CHAPTER 4

SENSITIVITY ANALYSIS AND THE SIMPLEX METHOD

1. Introduction

. Sensitivity analysis in linear programming is concerned with examining how sensitive the optimal solution
and the optimal objective function value are to changes in such parameters as the objective function
coefficients (unit selling price, unit cost, etc.) and the right-hand-side values of the constraints (amount of
resource available, minimum production requirement, etc.) without solving the new LP.

. Some useful concepts in analyzing the output reports from Solver follow:

(1) Constraints -

(a) Binding (or tight): When a structural constraint is evaluated at the optimal solution, it is said to be
binding if its left-hand-side (LHS) value is equal to the right-had-side (RHS) value.

(b) Nonbinding (or loose): When a structural constraint is evaluated at the optimal solution, it is said
to be nonbinding if its left-hand-side (LHS) value is not equal to the right-had-side (RHS) value.

(c) Slack: When a structural constraint is evaluated at the optimal solution, the absolute difference
between its LHS value and RHS value is called the slack.

Remark: The slack for a binding constraint is 0 and that for a nonbinding constraint is positive.

(d) Shadow price: The shadow price of the resource associated with a structural constraint is the effect
on the objective function value as a result of the change (either increase or decrease) in the RHS
value of the constraint by one unit.

Remark: The shadow price of the resource associated with a binding structural constraint is
nonzero, and it is 0 for a nonbinding structural constraint.

(e) Allowable increase: This is the maximum amount by which the RHS value of a structural
constraint may be increased without affecting the current shadow price.

(f) Allowable decrease: This is the maximum amount by which the RHS value of a structural
constraint may be decreased without affecting the current shadow price.

(2) Objective function -

(a) Reduced cost: In the objective function, the absolute value of the reduced cost of a decision
variable is the minimum amount by which the coefficient of the variable must be improved in order
for it to take on a positive value in the optimal solution.

(b) Allowable increase: This is the maximum amount by which the coefficient of a decision variable
may be increased without affecting the current optimal solution.

(c) Allowable decrease: This is the maximum amount by which the coefficient of a decision variable
may be decreased without affecting the current optimal solution.

. The essence of sensitivity analysis in linear programming may be summarized below:

(1) Constraints - The current shadow price of the resource associated with a structural constraint remains
valid as long as the new RHS value falls within the range of [current RHS value - allowable decrease,
current RHS value + allowable increase].
 2

(2) Objective function - The current optimal solution to the LP remains valid as long as the new objective
function coefficient falls within the range of [current objective coefficient - allowable decrease, current
objective coefficient + allowable increase].

(3) If the new RHS value or the new objective coefficient falls outside the corresponding range, the new
LP has to be solved to find the new optimal solution and the new objective function value.

2. Basic Sensitivity Analysis

. Example 4.1: A small furniture company in Richardson, TX, produces wooden chairs and tables for sale.
Each chair requires 1 hour of labor time and 4 pounds of wood while each table requires 3 hours of labor
time and 3 pounds of wood. The unit profit contributions of the chairs and tables are $4 and $5,
respectively. There are 13 hours of labor time and 25 pounds of wood are available in the company.

Let x be the number of chairs to be produced and y be the number of tables to be produced. A linear
program for determining the product mix to maximize the total profit is shown below:

Maximize Z = 4x + 5y
subject to: x + 3y  13
4x + 3y  25
x , y  0

(1) Run Solver to solve the LP and present the Answer Report as well as the Sensitivity Report.

(2) Respond to the following questions based on the Answer Report:

(a) Summarize and interpret the optimal solution and the optimal objective function value.

(b) Is any of the constraints binding at optimality?

(c) How many hours of labor time are used at optimality? What is the slack?

(3) Respond to the following questions based on the Sensitivity Report:

(a) What will the total profit be if the labor time available is reduced by 1 hour due to routine training?

(b) Verify your answer to Part (a) above by reducing the number of hours of labor time available from
13 to 12 hours and running Solver to solve the new LP.

(c) If additional wood is available, what is the maximum amount that you are willing to pay for one
more pound of it? Why?

(d) Verify your answer to Part (c) above by changing the availability of the wood from 25 to 26
pounds and running Solver to solve the new LP.

(e) What the total profit is going to be if the amount of the wood available decreases from 25 to 15
pounds?

(f) Verify your answer to Part (e) above by changing the availability of the wood from 25 to 15
pounds and running Solver to solve the new LP.

