Prevents a circumvention of privacy protections by setting min == max. The bounds for the statistic are not just important for calculating the maximum potential number of sampling iterations to run. In the case that min == max, then zero sampling rounds will be conducted, resulting in variance zero noise. This is clearly only DP if the distribution of the statistic is constant. More generally, this clamp is necessary anytime the range of derived bounds is greater than the range of manually provided bounds.
ecowan
commented
3 years ago
Prevents a circumvention of privacy protections by setting
min == max. The bounds for the statistic are not just important for calculating the maximum potential number of sampling iterations to run. In the case thatmin == max, then zero sampling rounds will be conducted, resulting in variance zero noise. This is clearly only DP if the distribution of the statistic is constant. More generally, this clamp is necessary anytime the range of derived bounds is greater than the range of manually provided bounds.