transitive reduction algorithm: pseudocode? - algorithm

Question

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 track 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?

Answer 1

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;

Answer 2

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

Answer 3

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

Answer 4

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.

Answer 5

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.

Answer 6

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.

Answer 7

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