Graphs and Complexity

6) Shortest Paths Problem

[email protected]
LS2N, bât. 34, bureau 303
tél. 02.51.12.58.16
 Plan

 Introduction
 Variant « Paths with the same origin »
 Dijkstra’s algorithm
 Bellman-Ford algorithm
 Particular case: DAGs
 Variant « All pairs of vertices »
 Johnson’s algorithm
 Floyd-Warshall algorithm

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 2
 Introduction

 Find a « best » path between two

cities/villages A and B.

 « Best » may be:

 Shortest (km)
 Fastest (minutes)
 Chepeast (€)
 With a minimum total difference in
height (meters)
© Anders Elias, www.data.gouv.fr, m.a.j. 25
mai 2023  With a minimum fuel consumption
(liters).
 …
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 3
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 4
 Convention/Terminology

G = (S, A, v) with weight function v : A → R

G is assumed to be directed.

Note. Every algorithm for paths works for undirected graphs as well.

Why ?!

Weight of a path c = ( (s0,s1), (s1,s2), …, (sk-1,sk) ):

𝑘

𝑝 𝑐 = ෍ 𝑣(si−1𝑠𝑖 )
𝑖=1
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 5
 Terminology

(Weighted) Distance from s to t :

𝑑 𝑠, 𝑡 = min 𝑝 𝑐 , 𝑐 path from 𝑠 to 𝑡 ∪ +∞

Shortest path from s to t : when 𝑑 𝑠, 𝑡 ≠ +∞, path c from s to t such that

𝑝 𝑐 = 𝑑 𝑠, 𝑡 .

a 5 b a 5 b
3 3
5 5
s 1 1 1 3 s 1 1 1 3
5 5
c 5 t c 5 t
3 3 9
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 6
 Existence of shortest paths

Property
Let 𝑠 ∈ 𝑆. We have d(s, t) > -∞ for all 𝑡 ∈ 𝑆 iff
G has no circuit of weight < 0 that is accessible from s.

5
a b
3
5
s 1 1
1 -3
5
5
c d
3

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 7
 Problem and variants
(Generic) Problem : given s, t ∈ S compute d(s, t)

Variants
Paths with the same origin (or same destination/target):
(according to the hypothesis)
Dijkstra’s algorithm ϴ(n2) or ϴ((n+m)logn)
Bellman-Ford algorithm ϴ(nm)

All pairs of vertices

Johnson’s algorithm ϴ(n2logn+nm) or ϴ(nmlog n)
(sparse graphs)

Floyd-Warshall algorithm ϴ(n3)

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 8
 Plan

 Introduction
 Variant « Paths with the same origin »
 Dijkstra’s algorithm
 Bellman-Ford algorithm
 Particular case: DAGs
 Variant « All pairs of vertices »
 Johnson’s algorithm
 Floyd-Warshall algorithm

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 9
 Paths with the same origin
Weighted digraph G = (S, A, v) with v : A → R
s fixed vertex, s∈ S

Problem : for each t  S compute

d(s, t) = min { p(c) ; c path from s to t }  {+}

5 3 5 8
a b a b
3 3
5 0 5
s 1 1 1 3 s 1 1 1 3
5 5
c 5 d c 5 d
3 3 4 9

Result : a shortest paths tree, i.e. a tree whose branches are the shortest paths.
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 10
 Shortest paths tree

Tree with root s whose branches are the shortest paths

3 8 3 8
5 5
a b a b
3 3
0 5 0 5
s 1 1 s 1 1
1 3 1 3
5 5
5 5
c d c d
3 4 9 3 4 9

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 11
 Basic properties

Property: G = (S, A, v)
Let c be a shortest path from p to r whose next-to-last vertex is q.
Then d(p, r) = d(p, q) + v(q, r).

p q r

Property: G = (S, A, v)
Let c be a path from p to r whose next-to-last vertex is q.
Then d(p, r) ≤ d(p, q) + v(q, r).

