tree 9b59d1bcaf09f1f1127ef4d4e3d28d604ac47ed7
parent 22648a237d150d648d76a68e8f30a4623e0b1a1d
author Differential Privacy Team <noreply@google.com> 1670917743 -0800
committer Kulankhina <oleksandra.kulankhina@gmail.com> 1671123191 +0100

Incorporating the "Connect the Dots" algorithm in PLD accounting library.

Reference: https://arxiv.org/abs/2207.04380

Main changes are as follows:
- Allow vectorized calls to get_delta_for_epsilon,
- Construct pessimistic PLDPmf using Connect-the-Dots algorithm,
- Support Connect-the-Dots algorithm for privacy loss distributions of additive noise mechanisms.

Some other minor changes included as follows:
- Add checks to make sure that truncation bounds in self convolution does not result in truncation of the input list,
- Use more numerically stable scipy.special.logsumexp for moment generating function computation,
- Use more efficient np.roll for shifting the output result for self convolution.

PiperOrigin-RevId: 494208119
Change-Id: I1c50896b3252c19c8802b1e1ebfd6f0ba94c3be2
GitOrigin-RevId: fca6f9ba22ab0a7b3d020d3c5bc95d30405261df