Primal-Dual Algorithm

Self-reducibility Incremental

Dynamic Program

Divide-and-Conquer Primal Primal-Dual

Local
Greedy
Ratio
 Primal-Dual
 An optimality condition comes from the
primal-dual theorem e.g., max-flow=min-cut.
 Start from a pre-solution satisfying the
optimality condition.
 A pre-solution is not feasible, but
approaching to feasibility step by step.
 Eventually, the pre-solution becomes a
feasible solution which is optimal.
 Primal-Feasible
Consider a maximization problem

Optimality condition:
Max Primal = Min Dual
Optimal solution

Primal feasible solutions

 Dual Feasible
Consider a maximization problem

Dual feasible solutions coincide with pre-primal solution

Optimal solution
Optimality condition:
Max Primal = Min Dual
 Lecture 16

Max Flows 3
Preflow-Relable Algorithms
Review
 The Ford Fulkerson Maximum Flow Algorithm

Begin x := 0;
create the residual network G(x);
while there is some directed path from s to t in G(x)
do
begin
let P be a path from s to t in G(x);
∆:= δ(P);
send ∆ units of flow along P;
update the r's;
End
end {the flow x is now maximum}.
The two properties of distance label and this lemma
hold during the whole process of the algorithm
and will be proved after descriptio n of the algorithm.
Push: Saturating

4 3
2 1

1 3 3
2 1
Push: Non-Saturating

3 4
2 1

0 3
1
2 1

3
 Relabel

2 3
4 Min{3,5}+1=4
3
2 5
1
1
1

2 3
4
3
4 5
1
1
1
e(3)=1
Exact distance: d(i) equals the length of the shortest
path from node i to node t in G(x)
The Goldberg-Tarjan Preflow
Push Algorithm for the
Maximum Flow Problem
 Preflow Push

4
2 5
3 1 1 1
2 4
s 4 t
3 2
1

3
This is the original network,
and the original residual
network.
 Initialize Distances

4
5
2 5
1
3 1 1 1
4
2 4
3
s
2 4
1 0t
2 2 s 3 2
1
1 3 4 5
0 t 3
1

The node label henceforth will be the distance label.

d(j) is at most the distance of j to t in G(x)
 Saturate Arcs out of node s

s 3
4
5
2 5
1
3 1 1 1
4
2 4
3
s
2
6 4
1 0t
2 2 s 3 2 2
1
1 3 4 5
0 t 3
1
3

Saturate arcs out of node s.

Move excess to the adjacent arcs
Relabel node s after all incident arcs have been saturated.
 Select, then relabel/push

s 3
4
5
2 5
1
3 1 1 1
4
2 4
3
s
2
6 4
1 0t
2 2 s 3 2 2
1
1 3 4 5
0 t 3
1
3
1

Select an active node, that is, one with excess

Push/Relabel
Update excess after a push
 Select, then relabel/push

s 3
4
5
2 5
1
3 1 1 1
4
2 4
3
s
2
6 4
1 0t
2 2 3
s 3 2
1
1 3 4 5 2
0 t 3
1
2
1

Select an active node, that is, one with excess

No arc incident to the selected node is admissible. So

relabel.
 Select, then relabel/push

s 3
4
5
2 5
1
3 1 1 1
4
2 4
3
s
2
6 4
1 0t
2 2 3
s 3 2
3
1
1 4 5 2
0 t 2
1

Select an active node, that is, one with excess

Push/Relabel
 Select, then relabel/push

s 3 3
4
1
5
2 3
5
1
3 1 1 1
4
2 4
3
s
2
6 4
1 0t
2 2 3
s 3 3
2
1 4 5 1
2
0 t 2
1

Select an active node.

Push/Relabel
 Select, then relabel/push

s 3
2
4
1
5
2 3
5
1
3 1 1 1
4
2 4
3
s
2
6 4
1 0t
2 2 3
s 3 3
2
1 4 5 1
2
0 t 2
1

Select an active node.

Push/Relabel
 Select, then relabel/push

s 3
2
4
1
5
2 3
5
2
1
3 1 1
4 1
2 4
3
s
2
6 4
1 0t
2 2 3
s 5 3 3
2
1 4 5 1
2
0 t 2
1

Select an active node.

