Business Optimization 2024

Course Summary By:

Asma Ben Yaghlene Riahi
Roua Ben Ammar
Fedia Daami

1
Content
1. Linear Programming Assumptions
2. Formulation
3. Graphical Solution
4. Algebraic solution: The Simplex Method
5. Duality Theorem
6. Sensitivity Analysis

2
 Assumptions of LPs
1. Linearity: both the objective function and the constraints are linear. This
means that each decision variable is multiplied by a constant coefficient, and
these products are either added or subtracted. There are no quadratic,
exponential, or other non-linear terms in the equations.
2. Certainty: It means that numbers in the objective and constraints are known
with certainty and do not change during the period being studied
3. Additivity: This assumption states that the total value of the objective
function and each constraint is the sum of the values of the individual terms. In
other words, the effect of changing a decision variable is independent of the
levels of other variables.
4. Divisibility: The divisibility assumption implies that decision variables can
take on any value, including fractional values. This means that solutions to LP
problems can include non-integer values. When integer decision variables must
be used in a program , we call it an integer linear program

3
Assumptions of LPs
5. Proportionality: the contribution of each decision variable to the total value
of the objective function, and each constraint, is directly proportional to the value
of the variable.

6. Non-negativity: all decision variables must be greater than or equal to

zero. This assumption is based on the practical consideration that negative
amounts of resources or activities usually do not make sense (e.g., you cannot
produce a negative amount of a product).

7. Finite number of variables and constraints: Although not always

explicitly stated, there's an underlying assumption that the LP problem is
described by a finite number of variables and constraints. This is necessary for
computational feasibility.

4
Formulation
› Formulating a problem corresponds to translating the information given in its description into a
system of linear equations. The formulation process can be summarized in the following steps:
• Identify the objective
• The objective function can be to Maximize (e.g. Max profit) or to
1 Minimize (e.g. Min cost or distance)

• Identify the decision variables (The most crucial step!)

• These are the entities you can control in the situation
2

• Identify the constraints: (≤ , ≥ or = )

• A constraint can be raw material, budget, labor hours, space, marketing
requirements,…
3 • DO NOT FORGET THE IMPLICIT CONSTRAINTS SUCH AS THE
NON-NEGATIVITY ONES !

5
Formulation-continued
› Examples of constraints:

∙ At least 2 among the four projects must be chosen: X1+X2+X3+X4 ≥ 2

∙ Projects 1 and 4 are mutually exclusive: X1+X4 ≤ 1

∙ Project 3 cannot be selected unless project 2 is also selected: X3 ≤ X2

6
 Graphical Solution
› Any point that satisfies all the constraints of the model is called a feasible solution. It can be an interior point [inside the feasible
region], a boundary point [belongs to the line of a constraint but it is not the point of intersection point with another constraint], or
an extreme (corner) one [intersection point of two constraints]. The set of all feasible solutions is called the feasible region. A
feasible solution that makes the value of the objective function an optimum (maximum or minimum) is called an optimal
solution. It is one of the extreme (corner) points.

• Point E is called an interior point

• Poin F is called a boundary point
• Points A, B, C and D are corner /extreme points
• Any point whitin the region ABCD represents a feasible
solution (this includes all points on segments [DC] , [AB],
[BC] and [AD] and all the points inside the FR such as
point E)
• The region ABCD is the feasible region
• The optimal solution could either be A, B, C or D
(depending on the objective value)

7
Graphical Solution- continued
› To solve a problem graphically, we have to follow the following steps:
1. Rewrite the constraints of the LP as equality ones. To do that, you have to add the slack variables for
the constraints of type “≤” and the surplus variables for the constraints of type “≥”

2. Draw the axis that represent the non-negativity constraints ( this is a step that is often forgotten but
is actually crucial! , when you have non-negativity constraints, the FR is located in the upper right
side of the xy plane !)
3. Draw the 1st constraint by finding 2 points that satisfy the inequation of the constraint ( Choose a
value for one variable and deduce the other, always go for easy values first for ex x=0)

8
Graphical Solution- continued
3. Test the (0,0) point to know where is the feasible region and color it. We plug-in the point
O(0,0) in C1’s equation: 0+2*0<=0 the region limited (constrained by) C1 is OBG

4. Draw the 2nd constraint. Test the (0,0) point to locate the feasible region (that satisfies both
constraints) and color it.
5. Draw the nth constraint. Test the (0,0) point to locate the feasible region (that satisfies all constraints)
and color it.

