Design Technique: Dynamic

Programming
Dynamic Programming
• Dynamic programming (DP) is an algorithm design technique (like divide and
conquer) used to solve a wide variety of optimization problems such as
scheduling, travelling salesman, packaging, and inventory management, etc.

• In divide and conquer approach, problem is divided into independent sub-

problems which are recursively solved, while in DP, sub-problems are not
independent, e.g., they share the same sub-problems/ repeating subproblems.

• DP is a useful technique for solving optimization problems that can be divided

into smaller subproblems with optimal substructure and overlapping
subproblems.

• DP solves each sub-problems just once and stores the result in a table so that it
can be repeatedly retrieved if needed again.

• It is called dynamic since it decide dynamically whether to call function or use

table.
DAA (CS-204) PDPM IIITDM, Jabalpur 2
Development of Dynamic Programming Algorithm
Dynamic Programming works when a problem has the following features:-
Optimal Substructure: If an optimal solution contains optimal sub solutions, then a
problem exhibits optimal substructure.
Overlapping sub-problems: When a recursive algorithm would visit the same sub-
problems repeatedly, then a problem has overlapping sub-problems.
Elements of Dynamic Programming-
1. Substructure: Decompose the given problem into smaller sub-problems. Express the
solution of the original problem in terms of the solution for smaller problems.
2. Table Structure: After solving the sub-problems, store the results of the sub problems
in a table/ array. This is done because sub-problem solutions are reused many times,
and we do not want to repeatedly solve the same problem over and over again.
There are two ways to attain the above properties-
• Memoization
• Tabulation
DP will only be applicable to any problem, if we can represent problem recursively. Let
us understand the concept of DP using Fibonacci series problem.

DAA (CS-204) PDPM IIITDM, Jabalpur 3

 Fibonacci sequence using Recursion
The Fibonacci numbers Fn are defined as follows:
F0 = 0 FIB (n)
F1 = 1 1.If (n < 2)
Fn = F(n-1)+ F(n-2) 2.then return n
3.else return FIB (n - 1) + FIB (n - 2)

Figure: Recursive calls during computation of Fibonacci number

Time complexity =?

4
 Fibonacci sequence using Recursion
F0 = 0 FIB (n)
F1 = 1 1.If (n < 2)
Fn = F(n-1)+ F(n-2) 2.then return n
3.else return FIB (n - 1) + FIB (n - 2)

Figure: Recursive calls during computation of Fibonacci number

Recurrence relation: T(n) = 2T(n-1) +1,

So, the number of function calls made to compute Fib(n) = O(2n /2) = O(2n) using the
master method. There are also many repeating/ overlapping subproblems.
5
Fibonacci sequence using Recursion

How to reduce number of function call and time complexity?

DAA (CS-204) PDPM IIITDM, Jabalpur 6

Fibonacci sequence using DP
How to reduce number of function call and time complexity?
DP can be used to reduce the time complexity. DP will only be applicable, if
we can find optimal substructure and overlapping subproblems.

It has been observed from the Fibonacci example that larger problem can
only be solved if we have solution of smaller subproblems. It means, optimal
substructure has been found.

Besides, there are many overlapping subproblems. So, We can use DP. In
DP, solution of subproblems are stored in table/array.

Two ways the solution of sub-problem can be stored so that it can be reused-

• Memoization Method
• Tabulation Method

DAA (CS-204) PDPM IIITDM, Jabalpur 7

Fibonacci sequence : Memoization
Memorization:
In memorization, a global array of size n+1 is defined to store results of
subproblems. The array is initialized with -1. In this approach, recursion tree is
traversed in top-down manner to fill the array A[n+1]. The steps of memorization
approach is given below-
If we trace through the recursive calls to MEMOFIB, we find that array A [n] gets
filled from top down.
Algorithm with memorization

MEMOFIB (n)
if (n < 2)
return n
if (A[n] !=-1)
return A[n]
else
return A[n] = MEMOFIB (n - 1) + MEMOFIB (n - 2)

8
 Fibonacci sequence : Memoization
Memoization method
This algorithm clearly takes only O(n) time to compute Fib (n).

Memoization approach reduces the time complexity from 2^n to n.

Figure: Recursive calls during computation of Fibonacci number

9
Fibonacci sequence : Tabulation Method
• This approach defines a table/array of size n+1 and fills the table/array
from its starting index i.e., 0th index to the last index. It means the base
cases are first filled and then the next higher indexes are filled.

• This approach is also known as bottom-up method for solving DP

problems, since it first fills the values of function calls present at the last
level.
• In the tabulation method, the problem is iteratively defined to compute the
overlapping subproblems only once.

F0 = 0, if n=0
F1=1 if n=1
Fn=F(n-1)+ F(n-2), if n>=1

DAA (CS-204) PDPM IIITDM, Jabalpur 10

 Fibonacci sequence: Tabulation Method
Tabulation Method:
int A[n+1] ={-1};
int FIB_tab (int n)
{
A[0] = 0;
A[1] = 1;
for (int i=2; i<=n; i++)
{
A[i] = A[i-1]+A[i-2];
}
return A[n]
}

There exist only n+1 unique subproblems for the Fibonacci series. So, the
number of subproblems called will be O(n+1). Thus, this algorithm takes only
O(n) time to compute Fib (n).

11
Memoization or Tabulation : Which one is better
Tabulation is often faster than memoization because it is iterative
and solving subproblems requires no overhead.

Memoization algorithms are easier to understand and implement

but they can cause the stack overflow (SO) error.

Therefore, Tabulation method is normally used in DP approach.

Multistage Graph (Shortest Path)
A multistage graph G=(V,E) is a directed graph, in which all the vertices are partitioned
into the k stages where k>=2. There is no edge between vertices of within same stage
and from a vertex of current stage to previous stage.

Source Destination

DAA (CS-204) PDPM IIITDM, Jabalpur 13

Multistage Graph (Shortest Path)
Find the shortest path from source to sink/ destination vertex for a multistage
graph G=(V,E).

Source Destination

DAA (CS-204) PDPM IIITDM, Jabalpur 14

Multistage Graph (Shortest Path)
Problem: Find the shortest path from source to sink/ destination vertex for a
multistage graph G=(V,E).
• In the brute force approach, all paths from source to destination are identified and
their costs are computed to find the shortest path. This approach would be
inefficient because there will be 2^n possibilities for a graph having n vertices.
• Dijkstra algorithm can also be one way to find the shortest path, but Dijkstra
algorithm will generate shortest path for all vertices, which is not our objective.
• Is there any other efficient approach that can be used to find the optimal
solution?
• DP may be used, however, DP will only applicable if we should-
• Able to find the optimal substructure / Able to write recursive equation
• Able to find overlapping subproblems
• So, Let us define a recursive equation to find the optimal substructure and
repeating/ overlapping subproblems.

DAA (CS-204) PDPM IIITDM, Jabalpur 15

Multistage Graph -Shortest Path
• If we want to find the shortest path from 1 to 8 then, problem can be broken into
subproblems as follows-
• We can go from 1 to 2 and find the shortest path from 2 to 8, or
• We can go from 1 to 3 and find the shortest path from 3 to 8, or
• We can go from 1 to 4 and find the shortest path from 4 to 8
• The above statements can be formulated as follows-

1/S-> 2 + min (2, 2)

min(n-1, i) = cost(i->T), where T is
min(1, 1/S) = min 1/S ->3 + min (2, 3) sink/target, n is no. of stages
1/S->4 + min (2, 4)
Base case
Min. cost to go Recursive step
from stage 1 and
node 1 to sink 4
2 5
11 9 18
1
Using the formula given above, the
5 13
stage1 problem can be converted into 2 3 6 8
stage 2 and stage 2 to stage 3 and ... 1
16
recursively. 5 2
2
4 7

DAA (CS-204) PDPM IIITDM, Jabalpur 16

Multistage Graph -Shortest Path
• We know that DP will only be beneficial if problem has optimal sub-structure and
overlapping sub-problem.
• So, lets use the following formula and draw the recursion tree to identify the
overlapping sub-problems.