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 12
 Relaxation
Compute the values d(s,t) by successive approximations
d’(r)
tS d’(t) = estimation of d(s, t)
(t) = predecessor of t, s q r
i.e. next-to-last vertex of a path
from s to t whose weight is d’(t) d’(q)

Initialisation of d’ and  function INIT

for each t  S do
d’(t)  ;(t)  -1;
endfor
d’(s)  0;

Relaxation of the arc (q, r) function RELAX(q, r)

if d’(q) + v(q, r) < d’(r) then
d’(r)  d’(q) + v(q, r); (r)  q;
endif
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 13
 Plan

 Introduction
 Variant « Paths with the same origin »
 Dijkstra’s algorithm
 Bellman-Ford algorithm
 Particular case: DAGs
 Variant « All pairs of vertices »
 Johnson’s algorithm
 Floyd-Warshall algorithm

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 14
 Dijkstra’s algorithm

Condition : v(q, r)  0 for each arc (q, r)

begin
INIT;
Q  S;
while Q   do
let q satisfy d’(q)= min{d’(v), v ∈ Q} ;
Q  Q - {q};
for each successor r of q such that r ∈ Q do
RELAX(q, r) ;
endfor
endwhile
end

This is a greedy algorithm.

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 15
 Running time

Using an adjacency matrix

global running time ϴ(n2)

Using adjacency list(s)

Q : implemented as a balanced search tree (AVL, fr.), ordered
according to the values d’(.)
|S| operations min : ϴ(n log n)
|A| operations RELAX : ϴ(m log n)
Global running time ϴ((n+m) log n)

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 16
 Variant: Dijkstra-Moore algorithm
Condition : none

begin
INIT;
QS;
while Q   do
let q satisfy d’(q)= min{d’(v), v ∈ Q} ;
Q  Q - {q};
for each successor r of q do
old  d’(r)
RELAX(q, r) ;
if old ≠ d’(r) then
Q  Q ∪ {r}
endif
endfor Running time:
endwhile Exponential !
end
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 17
 Plan

 Introduction
 Variant « Paths with the same origin »
 Dijkstra’s algorithm
 Bellman-Ford algorithm
 Particular case: DAGs
 Variant « All pairs of vertices »
 Johnson’s algorithm
 Floyd-Warshall algorithm

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 18
 Approach

 Apply RELAX once (1st) to each arc = visit (and relax along) all the paths
of length 1 of the graph.
 Apply RELAX again (2nd) to each arc = visit (and relax along) all the paths
of length 2 of the graph.
 …
 Apply RELAX again (n-1 –th) to each arc = visit (and relax along) all the
paths of length n-1 of the graph.

The longest path without a cycle has n-1 arcs.

→ The n-th call of RELAX cannot improve the results, unless a negative cycle
exists.
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 19
 Bellman-Ford algorithm
No condition: for each arc (q, r), v(q, r)  R

begin
INIT;
QS;
for i  1 to |S| -1 do
for each (q, r)  A do
RELAX(q, r) ;
endfor
endfor
for each (q, r)  A do
if d’(q) + v(q, r) < d’(r) then
return « the graph contains a negative cycle »
endif
endfor
Running time:
return « the weights are computed in d’() » ϴ(nm)
end
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 20
 Plan

 Introduction
 Variant « Paths with the same origin »
 Dijkstra’s algorithm
 Bellman-Ford algorithm
 Particular case: DAGs
 Variant « All pairs of vertices »
 Johnson’s algorithm
 Floyd-Warshall algorithm

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 21
 When G is a DAG

No condition : for each arc (q, r), v(q, r)  R

Assume the topological order has been computed, and we want the
sortest paths with origin=the first vertex in the order.

begin
INIT;
for each q  S in topological order do Running time:
for each successor r of q do ϴ(n+m) !
RELAX(q, r) ;
endfor
endfor
end

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 22
 Example
-6
a b
3
5
s -1
-1
5
Topological order
5
c d c, s, a, b, d
-3

5
-1
c 5 s 3 a -6 b -1 d

-3 5
What to do if our initial vertex is not c ?
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 23
 Plan

 Introduction
 Variant « Paths with the same origin »
 Dijkstra’s algorithm
 Bellman-Ford algorithm
 Particular case: DAGs
 Variant « All pairs of vertices »
 Johnson’s algorithm
 Floyd-Warshall algorithm

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 24
 Variant « All pairs of vertices »