There is no incident admissible arc. So Relabel.

 Select, then relabel/push

s 3
2
1
4
1
5
2 3
5
1
2
3 1 1
4 1
2 4
3
s
2
6 4
1 0t
2 2 3
s 5 3 4
2
3
1 4 1
2
0 t 2
1

Select an active node.

Push/Relabel
 Select, then relabel/push

s 3
2
1
4
1
5
2 3
5
1
2
3
3 1 1
4 1
2 4
3 5
s
2
6 4
1 0t
2 2 3
s 5 3 4
2
3
1 4 1
2
0 t 2
1

Select an active node.

There is no incident admissible arc. So relabel.

 Select, then relabel/push

s 3
2
1
4
1
5
2 3
5
1
2
3
3 1 1
4 1
2 4
3 5
s
2
6 4
1 0t
2 2 3
s 3 4
2
3
1 4 5 1
2
0 t 2
1

Select an active node.

Push/Relabel
 Select, then relabel/push

s 1
4
2
5
2 2
5
1
2
3
3 1 1
4 1
2 4
3 5
s
2
6 4
1 0t
2 2 3
s 3 2
3
4
1 4 5 1
2
0 t 2
1

Select an active node.

Push/Relabel
 Select, then relabel/push

s 1
4
2
5
4
2 2
5
1
2
3
3 1 1
4 2 1
2 4
3 5
s
2
6 4
1 0t
2 2 3
s 3 2
3
4
1 4 5 1
2
0 t 2
1

Select an active node.

Push/Relabel
 Select, then relabel/push

s 1
4
2
5
2
4 2
5
1
2
3
3 1 1
4 2 1
2 4
3 5
s
2
6 4
1 0t
2 3 s 3 2
3
4
1 4 5 1
2
0 t 2
1

Select an active node.

Push/Relabel
 Select, then relabel/push

s 1
4
1
5 5 2
4 3
1
2
5
3
3 1 1
4 2 1
2 4
3 5
s
2
6 4
1 0t
2 3 s 3 2
3
4
1 4 5 1
2
0 t 2
1

Select an active node.

Push/Relabel
 Select, then relabel/push

s 1
4
1
5 5 2
4 3
1
2
5
3
3 1 1
4 2 1
2 4
3 5
s
2
6 4
1 0t
2 3 s 3 2
3
4
1 4 5 1
2
0 t 2
1

One can keep pushing flow between nodes 2 and 5 until

eventually all flow returns to node s.
There are no paths from nodes 2 and 5 to t, and there are
ways to speed up the last iterations.
Prove this lemma.
Push: Saturating

4 3
2 1

1 3 3
2 1
Push: Non-Saturating

3 4
2 1

0 3
1
2 1

3
 Relabel

2 3
4 Min{3,5}+1=4
3
2 5
1
1
1

2 3
4
3
4 5
1
1
1
 Relabel

2 3
4
3
2 5

Relabel will increase F and also increase  d (i ). 1

1
iV
1
(The total value that the relabel increases on F )

(The total value that the relabel increases on  d (i )) 2 3
iV

 2n 2 . 4
3
4 5
1
1
Min{3,5}+1=4 1
 Push: Non-Saturating

F decreases at least 1.

Case 1 Case 2
3 4 1
3 4
2 1
2 1

0 3
1 0 4
2 1 1
2 1

3
3
Why there are at most nm saturating Pushes?

4 3 Consider an arc (i, j ).

d 1 d
After a saturating push, (i, j ) disappears
4 and ( j , i ) appears.
1 3 3
d 1 d
To make (i, j ) appear again, we have to
1 3 have a push at ( j , i ).
3
e e 1

For any node i, distance label d (i ) never

4 3 decreases. Thus, e  d  1, f  e  1 and
f 1 f
f  1  d  3.

Between two saturating pushes along (i, j ), d (i ) increases by value

at least 2. Moreover, d (i )  2n  1. Thus, there are at most n saturating
pushes on arc (i, j ).
 Assignments

Read sections 7.3 and 7.6

Exercises
7.4