Part-3: Operations Research

Module-1: Linear programming

Topics: Formulations, Graphical method, Sensitivity analysis

Formulations
What is Operations Research? It is the application of mathematical methods and computing
in decision making arising in the domains of management, engineering, policy making, etc.
Major stages in an Operations Research exercise involves: a) understanding the system, b)
formulating a mathematical model, c) finding a good solution, and d) implementation. In the
classroom setting, our focus will be on the second and third stages.

Let us consider an example. A carpenter is receiving many orders for chairs and tables, more
that what it can fulfill. What should it do, as in which orders to accept and which orders to
reject or delay? A careful look into the carpenter’s operations reveals the following: A table
requires 5 units of timber and 9 units of labor, whereas a chair requires 2 units of timber and
6 units of labor. Selling price is ₹7000 for a table and ₹4000 for a chair. 40 units of timber
and 90 units of labor are available with the carpenter.

The first step in formulating a mathematical model is to introduce decision variables. Here,
it’s production quantities of tables and chairs, which we denote by 𝑥 and 𝑦. Note that 𝑥 and 𝑦
should be non-negative integers. However, we can consider them to take non-integer values
as well, as the carpenter’s business will continue after the current set of tables and chairs are
sold. A fractional value simply means a work-in-progress that will be completed and sold in
future. It is always easier to solve a mathematical model involving continuous variables than
those involving integer variables. So, we must try to formulate mathematical models using
continuous variables as much as we can. Thus, we take 𝑥, 𝑦 ≥ 0.

Once we finalize the decision variables, we capture the restrictions on them using equality or
inequality relations. Here, the production quantities must be such that the timber and labor
used do not exceed their availabilities. With 𝑥 units of table and 𝑦 units of chairs produced,
the timber used is 5𝑥 + 2𝑦, which must not exceed 40, and the labor used is 9𝑥 + 6𝑦, which
must not exceed 90. Now, we are left with formulating the objective (of the decision maker).
Here, maximizing revenue, which is 7𝑥 + 4𝑦 (in thousand rupee), seems appropriate. Now,
we write the entire formulation together in the following manner.

Max 7𝑥 + 4𝑦 (Objective)
s. t. 5𝑥 + 2𝑦 ≤ 40 (Constraint-1)
9𝑥 + 6𝑦 ≤ 90 (Constraint-2)
and 𝑥, 𝑦 ≥ 0 (Variable type)

Observe that all the expressions in the above formulation are linear functions of the decision
variables. We refer to this kind of Operations Research problems as the linear programming

MBA651 Notes prepared by Dr. Avijit Khanra

problem. It is the simplest to solve and the most widely used Operations Research model. In
this module, we focus on linear programming. In the next module, integer programming and
network models will be introduced. In the last module, we will have a glimpse of stochastic
Operations Research problems. Some of the important topics that we skip include dynamic
programming, non-linear programming, and multi-objective optimization. Let us see some
linear programming formulations before attempting to solve it.

Diet problem: There are 𝑖 = 1,2, … , 𝑚 food items and 𝑗 = 1,2, … , 𝑛 nutrients. Let 𝑎𝑖𝑗 denote
the amount of nutrient-𝑗 present in one unit of food-𝑖. Every day an army jawan needs at least
𝑏𝑗 amount of nutrient-𝑗. One unit of food -𝑖 costs 𝑐𝑖 amount of money. We need to prepare a
cost-effective diet plan to fulfill the nutrient requirements. Note that 𝑎𝑖𝑗 , 𝑏𝑗 , and 𝑐𝑖 are known
numbers. These are represented in a tabular form below.

Nutrient-1 Nutrient-2 ⋯ Nutrient-𝑛 Unit cost Quantity

Food-1 𝑎11 𝑎12 ⋯ 𝑎1𝑛 𝑐1 𝑥1
Food-2 𝑎21 𝑎22 ⋯ 𝑎2𝑛 𝑐2 𝑥2
⋮ ⋮ ⋮ ⋱ ⋮ ⋮ ⋮
Food-𝑚 𝑎𝑚1 𝑎𝑚2 ⋯ 𝑎𝑚𝑛 𝑐𝑚 𝑥𝑚
Requirement 𝑏1 𝑏2 ⋯ 𝑏𝑛

The above table also shows the decision variables 𝑥1 , 𝑥2 , … , 𝑥𝑚 denoting quantity of the food
items to be included in the diet. The mathematical model is given below.