G = (S, A, v) weighted digraph

S = {1,2,…,n} v : A → R

Matrix of weights: W = (W [i,j]) with

0 if i = j
W [i,j] = v (i,j) if (i, j)  A
 otherwise

Problem. Compute the matrix of distances:

D = ( d(i, j) | 1 ≤ i, j ≤ n )
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 25
 Johnson’s algorithm

 Returns the matrix of distances, or

 Indicates that the input graph contains a negative cycle

Approach :
❖ define a new weight ≥ 0 on the arcs, paying attention to keep
the shortest paths unchanged
❖ apply Dijkstra’s algorithm for each vertex

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 26
 Reweighting
Property. G=(S,A,v) weighted digraph v: A→ R. Let h: S→ R be an arbitrary
function.
Define v’: A→ R as v’(s,t)=v(s,t)+h(s)-h(t).

Then for every path c = ( (s0,s1), (s1,s2), …, (sk-1,sk) ) we have

p’(c)=p(c)+h(s0) - h(sk), where p’ is the path weight obtained with v’.

+ -
+ -
+ -
+ -
+ -
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 27
 Conserved features

 The relative lenghts of the paths

if c, c’ are paths from s0 to sk such that p(c)≤ p(c’)
then p’(c)≤ p’(c’)
The shortest paths are the same!

 The circuits with negative weight

in this case s0=sk thus p(c)=p(c’)
The nature of the graph doesn’t change!

Goal: Find a function h that yields positive or 0 weights v’, and then look
for the shortest paths with respect to v’.
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 28
 Positive or 0 weights

G=(S,A,v) weighted digraph

Property. Let w be a new vertex, and let G’=(S∪{w},A∪ {ws | s∈ S},v),

where by definition v(ws)=0 for each s∈ S.
Then G has a negative cycle if and only if G’ has a negative cycle.
(deduce what this means in terms of shortest paths!)

Let h(s)=d(w,s) for each s∈ S.

→ h(t) ≤ h(s) + v(s,t),  s,t (triangular inequality)

→ v’(s,t)=v(s,t)+h(s)-h(t)≥ 0,  s,t
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 29
 Example

0
0

1 0 2
1/-1 1 2/0 1
0
7 2
8 2 w 0 0 3
w -1 -4 7 0
0
4 5 3
9
4/0 3/-4

4
0

Weight vv Weight v’

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 30
Algorithm Johnson(G=(S,A,v))
Johnson’s algorithm
begin
G’← (S∪ {w}, A∪ {ws | s∈ S} ) ;
for each s ∈ S do v(w,s)  0 endfor
Bellman-Ford (G’,w,v) // detects a negative cycle or computes the distances
if there exists a negative cycle then stop
else
for each vertex s∈ S∪ {w} do
h(s) ← d(w,s) endfor // computed above using Bellman-Ford
for each arc (s,t)∈ A∪ {ws | s∈ S} do
v’(s,t)← v(s,t)+h(s)-h(t) endfor //reweighting
for each vertex s∈ S do
Dijkstra(G,s,v’) // obtains the distances d’(s,u), u ∈ S, with respect to v’
for each vertex u ∈ S do
d(s,u)=d’(s,u)-h(s)+h(u) // computes the distances with respect to v
endfor endfor
endif
end
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 31
 Running time

Using adjacency (successors) list(s)

Bellman-Ford : ϴ(nm)
Dijkstra : ϴ((n+m)log n) repeated n times

Total running time : ϴ(nm logn)

May be implemented in ϴ(n2log n + nm).

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 32
 Plan

 Introduction
 Variant « Paths with the same origin »
 Dijkstra’s algorithm
 Bellman-Ford algorithm
 Particular case: DAGs
 Variant « All pairs of vertices »
 Johnson’s algorithm
 Floyd-Warshall algorithm

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 33
 Dynamic programming

 In general, for optimisation problems (min/max of a cost function)

 « Divide and conquer », but:

 Separate the search of the min/max cost and the identification of a solution
reaching the min/max cost.
 Express the min/max of the problem as a recursive formula that uses the costs
obtained for one or several subproblems.
 Solve the easiest subproblems first, and store their resulting costs in a
matrix/vector.
 Record the information necessary to construct the final solution, once the best cost
is found.
 Use iterative algorithms/programming, not recursive ones.

