The algorithm is:
We will analyze the influence of the background on optimal extraction for the segment Lif 1a of the object: SK-66D172, ID: P117-22.
The average flux by pixel on the extraction area is 0.231658 counts/pixels.
Here are the weights for the real background which is 2.137617E-03 counts/pixels (a nearly 2ks exposure). The green correspond to 
1, and the red to nearly 1.5:
Now, let's have a look of the different 
with different S/N ratios:
With a background of 0.0001 counts/pixels ( 
):
With a background of 0.001 counts/pixels ( 
):
With a background of 0.01 counts/pixels ( 
):
With a background of 0.1 counts/pixels ( 
):
You can see clearly that for a given flux, the larger the background, the more the extraction is restricted to a small number of pixels.
You can also see that for a low background, the weights (well, most of them) are equal to 1 (green), which mean an extraction on the entire segment. This happen when you have a bright target, so already a good S/N ratio. But anyway, it is still unlogic to extract the spectrum on pixels that we known have only background. So we modified the 
calibration file and put all probability which are below 0.05% equal to 0. Here are the new weights for the real background (2.137617E-03 counts/pixels), after doing the modifications:
