This PR changes the way the artificial CXDR rates are computed:
Rather than using [median_rate-d, median_rate, median_rate+d], choosing d such that we get the right variance, the rates are now [median_rate, median_rate, median_rate+d'] for a d' that results in the same variance.
As a result, we can no longer get a rate 0 when the variance is large.
This PR changes the way the artificial CXDR rates are computed: Rather than using
[median_rate-d, median_rate, median_rate+d]
, choosingd
such that we get the right variance, the rates are now[median_rate, median_rate, median_rate+d']
for ad'
that results in the same variance.As a result, we can no longer get a rate
0
when the variance is large.