1/S-> 2 + min (2, 2)

min(1, 1/S) = min 1/S ->3 + min (2, 3)
1/S->4 + min (2, 4)

Recursive step

4
2 5
11 18
1 9

13
2 5
3 6 8
1
16
5 2
2
4 7

DAA (CS-204) PDPM IIITDM, Jabalpur 17

Multistage Graph -Shortest Path
4 5
2
11 9 18
1
5 13
2 3 6 8
1
16
5 2
2
4 7

Recursion tree

Cost (5->8)=18 13 18 13 2 2

DAA (CS-204) PDPM IIITDM, Jabalpur 18

Multistage Graph -Shortest Path

• The depth of the tree is equal to the number of stages, i.e., k. So, space complexity will
O(K), but at each level there may be n vertices and total there could be kn vertices in
worst case. So, time complexity will be exponential.
• If we explore the recursion tree then we could notice, there are many overlapping
subproblems. So, we can use DP to save the results in a table so that overlapping
subproblems cannot be computed in each call and it reduces the time complexity.
• In DP, only distinct subproblems are computed and their result are stored in a table. For
this problems, number of distinct sub-problems are equal to number of vertices. So, we
will define an array whose size will be equal to number of vertices to store cost from a
source vertex to the target.
DAA (CS-204) PDPM IIITDM, Jabalpur 19
Multistage Graph -Shortest Path
4
2 5
11 9 18
1

5 13
2 3 6 8
1
16
5 2
2
4 7
Recursion tree Overlapping
sub-problems

18 13 18 13 2 2

DAA (CS-204) PDPM IIITDM, Jabalpur 20

Multistage Graph -Shortest Path-bottom up Approach
Problem: Find the shortest path from source to sink/ destination vertex for a
multistage graph G=(V,E).
T[i] = min of all
j= (i+1) to n Edge cost (i, j) + T[j]

4 Define an array T of size |v| and

2 5
11 18 initialize T[8] =0, since cost (8->8) =0
1 9

5 13 0
2 3 6 8
1 1 2 3 4 5 6 7 8
16
5 2 Compute T[7]
2
4 7
T[7] = min{( Edge cost (7, 8) + T[8])}
T[7] = min{(2+0)} =2

2 0
1 2 3 4 5 6 7 8

DAA (CS-204) PDPM IIITDM, Jabalpur 21

Multistage Graph -Shortest Path-bottom up Approach
Problem: Find the shortest path from source to sink/ destination vertex for a
multistage graph G=(V,E).
T[i] = min of all
j= (i+1) to n Edge cost (i, j) + T[j]
Compute T[6]
4 T[6] = min{(Edge cost (6, 7) + T[7],
2 5
11 18 Edge cost (6, 8) + T[8])}
1 9
T[6] = min{(∞+2, 13+0)} =13
5 13
2 3 6 8
1 13 2 0
16
5 2 1 2 3 4 5 6 7 8
2
4 7 Compute T[5]
T[5] = min{( Edge cost (5, 6) + T[6]),
Edge cost (5, 7) + T[7]), Edge cost (5,
8) + T[8])}
T[5] = min{(∞+13, ∞+2, 18+0)} =18
18 13 2 0
1 2 3 4 5 6 7 8
DAA (CS-204) PDPM IIITDM, Jabalpur 22
Multistage Graph -Shortest Path-bottom up Approach
Problem: Find the shortest path from source to sink/ destination vertex for a
multistage graph G=(V,E).
T[i] = min of all j= Edge cost (i, j) + T[j]
(i+1) to n
4
2 5 Compute T[4]
11 9 18 T[4] = min{( Edge cost (4, 7) + T[7]), (
1
13 Edge cost (4, 6) + T[6])}
2 5 6 8
1 3 T[4] = min{(2+2), (∞+13) } =4
16
5 2 4 18 13 2 0
2
4 7
1 2 3 4 5 6 7 8
Compute T[3]
T[3] = min{( Edge cost (3, 7) + T[7]), (Edge cost (3, 6) + T[6]), (Edge cost (3, 5) + T[5]) }
T[5] = min{(16+2), (5+13), (9+18) } =18

18 4 18 13 2 0
1 2 3 4 5 6 7 8

DAA (CS-204) PDPM IIITDM, Jabalpur 23

Multistage Graph -Shortest Path-bottom up Approach
Problem: Find the shortest path from source to sink/ destination vertex for a
multistage graph G=(V,E).
T[i] = min of all j= Edge cost (i, j) + T[j]
4 (i+1) to n
2 5
11 9 18
1
5 13
2 3 6 8
1
16
5 2
2
4 7
Compute T[2]
T[2] = min{( Edge cost (2, 5) + T[5]), (Edge cost (2, 6) + T[6])}
T[2] = min{(4+18), (11+13) } =22

Compute T[1]
T[1] = min{( Edge cost (1, 2) + T[2]), (Edge cost (1, 3) + T[3]), (Edge cost (1, 4) + T[4])}
T[1] = min{(1+22), (2+18), (5+4)} =9
9 22 18 4 18 13 2 0
So, minimum cost to reach 8 from 1 will be
9. 1 2 3 4 5 6 7 8
DAA (CS-204) PDPM IIITDM, Jabalpur 24
Multistage Graph -Shortest Path-bottom up Approach
Time complexity:
Total time/ work done = Number of subproblems * time to solve each
subproblems
Number of subproblems solved to get the shortest path = n
Overall work done = number of times functions will be evaluated at each
stage
9 22 18 4 18 13 2 0
1 2 3 4 5 6 7 8
O(7) O(6) O(5) O(4) O(3) O(2) O(1) O(1)

2 4 5
1 11 9 18 Overall work done=
13 1+1+2+3+…+n-1
2 3 5 6 8
1 n-1 (n-2)/2 = n2
16
5
O(n2) = O(V2) = O(E)
2 2
4 7
0/1 Knapsack Problem

Given a set of items, each having different weight and value/ profit associated with it.
Find the set of items such that the total weight is less than or equal to a capacity of
the knapsack and the total value/ profit earned is as large as possible.

DAA (CS-204) PDPM IIITDM, Jabalpur 26

0/1 Knapsack Problem
Given a set of items, each having different weight and value, or profit associated with
it. Find the set of items such that the total weight is less than or equal to a capacity of
the knapsack and the total value earned is as large as possible.

Suppose we have n items:

1, 2, 3, 4, 5, …, n
Then there will 2 choices for each item whether it can be included or not in the
knapsack. So, total number of possible solutions would be = 2x2x2x2x … x2 = 2n
Total 2n solutions need to be explored to find the optimal solution using brute force
approach.
Is there any other efficient approach that can be used to find the optimal
solution?
• DP may be used, however, DP will only applicable if we should-
• Able to find the optimal substructure / Able to write recursive equation
• Able to find overlapping subproblems
• So, Let us define a recursive equation to find the optimal substructure and
repeating/ overlapping subproblems.

DAA (CS-204) PDPM IIITDM, Jabalpur 27

 0/1 Knapsack Problem
If there is no elements i=0 or
knapsack capacity j= 0, nothing can
be included.

In above equation,
• wt [i] is the capacity of the object i and,
• j is the remaining/ available capacity.

There will be two possibility for each item, whether it is included in knapsack or not.

