Introduction
 Note
 This documentation is focused on the developer of pgr_contraction.

This documentation might include documentation used for the user.

The developer should also read the documentation for the user.
Contracting a graph becomes a crucial operation when talking about big graphs like the graphs involved in routing across cities, countries, continents or the whole world.
The contraction level and contraction operations can become very complex, as the complexity of the graphs grows.
For this proposal, we are making our contraction algorithm simple as possible so that more contraction operations can be added in the future.
We are not aiming with this work to implement all the possible contraction operations but to give a framework such that adding a contraction operation can be easily achieved.
For this contraction implementation, there are two operations:
 dead end contraction: vertices have one incoming edge
 linear contraction: vertices have one incoming and one outgoing edge
And with the additional characteristics:
 The user can forbid to contract a particular set of nodes or edges.
 The user can decide how many times the cycle can be done.
 If possible, the user can decide the order of the operations on a cycle.
The contraction skeleton
In general we have an initial set up that may involve analizing the graph given as input and setting the non contractable nodes or edges. We have a cycle that will go and perform a contraction operation until while possible, and then move to the next contraction operation. Adding a new operation then becomes an "easy" task but more things might be involved, because the charachteristics of the graph change each time its contracted, so some interaction between contractions has to be implemented also.
Procedure
For contracting, we are going to cycle as follows
input: G(V,E)
removed_vertices = {}
<initial set up>
do N times {
while ( <conditions for 1> ) {
< contraction operation 1 >
}
while ( <conditions for 2> ) {
< contraction operation 2>
}
.....
}
output: G'(V',E'), removed_vertices
Notation for this documentation
 V: is the set of vertices
 E: is the set of edges
 G: is the graph
 V1: is the set of dead end vertices
 V2: is the set of linear vertices
 removed_vertices: is the set of removed vertices
The contracted graph will be represented with two parameters, the modified Graph, and the removed_vertices set.
removed_vertices = {(v,1):{2}, (e,1):{3}}.
The above notation indicates:
 Vertex 2 is removed, and belongs to vertex 1 subgraph
 Vertex 3 is removed, and belongs to edge 1 subgraph
Dead End Contraction
Characteristics:
 Contraction code: 1
V1
: set of vertices with 1 incoming edge in increasing order of id:
 Edges with the same identifier are considered the same edge and if it has the
reverse_cost
valid the outgoing edge is ignored
while ( V1 is not empty ) {
the deleted
vertex add it to removed_vertices
vertex that leads to removed
vertex, inherits the removed vertex
<adjust any conditions that might affect other contraction operation>
}
Linear Contraction
Characteristics:
 Contraction code: 2
V2
: vertex with 1 incoming edge and 1 outgoing edge:
 The outgoing edge must have different identifier of the incoming edge
while ( V2 is not empty ) {
the deleted
vertex add it to removed_vertices
<adjust any conditions that might affect other contraction operations>
}
Contraction examples
Dead End Contraction
 For simplicity all the edges in the examples have unit weight.
 Perform dead end contraction
 One cycle of contractions.
Input:
G = {E:{1(1, 2)}, V:{1, 2} }
Visually the graph is
 Todo:
 (Rohith) practice a little graphviz :)
Initial set up:
removed_vertices={}
V1 = {2}
V2 = {}
Procedure:
V1 = {2} is not empty
V1 = {}
V2 = {}
G = {V:{1}, E:{}}
removed_vertices = {(v, 1):{2}}.
V1 is empty
Results:
G = {V:{1}, E:{}}
removed_vertices = {(v, 1):{2}}
~~~~{.c}
Visually the graph is
@image html ../doc/images/twoNodesoneEdge_b.png
@subsection contraction_examples_linear Linear Contraction
@subsection contraction_examples_linear_dead_end Linear and Dead End Contraction
 For simplicity all the
edges in the example have unit weight.
 A weight different than 1 will be expressed explicitly.
Perform linear contraction operation followed by a dead end operation.
1 cycle of contraction.