9
Graphical Solution- continued
6. Draw Z1: Choose a value for your objective function (for example Z=0) then find two points which
coordinates satisfy this value and draw your Z1 line :
Exp : Min Z= -6x+3y , let Z= 0 , two points that satisfy this equation are: (0, 0) and (1,2)
7. Draw Z2: Choose a second constant equals to your objective function. If your objective is to maximize
the function, the second constant must be greater than the first one. However, if your objective is to
minimize the function, the second constant must be less than the first one. Then, find two points and
draw your Z2 line.
8. Now that you know the direction of your objective function, pick up a ruler and put it on the Z2 line.
Slide it in a parallel manner until it touches the last corner point before quitting the feasible region. This
point is your optimal solution.

10
Graphical Solution- continued
› Note that if the Z line touches two points simultaneously before quitting the feasible region, we have
a multiple solutions case. Here the segment containing these two optimal points is composed of an
infinite number of optimal solutions. Any weighted average of optimal solutions is also an optimal
solution. The case of multiple solutions happens when the objective function is parallel to one of the
constraints.

11
Graphical Solution- continued

› The problem can be unbounded, which can be determined graphically. The feasible region
would be “open” and the direction of the Z line follows this opening

The problem can also be unfeasible, where we cannot find a feasible region. This means that we do
not have any point that satisfies all the constraints.

12
Graphical Solution- continued
How to know if the constraint is binding (active)?

› A binding constraint is the one that when relaxing it the optimal solution changes. That is, the optimal solution is
the intersection between at least two active constraints. Note that the slack of a binding constraint is equal to zero
and any additional unit of this “resource” will improve the Z value.

13
Graphical Solution- continued
›

14
Algebraic solution: Simplex method
∙ The simplex algorithm starts from a basic feasible solution (corner point) and moves to a better
adjacent basic feasible solution, performing an optimality test at each “move” (iteration) until
optimality is reached.
∙ If an optimal solution of a LP exists, then at least one of the corner points is optimal.
∙ In order to determine analytically the basic feasible solutions of a linear problem, we start by solving
the system of binding constraints written in standard form. A basic solution is formed through a
system for which the number of equations is equal to the number of variables

15
Algebraic solution: Simplex method

› To solve a linear problem using the simplex method, follow these steps:

1) Write your LP in the standard form. You just need to add the necessary slack/surplus variables to
obtain equality constraints.

16
Algebraic solution: Simplex method
2) Write your LP in a tabular form (initial simplex tableau)

17
Algebraic solution: Simplex method
3) If your objective function is a maximization one, choose the variable having the largest ∆j value to
enter the basis. However, if your objective is a minimization one, you will choose the variable having the
most negative Cj-Zj value to enter the basis. For both cases, this variable is called the entering variable
and its column is the pivot column.

4) Perform the ratio test: divide the R.H.S value by that of the same row of the pivot column.
Independently if you are maximizing or minimizing your objective function, you will choose the variable
having the smallest positive value in the ratio test column. Note that dividing by 0 or a negative value is
not possible. The chosen variable is the leaving one and its row is called the pivot row. The intersection
between the pivot row and the pivot column is called the pivot element.

18
Algebraic solution: Simplex method

We can now move to the next simplex tableau

19
Algebraic solution: Simplex method
›

20
Algebraic solution: Simplex method

21
Algebraic solution: Simplex method
5) Once all values are found, check if you reached the optimal tableau. For this, check the ∆j values. If you are
maximizing your objective, you reach the optimal tableau when all values are 0 or negative. However, if you are
minimizing your objective, all ∆j values must be zero or positive to say that the tableau is optimal.If optimality is
reached, stop and provide the optimal solution. Else, find the pivot column, the pivot row, and the pivot element
and move to the next iteration. Then perform all iterations needed till optimality is reached.

How to know what is the value of each variable in a given tableau?

› If the variable is not basic, then its value is equal to zero.

› If the variable is basic, then its value is equal to the one in the R.H.S column.

› The Z value is equal to the value corresponding to the intersection between the Z row and the R.H.S
column.

22
Simplex method (special cases)
› How to know if the problem has alternate (multiple solutions) or if it is unbounded?
› If we have an unbounded solution : We only have zero or negative values in the pivot column and
hence the ratio test fails the entering variable can improve the objective value with no stop

› If we have multiple (alternate) solutions : there will be a non-basic variable with a cbar equal to zero
when optimality is reached:

23
Duality Theory
› The notion of duality within LP asserts that every LP has an associated LP called the dual. The
original problem in relation to its dual is termed the primal

› The essence of duality theory is the relationship between the primal and its dual, both on a
mathematical and economic level.

› To find the dual of a LP, there is an important notion to keep in mind which is the canonical form:

› For a maximization problem : the ≤ constraints are said to be following the canonical form. To be
able to memorize this, you can think of the canonical form as the « intuitive » or « logical » form
(despite that is not what it actually means). When we are maximizing, it is « logical » that the
constraints represent a ceiling to your objective function and hence they should be ≤

24
Duality Theory
› For a minimization problem: the ≥ constraints are said to be following the canonical form.
You can memorize using the same logic. When we are minimizing, it is « logical » that your
constraints represent a floor to your objective function and hence they should be ≥

› Remarks:

∙ Do not forget that you can use the usual trick of multiplying the constraint by -1to have the
canonical form and hence a positive dual variable.