Min 𝑐1 𝑥1 + 𝑐2 𝑥2 + ⋯ + 𝑐𝑚 𝑥𝑚
𝑠. 𝑡. 𝑎11 𝑥1 + 𝑎21 𝑥2 + ⋯ + 𝑎𝑚1 𝑥𝑚 ≥ 𝑏1
𝑎12 𝑥1 + 𝑎22 𝑥2 + ⋯ + 𝑎𝑚2 𝑥𝑚 ≥ 𝑏2
⋮
𝑎1𝑛 𝑥1 + 𝑎2𝑛 𝑥2 + ⋯ + 𝑎𝑚𝑛 𝑥𝑚 ≥ 𝑏𝑛
and 𝑥1 , 𝑥2 , … , 𝑥𝑚 ≥ 0

Blending problem: Suppose that three types of oil are produced by blending three types of
crude. Features of crudes and requirements of oils are given in the table below.

Octane Sulphur Cost Oil-1 Oil-2 Oil-3

Crude-1 12 0.5% 45/L Min Octane 10 8 6
Crude-2 6 2.0% 35/L Max Sulphur 1.0% 2.0% 1.5%
Crude-3 8 3.0% 25/L Price 70/L 60/L 50/L

Availability of each crude is 5000L and any amount of oil produced can be sold. Considering
cost of crudes alone, we need to a prepare a blending plan to maximize profit. The following
table presents the decision variables and some calculations for model building.

Variables Oil-1 Oil-2 Oil-3 Crude used

Crude-1 𝑥11 𝑥12 𝑥13 𝑥11 + 𝑥12 + 𝑥13

MBA651 Notes prepared by Dr. Avijit Khanra

Crude-2 𝑥21 𝑥22 𝑥23 𝑥21 + 𝑥22 + 𝑥23
Crude-3 𝑥31 𝑥32 𝑥33 𝑥31 + 𝑥32 + 𝑥33
Oil produced 𝑥11 + 𝑥21 + 𝑥31 𝑥12 + 𝑥22 + 𝑥32 𝑥13 + 𝑥23 + 𝑥33 Requirements
12𝑥11 + 6𝑥21 + 8𝑥31 12𝑥12 + 6𝑥22 + 8𝑥32 12𝑥13 + 6𝑥23 + 8𝑥33 ≥ 10,8,6,
Octane
𝑥11 + 𝑥21 + 𝑥31 𝑥12 + 𝑥22 + 𝑥32 𝑥13 + 𝑥23 + 𝑥33 respectively
0.5𝑥11 + 2𝑥21 + 3𝑥31 0.5𝑥12 + 2𝑥22 + 3𝑥32 0.5𝑥13 + 2𝑥23 + 3𝑥33 ≤ 1,2,1.5,
Sulphur
𝑥11 + 𝑥21 + 𝑥31 𝑥12 + 𝑥22 + 𝑥32 𝑥13 + 𝑥23 + 𝑥33 respectively

Now, the mathematical model can be constructed as follows:

Max 70(𝑥11 + 𝑥21 + 𝑥31 ) + 60(𝑥12 + 𝑥22 + 𝑥32 ) + 50(𝑥13 + 𝑥23 + 𝑥33 )
− 45(𝑥11 + 𝑥12 + 𝑥13 ) − 35(𝑥21 + 𝑥22 + 𝑥23 ) − 25(𝑥31 + 𝑥32 + 𝑥33 )
𝑠. 𝑡. 𝑥11 + 𝑥12 + 𝑥13 ≤ 5000 ∼ 𝐶𝑟𝑢𝑑𝑒 1 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑖𝑙𝑖𝑡𝑦
𝑥21 + 𝑥22 + 𝑥23 ≤ 5000 ∼ 𝐶𝑟𝑢𝑑𝑒 2 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑖𝑙𝑖𝑡𝑦
𝑥31 + 𝑥32 + 𝑥33 ≤ 5000 ∼ 𝐶𝑟𝑢𝑑𝑒 3 𝑎𝑣𝑎𝑖𝑙𝑎𝑏𝑖𝑙𝑖𝑡𝑦
12𝑥11 + 6𝑥21 + 8𝑥31 ≥ 10(𝑥11 + 𝑥21 + 𝑥31 ) ∼ 𝑂𝑖𝑙 1 𝑚𝑖𝑛 𝑜𝑐𝑡𝑎𝑛𝑒
12𝑥12 + 6𝑥22 + 8𝑥32 ≥ 8(𝑥12 + 𝑥22 + 𝑥32 ) ∼ 𝑂𝑖𝑙 2 𝑚𝑖𝑛 𝑜𝑐𝑡𝑎𝑛𝑒
12𝑥13 + 6𝑥23 + 8𝑥33 ≥ 6(𝑥13 + 𝑥23 + 𝑥33 ) ∼ 𝑂𝑖𝑙 3 𝑚𝑖𝑛 𝑜𝑐𝑡𝑎𝑛𝑒
0.5𝑥11 + 2𝑥21 + 3𝑥31 ≤ 1(𝑥11 + 𝑥21 + 𝑥31 ) ∼ 𝑂𝑖𝑙 1 𝑚𝑎𝑥 𝑠𝑢𝑙𝑝ℎ𝑢𝑟
0.5𝑥12 + 2𝑥22 + 3𝑥32 ≤ 2(𝑥12 + 𝑥22 + 𝑥32 ) ∼ 𝑂𝑖𝑙 2 𝑚𝑎𝑥 𝑠𝑢𝑙𝑝ℎ𝑢𝑟
0.5𝑥13 + 2𝑥23 + 3𝑥33 ≤ 1.5(𝑥13 + 𝑥23 + 𝑥33 ) ∼ 𝑂𝑖𝑙 3 𝑚𝑎𝑥 𝑠𝑢𝑙𝑝ℎ𝑢𝑟
𝑎𝑛𝑑 𝑥𝑖𝑗 ≥ 0 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑖, 𝑗

Production process problem: Let us consider the perfume manufacturing process illustrated
in the diagram below. It shows processing cost and time per liter of input of three processes.
Process-A takes raw material as input and outputs regular male and regular female perfumes.
Processes-B and C take the regular perfumes as input and output premium perfumes. A part
of the input is wasted during processing (leading to total output less than 100%).

4000 liter raw material and 6000 hour of time (for all processes together) are available in a
year. Maximum demand for regular male and regular female perfumes are 800 and 1200 liter,
respectively. For premium perfume, the maximum demands are 200 and 300 liter. We need to
find the best production plan, which can be captured by the raw material processed and the
regular perfumes processed further or sold in the market. Let 𝑥 denote the amount of raw

MBA651 Notes prepared by Dr. Avijit Khanra

material processes and 𝑦 and 𝑧 denote the regular male and regular female perfume sold. All
numbers are in liter. The corresponding production plan is shown below.

Our aim is to determine 𝑥, 𝑦, 𝑧 that maximizes profit while satisfying the restrictions on raw
material, processing time, and demand as well as the “internal restrictions” inherent to the
processes. The mathematical model (non-compact version) is presented below.

Max 7𝑦 + 6𝑧 + 15 × 0.7(0.3𝑥 − 𝑦) + 13 × 0.8(0.4𝑥 − 𝑧)

− 3𝑥 − 𝑥 − 4(0.3𝑥 − 𝑦) − 3(0.4𝑥 − 𝑧)
𝑠. 𝑡. 𝑥 ≤ 4000 ∼ 𝑅𝑎𝑤 𝑚𝑎𝑡𝑒𝑟𝑖𝑎𝑙
𝑥 + 3(0.3𝑥 − 𝑦) + 2(0.4𝑥 − 𝑧) ≤ 6000 ∼ 𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔 𝑡𝑖𝑚𝑒
𝑦 ≤ 800 ∼ 𝐷𝑒𝑚𝑎𝑛𝑑 𝑟𝑒𝑔𝑢𝑙𝑎𝑟 𝑚𝑎𝑙𝑒
𝑧 ≤ 1200 ∼ 𝐷𝑒𝑚𝑎𝑛𝑑 𝑟𝑒𝑔𝑢𝑙𝑎𝑟 𝑓𝑒𝑚𝑎𝑙𝑒
0.7(0.3𝑥 − 𝑦) ≤ 200 ∼ 𝐷𝑒𝑚𝑎𝑛𝑑 𝑝𝑟𝑒𝑚𝑖𝑢𝑚 𝑚𝑎𝑙𝑒
0.8(0.4𝑥 − 𝑧) ≤ 300 ∼ 𝐷𝑒𝑚𝑎𝑛𝑑 𝑝𝑟𝑒𝑚𝑖𝑢𝑚 𝑓𝑒𝑚𝑎𝑙𝑒
𝑦 ≤ 0.3𝑥 ∼ 𝐼𝑛𝑡𝑒𝑟𝑛𝑎𝑙 𝑙𝑖𝑚𝑖𝑡 𝑜𝑛 𝑦
𝑧 ≤ 0.4𝑥 ∼ 𝐼𝑛𝑡𝑒𝑟𝑛𝑎𝑙 𝑙𝑖𝑚𝑖𝑡 𝑜𝑛 𝑧
𝑎𝑛𝑑 𝑥, 𝑦, 𝑧 ≥ 0

