Shortest Path Algorithm
The Problem Statement
• Given a directed graph G=(V,E) such that each
edge is having a weight possibly negative but
no negative cycle, then find the shortest path
from a vertex s to sink vertex t.
• Dijkstra’s algorithm worked for Shortest Path
Problem when there are no negative costs.
Is negative edge required
• Yes
• In many applications the negative edge makes
sense.
• For example, the nodes may represent agents in a
financial setting and 𝑐𝑖𝑗 represents the cost of a
transaction in which we buy from agent 𝑖 and
immediately sell to agent 𝑗. So a negative weight
indicates profitable transactions.
Failure of Dijkstra’s Algorithm
• If we run Dijkstra’s algorithm, first u
visited vertex will be {s}. The next 2 3
unvisited vertex with least distance s t
from s is v.
1 v -6
• As per Dijkstra’s algorithm the
shortest path from s to v is 1 with the
single edge {s,v}.
• The path {s,v} is clearly the shortest to
v if there are no negative edge costs:
any other path from s to v would have
to start on an edge out of s that is at
least as expensive as edge (s,v).
Modify the cost of edge
• A natural idea is to increase 2
u
2
the cost of all edges by some s t
large constant M. Then all 3 3
v v
costs will be positive. -3
• This does not work. Add 3 to
all the edge weights and
check!
Dynamic Programming Approach
• Important Observation:
If G has no negative cycles, then there is a shortest path from s to t that is simple
(i.e. does not repeat nodes), and hence has at most n-1 edges.
Proof: Since every cycle has non-negative cost, the
shortest path P from s to t with the fewest number
of edges does not repeat any vertex v. For if P did
repeat a vertex, we could remove the portion of P
between consecutive visits to v, resulting in a path of
no greater cost and fewer edges.
Opt(i,v)
• Let 𝑂𝑃𝑇(𝑖, 𝑣) to denote the minimum cost of
a 𝑣 − 𝑡 path using at most 𝑖 edges.
• Thus we have to compute 𝑂𝑃𝑇(𝑛 − 1, 𝑠).
v t
w
• Let’s fix an optimal path P representing 𝑂𝑃𝑇 𝑖, 𝑣
as shown above.
• If the path P uses at most 𝑖 − 1 edges, then
𝑂𝑃𝑇 𝑖, 𝑣 = 𝑂𝑃𝑇 𝑖 − 1, 𝑣
• If the path P uses 𝑖 edges, and the first edge is
𝑣, 𝑤 , then 𝑂𝑃𝑇 𝑖, 𝑣 = 𝑐𝑣𝑤 + 𝑂𝑃𝑇(𝑖 − 1, 𝑤).
Recursion
If i>0 then
𝑂𝑃𝑇 𝑖, 𝑣 = min{𝑂𝑃𝑇 𝑖 − 1, 𝑣 , min(𝑐𝑣𝑤 + 𝑂𝑃𝑇(𝑖 − 1, 𝑤))}
Pseudo code
Shortest-Path (G,s,t)
n = number of nodes in G
Array M[0…n-1, V]
Define M[0,t] = 0 and M[0,v] = inf for all other vertices
For i = 1 to n-1
for 𝑣𝜖𝑉 in any order
Compute 𝑀[𝑖, 𝑣] using the given recurrence.
Endfor
Endfor
Return 𝑀[𝑛 − 1, 𝑠]
Running Time
• A loose analysis says that this algorithm runs in 𝑂(𝑛3 ) time
where there are quadratic entries and each entry requires
to compute the recurrence which is linear in time.
• 𝑀[ 𝑖, 𝑣 ] = min{𝑀[ 𝑖 − 1, 𝑣 ], min(𝑐𝑣𝑤 + 𝑀[(𝑖 − 1, 𝑤)]}
• Note that the inner minimum is computed only for
neighbors w of v. So it takes 𝑛𝑣 , where that is the number
of outgoing edges at 𝑣.
• So it takes 𝑂(𝑛𝑣 ) to compute the array entry 𝑀 𝑖, 𝑣 .
• We have to compute an entry for every v and every index
0 ≤ 𝑖 ≤ 𝑛 − 1, so running time bound is
• 𝑂 𝑛 σ𝑣𝜖𝑉 𝑛𝑣 = 𝑂(𝑚𝑛)
Pointer Graph
• To help with recovering the shortest path,
each node v maintains the first node (after
itself) on its path to the destination t;
• We will call this vector 𝑓𝑖𝑟𝑠𝑡[𝑣].
• Let P denote the directed “pointer graph”
whose nodes are V and whose edges are {(𝑣,
𝑓𝑖𝑟𝑠𝑡[𝑣])}.
Detecting a cycle
• If the pointer graph P contains a cycle C, then this
cycle must have negative cost.
• Proof: Notice that if 𝑓𝑖𝑟𝑠𝑡 𝑣 = w, then 𝑀 𝑣 ≥
𝑐𝑣𝑤 + 𝑀[𝑤].
• Let 𝑣1 , 𝑣2 , … , 𝑣𝑘 be the nodes along the cycle in
the pointer graph and assume that (𝑣𝑘 , 𝑣1 ) is the
last edge to have been added. Now 𝑀 𝑣𝑖 ≥
𝑐𝑣𝑖 𝑣𝑖+1 + 𝑀[𝑣𝑖+1 ] for all 𝑖 = 1, … , 𝑘 − 1, and also
have 𝑀 𝑣𝑘 ≥ 𝑐𝑣𝑘𝑣1 + 𝑀[𝑣1 ] . Adding all the
equations we get 0 > σ𝑘−1 𝑖=1 𝑐𝑣𝑖 𝑣𝑖+1 + 𝑐𝑣𝑘 𝑣1 . Thus
it is a negative cycle.
Reference
• Algorithm Design by Eva Tardos and Jon
Kleinberg.