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 34
 Basic remarks

Condition: no negative cycle in G

Property. If ( (s0,s1), …, (si ,si+1), …, (sj -1, si), …, (sk-1,sk) ) is a

shortest path from s0 to sk in G, then
( (si ,si+1), …, (sj-1,sj) ) is a shortest path from si to sj in G.

shortest

s0 si sj sk

shortest

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 35
 Approach
Notation
Dk = (Dk [i, j] | 1  i, j  n ) with

Dk [i, j] = min{ W(c) | c path from i to j whose

intermediate vertices have numbers  k }

D0 = W
Dn = matrix of distances for G
(= the matrix D we want to compute) i k j

no k no k
Property. For each k  1,
Dk [i,j] = min{ Dk-1 [i,j] , Dk-1 [i,k] + Dk -1 [k,j] }

Computation :
of Dk using Dk-1 in ϴ(n2)
of D = Dn in ϴ(n3)
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 36
 Floyd-Warshall algorithm

for k  1 to n do
for i  1 to n do
for j  1 to n do
D[i, j]  min { D[i, j] , D[i, k] + D[k, j] } endfor …

k j

k c
Dk = min { a, b + c }
i b a

a
i j
k
b c
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 37
Example
1
1 2 0 1 8
8 2 D0 = W =  0 4
0 4 7  7 0 9
0 2 0
4 3
9 0 18
D1 =  0 4
 7 0 9
0 1 0

0 1 58
D2 =  0 4
 7 0 9
0 1 5 0

0 1 5 8
D3 =  0 4 13
 7 0 9
0 1 5 0

0 1 5 8
D4 = 13 0 4 13
9 7 0 9
0 1 5 0

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 38
 Recovering the paths

Record all the paths from i to j in an explicit way, for 1 ≤ i, j ≤ n:

n2 paths of maximal length n-1 : O(n3) memory space

Matrix recording the predecessors: ϴ(n2 ) memory space

Pk = (Pk [i, j] | 1 ≤ i, j ≤ n ) with
Pk [i, j] = predecessor of j on a shortest path from i to j whose
intermediate vertices have numbers  k
Dk-1[i, j]
Computation: i s
j
k t
i if (i, j)  A Dk-1[i, j] Dk-1[i, j]
P0 [i, j] = - otherwise

Pk [i, j] = Pk-1 [i, j] if Dk-1[i, j]  Dk-1[i, k] + Dk-1[k, j]

Pk-1 [k, j] otherwise
_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 39
 1
1 2 0 18 - 1 - 1
D0 = W = 0 4 - - 2 -
0 8 2 7 0 9 P0 =
4 7 - 3 - 3
0 2 0 4 4 - -
4 3
9 0 18
0 4 - 1 - 1
D1 = P1 = - - 2 -
7 0 9
0 1 0 - 3 - 3
4 1 - -
0 1 58
0 4 - 1 2 1
D2 = P2 = - - 2 -
7 0 9
0 1 5 0 - 3 - 3
4 1 2 -
0 1 5 8 - 1 2 1
D3 = 0 4 13 - - 2 3
7 0 9 P3 =
- 3 - 3
0 1 5 0 4 1 2 -

0 1 5 8 - 1 2 1
D4 = 13 0 4 13 4 - 2 3
9 7 0 9 P4 = 4 3 - 3
0 1 5 0 4 1 2 -

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 40
 1
1 2
0 8 2 0 1 5 8 - 1 2 1
4 7 D4 = 13 0 4 13 4 - 2 3
9 7 0 9 P4 = 4 3 - 3
4 3 0 1 5 0 4 1 2 -
9

Example of a path
distance from 2 to 1 : D4[2,1] = 13
P4[2,1] = 4 ; P4[2,4] = 3 ; P4[2,3] = 2 ;

4 9 0
2 3 4 1

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 41
 In conclusion

 The shortest path problem has many applications :

 In real-life (all networks)
 In the modelisation of other problems (reconstruction of biological sequences
from segments, for instance)

 The algorithms are simple and as efficient as possible

_______________________________________________________________________________________________________________________________
M1 Informatique – Graphs and Complexity 42