In theoretical computer science, it is a common practice to show existential lower bounds for problems, meaning there is a family of pathological inputs on which no algorithm can do better. However, most inputs of interest can be solved much more efficiently, giving rise to the notion of universally optimal algorithms, which run as fast as possible on every input. Questions on the existence of universally optimal algorithms were first raised by Garay, Kutten, and Peleg in FOCS '93. This research direction reemerged recently through a series of works, including the influential work of Haeupler, Wajc, and Zuzic in STOC '21, which resolves some of these decades-old questions in the supported CONGEST model. We work in the HYBRID distributed model, which analyzes networks combining both global and local communication. Much attention has recently been devoted to solving distance related problems, such as All-Pairs Shortest Paths (APSP) in HYBRID, culminating in a $\tilde \Theta(n^{1/2})$ round algorithm for exact APSP. However, by definition, every problem in HYBRID is solvable in $D$ (diameter) rounds, showing that it is far from universally optimal. We show the first universally optimal algorithms in HYBRID, by presenting a fundamental tool that solves any broadcasting problem in a universally optimal number of rounds, deterministically. Specifically, we consider the problem in a graph $G$ where a set of $k$ messages $M$ distributed arbitrarily across $G$, requires every node to learn all of $M$. We show a universal lower bound and a matching, deterministic upper bound, for any graph $G$, any value $k$, and any distribution of $M$ across $G$. This broadcasting tool opens a new exciting direction of research into showing universally optimal algorithms in HYBRID. As an example, we use it to obtain algorithms for approximate and exact APSP in general and sparse graphs.
翻译:HYBRID分布式模型中的通用最优确定性广播
翻译后的摘要:
在理论计算机科学中,展示问题的存在性下界是一种常见的做法,意味着存在一组病态输入,无论如何算法都不能表现更好。然而,大多数有趣的输入可以更有效地解决,从而引出了通用最优算法的概念,这些算法在每个输入上都尽可能快地运行。Garay,Kutten和Peleg在FOCS’93中首次提出了通用最优算法的存在性问题。这个研究方向最近通过一系列的作品重新出现,包括Haeupler,Wajc和Zuzic在STOC’21中的有影响力的作品,解决了一些这些数十年前的问题,支持CONGEST模型。我们在HYBRID分布式模型中工作,该模型分析同时结合全局和局部通信的网络。近年来,越来越多的关注点集中在解决与距离相关的问题,例如在HYBRID中的全对最短路径(APSP),达到了精确的APSP的倒数$\tilde \Theta(n^{1/2})$轮算法。但是,按定义,HYBRID中的每个问题都可以在$D$(直径)轮内解决,表明它远非通用最优。我们展示了HYBRID中的第一个通用最优算法,这是通过提供解决任何广播问题的基本工具在一定的轮数内(确定性地)实现的。具体而言,我们在一个图$G$中考虑问题,其中分布在$G$上的一组$k$条消息$M$需要使每个节点了解所有$M$。我们展示了任何图$G$、任何值$k$和$M$在$G$上的任何分布的通用下界和匹配的确定性上界。这种广播工具开启了一个新的令人兴奋的方向,即在HYBRID中展示通用最优算法。例如,我们使用它得到了一些在一般和稀疏图上实现近似和精确APSP的算法。