Convolve incorrectly handles large gaps in Q values

[bmaranville]

In a dataset with a gap in Q (from masking, for instance) the current convolve_gaussian algorithm finds the left edge of the convolution window by moving forward through the x-values until x >= (xo - limit), then steps backwards while (x > xo - limit). This has the effect of always selecting one point beyond the end of (xo-limit), even if that point is very far away from the desired window. I think we'd be better served by dropping the second loop, and just keeping points where x >= (xo - limit).

[bmaranville]

Why is the limit of Gaussian convolution set by this?

LOG_RESLIMIT = -6.90775527898213703123
limit = sqrt(-2.*sigma*sigma * LOG_RESLIMIT)

The result of this is that

limit == 3.716922 * sigma

while I thought in other communications it had been agreed to settle on 3-sigma as the cutoff?

[pkienzle]

Confused reasoning on my part.

As mentioned in the comment above the code, this is 0.001 of the peak value of the normal distribution. At the time I was trying to do a good approximation of the full Gaussian convolution and chose that as the cutoff. The code I was modifying used a pdf of 0.001 as the cutoff.

If we believe that resolution on reflectometers have truncated tails, then 3σ is a better choice. Feel free to change it.