transitive reduction algorithm: pseudocode?

I have been looking for an algorithm to perform a transitive reduction on a graph, but without success. There's nothing in my algorithms bible (Introduction To Algorithms by Cormen et al) and whilst I've seen plenty of transitive closure pseudocode, I haven't been able to t开发者_StackOverflowrack down anything for a reduction. The closest I've got is that there is one in "Algorithmische Graphentheorie" by Volker Turau (ISBN:978-3-486-59057-9), but unfortunately I don't have access to this book! Wikipedia is unhelpful and Google is yet to turn up anything. :^(

Does anyone know of an algorithm for performing a transitive reduction?

See Harry Hsu. "An algorithm for finding a minimal equivalent graph of a digraph.", Journal of the ACM, 22(1):11-16, January 1975. The simple cubic algorithm below (using an N x N path matrix) suffices for DAGs, but Hsu generalizes it to cyclic graphs.

// reflexive reduction
for (int i = 0; i < N; ++i)
  m[i][i] = false;

// transitive reduction
for (int j = 0; j < N; ++j)
  for (int i = 0; i < N; ++i)
    if (m[i][j])
      for (int k = 0; k < N; ++k)
        if (m[j][k])
          m[i][k] = false;

The basic gist of the transitive reduction algorithm I used is

foreach x in graph.vertices
   foreach y in graph.vertices
      foreach z in graph.vertices
         delete edge xz if edges xy and yz exist

The transitive closure algorithm I used in the same script is very similar but the last line is

         add edge xz if edges xy and yz OR edge xz exist

Based on the reference provided by Alan Donovan, which says you should use the path matrix (which has a 1 if there is a path from node i to node j) instead of the adjacency matrix (which has a 1 only if there is an edge from node i to node j).

Some sample python code follows below to show the differences between the solutions

def prima(m, title=None):
    """ Prints a matrix to the terminal """
    if title:
        print title
    for row in m:
        print ', '.join([str(x) for x in row])
    print ''

def path(m):
    """ Returns a path matrix """
    p = [list(row) for row in m]
    n = len(p)
    for i in xrange(0, n):
        for j in xrange(0, n):
            if i == j:
                continue
            if p[j][i]:
                for k in xrange(0, n):
                    if p[j][k] == 0:
                        p[j][k] = p[i][k]
    return p

def hsu(m):
    """ Transforms a given directed acyclic graph into its minimal equivalent """
    n = len(m)
    for j in xrange(n):
        for i in xrange(n):
            if m[i][j]:
                for k in xrange(n):
                    if m[j][k]:
                        m[i][k] = 0

m = [   [0, 1, 1, 0, 0],
        [0, 0, 0, 0, 0],
        [0, 0, 0, 1, 1],
        [0, 0, 0, 0, 1],
        [0, 1, 0, 0, 0]]

prima(m, 'Original matrix')
hsu(m)
prima(m, 'After Hsu')

p = path(m)
prima(p, 'Path matrix')
hsu(p)
prima(p, 'After Hsu')

Output:

Adjacency matrix
0, 1, 1, 0, 0
0, 0, 0, 0, 0
0, 0, 0, 1, 1
0, 0, 0, 0, 1
0, 1, 0, 0, 0

After Hsu
0, 1, 1, 0, 0
0, 0, 0, 0, 0
0, 0, 0, 1, 0
0, 0, 0, 0, 1
0, 1, 0, 0, 0

Path matrix
0, 1, 1, 1, 1
0, 0, 0, 0, 0
0, 1, 0, 1, 1
0, 1, 0, 0, 1
0, 1, 0, 0, 0

After Hsu
0, 0, 1, 0, 0
0, 0, 0, 0, 0
0, 0, 0, 1, 0
0, 0, 0, 0, 1
0, 1, 0, 0, 0

The Wikipedia article on transitive reduction points to an implementation within GraphViz (which is open source). Not exactly pseudocode, but maybe someplace to start?

LEDA includes a transitive reduction algorithm. I don't have a copy of the LEDA book anymore, and this function might have been added after the book was published. But if it's in there, then there will be a good description of the algorithm.

Google points to an algorithm that somebody suggested for inclusion in Boost. I didn't try to read it, so maybe not correct?

Also, this might be worth a look.

The algorithm of "girlwithglasses" forgets that a redundant edge could span a chain of three edges. To correct, compute Q = R x R+ where R+ is the transitive closure and then delete all edges from R that show up in Q. See also the Wikipedia article.

Depth-first algorithm in pseudo-python:

for vertex0 in vertices:
    done = set()
    for child in vertex0.children:
        df(edges, vertex0, child, done)

df = function(edges, vertex0, child0, done)
    if child0 in done:
        return
    for child in child0.children:
        edge.discard((vertex0, child))
        df(edges, vertex0, child, done)
    done.add(child0)

The algorithm is sub-optimal, but deals with the multi-edge-span problem of the previous solutions. The results are very similar to what tred from graphviz produces.

ported to java / jgrapht, the python sample on this page from @Michael Clerx:

import java.util.ArrayList;
import java.util.List;
import java.util.Set;

import org.jgrapht.DirectedGraph;

public class TransitiveReduction<V, E> {

    final private List<V> vertices;
    final private int [][] pathMatrix;
    
    private final DirectedGraph<V, E> graph;
    
    public TransitiveReduction(DirectedGraph<V, E> graph) {
        super();
        this.graph = graph;
        this.vertices = new ArrayList<V>(graph.vertexSet());
        int n = vertices.size();
        int[][] original = new int[n][n];

        // initialize matrix with zeros
        // --> 0 is the default value for int arrays
        
        // initialize matrix with edges
        Set<E> edges = graph.edgeSet();
        for (E edge : edges) {
            V v1 = graph.getEdgeSource(edge);
            V v2 = graph.getEdgeTarget(edge);

            int v_1 = vertices.indexOf(v1);
            int v_2 = vertices.indexOf(v2);

            original[v_1][v_2] = 1;
        }
        
        this.pathMatrix = original;
        transformToPathMatrix(this.pathMatrix);
    }
    
    // (package visible for unit testing)
    static void transformToPathMatrix(int[][] matrix) {
        // compute path matrix 
        for (int i = 0; i < matrix.length; i++) {
            for (int j = 0; j < matrix.length; j++) { 
                if (i == j) {
                    continue;
                }
                if (matrix[j][i] > 0 ){
                    for (int k = 0; k < matrix.length; k++) {
                        if (matrix[j][k] == 0) {
                            matrix[j][k] = matrix[i][k];
                        }
                    }
                }
            }
        }
    }

    // (package visible for unit testing)
    static void transitiveReduction(int[][] pathMatrix) {
        // transitively reduce
        for (int j = 0; j < pathMatrix.length; j++) { 
            for (int i = 0; i < pathMatrix.length; i++) {
                if (pathMatrix[i][j] > 0){
                    for (int k = 0; k < pathMatrix.length; k++) {
                        if (pathMatrix[j][k] > 0) {
                            pathMatrix[i][k] = 0;
                        }
                    }
                }
            }
        }
    }

    public void reduce() {
        
        int n = pathMatrix.length;
        int[][] transitivelyReducedMatrix = new int[n][n];
        System.arraycopy(pathMatrix, 0, transitivelyReducedMatrix, 0, pathMatrix.length);
        transitiveReduction(transitivelyReducedMatrix);
        
        for (int i = 0; i <n; i++) {
            for (int j = 0; j < n; j++) { 
                if (transitivelyReducedMatrix[i][j] == 0) {
                    // System.out.println("removing "+vertices.get(i)+" -> "+vertices.get(j));
                    graph.removeEdge(graph.getEdge(vertices.get(i), vertices.get(j)));
                }
            }
        }
    }
}

unit test :

import java.util.Arrays;

import org.junit.Assert;
import org.junit.Test;

public class TransitiveReductionTest {

    @Test
    public void test() {

        int[][] matrix = new int[][] {
            {0, 1, 1, 0, 0},
            {0, 0, 0, 0, 0},
            {0, 0, 0, 1, 1},
            {0, 0, 0, 0, 1},
            {0, 1, 0, 0, 0}
        };
        
        int[][] expected_path_matrix = new int[][] {
            {0, 1, 1, 1, 1},
            {0, 0, 0, 0, 0},
            {0, 1, 0, 1, 1},
            {0, 1, 0, 0, 1},
            {0, 1, 0, 0, 0}
        };
        
        int[][] expected_transitively_reduced_matrix = new int[][] {
            {0, 0, 1, 0, 0},
            {0, 0, 0, 0, 0},
            {0, 0, 0, 1, 0},
            {0, 0, 0, 0, 1},
            {0, 1, 0, 0, 0}
        };
        
        System.out.println(Arrays.deepToString(matrix) + " original matrix");

        int n = matrix.length;

        // calc path matrix
        int[][] path_matrix = new int[n][n];
        {
            System.arraycopy(matrix, 0, path_matrix, 0, matrix.length);
            
            TransitiveReduction.transformToPathMatrix(path_matrix);
            System.out.println(Arrays.deepToString(path_matrix) + " path matrix");
            Assert.assertArrayEquals(expected_path_matrix, path_matrix);
        }

        // calc transitive reduction
        {
            int[][] transitively_reduced_matrix = new int[n][n];
            System.arraycopy(path_matrix, 0, transitively_reduced_matrix, 0, matrix.length);
            
            TransitiveReduction.transitiveReduction(transitively_reduced_matrix);
            System.out.println(Arrays.deepToString(transitively_reduced_matrix) + " transitive reduction");
            Assert.assertArrayEquals(expected_transitively_reduced_matrix, transitively_reduced_matrix);
        }
    }
}

test ouput

[[0, 1, 1, 0, 0], [0, 0, 0, 0, 0], [0, 0, 0, 1, 1], [0, 0, 0, 0, 1], [0, 1, 0, 0, 0]] original matrix
[[0, 1, 1, 1, 1], [0, 0, 0, 0, 0], [0, 1, 0, 1, 1], [0, 1, 0, 0, 1], [0, 1, 0, 0, 0]] path matrix
[[0, 0, 1, 0, 0], [0, 0, 0, 0, 0], [0, 0, 0, 1, 0], [0, 0, 0, 0, 1], [0, 1, 0, 0, 0]] transitive reduction