• If an object i is chosen, then profit P[i] is added and for the remaining i-1 elements,
max profit need to be identified. So, there will be a recursive call for remaining i-1
elements with remaining knapsack capacity i.e., j- wt [i].

• If i is not chosen, then profit will not increase and object will not be included in
knapsack, so there will be a recursive call for remaining i-1 elements. The choice
that gives the max profit will be selected.

DAA (CS-204) PDPM IIITDM, Jabalpur 28

 0/1 Knapsack Problem
Consider there are 10 items, and the weight
of each item is 1 kg. The knapsack capacity
is 10 kg. There will two possibilities for each
item, whether it will be included in the
knapsack or not. Based on this, the
recursion tree can be constructed as follows-
include
T (10, 10) Not include
From the recursion tree, we
T (9, 9) T (9, 10) can see that the depth of the
recursion tree could be n in
T (8, 8) T (8, 9)
T (8, 9) T (8, 10) worst case. If the depth of a
tree is O(n), then there could
. .
T (7, 7) T (7, 8) T (7, 8) T (7, 9) be 2n/2 number of nodes (for
. . . . .
. half filled complete binary
. . . . .
. tree). So, total number of
. . . . function call will be O(2n).
T (0, 0) Recursion Tree
From the tree, it is identified that there are overlapping sub-problems, and it also has optimal
substructure. So, DP can be used to solve this problem. For DP, unique sub-problems need
to be identified.
Let us find unique subproblems for the 0/1 knapsack with n items and knapsack capacity w.
DAA (CS-204) PDPM IIITDM, Jabalpur 29
0/1 Knapsack Problem
How many unique subproblems we have for the 0/1 knapsack with n items and
knapsack capacity w?

• There will be n x w unique subproblems in worst case.

• So, to solve this problem using DP will need to create a table/matrix of size
n+1 x w+1, 1 extra row and column will be created to store the base case
information (0 value). After that, table is filled one by one using the following
equation.

DAA (CS-204) PDPM IIITDM, Jabalpur 30

0/1 Knapsack Problem
Problem : A thief enters a house for robbing it. He can carry a maximal weight of 7
kg into his bag. There are 4 items in the house with the following weights and values.
What items should thief take if he either takes the item completely or leaves it
completely?
Given-
Weight Knapsack capacity (W) = 7 kg
Item Profit ($)
(kg) •Number of items (n) = 4
Mirror 1 1
Silver
3 4
nugget
Painting 4 5
Vase 5 7
Step-01:
Draw a table say ‘T’ with (N+1) = 4 + 1 = 5 number of rows and (W+1) = 7 + 1 = 8 number of
columns.
•Fill all the cells of 0th row and 0th column with 0 [base case].

DAA (CS-204) PDPM IIITDM, Jabalpur 31

0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

4, 3, 2, 1, 0
No. of items
0
0
0
Given- 0
Knapsack capacity (W) = 7 kg
Number of items (n) = 4

Step-02:
Start filling the table row wise top to bottom from left to right using the formula-

0 i=0 or j=0 Where i= 0, 1, …, N and j =0, 1,

…, W.
T[i, j] = T[i-1, j] wt[i] > j
W is knapsack capacity and N
max(T[i-1, j-wt[i]]+ P[i], T[i-1, j]) otherwise = number of items

DAA (CS-204) PDPM IIITDM, Jabalpur 32

 0/1 Knapsack Problem
Weights (j)
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

No. of items (i)

4 3 2 1 0
0 1 1
0
0
0
Knapsack capacity (W) = 7 kg
Finding T(1,1)
Step-02: Since weight of item wt[i] =1 and
the capacity of bag is j=1. so, we
Start filling the table row wise top to bottom from left to can place it in the bag.
right using the formula- T[1,1]= max(T[0, 0]+ 1, T[0, 1])
=1
Finding T[1,2]
weight of item wt[i] =1
the capacity of bag is j=2. so, we
can place it in the bag.
T[1,2]= max(T[0, 1]+ 1, T[0, 2])
=1
DAA (CS-204) PDPM IIITDM, Jabalpur 33
 0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

4, 3, 2, 1, 0
No. of items
0 1 1 1 1 1 1 1
0 1 1
0
0
Knapsack capacity (W) = 7 kg
Similarly, all columns of row 1 will
Step-02: be 1, since we are placing only one
item of weight 1 in the bag of
Start filling the table row wise top to bottom from left to varying capacity (j= 1, 2, … 7).
right using the formula-
Finding T[2,1] and T[2,2]
weight of item wt[i] = 3
the capacity of bag is j=1 i.e.,
wt[2]>j,
T[2,1]= T[i-1, j] = T[1, 1] =1
T[2,2]= T[1, 2] =1

DAA (CS-204) PDPM IIITDM, Jabalpur 34

 0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

4, 3, 2, 1, 0
No. of items
0 1 1 1 1 1 1 1
0 1 1 4 5
0
0
Knapsack capacity (W) = 7 kg
Finding T[2,3]
Step-02: weight of item wt[i] = 3
the capacity of bag is j=3.
Start filling the table row wise top to bottom from left to T[2,3]= max(T[1, 3-3]+ 4, T[1, 3])
right using the formula- =4
Finding T[2,4]
weight of item wt[i] = 3
the capacity of bag is j=4.
T[2,4]= max(T[1, 4-3]+ 4, T[1, 4])
=5 [Item 1 and 2 both can be
kept in bag]

DAA (CS-204) PDPM IIITDM, Jabalpur 35

 0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

4, 3, 2, 1, 0
No. of items
0 1 1 1 1 1 1 1
0 1 1 4 5 5 5 5
0 1 1 4
0
Knapsack capacity (W) = 7 kg
All remaining values of this row will be
Step-02: 5 since we have two items, and both
can be placed in bag.
Start filling the table row wise top to bottom from left to Finding T[3,1]
right using the formula- weight of item wt[i] = 4,
the capacity of bag is j=1, i.e., wt[i] >j
T[3,1]= T[2, 1]
=1 [only Item 1]
All places before T[3,3] will be 1 since
weight of item is larger than capacity.
Finding T[3,3)
T[3,3]= T[2, 3]
=4 [only Item 2]
DAA (CS-204) PDPM IIITDM, Jabalpur 36
 0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

4, 3, 2, 1, 0
No. of items
0 1 1 1 1 1 1 1
0 1 1 4 5 5 5 5
0 1 1 4 5 6
0
Knapsack capacity (W) = 7 kg
Finding T[3,4]
Step-02: weight of item wt[i] = 4
the capacity of bag is j=4.
Start filling the table row wise top to bottom from left to T[3,4]= max(T[2, 4-4]+ 5, T[2, 4])
right using the formula- =5 [3rd object can be placed in bag]
Finding T[3,5]
weight of item wt[i] = 4
capacity of bag is j=4.
T[3,5]= max(T[2, 5-4]+ 5, T[2, 5])
=6
[3rd and 1st object can be placed in bag]

DAA (CS-204) PDPM IIITDM, Jabalpur 37

 0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

4, 3, 2, 1, 0
No. of items
0 1 1 1 1 1 1 1
0 1 1 4 5 5 5 5
0 1 1 4 5 6 6 9
0
Knapsack capacity (W) = 7 kg
Finding T[3,6]
Step-02: weight of item wt[i] = 4
capacity of bag is j=6
Start filling the table row wise top to bottom from left to T[3,6]= max(T[2, 6-4]+ 5, T[2, 6])
right using the formula- =6
Finding T[3,7]
weight of item wt[i] = 4
capacity of bag is j=7
T[3,7]= max(T[2, 7-4]+ 5, T[2, 7])
=9
[3rd and 2nd object can be placed in
bag]
DAA (CS-204) PDPM IIITDM, Jabalpur 38
 0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

4, 3, 2, 1, 0
No. of items
0 1 1 1 1 1 1 1
0 1 1 4 5 5 5 5
0 1 1 4 5 6 6 9
0 1 1 4 5 7 8
Knapsack capacity (W) = 7 kg
Finding T[4,5]
Step-02: weight of item wt[i] = 5
capacity of bag is j=5
Start filling the table row wise top to bottom from left to T[4,5]= max(T[3, 5-5]+ 7, T[3, 5]
right using the formula- =7
Finding T[4,6]
weight of item wt[i] = 5
capacity of bag is j=6
T[4,6]= max(T[3, 6-5]+ 7, T[3, 6]
=8
[1st and 4th object can be placed in
bag]
DAA (CS-204) PDPM IIITDM, Jabalpur 39
 0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0

4, 3, 2, 1, 0
No. of items
0 1 1 1 1 1 1 1
0 1 1 4 5 5 5 5
0 1 1 4 5 6 6 9
0 1 1 4 5 7 8 9
Knapsack capacity (W) = 7 kg
Finding T[4,7]
Step-02: weight of item wt[i] = 5
capacity of bag is j=7
Start filling the table row wise top to bottom from left to T[4,7]= max(T[3, 7-5]+ 7, T[3, 7]
right using the formula- =9
[3rd and 2nd object can be placed in
bag]

DAA (CS-204) PDPM IIITDM, Jabalpur 40

 0/1 Knapsack Problem
Weights
w0 w1 w2 w3 w4 w5 w6 w7
0 0 0 0 0 0 0 0
4, 3, 2, 1, 0
No. of items

0 1 1 1 1 1 1 1
0 1 1 4 5 5 5 5
0 1 1 4 5 6 6 9
0 1 1 4 5 7 8 9
Knapsack capacity (W) = 7 kg
Find the object that can be placed in the bag to maximize the profit-
• Go to last row last column i.e., T[4][7] = 9 and check if the previous row with same
column i.e., T[3][7]. If the values of T[3][7] = T[4][7], ignore 4th row and include 3rd
row’s weight and profit in the result set.
• Subtract its profit from total (9), 9-5=4. check for 4 in 2nd row. if it is present include
object’s weight and profit in result set else go to its previous row.
• As 4 is in 2nd row include it in result set and subtract its profit from the remaining
profit. Profit =4-4=0
Since, profit is now 0 means we reached to the final solution.
The bag/ result set will include item 2 and 3 and profit will 9.

DAA (CS-204) PDPM IIITDM, Jabalpur 41

0/1 Knapsack Problem

Time Complexity
Time complexity for 0/1 Knapsack problem solved using DP is
O(N*W) where N denotes number of items available, and W
denotes the capacity of the knapsack
Longest Common Subsequence (LCS)
Given two strings, the task is to find the longest common subsequence (LCS)
present in the given strings in the same order.
A subsequence of a string is a new string generated from the original string
with some characters (can be none) deleted without changing the relative
order of the remaining characters. Please note, in substring, characters are
selected in contiguous manner, while in subsequence characters may not be
contiguous.
For example, "ace" is a subsequence of "abcde". LCS :- cdgi
Suppose we have string s1= “a b c d e f g h i j” and
s2= “e c d g i”
The substring “ecdgi” is also present in s1 but its characters are not in the increasing
order. Please note that matching character need not to be contiguous, but they must
in the increasing order by their index.
s1= “a b c d e f g h i j”
c is before e in string s1 so, it will not
s2= “e c d g i” be considered as a subsequence.
Longest Common Subsequence (LCS)
string s1= “a b c d e f g h i j” (size m) → For generating a subsequence, a character
may be either taken or not taken. So, for
s1= “a b c d e f g h i j” (size m) each character, there will be 2 choices. So,
s2= “e c d g i” (size n) total no. of subsequence could be = 2m

To find the LCS, each subsequence of s1 must be checked in S2 by scanning the

entire sequence. However, subsequence of s1 is may be only checked in s2, when
size of subsequence of s1 is smaller or equal to s2. The LCS is identified after finding
all subsequences which are common in both strings.
So, time complexity of the above approach will include
• Time required to find all subsequences of s1 = O(2m)
• For each, subsequence of s1, find all subsequences which are also present in s2=
O(n* 2m)
• Finding the longest common subsequence = O(2m)
Time complexity= O(2m) + O(n* 2m) + O(2m) = O(n* 2m)
Time complexity is exponential. Is there any other efficient approach that can be
used?
Longest Common Subsequence (LCS)
Is there any other efficient approach that can be used to find the
optimal solution?
DP may be used, however, DP will only applicable if we should-
• Able to find the optimal substructure / Able to write
recursive equation
• Able to find overlapping subproblems

So, Let us define a recursive equation to find the optimal

substructure and repeating/ overlapping subproblems.
Longest Common Subsequence (LCS)
In optimal substructure, the solution of the smaller subproblems is going to be the
solution of the main problems. The optimal substructure can easily be identified
using recursive formula.
Suppose we have two strings, X= x1 x2 x3…xi and Y= y1 y2 y3…yj. Then to find the longest
subsequence,
1. Define a recursive function that first compares xn and yn characters of X and Y, and if
they are equal, call the same recursive function for x1 x2 x3…xi-1 and y1 y2 y3…yj-1.
2. If xn and yn are not equal, then, call two recursive function, one for x1x2x3…xi-1 and
y1 y2 y3…yj and other for x1 x2 x3…xi and y1 y2 y3…yj-1 and return the max of their
result.
Let us assume C(i, j) is a LCS for string X and Y, where X = 1 to i and Y = 1 to j. Then, we
can use the following formula to compute the LCS.

0 if i=0 or j=0
C(i, j) = 1+C(i-1, j-1) if i>0, j>0, and Xi=Yj
max[C(i, j-1), C(i-1, j)] if i>0, j>0, and Xi!=Yj
Longest Common Subsequence (LCS)
Suppose we have two string X= {A, A, A, A} and Y= {B, B, B,B}. Then following tree can be
constructed for the following recursive formula.
X= {A, A, A, A}
No match
Y= {B, B, B,B}

C(4, 4)

C(4, 3) C(3, 4)

C(3, 3) C(4, 2) C(2, 4) C(3, 3)

C(2, 3) C(3, 2) C(1, 4) C(2, 3)

C(1, 3) C(2, 2)

C(0, 3) C(1, 2)
 Longest Common Subsequence (LCS)
The maximum depth of the tree could be m+n, since longest path may include the
worst case, that decrease n by 1 in one step and m by 1 in another step as given in the
following figure.
X= {A, A, A, A}
C(n, m)
No match
C(n-1, m) Y= {B, B, B,B}
C(n-2, m-1)
C(n-2, m-2)
.
.
.
C(0, --) or C(--, 0)

Since, the depth of the root is m+n so, total number of function will be O(2m+n/2), if
tree will be half filled
 Longest Common Subsequence (LCS)
Since, the depth of the root is m+n so, total number of function will be O(2m+n/2), if tree
will be half filled. So, can we reduce time complexity using DP? DP will only applicable,
if we can find optimal substructure and overlapping subproblems.
There are many subproblems in the following tree which are overlapping. In DP, only
unique subproblems are evaluated. Let us find unique subproblems for the LCS
problems.
The unique subproblems could be nearly O(mxn), if string X contains m characters and
string Y contains n characters.

Overlapping
subproblems
 Longest Common Subsequence (LCS)
There are many subproblems in the following tree which are overlapping. In DP, only unique
subproblems are evaluated. Let us find unique subproblems for the LCS problems.
C(3, 3)

C(3, 2)

C(2, 3) C(2, 2) C(3, 1)

C(2, 1) C(3, 0)
C(1, 3) C(2, 2)

C(1, 2) C(2, 1)
C(0, 3) C(1, 2)

C(0, 2) C(1, 1) C[0, 2] C(1, 1) C(1, 1) C(2, 0)

C(0, 1) C(1, 0) C(0, 1) C(1, 0) C(0, 1) C(1, 0)

Non-overlapping subproblems are- C[3,3], C[3,2], C[3,1], C[2,3], C[2,2], C[2,1], C[1,3],
C[1,2], C[1,1], C[0,3], C[0,2], C[0,1], [3,0], C[2,0], C[1,0], C[0,0] (9+7=16 sub problems
including 7 subproblems with 0 base cases).
To solve this problem using DP will require to create a table of size (m+1xn +1) [1 row and
column to store 0 (base case) ].
 Longest Common Subsequence (LCS)
To find the LCS for X= “AAB” and Y =“ACA”, first a table of size (X+1) and (Y+1)
is created.

The value at any place will depends on its left diagonal, its previous column, and
previous row with same column.
For e.g.
C[1,1] = 1+c[0,0], or max (C[1,0], C[0,1])
So, the value of C[3,3] can only be computed, if we have values of C[2,2] (in case of
match) or C[2,1] and C[1,2].
The table should be filled either row wise or column wise to efficiently compute values.
0 1 2 3
0 00 01 02 03
1 10 11 12 13
2 20 21 22 23

3 30 31 32 33
 Longest Common Subsequence (LCS)
To find the LCS for X= “AAB” and Y =“ACA”, first a table of size (X+1) and (Y+1)
are created.

Step 1: Fill 0 in 0th row and column since, if any of the string will null the LCS will also
be NULL.
Step 2: Compute C[1,1] using the formula
C[1,1]= 1+C[0,0], since Xi = Yi
= 1+0 =1
C[1,2] = max(C[1,1], C[0,2]), since Xi != Yi
= max(1,0) =1
0 A C A
C[1,3] = 1+C[0, 2], since Xi = Yi
0 1 2 3
= 1+0 =1
0 0 0 0 0 0
C[2,1] = 1+C[1, 0], since Xi = Yi
= 1+0 =1 0 1 1 1
A 1
C[2,2] = max(C[2,1], C[1,2]), since Xi != Yi
= max(1,1) =1 A 2 0 1 1 2

C[2,3] = 1+C[1, 2], since Xi = Yi B 3 0

= 1+1 =2
 Longest Common Subsequence (LCS)
To find the LCS for X= “AAB” and Y =“ACA”, first a table of sizeof (X+1) and
(Y+1) are created.

C[3,1] = max(C[2,1], C[3,0]), since Xi != Yi

= max(1,0) =1

C[3,2] = max(C[2,2], C[3,1]), since Xi != Yi

= max(1,1) =1

C[3,3] = max(C[2,3], C[3,2]), since Xi != Yi

= max(2,1) =2 0 A C A
0 1 2 3
This is the length of longest common 0 0 0 0 0 0
subsequence for the given problem. 0 1 1 1
A 1
A 2 0 1 1 2

B 3 0 1 1 2
 Longest Common Subsequence (LCS)
To find the LCS for X= “AAB” and Y =“ACA”, first a table of sizeof (X+1) and
(Y+1) are created.

To find the LCS, check the places where values of the cells an have been
updated, which means taken from diagonal and incremented. In this example,
values at C[1,1] and C[2,3] have been incremented. The LCS will include
characters at X[1] and X[2].

0 A C A
LCS = AA.
0 1 2 3
0 0 0 0 0 0

A 1 0 1 1 1

A 2 0 1 1 2

B 3 0 1 1 2
 Longest Common Subsequence (LCS)
Find the LCS for X= “ABCBDAB” and Y =“BDCABA” in polynomial time.

0 A B C B D A B
0
B

D
C

A
Subset Sum Problem
Given an array of non-negative integers and an integer sum. We have
to tell whether there exists any subset in an array whose sum is equal
to the given integer sum.
Example:
Input: arr [] = {3, 34, 4, 12, 3, 2}, sum = 7
Output: True
Explanation: There is a subset (4, 3) with sum 7.
Subset Sum Problem
Given an array of non-negative integers and an integer sum. We have to tell
whether there exists any subset in an array whose sum is equal to the given
integer sum.
Example:
Input: arr[] = {3, 34, 4, 12, 3, 2}, sum = 7
Output: True
Explanation: There is a subset (4, 3) with sum 7.
Let us try brute force approach to solve the problem

In brute force all possible subsets are identified. In a subset, element will
either present or not present. So, there will be 2 possibilities for each
element. For n elements, the number of possible subsets could be 2n. So,
this approach will be computationally very expensive.

Can we reduce time complexity?

Longest Common Subsequence (LCS)
Is there any other approach that find the subsets in efficiently.
DP may be used, however, DP will only applicable if we should-
• Able to find the optimal substructure / Able to write
recursive equation
• Able to find overlapping subproblems

So, Let us define a recursive equation to find the optimal

substructure and repeating/ overlapping subproblems.
Subset Sum Problem
Let us define recursive formula. Suppose there are 1,2,3, …, i elements and
we have to find the subset from i elements whose sum is equal to S.
A= { a1, a2, a3, …., ai} Sum =S
This problem is very similar to 0/1 knapsack problem and two conditions are
similar to that.

SS(i-1, S) S<ai
SS(i-1, S-ai) OR SS(i-1, S) S>=ai
SS(i, S) = True S=0
False i=0, S!=0

A= { a1, a2, a3, …., ai} Sum =S

Let us draw the recursion tree to identify the overlapping solutions.

 Subset Sum Problem
A= { 1, 1, 1, 1, 1} Sum =5

SS (5, 5)

SS (4, 4) SS (4, 5)

SS (3, 3) SS (3, 4) SS (2, 4)

SS (3, 4)

SS (2, 2) SS (1, 3) SS (1, 4)

SS (2, 3)

SS (1, 1) SS (1, 2) SS (1, 2) SS (1, 3) SS (0, 3) SS (0, 4)

SS (0, 1) SS (0, 2) SS (0, 2) SS (0, 3)

SS (0, 0) SS (0, 1)
In the worst case, the depth of the tree will be n and there could be O(2n) number of
calls. So, this approach is very expensive. It is also observed from the tree that many
subproblems are overlapping. Hence, DP will certainly reduce time complexity.
 Subset Sum Problem
A= { 1, 1, 1, 1, 1} Sum =5

SS (5, 5)

SS (4, 4) SS (4, 5)

SS (3, 3) SS (3, 4) SS (2, 4)

SS (3, 4)

SS (2, 2) SS (1, 3) SS (1, 4)

SS (2, 3)

SS (1, 1) SS (1, 2) SS (1, 2) SS (1, 3) SS (0, 3) SS (0, 4)

SS (0, 1) SS (0, 2) SS (0, 2) SS (0, 3)

SS (0, 0) SS (0, 1)
Overlapping subproblems
There will (i+1, S+1) number of unique solution/ calls for the problem having i elements
with subset sum S.
 Subset Sum Problem
SS[0,0] = S=0 is possible, if there is no
element,
SS[0,0] = T
For SS[0,1], S!=0 not possible because
there is no element
SS[0,1] = F A= { 6, 3, 2, 1} S =5
Similarly, all other positions of 0th row will
be false. Sum
0 1 2 3 4 5
SS[1,0] = T, (If we have one element and
we don’t choose that then S=0 is 0 T F F F F F
possible.) Similarly, all other rows of 1 T F F F F F
column 0 will also T. Items
SS[1,1] = SS[0, 1], since s< ai (1<6) 2 T
SS[1,1] = F , 3 T
It cannot be included in the subset since, the value of
1st item is 6 which will never be equal to 1. (greater 4 T
than current sum).
SS[1,2] = F, value of 2nd item will never equal to 2.
All column of this row will be false, since sum =1,2, …,5 will always smaller than item
value i.e., 6.
 Subset Sum Problem
A= { 6, 3, 2, 1} S =5
SS[2,1] = SS[1,1] since 1<3
SS[2,1] = F
S[2,2] = SS[1,2] since 2<3
SS[2,2] = F
S[2,3] = SS[1,3-3] OR SS[1, 3] Sum
SS[2,3] = SS[1,0] OR SS[1, 3] 0 1 2 3 4 5
SS[2,3] = T 0 T F F F F F
From first two elements, sum 3 is
possible. (include 2nd item) 1 T F F F F F
Items
S[2,4] = SS[1,4-3] OR SS[1, 4] 2 T F F T F F
SS[2,4] = SS[1,1] OR SS[1, 4] 3 T
SS[2,4] = F
From first two elements, sum 4 is possible. 4 T

S[2,5] = SS[1,5-3] OR SS[1, 5]

SS[2,3] = SS[1,2] OR SS[1, 5]
SS[2,3] = F,
From first two elements, sum 5 is not possible.
 Subset Sum Problem
A= { 6, 3, 2, 1} S =5
SS[3,1] = SS[2,1] since 1<2
SS[3,1] = F
S[3,2] = SS[2,2-2] OR SS[2, 2]
SS[3,2] = SS[2,0] OR SS[2, 2]
SS[3,2] = T Sum
From first three elements, sum 2 is 0 1 2 3 4 5
possible. (include 3rd item)
0 T F F F F F
S[3,3] = SS[2,3-2] OR SS[2, 3]
SS[3,3] = SS[2,1] OR SS[2, 3] 1 T F F F F F
SS[3,3] = T Items
2 T F F T F F
From first three elements, sum 3 is
possible. (include 2nd item) 3 T F T T F

S[3,4] = SS[2,4-2] OR SS[2, 4] 4 T

SS[3,4] = SS(2,2) OR SS[2, 4]
SS[3,4] = F
From first three elements, sum 4 is not
possible.
 Subset Sum Problem
A= { 6, 3, 2, 1} S =5
S[3,5] = SS[2,5-2] OR SS[2, 5]
SS[3,5] = SS[2,3] OR SS[2, 5]
SS[3,5] = T
From first three elements, sum 5 is
possible (include 2nd and 3rd item.)
Sum
S[4,1] = SS[3,1-1] OR SS[3, 1] 0 1 2 3 4 5
SS[4,1] = T
From first four elements, sum 1 is possible 0 T F F F F F
(include 4th item.)
1 T F F F F F
Items
S[4,2] = SS[3,2-1] OR SS[3, 2] 2 T F F T F F
SS[4,2] = T
From first four elements, sum 2 is 3 T F T T F T
possible. (include 3rd item.)
4 T T T T
S[4,3] = SS[3,3-1] OR SS[3, 3]
SS[4,3] = T
From first four elements, sum 3 is
possible. (include 2nd item.)
 Subset Sum Problem
A= { 6, 3, 2, 1} S =5

S[4,4] = SS[3,4-1] OR SS[3, 4]

SS[4,4] = T
From first four elements, sum 4 is possible
(include 2nd and 4th item.)
Sum
S[4,5] = SS[3,5-1] OR SS[3, 5] 0 1 2 3 4 5
SS[4,5] = T
From first four elements, sum 5 is possible 0 T F F F F F
(include 2nd and 3rd item.) 1 T F F F F F
Items
Time complexity: 2 T F F T F F
3 T F T T F T
O(nxw), if n represents the number of items and
w is the weight/ sum. 4 T T T T T T
If w> 2n, then brute force approach will the best
method to find the subset sum.
Matrix Chain Multiplication
Two matrices A1 and A2 of dimensions [p x q] and [r x s] can only be multiplied if
and only if q=r.

The total number of multiplications required to multiply A1 and A2 are p*q*s

Matrix Chain Multiplication
Matrix chain multiplication (or Matrix Chain Ordering Problem, MCOP) is an
optimization problem that to find the most efficient way to multiply a given
sequence of matrices.
The problem is not actually to perform the multiplications but just to decide the
sequence of the matrix multiplications involved.

The matrix multiplication is associative as no matter how the product is

parenthesized, the result obtained will remain the same. For example, for four
matrices A, B, C, and D, we would have:

((AB)C)D = ((A(BC))D) = (AB)(CD) = A((BC)D) = A(B(CD))

Matrix Chain Multiplication
Let's assume we have three matrices M1​,M2​,M3 of
dimensions 5×105×10, 10×810×8, and 8×58×5 respectively and for some reason we
are interested in multiplying all of them. There are two possible orders of
multiplication –

1. M1×(M2×M3​)
Steps required in M2×M3 will be 10×8×5 = 400.
Dimensions of M23 will be 10×5.
Steps required in M1×M23 will be 5×10×5 = 250.
Total Steps =400+250 = 650

2. (M1×M2​)×M3
Steps required in M1×M2 will be 5×10×8 = 400.
Dimensions of M12 will be 5×8.
Steps required in M12×M3 will be 5×8×5 = 200.
Total Steps = 400+200 = 600
Matrix Chain Multiplication
 Matrix Chain Multiplication
A1 x A2 X A3
d0 d1 d1 d2 d2 d3
2 3 3 4 4 2

(A1 x A2) X A3 A1 x (A2 X A3)

2 3 3 4 4 2 2 3 3 4 4 2
No. of multiplications = 2*3*4=24 No. of multiplications = 3*4*2=24
[2 4] [4 2] [2 3] [3 2]

No. of multiplications = 2*4*2=16 +24 No. of multiplications = 2*3*2=12+ 24

=40 =36

For this example, 2nd parenthesizing will be better. So, to minimize the number of
multiplication, their sequence must be identified in advance. For larger number of
matrices, a greater number of possibilities exists. So, trying all will increase time.
Matrix Chain Multiplication
In matrix chain multiplication, the number of ways matrices can be multiplied
depends on their size. If there will 3 matrices, the number of way they will be
multiplied is 2 while for 4 matrices, they can be multiplied in 5 ways. For 5 matrices,
they can be multiplied in

(2n)!/(n+1)! n!, where n= m-1 (number of matrices-1)

For 5 matrices (2*4)!/ (5)! (4)! = 8!/ 5! 4! = 8*7*6/2*3*4 = 14 ways

So, there will 14 ways to multiply 5 matrices.

The complexity of identifying the optimal sequence involve computations of all

possible sequences. So, the naïve approach would be very expensive. Let us check
the time complexity of naïve recursive approach.
Matrix Chain Multiplication
Le us derive recursive formula. C[1,3]

C[1,2] =24 C[3,3] =0 C[1,1] =0 C[2,3] =24

d0 d2 d3
d0 d1 d3
k i j k+1 k i j k+1

C[1,3] =C[1,2]+ C[3,3]+d0xd2xd3 C[1,3] =C[1,1]+ C[2, 3]+d0xd1xd3

= 24+0+16 =40 =0+ 24+12 =36
i-1 j
i-1 j

0 i=j
C[i,j] = min[(C[i,k] +C[k+1,j] +di-1xdkxdj)]
for i<=k<j, since splitting can
be anywhere b/w i and j
 Matrix Chain Multiplication
Le us draw a recursion tree for four matrices-
A1A2A3A4. The cost can be represented by
C[1,4].

C[1,4]
Split point A1A2A3A4 Split point A1A2A3A4 Split point A1A2A3A4

C[1,1] C[2,4] C[1,2] C[3,4] C[1,3] C[4,4]

C[2,2] C[3,4] C[2,3] C[4,4] C[1,1] C[2,2] C[1,1] C[2,3] C[1,2] C[3,3]

C[3,3] C[4,4] overlapping subproblems

From the recursion tree, the depth in worst case will be O(n). So, if the tree will half fill, then
too it will have exponential function call O(2n) (as per the property of the complete binary tree
and this tree is bigger than a binary tree). As, the recursion tree has many overlapping
subproblems. The time complexity can be reduced using DP. DP uses bottom-up
computation and stores unique results in a table. Let us find the unique results.
Matrix Chain Multiplication
For the previous example, the number of unique solutions are-
Unique function calls No. of subproblems Size
C(1, 1), C(2, 2), C(3, 3), C(4,4) 4 1
C(1, 2), C(2, 3), C(3,4) 3 2
C(1, 3), C(2,4) 2 3
C(1,4) 1 4
Total no, of function calls/ subproblems - 1+2+3+4= 10
Suppose we have n matrices A1 A2 A3, …, An, then
Subproblems Size
No. of subproblems= 1+2+3+…+n
n 1
= n(n+1)/2 = O(n2)
n-1 2
n-2 3 Therefore, for 4 matrices, a table with 4 row
and 4 column need to be created.
. .
We create one more table to store the k
1 n value for which minimum cost has been
found.
Matrix Chain Multiplication
Find the best order to multiply following matrices- 1 2 3 4
1 0 24
A1 A2 A3 A4
d0x d1 d1x d2 d2x d3 d3x d4 2 0 16
3x2 2x4 4x2 2x5
3 0
Step 1: Fill 0 to the matrices of size 1 4 0
C[1,1], C[2,2], C[3,3], C[4,4] =0 C table

Step 2: Find the value of matrices of size 2 1 2 3 4

C[1,2] = min { C[1,1]+C[2,2]+d0xd1xd2} 1 1
1<=k<2 2
2
=min { 0+0+3x2x4} = 24
1<=k<2 3
We got C[1,2] for k value 1. so, put 1 in k table at (1,2). 4
C[2,3] = min { C[2,2]+C[3,3]+d1xd2xd3} K table
2<=k<3
=min { 0+0+2x4x2} = 16
2<=k<3
Put 2 in k table at (2,3).
Matrix Chain Multiplication
Find the best order to multiply following matrices- 1 2 3 4
1 0 24 28
A1 A2 A3 A4
d0x d1 d1x d2 d2x d3 d3x d4 2 0 16
3x2 2x4 4x2 2x5
3 0 40
C[3,4] = min { C[3,3]+C[4,4]+d2xd3xd4} 4 0
3<=k<4 C table
=min { 0+0+4x2x5} = 40
3<=k<4 1 2 3 4
Put 3 in k table at (3,4). 1 1 1

2 2
Step 3: Find the value of matrices of size 3
C[1,3] = min 3
K =1 C[1,1]+C[2,3]+d0xd1xd3, 3
1<=k<3
K =2 C[1,2]+C[3,3]+d0xd2xd3 4
= min { 0+16+3x2x2, 24+0+3x4x2}
1<=k<3 K table
= min {28, 48} = 28 (for k=1)
Matrix Chain Multiplication
Find the best order to multiply following matrices- 1 2 3 4
1 0 24 28
A1 A2 A3 A4
d0x d1 d1x d2 d2x d3 d3x d4 2 0 16 36
3x2 2x4 4x2 2x5
3 0 40
C[2,4] = min K =2 C[2,2]+C[3,4]+d1xd2xd4, 4 0
2<=k<4 K =3 C[2,3]+C[4,4]+d1xd3xd4 C table

= min { 0+40+2x2x5, 16+0+2x2x5} 1 2 3 4

2<=k<4 1 1 1
= min {60, 36} = 28 (for k=3) 2 3
2
Step 4: Find the value of matrices of size 4 3
3
4
K table
Matrix Chain Multiplication
Find the best order to multiply following matrices- 1 2 3 4
1 0 24 28 58
A1 A2 A3 A4
d0x d1 d1x d2 d2x d3 d3x d4 2 0 16 36
3x2 2x4 4x2 2x5
3 0 40
Step 4: Find the value of matrices of size 4 0
4
C table
C[1,4] = min K =1 C[1,1]+C[2,4]+d0xd1xd4,
1<=k<4 K =2 C[1,2]+C[3,4]+d0xd2xd4, 1 2 3 4
K =3 C[1,3]+C[4,4]+d0xd3xd4, 1 1 3
1
2 3
C[1,4] = min {C[1,1]+C[2,4]+d0xd1xd4, C[1,2]+C[3,4]+d0xd2xd4, 2
C[2,3]+C[4,4]+d0xd3xd4} 3
3
= min {0+36+3x2x5, 24+40+3x4x5, 28+0+3x2x5}
= min {66, 124, 58} 4
= 58 (for k= 3) K table
So, to multiply 4 matrices, the minimum
Number (optimal) of scaler multiplication will
be 58. So, how to parenthesize the matrices to
get this cost?
Matrix Chain Multiplication
How to parenthesize the matrices to get the cost? 1 2 3 4
The K table tells how to parenthesize matrices to get the minimum 1 0 24 28 58
cost.
1. The cost C[1,4] is found for k=3, so, first parenthesis must be place 2 0 16 36
after third matrix. 3 0 40
(A1 A2 A3) (A4)
2. Now, matrices A1A2A3 must be parenthesize in such way so that 4 0
cost could be minimized. Check the minimum cost found for C table
matrices C[1,3]. It is found for k=1. so, second parenthesis must be 1 2 3 4
place after first matrix.
((A1) (A2 A3)) (A4) 1 1 1 3
The above parenthesize is the final and it tell that first matrix A2 and 2 2 3
A3 must be multiplied, and their resultant matrix is multiplied with
A1 and their result will finally be multiplied with A4 to get the final 3
3
result. 4
Time Complexity: K table
Time taken to find the parenthesizing will be equal to the number of values to be filled in the
C table = n(n+1)/2 =4*5/2=10 = O(n2), for n matrices. However, to find any value, such as
C[1,4], all possible values of k have been tried out and k = i, i+1, … j. where, j is closed to n.
So, to compute any value, it will take O(n) time in worst case. So, overall complexity = O(n*
n2) = O(n3).
Travelling Salesman Problem
Given a set of cities and the distance between every pair of cities, the problem is to
find the shortest possible route that visits every city exactly once and returns to the
starting point.
Brute-force Approach: The brute-force approach evaluate every possible path and
select the best one. For n number of vertices in a graph, if start and end vertices are
fixed, there will be (n - 1)! number of possibilities. So, this approach will
computationally very expensive.

Let us check, can we use DP?

DP will only applicable, if we can find optimal substructure and overlapping
subproblems.
 Travelling Salesman Problem
Let us take an example to check whether we can find optimal substructure and
overlapping subproblems.
1 2 3 4
1 2 1 0 1 2 3
2 1 0 4 2

3 1 2 0 5
3 4 4 3 4 1 0
Cost Matrix
If the salesperson starts from city 1 and visit all other cities, then the problem can be
formulated as follows: Travel 2 from 1 and continue tour from
(1, 2) + T(2, {3, 4}) 2, visit all nodes and get back to 1
Travel 3 from 1 and continue tour from
T(1, {2, 3, 4}) = (1,3) + T(3, {2, 4}) 3, visit all nodes and get back to 1
min
Travel all nodes (1,4) + T(4, {2, 3}) Travel 4 from 1 and continue tour from
and get back to 1 3, visit all nodes and get back to 1

From the above, it has been observed that main problem can be broken into smaller
subproblems. So, it has substructure.
 Travelling Salesman Problem
(1, 2) + T(2, {3, 4})
T(1, {2, 3, 4}) = (1,3) + T(3, {2, 4})
min (1,4) + T(4, {2, 3})
E(1, 2) = 1, E(1, 3) = 2, E(1, 4) = 3
T(2, {3, 4}), T(3, {2, 4}), T(4, {2, 3}) are unknown.