1. Input: G = {V:{1, 2, 3}, E:{(1, 2), (2, 3)}}
Visually the graph is
@image html ../doc/images/threeNodestwoEdges_a.png
2. initial set up
~~~~{.c}
removed_vertices={}
V1 = {3}
V2 = {2}
 Procedure
V2 = {2} is not empty
V1 = {3}
removed_vertices = {(e, 1):{2}}
V2 = {}
G = {V:{1, 3}, E:{1(1,3,c=2)}}
V2 is empty
Visually the graph is
 Since V2 is empty we go on to the next contraction operation
V1 = {3} is not empty
V1 = {}
V2 = {}
removed_vertices = {(v, 1):{3, 2}}.
G = {V:{1}, E:{}}
V1 is empty
 Results
removed_vertices = {(v, 1):{3, 2}}.
G = {V:{1}, E:{}}
Visually the graph is
Contraction of Sample Data
 A weight different than 1 will be expressed explicitly. Perform dead end operation follwed with a linear operation. 1 cycle of contraction.
 initial set up
G = { V:{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17},
E:{ 1(1, 2), 2(2,3), 3(3,4), 4(2,5),
5(3,6), 6(7,8), 7(8,5), 8(5,6),
9(6,9), 10(5,10), 11(6,11), 12(10,11),
13(11,12), 14(10,13), 15(9,12), 16(4,9),
17(14,15), 18(16,17) }
 Todo:
 (Rohith) add the id of each edge to the image
Visually the graph is
 Initial set up:
removed_vertices={}
V1 = {1,7,13,14,15,16,17}
V2 = {4,8,12}
 Dead end Contraction
The dead end vertices can visually be seen in the next graph.
 The contraction is done in ascending order of the elements of V1
 Todo:
 (Rohith): add the id of each edge to the image, put a color to the dead end vertices
V1 = {1,7,13,14,15,16,17}
V1 = {7,13,14,15,16,17}
V2 = {2,4,8,12}
G = {V:{2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17},
E:{2(2,3), 3(3,4), 4(2,5), 5(3,6), 6(7,8), 7(8,5), 8(5,6), 9(6,9),
10(5,10), 11(6,11), 12(10,11), 13(11,12), 14(10,13), 15(9,12), 16(4,9), 17(14,15), 18(16,17)}}
removed_vertices = {(v, 2):{1}}.
V1 = {7,13,14,15,16,17}
V1 = {8,13,14,15,16,17}
V2 = {2,4,12}
G = {V:{2, 3, 4, 5, 6, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17},
E:{2(2,3), 3(3,4), 4(2,5), 5(3,6), 7(8,5), 8(5,6), 9(6,9), 10(5,10),
11(6,11), 12(10,11), 13(11,12), 14(10,13), 15(9,12), 16(4,9), 17(14,15), 18(16,17)}}
removed_vertices = {(v, 2):{1}, (v,8):{7}}.
V1 = {8,13,14,15,16,17}
V1 = {13,14,15,16,17}
V2 = {2,4,12}
G = {V:{2, 3, 4, 5, 6, 9, 10, 11, 12, 13, 14, 15, 16, 17},
E:{2(2,3), 3(3,4), 4(2,5), 5(3,6), 8(5,6), 9(6,9), 10(5,10),
11(6,11), 12(10,11), 13(11,12), 14(10,13), 15(9,12), 16(4,9), 17(14,15), 18(16,17)}}
removed_vertices = {(v, 2):{1}, (v,5):{8,7}}.
V1 = {13,14,15,16,17}
V1 = {14,15,16,17}
V2 = {2,4,10,12}
G = {V:{2, 3, 4, 5, 6, 9, 10, 11, 12, 14, 15, 16, 17},
E:{2(2,3), 3(3,4), 4(2,5), 5(3,6), 8(5,6), 9(6,9), 10(5,10),
11(6,11), 12(10,11), 13(11,12), 15(9,12), 16(4,9), 17(14,15), 18(16,17)}}
removed_vertices = {(v, 2):{1}, (v,5):{8,7}, (v,10):{13}}.
