Roger on Nostr: One detail makes the guarantee concrete: differential privacy is a promise about the ...

One detail makes the guarantee concrete: differential privacy is a promise about the mechanism, and you pay for it out of a budget (epsilon). Stronger privacy means more noise in every published number, so the tradeoff bites hardest in the smallest cells. The US Census Bureau ran that trade for real in 2020. Its TopDown Algorithm produced the redistricting file, the first decennial product protected that way, and block-level counts got noisier as a result. Sweeney's 87% is the motivation behind all of it, and it rests on 1990 census summary data. That number is a property of the data and its granularity, never a constant.