(2, 3) + T(3, {4})

T(2, {3, 4}) = (2,4) + T(4, {3}) T(3, {4}) = (3,4) + T(4, {Φ})
min T(4, {3}) = (4,3) + T(3, {Φ})
E(2, 3) = 4, E(2, 4) = 2, while T(3, {4}), T(4, {3}) T(2, {4}) = (2,4) + T(4, {Φ})
are unknown.
T(4, {2}) = (4,2) + T(2, {Φ})
(3, 2) + T(2, {4}) T(2, {3}) = (2,3) + T(3, {Φ})
T(3, {2, 4}) = (3, 4) + T(4, {2})
min T(3, {2}) = (3,2) + T(2, {Φ})
E(3, 2) = 2, E(3, 4) = 5, while T(2, {4}), T(4, {2}) are E(3, 4) = 5, E(4, 3) = 1, E(2, 4) = 2, E(4, 2) = 4, E(2,
unknown. 3) = 4, E(3, 2) = 2,
(4, 2) + T(2, {3}) T(4, {Φ}) = E(4,1 ) = 3, T(3, {Φ}) = E(3, 1) =1, and
T(4, {2, 3}) = (4,3) + T(3, {2}) T(2, {Φ}) = E(2, 1) = 1
min
E(4, 2) = 4, E(4, 3) = 1, while T(2, {3}), T(3, {2}) are
unknown.
 Travelling Salesman Problem
(1, 2) + T(2, {3, 4}) => 1+4 =5
T(1, {2, 3, 4}) = (1,3) + T(3, {2, 4}) => 2+5 =7
min (1,4) + T(4, {2, 3}) => 3+4 =7
E(1, 2) = 1, E(1, 3) = 2, E(1, 4) = 3
T(1, {2, 3, 4}) = min (5, 7, 7)= 5
Min tour cost = 5
(2, 3) + T(3, {4}) =>4+8 = 12 T(3, {4}) = (3,4) + T(4, {Φ}) = 5+3 = 8
T(2, {3, 4}) = (2,4) + T(4, {3}) => 2+2 = 4 T(4, {3}) = (4,3) + T(3, {Φ}) = 1+1 = 2
min E(2, 3) = 4, E(2, 4) = 2
T(2, {4}) = (2,4) + T(4, {Φ}) = 2+3 = 5
T(2, {3, 4}) = min (12,4)=>4
T(4, {2}) = (4,2) + T(2, {Φ}) = 4+1 = 5
(3, 2) + T(2, {4}) => 2+5 = 7 T(2, {3}) = (2,3) + T(3, {Φ}) = 4+1 = 5
T(3, {2, 4}) = (3, 4) + T(4, {2}) => 5+5 = 10
min E(3, 2) = 2, E(3, 4) = 5
T(3, {2}) = (3,2) + T(2, {Φ}) = 2+1 = 3

T(3, {2, 4}) = min (7,10)=>7 E(3, 4) = 5, E(4, 3) = 1, E(2, 4) = 2, E(4,

2) = 4, E(2, 3) = 4, E(3, 2) = 2,
(4, 2) + T(2, {3}) => 4+5 =9
T(4, {2, 3}) = (4,3) + T(3, {2}) => 1+3 = 4 T(4, {Φ}) = E(4,1 ) = 3, T(3, {Φ}) = E(3,
1) =1, and T(2, {Φ}) = E(2, 1) = 1
min E(4, 2) = 4, E(4, 3) = 1
T(4, {2, 3}) = min (9,4)=>4
 Travelling Salesman Problem
(1, 2) + T(2, {3, 4}) => 1+4 =5
T(1, {2, 3, 4}) = (1,3) + T(3, {2, 4}) => 2+5 =7
min (1,4) + T(4, {2, 3}) => 3+4 =7
E(1, 2) = 1, E(1, 3) = 2, E(1, 4) = 3
T(1, {2, 3, 4}) = min (5, 7, 7)= 5
Min tour cost = 5
T(3, {4}) = (3,4) + T(4, {Φ}) = 5+3 = 8
(2, 3) + T(3, {4}) =>4+8 = 12 T(4, {3}) = (4,3) + T(3, {Φ}) = 1+1 = 2
T(2, {3, 4}) = (2,4) + T(4, {3}) => 2+2 = 4 T(2, {4}) = (2,4) + T(4, {Φ}) = 2+3 = 5
min E(2, 3) = 4, E(2, 4) = 2
T(4, {2}) = (4,2) + T(2, {Φ}) = 4+1 = 5
T(2, {3, 4}) = min (12,4)=>4 T(2, {3}) = (2,3) + T(3, {Φ}) = 4+1 = 5
T(3, {2}) = (3,2) + T(2, {Φ}) = 2+1 = 3
(3, 2) + T(2, {4}) => 2+5 = 7
Find tour
T(3, {2, 4}) = (3, 4) + T(4, {2}) => 5+5 = 10
T(1, {2, 3, 4}) = (1, 2) + T(2, {3, 4})
min E(3, 2) = 2, E(3, 4) = 5
T(2, {3, 4}) = (2, 4) + T(4, {3})
T(3, {2, 4}) = min (7,10)=>7 T(4, {3}) = (4,3) + T(3, {Φ})

(4, 2) + T(2, {3}) => 4+5 =9

T(4, {2, 3}) = (4,3) + T(3, {2}) => 1+3 = 4
min E(4, 2) = 4, E(4, 3) = 1
T(4, {2, 3}) = min (9,4)=>4
Travelling Salesman Problem
From the example, it is clear that main problem can be written in terms of small
problems and small problem is solved recursively until we reached to the base case.
However, the complexity of the used approach is very high. So, can we reduce time
complexity?
Complexity can be reduced using DP. DP will only applicable if we can write recursive
equation and find overlapping solutions.
(i, 1) S=Φ
T (i, S) = min( (i, j) + T(j, S-{j})) S != Φ
j∈S
Every vertex is
picked from S

If there will no edge b/w two vertices (vertices are not adjacent) cost will ∞.
Let us construct recursion tree for the previous example.
 Travelling Salesman Problem

T(1, {2,3,4})

T(2, {3, 4}) T(3, {2, 4}) T(4, {2, 3})

T(3, {4}) T(4, {3}) T(2, {4}) T(4, {2}) T(2, {3}) T(3, {2})

T(4, {Φ}) T(3, {Φ}) T(4, {Φ}) T(2, {Φ}) T(3, {Φ}) T(2, {Φ})

Overlapping subproblems
There are total 15 subproblems and out of which 3 are repeating. So, total unique
subproblems are 12, which is still very large. So, DP approach can also not be able reduce
complexity below exponential. So, it is not recommended to use DP for TSP problem.
Travelling Salesman Problem
At first level, i has only three values i.e., 2, 3, 4 and at second level i has only three
values i.e., 2, 3, 4. Third level also has the same number of values. So, we can say at
each level, i will have (n-1) values. Set S has (n-2) elements at level 2, (n-3) at level 2
and the last level has 0 elements.
Travelling Salesman Problem
The number of unique subproblems for n vertices will be equal to (n-
2) 2n , which is also exponential. The time complexity to find the
shortest path = O(n 2n).
The time complexity of the DP approach is lesser than the brute-force
approach O(n!). However, it is still very high, so it is normally not
recommended to use DP for TSP problem.