∙ For any primal linear program, the dual of the dual is the primal

25
 The table below summarizes the components of the primal LP and their equivalents in the dual LP
Primal Dual
Maximization Minimization
Minimization Maximization
Number of constraints Number of variables
constaint ≤ in case of maximization Variable positive or zero
constraint ≤ in case of minimization Variable negative or zero
constraint ≥ in case of maximization Variable negative or zero

constraint ≥ in case of minimization Variable positive or zero

constraint = Unconstrained variable
Number of variables Number of constraints
Variable positive or zero in case of maximization constraint ≥
Variable positive or zero in case of minimization constraint ≤
Variable negative or zero in case of maximization constraint ≤
Variable negative or zero in case of minimization constraint ≥
Unconstrained variable constraint =
Coefficient of the jth variable in the objective function RHS of the jth constraint
RHS of the ith constraint Coefficient of the ith variable in the objective function
Technological coefficient of the jth variable in the ith constraint Technological coefficient of the ith variable in the jth constraint

26
Duality Theory
› Example:

27
Duality Theory
› Theorems related to the duality theorem (without proof):
weak duality theorem:

• If x is primal feasible and y is dual feasible, then cTx ≤ bTy.

• If x* is primal feasible and y* is dual feasible and cTx* = bTy* then x* and y* are optimal

• If one of a pair of primal and dual problems has an optimal solution, then the other also has an optimal solution
and the optimal values of their objective functions are equal.

• If the primal is feasible and unbounded then the dual is unfeasible.

• If the dual is feasible and unbounded, then the primal is unfeasible.

28
 Duality Theory
• Complementary Slackness Theorem: A primal feasible solution x* and a
dual feasible solution y* are optimal if and only if (b-Ax*)Ty* = 0 and
(ATy* - c)Tx* = 0.

✔ If xj*>0, then the jth constraint of the dual is binding.

✔ If the jth constraint of the dual is not binding, then xj*=0.

✔ If yi*>0, then the ith constraint of the primal is binding.

✔ If the ith constraint of the primal is not binding, then yi*=0

29
Duality Theory
• The dual variable y* measures the change in the optimal profit due to a unit change in resource i. The fact that the
profit would increase by yi* if an additional unit of resource i was available imputes a value or price to resource i.
This value or price is called the shadow price. Thus, this latter is the amount of contribution of an additional unit of a
resource to total profit. It is also the maximum amount of money that we are willing to pay for one additional unit of
the resource.

• If the resource is not totally consumed at optimal exploitation, then the related constraint is inactive (not binding).
Therefore, its corresponding shadow price is equal to zero and its corresponding slack variable is basic in the
optimal simplex tableau (the corresponding slack is positive) which means that the resource is abundant and it is not
worth acquiring additional units of it. However, the corresponding shadow price of a binding (active) constraint is
positive. Therefore, its slack variable is not basic in the optimal simplex tableau. That is, it is worthwhile acquiring
additional units of the resource given that it is totally consumed at optimal exploitation, which means that the resource
is scarce.

30
Duality Theorem-continued
› CJ-ZJ is called the reduced cost of the corresponding primal decision variable. It is interpreted as the deficit for the
decision variable to become basic. Therefore, the reduced cost is the minimum amount of money by which the
coefficient of the decision variable must increase to become profitable.

Example

› We multiply the c-bar values by -1 to find the shadow prices (the values of y1, y2 and y3) and the reduced costs (the
values of E1, E2 and E3) because all the constraints in the primal LP are in the canonical form (≤ for maximization). If a
constraint does not follow the canonical form? the dual variable would be ≤ 0. Hence, we do not multiply by -1.

› If we had a minimization problem (c bar values 0s or positive numbers) and all the constraints follow the canonical
form for minimization (≥), we take the values as they are. If there is a constraint that does not follow the canonical
form, we multiply the cbar value of the variable by -1.

31
Post-optimality analysis
› Before starting the chapter, we cannot talk about post optimality if we do not have the optimal solution of the
original problem. Therefore, we need the optimal simplex tableau to perform sensitivity analysis.

• Sensitivity analysis is the process of determining the range of parameter(s) for which optimality or the optimal
basis remains unchanged.

• This chapter can be divided into the following sections:

1. Change in the objective function’s coefficients

2. Change in the R.H.S of a constraint

3. Introducing a new constraint

4. Deleting a constraint

5. Introducing a new variable

6. Deleting a variable

32
Post-optimality analysis: Change in the objective function’s
coefficients

› When changing the coefficient in the objective function of a given variable, we have to check if it is basic in
the optimal simplex tableau or not.