Minimum absolute deviation regression: In the simple linear regression, we find the straight
line 𝑦 = 𝑎 + 𝑏𝑥 that best fits a given data set of the form (𝑥𝑖 , 𝑦𝑖 ) for 𝑖 = 1,2, … , 𝑛. The best-
fit line is obtained by minimizing total squared deviation of the data from the regression line,
i.e., 𝑎 and 𝑏 are obtained by minimizing ∑𝑛𝑖=1(𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 )2. Deviation can be measured by
absolute distance, i.e., |𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 |, which is a “more natural” measure. It is not popular
because minimizing absolute value is difficult. Linear programming can help.

In minimum absolute deviation regression, we choose 𝑎 and 𝑏 soh that ∑𝑛𝑖=1|𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 | is
minimized. Note that |𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 | is non-linear (why?). We need to “linearize” it.
𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 if 𝑎 + 𝑏𝑥𝑖 ≤ 𝑦𝑖
Let 𝑣𝑖 = { for 𝑖 = 1,2, … 𝑛 ⋯ ⋯ (1)
𝑎 + 𝑏𝑥𝑖 − 𝑦𝑖 if 𝑎 + 𝑏𝑥𝑖 > 𝑦𝑖

Then minimizing ∑𝑛𝑖=1|𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 | is same as minimizing ∑𝑛𝑖=1 𝑣𝑖 subject to (1) as the
constraint, where 𝑣1 , 𝑣2 , … , 𝑣𝑛 are decision variables in addition to 𝑎 and 𝑏. The objective
function ∑𝑛𝑖=1 𝑣𝑖 is linear, but the constraint (1) is non-linear (why?).

MBA651 Notes prepared by Dr. Avijit Khanra

Let us introduce new decision variables 𝑢1 , 𝑢2 , … , 𝑢𝑛 and new constraints.

𝑢𝑖 ≥ 𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 for 𝑖 = 1,2, … 𝑛 ⋯ ⋯ (2)

𝑢𝑖 ≥ 𝑎 + 𝑏𝑥𝑖 − 𝑦𝑖 for 𝑖 = 1,2, … 𝑛 ⋯ ⋯ (3)

Note that 𝑣𝑖 captures the exact value of |𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 | through (1), whereas 𝑢𝑖 captures all
values greater than or equal to |𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 | through (2) and (3). Therefore, 𝑢𝑖 ≥ 𝑣𝑖 for all 𝑖
and they can be equal. Then minimizing ∑𝑛𝑖=1 𝑣𝑖 subject to (1) produces the same absolute
deviation as minimizing ∑𝑛𝑖=1 𝑢𝑖 subject to (2) and (3) does. Hence, the linear programming
model for mean absolute deviation regression is given as follows:

Min ∑𝑛𝑖=1 𝑢𝑖
𝑠. 𝑡. 𝑢𝑖 ≥ 𝑦𝑖 − 𝑎 − 𝑏𝑥𝑖 for 𝑖 = 1,2, … 𝑛
𝑢𝑖 ≥ 𝑎 + 𝑏𝑥𝑖 − 𝑦𝑖 for 𝑖 = 1,2, … 𝑛
𝑎𝑛𝑑 𝑎, 𝑏 ∈ ℝ, 𝑢𝑖 ≥ 0 𝑓𝑜𝑟 𝑎𝑙𝑙 𝑖

Graphical method
We begin solving linear programming programs with the two variables case. Why don’t we
start with the single-variable case? Let us consider the following problems.

