FIT2004 - Algorithms and Data Structures

Seminar 8 - Network Flow

Rafael Dowsley
28 April 2025
Agenda

Divide-and- Greedy Dynamic Network Data

Conquer Algorithms Programming Flow Structures
(W1-3) (W4-5) (W6-7) (W8-9) (W10-11)

1 Maximum Flow Problem

2 Ford-Fulkerson method

3 Min-cut Max-flow Theorem

Maximum Flow Problem
Flow networks

3
a c
3 3

s t

3
5 5
b d
5

• A flow network is a connected directed graph where:

I There is a single source vertex, often denoted by s, which only has

outgoing edges.

I There is a single sink/target vertex, often denoted by t, which only has

incoming edges.

I Each edge has a given non-negative capacity (usually integers) giving the
maximum amount/rate of flow that the edge can carry.

1/51
What are flow networks used for?

3
a c
3 3

s t

3
5 5
b d
5

• Flow networks can model many real-world problems, such as:

I Water flowing through an assembly of pipes.

I Electric current flowing through electrical circuits.
I Information flowing through communication networks.

• Can be applied to solve a large range of other combinatorial

problems unrelated to physical flows.

2/51
Flow

3/3
a c 0/
3
3/ 3

s t

3
3/
2/
5 5/5
b d
5/5

• For an edge e, its flow f (e) is an assignment of how much material

is flowing through it in the flow network given its capacity c(e).

• All vertices (except source and sink) conserve their flow:

I Let Ein (v ) denote the set of all incoming edges to a vertex v , and similarly
Eout (v ) denote its set of outgoing edges.
I The total amount flowing into any vertex (through incoming edges) is equal
to the total amount flowing out of that vertex (through outgoing edges).

3/51
Quiz time!
Properties of a flow network

A flow network must satisfy the following properties:

Capacity constraint
For every edge e, its flow is bounded by its capacity: 0 ≤ f (e) ≤ c(e).

Flow conservation
For every vertex v ∈ V \ {s, t}, it holds that
X X
f (ein ) = f (eout )
ein ∈Ein (v ) eout ∈Eout (v )

We only consider integer capacities.

4/51
Flow of the network

3/3 12/12
a c 0/3 a c 15/
3/3 16
11/ 20

s 3 t s 0/10 1/4 9 7/7 t

3/ 4/
2/5 8/1
5/5 3 4/4
b d b d
5/5 11/14

• Given that the flow network satisfies the capacity constraint and
flow conservation properties, the flow of the network is the total flow
out of the source vertex.

I Equivalently, this is the same as the total flow into sink vertex.

I What is the flow value in the left flow network? 5

I What is the flow value in the right flow network? 19

5/51
Maximum-flow problem

3/3 12/12
a c 0/3 a c 15/
3/3 16
11/ 20

s 3 t s 0/10 1/4 9 7/7 t

3/ 4/
2/5 8/1
5/5 3 4/4
b d b d
5/5 11/14

Maximum-flow problem
Given a flow network, determine the maximum value of the flow that
can be sent from source s to sink t without violating the capacity
constraint and flow conservation properties.

6/51
Ford-Fulkerson method
Ford-Fulkerson method

Ford-Fulkerson is a method for solving max-flow problems.

Lester Ford Jr. Delbert Fulkerson

7/51
Ford-Fulkerson intuition

0/3
a c 0/
3
0/ 3

s t

3
0/
0/
5 0/5
b d
0/5

How can we increase the flow in the above network?

8/51
Quiz time!
Ford-Fulkerson intuition

0/3
a c 0/
3
0/ 3

s t

3
0/
0/
5 0/5
b d
0/5

How can we increase the flow in the above network?

1. Choose a path from source to sink.

2. Increase flow along it as much as possible.

Seems easy enough!

9/51
Ford-Fulkerson intuition

3/3
a c 0/
3
3/ 3

s t

3
3/
2/
5 5/5
b d
5/5

Can we increase the flow in the above network?

10/51
Quiz time!
Ford-Fulkerson intuition

3/3
a c 0/
3
3/ 3

s t

3
3/
2/
5 5/5
b d
5/5

Can we increase the flow in the above network?

We can! But there is no path from source to sink with spare capacity. . .

11/51
Ford-Fulkerson intuition

3/3
a c 3/
3
3/ 3

s t

3
0/
2/
5 5/5
b d
5/5

• There is no path from s to t with spare capacity.

• Redirect the 3 units on the edge c → b to go to edge c → t.

• Problem: The flow through b is not conserved anymore.

12/51
Ford-Fulkerson intuition

3/3
a c 3/
3
3/ 3

s t

3
0/
5/
5 5/5
b d
5/5

• There is no path from s to t with spare capacity.

• Redirect the 3 units on the edge c → b to go to edge c → t.

• Problem: The flow through b is not conserved anymore.

• Solution: Send 3 more units along s → b.

13/51
What actually happened here?

3/3
a c 3/
3
3/ 3

s t

3
0/
5/
5 5/5
b d
5/5

• We increased the total flow by 3.

• We sent 3 additional units along s → b, c → t.

• We removed 3 units of flow from c → b.

• Our path was s → b → c → t, but we had a backwards edge. . .

14/51
Residual network

Flow network Residual network

0
3/3
a c 1/ a c 2
3 0 3
3/ 3
3 2 1
s t s 2 t

3
1 4
2/ 2/ 5
5 4/ 3 4 1
b d b d
4/5
1

• Residual network has the same vertices as the original network.

• For every directed edge u → v in flow network, we add two edges in

the residual network:
I Forward edge/residual edge: an edge in the same direction as u → v with
the residual/remaining capacity in the flow network.
I Backward edge/reversible flow edge: an edge in the direction v → u with
weight equal to the current flow of u → v in the flow network.

15/51
Augmenting path

Flow network Residual network

3/3
a c 1/ a c 2
3 3
3/ 3
3 2 1
s t s 2 t

3
1 4
2/ 2/ 5
5 4/ 3 4 1
b d b d
4/5
1

• Augmenting path is any simple path (a path without repeating

vertices) from source s to target t along edges with positive weight
in the residual network.

I We can omit edges with weight 0 from the residual network.

• Residual capacity is the minimum edge weight in the residual along

this augmenting path (e.g., 2 in the example).

16/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 0

3
0/3
a c 0/ a c 3
3 3
0/ 3

s t s t
3

3
0/

0/ 5
5 0/ 5 5
b d b d
0/5
5

17/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 0 Path capacity = 3

3
0/3
a c 0/ a c 3
3 3
0/ 3

s t s t
3

3
0/

0/ 5
5 0/ 5 5
b d b d
0/5
5

18/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 3 Path capacity = 3

3
3/3
a c 0/ a c 3
3 3
3/ 3

s t s t
3

3
3/

0/ 5
5 3/ 5 5
b d b d
3/5
5

19/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 3 Residual needs to be updated!

3/3
a c 0/ a c 3
3 3
3/ 3
3 3
s t s t
3

3
3/

0/ 5
5 3/ 5 3 2
b d b d
3/5
2

20/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 3 Path capacity = 3

3/3
a c 0/ a c 3
3 3
3/ 3
3 3
s t s t
3

3
3/

0/ 5
5 3/ 5 3 2
b d b d
3/5
2

21/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 6 Path capacity = 3

3/3
a c 3/ a c 3
3 3
3/ 3
3 3
s t s t
3

3
0/

3/ 5
5 3/ 5 3 2
b d b d
3/5
2

22/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 6 Residual needs to be updated!

3/3
a c 3/ a c
3 3
3/ 3
3 3
s t s 3 t
3

3 3
0/

3/ 5
5 3/ 2 3 2
b d b d
3/5
2

23/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 6 Path capacity = 2

3/3
a c 3/ a c
3 3
3/ 3
3 3
s t s 3 t
3

3 3
0/

3/ 5
5 3/ 2 3 2
b d b d
3/5
2

24/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 8 Path capacity = 2

3/3
a c 3/ a c
3 3
3/ 3
3 3
s t s 3 t
3

3 3
0/

5/ 5
5 5/ 2 3 2
b d b d
5/5
2

25/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 8 Residual needs to be updated!

3/3
a c 3/ a c
3 3
3/ 3
3 3
s t s 5 t
3

3 5
0/

5/ 5
5 5/ 5
b d b d
5/5

26/51
Ford-Fulkerson method

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

Flow of the network = 8 No further augmenting paths!

3/3
a c 3/ a c
3 3
3/ 3
3 3
s t s 5 t
3

3 5
0/

5/ 5
5 5/ 5
b d b d
5/5

27/51
Implementation details

• Edge “type” in the residual does not matter as we can send flow
along any of them.

I Forward edges can have flow sent along them because they have spare
capacity.

I Backward edges can have flow sent along them because reducing a flow in
one direction is the same as increasing the flow in the opposite direction.

• If a pair of vertices in the flow network have edges in both directions,

then in the residual network there are only 2 edges (not 4).

28/51
Quiz time!
Implementation details

Flow network Residual network

x x

a/b c/d (b − a) + c (d − c) + a

y y

• The amount of flow we can send from x to y consists of:

I The spare capacity (b − a).

I The existing flow from y to x which can be reversed (c units).

29/51
Time complexity

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

• Cost of finding an augmenting path: O(|V | + |E |) using BFS/DFS.

• Augmenting flow along a path: length of path ≤ |V | − 1, so O(|V |).

• Updating the residual: two edges per edge of the augmenting path, so O(|V |).

• Total work in one iteration of the loop: O(|V | + |E |) = O(|E |) as the graph is
connected.

30/51
Time complexity

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

• Total work in one iteration of the loop: O(|E |).

• How many iterations?

31/51
Quiz time!
Time complexity

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

• Total work in one iteration of the loop: O(|E |).

• How many iterations?

I Assuming integer capacities, Ford-Fulkerson flows are always

integer-valued, so flow grows by at least 1 in each iteration.
I If maximum flow in the network is F , then at most F iterations.

• Total work: O(|E | · F ).

32/51
Time complexity

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

• Total work: O(|E | · F ).

I This looks polynomial.

I But it isn’t because F is a number, so its value is exponential in the space
required to store it.

• It can be proven that the complexity is O(|V | · |E |2 ) when using BFS to find
augmenting paths, which is polynomial.

I These two bounds are incomparable.

33/51
Correctness

Ford-Fulkerson method
1: function max flow(G = (V , E ), s, t)
2: set initial flow f to 0 on all edges
3: while there exists an augmenting path p in the residual network Gf do
4: augment the flow f along the augmenting path p as much as possible
5: return f

• Does the algorithm terminate?

I Yes, assuming all capacities are integers.

I The flow always increases by at least 1 per iteration and there cannot be
any augmenting path if all source’s outgoing edges (similarly, if all sink’s
incoming edges) are saturated.

• In order to show that the algorithm terminates exactly once it finds a flow whose
value is the maximum possible value among all feasible flows, we will need to
study the Min-cut Max-flow Theorem.

34/51
Min-cut Max-flow Theorem
Cut

12/12
6 a c 15/
1 1/1 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

• A cut (S,T ) of a flow network partitions the vertices into two

disjoint sets S and T such that s ∈ S and t ∈ T .

• The cut-set of a cut (S,T ) is the set of edges that “cross” the cut,
i.e., each edge connects one vertex in S with another in T .

I Outgoing edges of the cut: from a vertex in S to a vertex in T .

I Incoming edges of the cut: from a vertex in T to a vertex in S.

35/51
Flow and capacity of a cut

12/12
a c 15/
16
11/ 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

• Capacity of a cut (S,T ) is the total capacity of its outgoing edges.

• Flow of a cut (S,T ) is equal to total flow of outgoing edges − total

flow of incoming edges.

• Note that the flow of a cut is always ≤ to the capacity of the cut.
I Flow of an edge ≤ capacity of an edge.
I Capacity of a cut does not subtract capacities for incoming edges.

36/51
Flow and capacity of a cut

12/12
a c 15/
16
11/ 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

• Capacity of a cut (S,T ) is the total capacity of its outgoing edges.

I What is the capacity of this cut?
I 26

• Flow of a cut (S,T ) is equal to total flow of outgoing edges − total

flow of incoming edges.
I What is the flow of this cut?
I 19

37/51
Another cut

12/12
a c 15/
16
11/ 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

• Capacity of a cut (S,T ) is the total capacity of its outgoing edges.

I What is the capacity of this cut?
I 29

• Flow of a cut (S,T ) is equal to total flow of outgoing edges − total

flow of incoming edges.
I What is the flow of this cut?
I 19

38/51
Yet another cut

12/12
a c 15/
16
11/ 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

• Capacity of a cut (S,T ) is the total capacity of its outgoing edges.

I What is the capacity of this cut?
I 31

• Flow of a cut (S,T ) is equal to total flow of outgoing edges − total

flow of incoming edges.
I What is the flow of this cut?
I 19

39/51
Are the flows of every cut the same?

12/12
a c 15/
16
11/ 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

• It seems that the flow of every cut is the same, let’s prove this!

• Let Fout (v ) denote the total flow going out of vertex v and Fin (v )
denote the total flow coming into vertex v .

• Flow conservation property: Fout (v ) − Fin (v ) = 0 for every vertex

v ∈ V \ {s, t}.

• Flow of the network = Fout (s).

40/51
Flow of every cut is the same

12/12
a c 15/
16
11/ 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

X
Fout (s) = Fout (s) + (Fout (v ) − Fin (v ))
v ∈S\s
X
= (Fout (v ) − Fin (v ))
v ∈S
X
= (Fout (v ) + Fout (v ) − Fin (v ) − Fin (v ))
v ∈S

1st equation: by flow conservation, 2nd equation: as source has no incoming edges and
so Fin (s) = 0, 3rd equation: just separates the flow through red and green arrows.

41/51
Flow of every cut is the same

12/12
a c 15/
16
11/ 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

X
Fout (s) = (Fout (v ) + Fout (v ) − Fin (v ) − Fin (v ))
v ∈S
X X
= (Fout (v ) − Fin (v )) + (Fout (v ) − Fin (v ))
v ∈S v ∈S
X
= (Fout (v ) − Fin (v ))
v ∈S

The last equation follows from the fact that each red edge appears once as an incoming
edge and once as an outgoing edge.

42/51
Flow of every cut is the same

12/12
a c 15/
16
11/ 20

s 0/10 1/4 9 7/7 t

4/
8/1
3 4/4
b d
11/14

X
Fout (s) = (Fout (v ) − Fin (v ))
v ∈S

We conclude that the flow of any cut is equal to the flow of the network.

43/51
Min-cut

• Min-cut of a flow network is the cut with the minimum

capacity.

• Flow of the network = flow of any cut ≤ capacity of that cut.

• Maximum possible flow of the network ≤ capacity of min-cut.

• What if we can find a pair of flow and cut such that the flow of the
network = capacity of the cut?

I The flow value cannot be increased any further as it would violate the
capacity of that cut, so it is the maximum flow.
I The cut is a min-cut of the flow network.

44/51
Min-cut Max-flow Theorem

Min-cut Max-flow Theorem

Maximum possible flow of a network = capacity of the min-cut.

We will show that the Ford-Fulkerson method always terminates and

outputs a flow with value equal to the capacity of the min-cut.

45/51
Proof of correctness

Flow network Residual network

3/3
a c 3/ a c 1
4 1 3
3/ 4
3 3
s t s 5 t

3
3 5
5/
6 0/ 5/
6 1 5 1
b d b d
5/5

• Suppose the Ford-Fulkerson method has terminated (i.e., there does

not exist any augmenting path in the residual network).

• We define a cut (S,T ) such that:

I S contains every vertex v that is reachable from s in the residual network.

I T contains every other vertex. Note t cannot be in S because it is not

reachable from s (there is no further augmenting paths).

46/51
Proof of correctness

Flow network Residual network

3/3
a c 3/ a c 1
4 1 3
3/ 4
3 3
s t s 5 t

3
3 5
5/
6 0/ 5/
6 1 5 1
b d b d
5/5

• The flow of this cut equals to the capacity of this cut:

I For each outgoing edge, e.g., a → c, its flow is equal to the capacity of the
edge. Otherwise, edge a → c would be in the residual network and c would
be reachable from s (but we know this is not the case as c 6∈ S).

I For each incoming edge, e.g., c → b, its flow is zero. Otherwise, there
would be an edge b → c in the residual network implying c is reachable
from s (but we know this is not the case as c 6∈ S).

47/51
Min-cut max-flow connection used in practice

Original application: US Air Force wanted to identify targets for air

strikes to effectively cut off Soviet supply chains in Eastern Europe in
case of a conflict during the Cold War.

48/51
Min-cut max-flow connection used in practice

• Manage traffic flow and preventing bottlenecks on roads.

• Identifying parts that can lead to catastrophic failures in the power

grid.

• Businesses use it to analyse critical parts of their supply chain and

operational networks.

• Social networks use it to analyse and optimise flow of information.

• It can help identifying potential vulnerabilities in complex systems

and planning to increase their resilience.

49/51
Reading

• Course Notes: Chapter 9

• You can also check algorithms’ textbooks for contents related to this
lecture, e.g.:

I CLRS: Sections 26.1 and 26.2

I KT: Sections 7.1, 7.2 and 7.3

50/51
Concluding remarks

• Take home message: maximum flow of a network is equal to the

capacity of its min-cut and both can be found using Ford-Fulkerson.

• Things to do: make sure you understand Ford-Fulkerson

algorithm, why it is correct, and how to use it to find
max-flows and min-cuts.

• Coming up next: Circulation with demands, applications of network

flow to combinatorial problems.

51/51