Gabow

Finds the strongly connected components of a directed graph with one depth-first search and two stacks, without the lowlink values of Tarjan's algorithm.

Time complexity O(V+E), gograph function connectivity.Gabow.

How it works

Gabow's algorithm finds the strongly connected components in one depth-first search, as Tarjan's does, but without lowlink values. It keeps two stacks. The first holds the vertices the search has reached that aren't in a component yet. The second holds the vertices on the current path that could still be the first vertex of a component.

When an edge leads to a vertex still on the first stack, every vertex on the path after it is in the same component, so the second stack is popped back to it. When the search finishes a vertex that is still on top of the second stack, that vertex starts a component. The first stack is popped down to it, and those vertices form the component.

It costs O(V+E). On the same graph, Tarjan's and Gabow's algorithms find the same components in the same order. They differ only in how they tell that the search has finished one.

Import cycles

Eighteen packages of a Python web app. Each edge is an import, and some packages import each other in a loop.

A directed graph with 18 vertices and 30 edges.

Which groups of packages import each other in a loop, so that none of them can load without the others?

The graph has 10 strongly connected components. The ones with more than one vertex are logging and utils; cache, db and models; email and templates; auth, billing, invoices, session and users. The other 6 components have one vertex each: config, cli, handlers, middleware, api and main.

Step by step

  1. Gabow's path-based algorithm searches depth-first from each vertex no earlier search has reached, in the graph's vertex order, and numbers the vertices in the order it reaches them. Stack S holds every vertex reached but not yet placed in a component. Stack P holds the vertices on the search path that can still be the root of a component. An edge to a vertex still on S closes a cycle, so P loses every vertex numbered above that vertex. When a vertex finishes on top of P, it is the root of a component: it and the vertices above it on S form one strongly connected component.
  2. Start at main, the first vertex. main gets number 0 and goes on S and P.
  3. Follow main to cli. cli gets number 1 and goes on S and P.
  4. Follow cli to config. config gets number 2 and goes on S and P.
  5. Follow config to utils. utils gets number 3 and goes on S and P.
  6. Follow utils to logging. logging gets number 4 and goes on S and P.

36 more steps lead to the last one:

  1. Done. Gabow found 10 strongly connected components, 6 of them with one vertex. Each component is listed after every component it has an edge into. Every vertex now shows the component it belongs to.

Use it in Go

for _, component := range connectivity.Gabow(g) {
	var names []string
	for _, v := range component {
		names = append(names, v.Label())
	}
	fmt.Println(strings.Join(names, ", "))
}

Example graphs

More in Connectivity

  • Tarjan: Finds the strongly connected components of a directed graph, the groups of vertices that can all reach each other, with one depth-first search.
  • Kosaraju: Finds the strongly connected components of a directed graph with two depth-first searches, the second on the graph with every edge reversed.
  • Condensation: Replaces each strongly connected component with one vertex. What is left is a DAG, so it can be sorted and scheduled.
  • Connected components: Splits a graph into its connected pieces, the groups of vertices joined to each other by some path, with one search per piece.
  • Cycle detection: Says whether a graph has a cycle, a way to follow its edges from a vertex back to the same vertex, and stops at the first one it finds.
  • Find a cycle: Returns the vertices of one cycle in order: the first one a depth-first search closes with an edge back to its search path.
  • Bridges and articulation points: Finds the edges and the vertices whose loss splits an undirected graph into more pieces, its single points of failure, from one depth-first search.

All algorithms