(1) Max 2𝑥 − 𝑦 (1a) Max 𝑥 − 𝑦

s.t. 𝑥 ≤ 3 ⋯ (𝑎) s.t. 𝑥 ≤ 3 ⋯ (𝑎)
𝑥 − 𝑦 ≤ 1 ⋯ (𝑏) 𝑥 − 𝑦 ≤ 1 ⋯ (𝑏)
𝑥 − 2𝑦 ≥ 0 ⋯ (𝑐) 𝑥 − 2𝑦 ≥ 0 ⋯ (𝑐)
and 𝑥, 𝑦 ≥ 0 and 𝑥, 𝑦 ≥ 0

Note that the problems have the

same constraints. We begin by
plotting the constraints in a 2D
plane. This helps us identify the
feasible solutions that are shown
by the shaded triangular region
in the diagram to the right. Any
point outside the shaded region
represents an infeasible solution.
We need to find the best among
these feasible solutions, decided
by the objective of the problems
(which is different for them).

In problem (1), we want to maximize 2𝑥 − 𝑦. Let us pick a feasible solution, say, (1, 1⁄4).
The objective function value at this point is 7⁄4. Let’s plot the line 2𝑥 − 𝑦 = 7⁄4. It passes
through (1, 1⁄4) and many other feasible solutions, as shown in the left-side diagram below.
All these solutions have the same objective function value. We refer to this line as iso-profit

MBA651 Notes prepared by Dr. Avijit Khanra

line. There are many iso- profit lines, and these are always parallel. Objective function values
are the same for the solutions on the same iso-profit line. However, solutions on two different
iso-profit lines cannot have the same objective function value. We must identify the highest-
valued iso-profit line passing through a feasible solution.

The number of distinct iso-profit lines passing through the feasible region is unlimited. So,
one cannot exhaust them by enumeration. However, we can identify the direction of improve-
ment after plotting one. In the left-side diagram above, we start with the iso-profit line 2𝑥 −
𝑦 = 7⁄4. Pick a point on either side of the line, say, (1,0). Objective function value at this
point is 2, which is better than 7⁄4. So, the direction of improvement is towards (1,0) from
the current iso-profit line. Since the objective function is linear, the direction of improvement
never changes. If we keep finding iso-profit lines in this direction, a point would come when
we exit the feasible region. Objective function value is highest there, and we have the optimal
solution. For problem (1), this happens at the point (2,1).

Note that (2,1) is a corner point of the feasible region of problem (1). This is no coincidence.
The feasible region of every two-variable linear programming problem is a polygon. If it is
bounded, then the iso-cost/profit line exits the feasible region, in the direction of improve-
ment, through a corner point. Verify this for problem (1a), which is depicted in the right-side
figure above. You shall be able to see that all the points on the line segment connecting (1,0)
and (2,1) are optimal. This observation allows us to identify optimal solutions simply by
comparing objective function values at the corner points, provided that the feasible region is
bounded. In case multiple corner points are optimal, then all the points connecting them are
also optimal. These principles remain valid for problems involving more variables.

Feasible region of a linear programming problem can be unbounded. Then the corner point
enumeration principle does not hold. Consider the following problems. Their constraints are
the same. The feasible region is unbounded, as shown in the diagram that follows.

(2) Max 2𝑥 − 𝑦 (2a) Max 𝑥 + 𝑦

s.t. 𝑥 ≤ 3 ⋯ (𝑎) s.t. 𝑥 ≤ 3 ⋯ (𝑎)
𝑥 − 𝑦 ≤ 1 ⋯ (𝑏) 𝑥 − 𝑦 ≤ 1 ⋯ (𝑏)
and 𝑥, 𝑦 ≥ 0 and 𝑥, 𝑦 ≥ 0

MBA651 Notes prepared by Dr. Avijit Khanra

The concepts of iso-cost/profit lines and the direction of improvement holds for all linear
programming problems. In the left-side figure, corresponding to problem (2), when we move
in the direction of improvement, the iso-profit line leaves the feasible region via the corner
point (3,2). Hence, it is the optimal solution. However, in the right-side figure, corresponding
to problem (2a), the movement along the direction of improvement never ends. In this case,
optimal solution is unbounded; technically, it does not exist. In general, if the feasible region
is unbounded, optimal solution may not exist; in case it exists, a corner point is optimal. This
principle remains valid for problems involving more variables.

It is possible that a linear prog-

raming problem does not have a
feasible solution. Problem (3) is
such a case, as illustrated in the
figure to the right.

(3) Max 2𝑥 − 𝑦
s.t. 𝑥 ≤ 3 ⋯ (𝑎)
𝑥 − 𝑦 ≤ 1 ⋯ (𝑏)
𝑥 − 2𝑦 ≥ 1.5 ⋯ (𝑐)
and 𝑥, 𝑦 ≥ 0

We can summarize the graphical method as follows:

(1) Plot the constraints in a 2D plane and identify the feasible region.
(2) If there is no feasible solution, declare that the problem is infeasible.
(3) If the feasible region is bounded, then compare the corner points in terms of objective
function value and declare the best corner point as optimal.
(4) If the feasible region is unbounded, then check if optimal solution exists or not.

MBA651 Notes prepared by Dr. Avijit Khanra

 (i) If optimal solution exists, then compare the corner points in terms of objective
function value and declare the best corner point as optimal.
(ii) If optimal solution does not exist, declare the same.

In step-3 and 4.i, it is possible to identify two corner points as optimal. Then all points on the
line segment connecting them are also optimal. For problems with more variables, graphical
method does not work, but its core ideas remain valid. Using those ideas, simplex algorithm
solves linear programming problems of any size. Microsoft Excel has a solver implementing
the simplex algorithm. It will be illustrated in the class.

Sensitivity analysis
Let us consider the carpenter’s problem with the following data.

Table Chair Availability Mathematical model is as follows:

Prod. Qty. 𝑥1 𝑥2 - Max 3𝑥1 + 2𝑥2
Price 3 (𝑐1) 2 (𝑐2 ) - s.t. 2𝑥1 + 𝑥2 ≤ 100
Lumbar 2 1 100 (𝑏1 ) 𝑥1 + 𝑥2 ≤ 80
Labor 1 1 80 (𝑏2 ) 𝑥1 ≤ 40
Max. Dem. 40 (𝑏3 ) ∞ - and 𝑥1 , 𝑥2 ≥ 0

The problem is solved by the graphical

method. Point B (20,60) is the optimal
solution. The prices and the availabilities
may change during the implementation
of the solution. Given this, the following
questions become relevant:
1. What is the impact of a change in the
objective function coefficient on optimal
solution of a problem?
2. What is the impact of a change in the
RHS of a constraint on optimal solution
of a problem?

If the impact of a change in model parameter on the optimal solution is less, then the model is
robust or less sensitive. Robustness is a desirable property in optimization models. In the
above example, sensitivity analysis is performed by visual inspection of the 2D plot. For
larger problems, output of the simplex algorithm contains all the answers.

Change in objective function coefficient: Change in 𝑐1 = 3 or 𝑐2 = 2 changes the slope of the

iso-profit line. This may change the optimal solution. Current slope is 𝑚 = − 𝑐1⁄𝑐2 = −1.5.
If the iso-profit line flattens and the slope becomes −1, the slope of the labor constraint, then

MBA651 Notes prepared by Dr. Avijit Khanra

point A (0,80) becomes optimal. On the other hand, if the iso-profit line gets steeper and the
slope becomes −2, the slope of the lumbar constraint, then point C (40,20) becomes optimal.
Point B remains optimal, if slope of the iso-profit line is in the range [−2, −1]. If 𝑐1 changes
to 3 + 𝛿 and − (3 + 𝛿)⁄𝑐2 ∈ [−2, −1] ⇒ −2 ≤ −(1.5 + 0.5𝛿) ≤ −1 ⇒ −0.5 ≤ −0.5𝛿 ≤
0.5 ⇒ −1 ≤ 𝛿 ≤ 1 ≡ 𝑐1 ∈ [2,4], the optimal solution do not change. Find out the range of
𝑐2 such that the optimal solution does not change.

Hypothetically, let us assume that 𝑐1 = 1. Then slope of the iso-profit line is −0.5, and point
A, i.e., 𝑥1 = 0, 𝑥2 = 80 is optimal. Clearly, producing tables is not attractive anymore. In this
context, one may ask what increase in 𝑐1 would make 𝑥1 > 0 in the optimal solution? If the
iso-profit line is steeper and the slope becomes −1, then B becomes optimal and 𝑥1 > 0. Let
𝛿 denote the required change in 𝑐1. The we need − (1 + 𝛿)⁄𝑐2 = −1 ⇒ (1 + 𝛿)⁄2 = 1 ⇒
𝛿 = 1. So, 𝑐1 should become 2 (or more) for 𝑥1 > 0 in the optimal solution. This 𝛿 = 1 is
known as the reduced cost of 𝑐1. Ideally, it should be called increased profit, but the name
comes from the minimization perspective. There, 𝑐1 must be reduced in order to make 𝑥1 > 0
in the optimal solution. Reduced costs are always non-negative. Note that the idea of reduced
cost is relevant when a variable is zero in the optimal solution.

