Random selection, leader election, and collective coin flipping are fundamental tasks in fault-tolerant distributed computing. We study these problems in the full-information model where despite decades of study, key gaps remain in our understanding of the trade-offs between round complexity, communication per player in each round, and adversarial resilience. We make progress by proving improved bounds for these problems. We first show that any $k$-round coin flipping protocol over $\ell$ players, each player sending one bit per round, can be biased by $O(\ell/\log^{(k)}(\ell))$ bad players. We obtain a similar lower bound for leader election. This strengthens prior best bounds [RSZ, SICOMP 2002] of $O(\ell/\log^{(2k-1)}(\ell))$ for coin flipping protocols and $O(\ell/\log^{(2k+1)}(\ell))$ for leader election protocols. Our result implies that any (1-bit per player) protocol tolerating linear fraction of bad players requires at least $\log^* \ell$ rounds, showing existing protocols [RZ, JCSS 2001; F, FOCS 1999] are near-optimal. We next initiate the study of one-round, (1-bit per player) random selection. We construct a protocol resilient to $\ell / (\log \ell)^2$ bad players that outputs $(\log \ell)^2 / (\log \log \ell)^2$ uniform random bits. This implies a one-round leader election protocol resilient to $\ell / (\log \ell)^2$ bad players, improving the prior best protocol [RZ, JCSS 2001] which was resilient to $\ell / (\log \ell)^3$ bad players. Our resilience matches that of the best one-round coin flipping protocol by Ajtai & Linial. We also obtain an almost matching lower bound: any protocol outputting $(\log \ell)^2 / (\log \log \ell)^2$ bits can be corrupted by $\ell (\log \log \ell)^2 / (\log \ell)^2$ bad players. To obtain our lower bound, we introduce multi-output influence, an extension of influence of boolean functions to the multi-output setting.
翻译:随机选择、领导者选举与集体硬币抛掷是容错分布式计算中的基础任务。我们在全信息模型下研究这些问题,尽管经过数十年的研究,对于轮次复杂度、每轮每位参与者的通信量以及对抗性容错能力之间的权衡关系,仍存在关键空白。我们通过证明这些问题的改进界取得了进展。首先,我们证明任何在ℓ位参与者间进行的k轮硬币抛掷协议(每位参与者每轮发送1比特),可被O(ℓ/log^(k)(ℓ))个恶意参与者偏置。对于领导者选举,我们获得了类似的下界。这强化了先前的最佳界[RSZ, SICOMP 2002]:对于硬币抛掷协议为O(ℓ/log^(2k-1)(ℓ)),对于领导者选举协议为O(ℓ/log^(2k+1)(ℓ))。我们的结果表明,任何(每位参与者1比特)容忍线性比例恶意参与者的协议至少需要log* ℓ轮,这表明现有协议[RZ, JCSS 2001; F, FOCS 1999]接近最优。其次,我们首次研究单轮(每位参与者1比特)随机选择问题。我们构建了一个能抵抗ℓ/(log ℓ)^2个恶意参与者的协议,该协议输出(log ℓ)^2/(log log ℓ)^2个均匀随机比特。这导出了一个能抵抗ℓ/(log ℓ)^2个恶意参与者的单轮领导者选举协议,改进了先前最佳协议[RZ, JCSS 2001](仅能抵抗ℓ/(log ℓ)^3个恶意参与者)。我们的容错能力与Ajtai & Linial的最佳单轮硬币抛掷协议相匹配。我们还获得了一个几乎匹配的下界:任何输出(log ℓ)^2/(log log ℓ)^2比特的协议,可被ℓ(log log ℓ)^2/(log ℓ)^2个恶意参与者破坏。为获得下界,我们引入了多输出影响力,这是布尔函数影响力向多输出场景的扩展。