V1 = {14,15,16,17}
V1 = {16,17}
V2 = {2,4,10,12}
G = {V:{2, 3, 4, 5, 6, 9, 10, 11, 12, 15, 16, 17},
E:{2(2,3), 3(3,4), 4(2,5), 5(3,6), 8(5,6), 9(6,9), 10(5,10),
11(6,11), 12(10,11), 13(11,12), 15(9,12), 16(4,9)}, 18(16,17)}
removed_vertices = {(v, 2):{1}, (v,5):{8,7}, (v,10):{13}, (v,15):{14}}.
V1 = {16,17}
V1 = {}
V2 = {2,4,10,12}
G = {V:{2, 3, 4, 5, 6, 9, 10, 11, 12, 15, 17},
E:{2(2,3), 3(3,4), 4(2,5), 5(3,6), 8(5,6), 9(6,9), 10(5,10),
11(6,11), 12(10,11), 13(11,12), 15(9,12), 16(4,9)}}
removed_vertices = {(v, 2):{1}, (v,5):{8,7}, (v,10):{13}, (v,15):{14}, (v,17):{16}}.
3 Since V1 is empty, linear contraction is performed
The linear vertices can visually be seen in the next graph.
 The contraction is done in ascending order of the elements of V2
 Todo:
 (Rohith) add the id of each edge to the image, put a color to the linear vertices
V2 = {2,4,10,12}
V1 = {}
V2 = {4,10,12}
G = {V:{3, 4, 5, 6, 9, 10, 11, 12, 15, 17},
E:{1(3,5), 3(3,4), 5(3,6), 8(5,6), 9(6,9), 10(5,10),
11(6,11), 12(10,11), 13(11,12), 15(9,12), 16(4,9)}}
removed_vertices = {(e, 1):{1,2}, (v, 2):{1}, (v,5):{8,7}, (v,10):{13}, (v,15):{14}, (v,17):{16}}.
V2 = {4,10,12}
V1 = {}
V2 = {10,12}
G = {V:{3, 5, 6, 9, 10, 11, 12, 15, 17},
E:{1(3,5),2(3,9), 5(3,6), 8(5,6), 9(6,9), 10(5,10),
11(6,11), 12(10,11), 13(11,12), 15(9,12)}}
removed_vertices = {(e, 1):{1,2}, (e, 2):{4}, (v, 2):{1}, (v,5):{8,7}, (v,10):{13}, (v,15):{14}, (v,17):{16}}.
V2 = {10,12}
V1 = {}
V2 = {12}
G = {V:{3, 5, 6, 9, 11, 12, 15, 17},
E:{1(3,5),2(3,9), 3(5,11), 5(3,6), 8(5,6), 9(6,9),
11(6,11), 13(11,12), 15(9,12)}}
removed_vertices = {(e, 1):{1,2}, (e, 2):{4}, (e,3):{10,13}, (v, 2):{1}, (v,5):{8,7}, (v,15):{14}, (v,17):{16}}.
V2 = {12}
V1 = {}
V2 = {}
G = {V:{3, 5, 6, 9, 11, 15, 17},
E:{1(3,5),2(3,9), 3(5,11), 4(9,11), 5(3,6), 8(5,6), 9(6,9), 11(6,11)}}
removed_vertices = {(e, 1):{1,2}, (e, 2):{4}, (e,3):{10,13}, (e, 4):{12},
(v, 2):{1}, (v,5):{8,7}, (v,15):{14}, (v,17):{16}}.
 Results:
G = {V:{3, 5, 6, 9, 11, 15, 17},
E:{1(3,5),2(3,9), 3(5,11), 4(9,11), 5(3,6), 8(5,6), 9(6,9), 11(6,11)}}
removed_vertices = {(e, 1):{1,2}, (e, 2):{4}, (e,3):{10,13}, (e, 4):{12},
(v, 2):{1}, (v,5):{8,7}, (v,15):{14}, (v,17):{16}}.
