting here, will get to you on that asap.