Change in RHS of a constraint: Change in 𝑏1 = 100 or 𝑏2 = 80 or 𝑏3 = 40 changes the

feasible region, which may change the optimal solution. Let us consider 𝑏1 first. Let 𝑏1 =
100 + Δ. If Δ is positive, the feasible region enlarges and the corner points B and C move
towards E. Since none of the slopes change, the point of intersection of the lumbar and labor
constraints, i.e., B, remains optimal with changed position. If Δ increases further, lumber
constraint becomes redundant and point E becomes a feasible solution as well as the optimal
solution. Note that E is the point of intersection of labor and demand constraints. The limiting
value of Δ such that B stays optimal (with changed position) is obtained from 2𝑥1 + 𝑥2 =
100 + Δ that passes through E (40,40). So, Δ = 2 × 40 + 40 − 100 = 20.

If Δ is negative, the feasible region shrinks and the corner points B and C move towards A
and D, respectively. Since none of the slopes change, again, B remains optimal with changed
position. As Δ decreases, B merges with A and C merges with D, simultaneously. This makes
the labor and demand constraints redundant. Then the triangle OAD becomes the feasible
region, with A and D denoting the intersection of the lumbar constraint with 𝑥2 and 𝑥1 -axis,
respectively. A becomes the new optimal, with a changed position. The limiting value of Δ
such that B stays optimal (with changed position) is obtained from 2𝑥1 + 𝑥2 = 100 + Δ that
passes through A (0,80). So, Δ = 2 × 0 + 80 − 100 = −20.

Combining the above observations, if Δ ∈ [−20,20], the point of intersection of the lumbar
and labor constraint (point B) remains optimal. However, the optimal solution changes, i.e.,
position of B changes. New position can be obtained by solving 2𝑥1 + 𝑥2 = 100 + Δ and
𝑥1 + 𝑥2 = 80. It is 𝑥1 = 20 + Δ, 𝑥2 = 60 − Δ, and the new objective function value is 3𝑥1 +
2𝑥2 = 180 + Δ. Find the range of 𝑏2 , which is currently 80, such that point B stays optimal.
Identify its new position and the new objective function value. Observe that the impact of

MBA651 Notes prepared by Dr. Avijit Khanra

changes in an objective function coefficient and the RHS of a constraint are quite different. In
the first case, the slope of the iso-cost/profit line changes whereas the feasible region does not
change. It is reversed in the second case.

To capture the relative importance of different constraints, we define shadow price as the rate
of improvement of the objective function value (which is increase for maximization problems
and decrease for minimization problems) as the RHS of a constraint increases, provided that
such a change is allowed without changing the character of the optimal solution (captured by
the constraints whose intersection gives the optimal solution). For 𝑏1 = 100 + Δ, where Δ ∈
[−20,20], character of the optimal solution does not change. The new value of the objective
function is 180 + Δ and the improvement is Δ. Then the shadow price is: limΔ→0 Δ⁄Δ = 1.
Find the shadow price of 𝑏2 .The constraint with the highest magnitude of the shadow price
influences the objective function the most.

Shadow price can be negative. Consider

an additional constraint 𝑥2 ≥ 65 (𝑏4 ) in
the carpenter’s problem. It changes the
feasible region to the triangle AFG, as
shown to the right. Point F (15,65) is
the new optimal solution with objective
function value 175. F is the intersection
of the labor and the new constraint. The
range of 𝑏4 for which F stays optimal
(with revised position) is [60,80]. With
𝑏4 = 65 + Δ, the new position of F is
(15 − Δ, 65 + Δ) with objective value
175 − Δ. Then the shadow price of 𝑏4
is: limΔ→0 −Δ⁄Δ = −1.

Practice problems
Book-3: Operations Research by Winston
Formulation
Book-3, Chapter-3, Review Problem 3, 4, 7, 13, 17
Graphical method
Book-3, Chapter-3, Review Problem 1, 2, 6, 8
Sensitivity analysis
Book-5, Chapter-6, Review Problem 7, 11 (basis refers to the intersecting constraints that
determine the optimal solution), 15 (find the shadow price of the 2nd constraint)

MBA651 Notes prepared by Dr. Avijit Khanra