For the night of 2455864 we have tested the effect of using a 5x5 patch on the sky vs an 11x11 patch. This is the patch in which the average is taken and through which average the model curve is forced to pass. We had noted that a 5x5 patch was small compared to some of the wiggles in the sky background so we tested the use of 11x11 patches.
We also test the effect of using 64-bit rather than 32-bit code - we do this with 'no change of compilers or compiler flags'.
Finally we tested the effect of taking smaller steps in alfa. The figure summarizes these three experiments:
The difference in detected alfa was small when we went from 5x5 to 11x11 patches. The average change was near 0, the standard deviation apeared to be about 1 alfa step (i.e. 0.001), the max and min of the change were -0.6 and 0.4 but only 2% of the 384 images suffered a change in alfa larger than 0.01 in absolute value (i.e. outside the plotting range). We conclude that there is a tiny indication of a low-bias in determined alfa when using 5x5 patches, so we will use 11x11 from, now on.
For the 32-vs-64 bit analysis we found that for step size in alfa of 0.001 the change in going from 32 to 64 bit was most often 0. A small fraction of images had a +/- change of 0.001 in alfa - i.e. the step size.
The effect of smaller steps was negligible.
We next looked at changes in the determined DS intensity.
For each filter and for each of the three modes of the EFM code run on the CRAY we extracted the median DS intensity in a box 11x11 pixels large positioned 2/3 of the radius on the dark side. We also extracted the BS from an 11x11 box positioned at the center of gravity of the image in the observed image.
Run1 was the test for the effect of using a 5x5 vs 11x11 patch on the sky and the BS to force the model through.
Run2 was the test for effect of changing from 32 to 64 bit code.
Run3 was the test of the effect of using smaller step sizes in alfa.
We found that:
For all filters there is a 6 to 7% effect on the BS/DS ratio of going from a 5x5 to 11x11 patch on the sky when fitting the model to the observed image.
The effect of changing from 32 to 64 bit code was negligible, as above - the largest change in BS/DS ratio was 0.02% in the VE1 and VE2 filters; the others were at the 0.007-0.009% level.
The effect of smaller step sizes when searching for the optimal alfa was a little larger - in the range from 0.19 - 0.9 %; largest in the VE2 filter.
We conclude that careful choice of the way in which the model is fitted to the sky and the BS is warranted.