ensure sampling from a proper discrete Gaussian #10
Description
your technique of sampling from a normal distribution and then truncating to an integer doesn’t recover a discrete Gaussian. It’s probably still fine (see for example this answer by Vadim Lyubashevesky https://crypto.stackexchange.com/questions/42475/in-r-lwe-is-there-any-advantage-to-generate-secret-from-normal-distribution-ins), and I’ve seen someone do it in a paper once before. More typically now, people sample from “centered binomial random variables”, e.g. of the form sum_i X_i - X_i’ for X_i, X_i’ Bern(1/2)’s. By the above answer, this is essentially fine as well, and much simpler to sample.
The above is from a cryptography expert.