Visually the graph is
Generating the contracted graph
Original Data:
.. literalinclude:: doccontraction.queries :startafter: – q1 :endbefore: – q2
:Addition of new columns:
.. codeblock:: sql
ALTER TABLE edge_table ADD is_contracted BOOLEAN DEFAULT false
ALTER TABLE edge_table ADD contracted_vertices BIGINT[]
Data After Adding Columns:
.. literalinclude:: doccontraction.queries
:startafter:  q4
:endbefore:  q5
Addition of new edges by the algorithm:
.. codeblock:: sql
INSERT INTO edge_table(id, source, target, cost, reverse_cost, is_contracted, contracted_vertices) VALUES (1, 3, 5, 2, 2, true, ARRAY[1, 2])
INSERT INTO edge_table(id, source, target, cost, reverse_cost, is_contracted, contracted_vertices) VALUES (2, 3, 9, 2, 2, true, ARRAY[4])
INSERT INTO edge_table(id, source, target, cost, reverse_cost, is_contracted, contracted_vertices) VALUES (3, 5, 11, 2, 2, true, ARRAY[10, 13])
INSERT INTO edge_table(id, source, target, cost, reverse_cost, is_contracted, contracted_vertices) VALUES (4, 9, 11, 2, 2, true, ARRAY[12])
Data after addition of new edges:
.. literalinclude:: doccontraction.queries :startafter: – q9 :endbefore: – q10
Contracted Graph Data:
.. literalinclude:: doccontraction.queries :startafter: – q10 :endbefore: – q11
Dijkstra on contracted graph
There are five cases which arise when calculating the shortest path between a given source and target,
Case 1**: Both source and target belong to the contracted graph.
Case 2**: Source belongs to a contracted graph, while target belongs to a vertex subgraph.
Case 3**: Source belongs to a contracted graph, while target belongs to an edge subgraph.
Case 4**: Source belongs to a vertex subgraph, while target belongs to an edge subgraph.
Case 5**: The path contains a new edge added by the contraction algorithm.
source and target belong to the contracted graph
Routing from 3 to 11. Since 3 and 11 both are in the contracted graph it is not necessary expand the graph.
.. literalinclude:: doccontraction.queries :startafter: – q11 :endbefore: – q12
Source belongs to a contracted graph, while target belongs to a vertex subgraph.
Routing from 3 to 7. Since 7 is in the contracted subgraph of vertex 5, it is necessary to expand that vertex by adding {7, 8} to the vertex set, so the vertex set becomes {3, 5, 6, 9, 11, 15, 17 , 7, 8}
.. literalinclude:: doccontraction.queries :startafter: – q12 :endbefore: – q13
Source belongs to a contracted graph, while target belongs to an edge subgraph.
Routing from 3 to 13. Since 13 is in the contracted subgraph of edge (5, 11), it is necessary to expand that edge by adding {10, 13} to the vertex set, so the vertex set becomes {3, 5, 6, 9, 10, 11, 13, 15, 17}.
.. literalinclude:: doccontraction.queries :startafter: – q13 :endbefore: – q14
Source belongs to a vertex subgraph, while target belongs to an edge subgraph.
Routing from 7 to 13.i
 Since 13 is in the contracted subgraph of edge (5, 11), it is necessary to expand that edge by adding {10, 13} to the vertex set,
 and since 7 is in the contracted subgraph of vertex 5, it is necessary to expand that vertex by adding {7, 8} vertex set,
 so the vertex set becomes {3, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17}
.. literalinclude:: doccontraction.queries :startafter: – q14 :endbefore: – q15
The path contains a new edge added by the contraction algorithm.
Routing from 3 to 9. Since 3 and 9 both are in the contracted graph it is not necessary expand the graph.
.. literalinclude:: doccontraction.queries :startafter: – q15 :endbefore: – q16
.. literalinclude:: doccontraction.queries :startafter: – q16 :endbefore: – q17
 This implies that it is a shortcut and should be expanded.
 The contracted subgraph of edge 20(3, 9, c=2) is {4}.
 It is necessary to expand the edge by adding {4} to the vertex set,
 so the vertex set becomes {3, 4, 5, 6, 9, 11, 15, 17}.
.. literalinclude:: doccontraction.queries :startafter: – q17 :endbefore: – q18
References
http://www.cs.cmu.edu/afs/cs/academic/class/15210f12/www/lectures/lecture16.pdf
http://algo2.iti.kit.edu/documents/routeplanning/geisberger_dipl.pdf