› If it is basic, we will do all necessary changes and then determine the optimality interval (the range at which
the coefficient of the variable in the objective function may vary while preserving an optimal simplex tableau)
in order to know if the change will affect the optimal solution or not. Note that even if optimality is preserved,
the Z value will change given that the change corresponds to the coefficient of a basic variable.

› If the variable is not basic in the optimal simplex tableau, the only value that will change is its cbar. Therefore,
to preserve optimality of the basis, it suffices that the new Cj-Zj value be negative or zero if we are
maximizing or positive or zero if we are minimizing. Hence, if optimality is preserved, the Z value will
remain the same.

33
Post-optimality analysis: Change in the R.H.S of a constraint
› The change in the RHS of the constraints would only affect the corresponding column in the simplex
tableau. It suffices to take the coefficients of Si in the optimal tableau as those of Δbi in column bi.

› Note that for the basis to remain optimal, the corresponding solution must be feasible (R.H.S zero or
positive). Therefore, we need to determine the feasibility interval (the range at which the R.H.S of a
constraint may vary while preserving the optimal basis of the last simplex tableau) to know by how
much we can increase or decrease the resource availability without affecting the feasibility of the
problem.

› We may equivalently obtain the new optimal objective value (Z old + Zj of Si). In fact, while keeping a
variation within the feasibility interval of the RHS, row Δj remains unchanged and hence optimality is
preserved.

34
Post-optimality analysis: Introducing a new constraint

› When we introduce a new constraint to the linear problem, we have to check if it is satisfied by the
optimal solution or not.

› If yes, we say that the constraint is redundant and optimality is preserved.

› If not, optimality is not preserved and the new optimal solution will be worse than the “old” one.

› Note that to find the new optimal solution, we have to introduce the new constraint to the original LP
and solve it a second time.

35
Post-optimality analysis: Deleting a constraint
› When deleting a constraint, we have to check if it is binding or not. In other words, we have to see if

its corresponding slack variable is basic in the optimal simplex tableau or not. If it is basic (slack

positive), it means that the constraint is inactive (not binding). Therefore, removing it will not affect

the optimal solution. Therefore, it suffices to remove the row and column of the corresponding slack

variable from the optimal simplex tableau and provide the same “old” optimal solution without the

value of Si. However, if the slack variable of the constraint that we want to delete is not basic (it is

equal to zero) at the optimal simplex tableau, it means that optimality will change given that the

constraint is binding. Therefore, we have to proceed as follows:

36
 Post-optimality analysis: Deleting a constraint
✔ Force the slack variable of the constraint that we want to delete to enter the basis (even it has
a negative ∆j value if we are maximizing or a positive ∆j value if we are minimizing).

✔ Perform the ratio test and determine the leaving variable as well as the pivot element.

✔ Perform an iteration. It will give a worse Z value but we have now a non-binding constraint
that we can delete by removing its corresponding row and column.

✔ Check if the obtained tableau is optimal. If yes, stop and provide the new optimal solution.
Else, perform all needed iterations until optimality is reached and then provide the new
optimal solution.

37
Post-optimality analysis: Introducing a new variable
› Introducing a new variable in the primal is equivalent to introducing a new dual constraint.

› If the constraint is satisfied by the dual optimal solution, then we will not introduce the new variable.

› Else, it is worthwhile introducing the new variable to the primal.

› This can be explained by the fact that the primal and dual are a sort of “opposites”. In the case of adding a constraint,
we said that if optimality is not preserved, the new optimal solution will be worse than the “old” one.

› We know that the primal and dual solutions do not coincide except at optimality. Therefore, if the dual optimal solution
is worse, then the primal one will be better. Hence, it is beneficial to introduce the new variable.

› However, if the dual optimal solution will not change because the introduced constraint is redundant, it is not worth to
introduce the new variable to the primal given that at optimality both primal and dual optimal solutions coincide and
hence the primal optimal solution will not change as well

38
Post-optimality analysis: Deleting a variable
› When we want to delete a variable from the linear program, we have to check if it is basic in the
optimal simplex tableau or not.

› If it is not basic, the optimal solution will not change, as this variable is already equal to zero in the
optimal tableau. However, if the variable that we want to delete is basic, we have to add the constraint
Xi=0 and solve the LP again. Here, it is obvious that the optimal solution will change and will be
worse than the “old” one (if the LP remains feasible) as we are adding a constraint.

39
Additional Resources
Video tutorials:
https://youtube.com/playlist?list=PLx8TC2zdR0F8MdGbkFbw9o0ZvyPBTpwab

Old class files: (if you want to check them out stick to the most recent ones)
https://drive.google.com/drive/folders/0B5BPmfTpDLYPQnR3U1U2bWxaN0E?resour
cekey=0-VWKpC_O1uVah6chrLxNdBg

40