1st Algorithmic Breakthrough in 40 years for solving the Minimum Spanning Tree (MST) Replacement Edges problem
One of the most studied algorithms in computer science is called “Minimum Spanning Tree” or MST. In this problem, one is given a graph comprised of vertices and weighted edges, and asked to find a subset of edges that connects all of the vertices, and the total sum of their weights is as small as possible. Many real-world optimization problems are solved by finding a minumum spanning tree, such as lowest cost for distribution on road networks where intersections are vertices and weights could be length of the road or time to drive that segment. In 1926, Czech scientist Otakar Borůvka was the first to design an MST algorithm. Other famous approaches to solving MST are often given by the name of the scientist who designed MST algorithm in the late 1950’s such as Prim, Kruskal, and Dijkstra.
Several important variations of MST are also used in real applications, including the replacement problem. Imagine a use case when an edge in the MST degrades and either has a significanly increased cost or is removed entirely. One must quickly find the lowest cost “replacement edge” that reconnects the spanning tree. Several algorithms are known for this MST replacement edge problem. The first algorithm, due to Spira and Pan in 1975, took cubic time in the number of vertices. They presented an $O(n^2)$ algorithm to update the MST when new vertices are added, and could find all replacement edges in $O(n^3)$ time, where $n$ is the number of vertices in the graph. This was improved by Chin and Houck in 1978 to a quadratic time algorithm, or $O(n^2)$, using a more efficient approach to insert and delete vertices from the graph. The best approach to date is due to Tarjan in 1979, who gave an $O(m \alpha(n, m))$ time algorithm using path compression, where $m$ is the number of edges in the graph and $\alpha(n, m)$ is the inverse Ackermann’s function. $\alpha()$ is a very slow growing function, usually a number around 3 or 4 in practice, but still a gap has remained if a better approach exists.
For the first time in 40 years, progress has been made on this important graph algorithm. With Paul Burkhardt, we’ve designed a simple algorithm that runs very fast in linear time and space, or $O(n+m)$ where $n$ and $m$ are the number of vertices and edges, respectively, in the graph. The paper entitled A Linear Time Algorithm for Finding Minimum Spanning Tree Replacement Edges is now available in Arxiv. The main result of this paper is the first linear-time algorithm for finding all replacement edges in the minimum spanning tree. Our linear time and space algorithm is an asymptotic improvement from all prior algorithms, uses only simple arrays and Gabow-Tarjan disjoint set union data structures, alleviates the need to use least common ancestor (LCA) algorithms, and is easy to implement.
Other important graph algorithms need to find replacement edges, a step often considered their bottleneck in performance. For example, the most vital edge of a connected, weighted graph $G$ is the edge whose removal causes the largest increase in the weight of the minimum spanning tree. When the graph contains bridges, the most vital edge is undefined. Several algorithms were designed in the early 1990’s for the most vital edge, including $O(m \log m)$ and $O(n^2)$ time from Hsu et al. in 1991, and improvements to this approach by Iwano and Katoh in 1993 with $O(m+n \log n)$ and $O(m \alpha(m,n))$ time algorithms. When using our new MST replacement edge algorithm, we now can find the most vital edge in linear time (or $O(n)$) by simply finding the tree edge with maximum difference in weight from its replacement edge. Thus, our approach is the first linear algorithm for finding the most vital edge of the minimum spanning tree.