Dijkstra’s Algorithm

SSSP, non-neg

Edge weights = w(x,y)

Final distances = d(x,y) = dw(x,y)

2 b 2

s 2 e
6 2
c
6 1

d
Dijkstra’s Algorithm
SSSP, non-neg

s b c d e
L(x) 0 ∞ ∞ ∞ ∞

2 b 2

s 2 e
6 2
c x <- extractmin

6 1 // L(x) is the distance of s to x

// mark x as final
d
"relax” all the edges out of x
L(y) <- min ( L(y), L(x) + w(x,y) )
Dijkstra’s Algorithm
SSSP, non-neg

s b c d e
L(x) 0 2 6 6 ∞

2 b 2

s 2 e
6 2
c x <- extractmin

6 1 // L(x) is the distance of s to x

// mark x as final
d
"relax” all the edges out of x
L(y) <- min ( L(y), L(x) + w(x,y) )
Dijkstra’s Algorithm
SSSP, non-neg

s b c d e
L(x) 0 2 4 6 4

2 b 2

s 2 e
6 2
c x <- extractmin

6 1 // L(x) is the distance of s to x

// mark x as final
d
"relax” all the edges out of x
L(y) <- min ( L(y), L(x) + w(x,y) )
Dijkstra’s Algorithm
SSSP, non-neg

s b c d e
L(x) 0 2 4 6 4

2 b 2

s 2 e
6 2
c x <- extractmin

6 1 // L(x) is the distance of s to x

// mark x as final
d
"relax” all the edges out of x
L(y) <- min ( L(y), L(x) + w(x,y) )
Dijkstra’s Algorithm
SSSP, non-neg

s b c d e
L(x) 0 2 4 5 4

2 b 2

s 2 e
6 2
c x <- extractmin

6 1 // L(x) is the distance of s to x

// mark x as final
d
"relax” all the edges out of x
L(y) <- min ( L(y), L(x) + w(x,y) )
Dijkstra’s Algorithm m decreasekeys
SSSP, non-neg n extractmins

Fib heap: O(m + n log n)

s b c d e
L(x) 0 2 4 5 4

2 b 2

s 2 e
6 2
c x <- extractmin

6 1 // L(x) is the distance of s to x

// mark x as final
d
"relax” all the edges out of x
L(y) <- min ( L(y), L(x) + w(x,y) )
Bellman-Ford-Moore Algorithm
SSSP, neg wts OK
Bellman-Ford-Moore Algorithm n rounds
SSSP, neg wts OK m time per round
O(mn) time

s b c d e
L(x) 0 ∞ ∞ ∞ ∞

// L(x) is an upper bound on d(s,x)

2 b 2

2 e for t = 1 to n-1
s
6 -2 for all vertices x
c // "relax” all the edges out of x
6 1 L(y) <- min ( L(y), L(x) + w(x,y) )

d Claim: if graph has non negative cycles, B-F-M is OK.

Proof: induction. At end of round t, L(y) is shortest
path using at most t edges.
Bellman-Ford-Moore Algorithm n rounds
SSSP, neg wts OK m time per round
O(mn) time

s b c d e
L(x) 0 2 2 5 4

// L(x) is an upper bound on d(s,x)

2 b 2

2 e for t = 1 to n-1
s
6 -2 for all vertices x
c // "relax” all the edges out of x
6 1 L(y) <- min ( L(y), L(x) + w(x,y) )

d Claim: if graph has non negative cycles, B-F-M is OK.

Proof: induction. At end of round t, L(y) is shortest
path using at most t edges.
Bellman-Ford-Moore Algorithm n rounds
SSSP, neg wts OK m time per round
O(mn) time

s b c d e
L(x) 0 2 2 5 4

// L(x) is an upper bound on w(s,x)

2 b 2

2 e for t = 1 to n-1
s
6 -2 for all vertices x
c // "relax” all the edges out of x
6 1 L(y) <- min ( L(y), L(x) + w(x,y) )

d Claim: If at end, some edge xy is “over-tight”

( it has L(y) > L(x) + w(x,y) )
then graph has negative cycle.
Lots more work!
E.g., extensions and implementations of (just) Dijkstra’s algorithm:

slide from Mikkel Thorup

All Pairs SP

Non-neg weights:
n Dijkstras O(mn + n2 log n)

Neg weights:
n B-F-M O(mn2)
Johnson’s Algorithm
APSP, neg wts OK

Find “feasible potentials” ©(x) such that

the “reduced weights”
ŵ(x,y) := ©(x) + w(x,y) - ©(y) ≥ 0

Fact: reduced weights ŵ non-neg

Claim: dw (x,y) = dŵ(x,y) + ©(y) - ©(x)
2 b 2
e So shortest paths don’t change, though
s 2
6 -2 their weights might.
c  suffices to find APSP in this
non-neg weights graph!
6 1
How to find these “feasible potentials”?
d
Johnson’s Algorithm
APSP, neg wts OK

Find “feasible potentials” ©(x) such that

the “reduced weights”
ŵ(x,y) := ©(x) + w(x,y) - ©(y) ≥ 0

Fact: reduced weights ŵ non-neg

Claim: dw (x,y) = dŵ(x,y) + ©(y) - ©(x)
2 b 2
e So shortest paths don’t change, though
s 2
6 -2 their weights might.
c  suffices to find APSP in this
non-neg weights graph!
6 1
0
0 How to find these “feasible potentials”?
d 0
0 Shortest paths from Z.
0 Z No negative cycles created, so B-F-M works.
All Pairs SP

Non-neg weights:
n Dijkstras O(mn + n2 log n)

Neg weights:
n B-F-M O(mn2)

Neg weights:
B-F-M + n Dijkstras O(mn + n2 log n)

Neg weights:
Floyd-Warshall O(n3)

for all vertices z

for all vertices x,y
d(x,y) <- min{ d(x,y), d(x,z) + d(z,y)
All Pairs SP

Non-neg weights:
n Dijkstras O(mn + n2 log n)

Neg weights:
n B-F-M O(mn2)

Neg weights:
B-F-M + n Dijkstras O(mn + n2 log n)

Neg weights:
Floyd-Warshall O(n3)

Neg weights:
Naïve Min-Sum-Product O(n3 log n)