Core Concepts
Any Las Vegas algorithm with locally certifiable failures can be converted into a zero-error Las Vegas algorithm that faithfully reproduces the correct output of the original algorithm in successful executions, with only polylogarithmic overhead in time complexity.
Abstract
The content presents a technique for perfectly simulating the output of Las Vegas algorithms in the LOCAL model of distributed computing. The key ideas are:
Warm-up: Under the assumption of correlation decay between distant variables, a simple sampling algorithm can be used to locally fix the random assignment around a failed node, while preserving the correct distribution.
LLL Augmentation: To handle the case without correlation decay, the authors introduce a technique to augment the Lovász Local Lemma (LLL) instance by adding a new "rare bad event". This enforces the desired correlation decay property, enabling the correct sampling.
Recursive Sampling: A recursive sampling framework is then developed to upgrade the previous sampling algorithms with bounded expected complexity to ones with exponentially convergent running time. This completes the proof of the main result.
The final algorithm can efficiently simulate any Las Vegas LOCAL algorithm with locally certifiable failures, producing the correct output distribution conditioned on no failure, with only polylogarithmic overhead in time complexity.