[Solution] (1) Both the Answer Report and the Sensitivity Report are shown below:
 3

(2) Based on the Answer Report:

(a) The optimal solution is (x*, y*) = (4, 3) and the optimal objective function value is Z*
= 31. In other words, 4 chairs and 3 tables should be produced to maximize the total
profit at $31.

(b) Both constraints are binding at optimality.

(c) Since the corresponding constraint (i.e., the first constraint) is binding and the slack
is 0 hours, all of the 13 hours of labor time are used up at optimality. To see why this
is the case, we plug x* = 4 and y* = 3 into the left-hand side of the first constraint (x +
3y  13) to find that x* + 3y* = 4 + 3(3) = 13 hours of labor time are used to make 4
chairs along with 3 tables. As a consequence, no machine time is left idle, the
constraint is binding, and hence the slack is 0.

(3) Based on the Sensitivity Report:

 4

(a) Since the allowable decrease for the labor time is 6.75 hours and 1 hour is smaller
than 6.75 hours, the current shadow price of the labor time, which is $0.8889/hour,
will remain unchanged under this new scenario. Consequently, the optimal total profit
will decrease from $31 to 31 - 0.8889 = $30.1111 if the labor time availability is
reduced by 1 hour due to routine training.

(b) Running Solver to solve the new LP with 12 hours (instead of 13 hours) of labor time
being available, we see from the following Answer Report that the new optimal
objective function value is Z* = 30.1111. This is consistent with the prediction given
in Part (a) above.

(c) Since the shadow price of the wood is $0.7778/pound, the allowable increase is 27
pounds, and 1 pound is smaller than 27 pounds, I will be willing to pay up to $0.7778
to buy one more pound of it. This is because the additional wood will lead to an
increase of the total profit by $0.7778.

(d) Running Solver to solve the new LP with 26 pounds (instead of 25 pounds) of wood
being available, we see from the following Answer Report that the new optimal
objective function value is Z* = 31.7778, which represents an increase in the total
profit by 31.7778 - 31 = $0.7778. Thus I will be willing to pay up to $0.7778 to buy
one more pound of the wood, which is consistent with the prediction given in Part (c)
above.
 5

(e) A decrease of 25 - 15 = 10 pounds of the wood is within the allowable decrease of 12

pounds. Therefore, the current shadow price of $0.7778/pound remains valid and the
total profit will become Z* = 31 - 0.7778(10) = $23.222.

(f) Running Solver to solve the new LP with 15 pounds (instead of 25 pounds) of wood
being available, we see from the following Answer Report that the new optimal
objective function value is Z* = 23.222. This is consistent with the prediction given in
Part (e) above.

. Example 4.2: Consider the Sensitivity Report generated by Solver for Example 4.1. What impact, if any,
does each of the following changes have on the optimal solution and the optimal objective function value?

(1) If the unit selling price of the table decreases from $5 to $4, what is the impact, if any, on the optimal
solution and/or the optimal objective function value?
 6

(2) Verify your answer to Part (1) above by changing the unit selling price of the table from $5 to $4 and
running Solver to solve the new LP.

(3) If the unit selling price of the chair increases from $4 to $8, what is the impact, if any, on the optimal
solution and/or the optimal objective function value?

(4) Verify your answer to Part (3) above by changing the unit selling price of the chair from $4 to $8 and
running Solver to solve the new LP.

[Solution] The Sensitivity Report is reproduced below:

(1) Observe that a decrease of $1 from $5 to $4 for the unit selling price of the table
(associated with the decision variable y) is smaller than the allowable decrease of $2. As
such, the current optimal solution of (x*, y*) = (4, 3) remains optimal under the new
scenario. However, the objective function value will become 4x* + 4y* = 4(4) + 4(3) = 28.

(2) As indicated in the following Answer Report, the optimal solution is still (x*, y*) = (4, 3)
but the optimal objective function value becomes 28. Clearly, this is consistent with the
prediction given in Part (1) above.
 7

(3) Note that an increase of $4 from $4 to $8 for the unit selling price of the chair (associated
with the decision variable x) is greater than the allowable increase of $2.6667. As such,
the current optimal solution of (x*, y*) = (4, 3) is no longer optimal under this new
scenario. Consequently, we need to solve the new LP to obtain the new optimal solution
as well as the new optimal objective function value.

(4) As indicated in the following Answer Report, the optimal solution turns out to be (x*, y*)
= (6.25, 0) and the new maximum objective function value is Z* = 50. Obviously, this is
consistent with the prediction given in Part (3) above.
 8

. Example 4.3: Lone Star Company in Little Rock, AR, is developing a new paint that must have a brilliance
rating of at least 15 degrees, a hue rating of at least 32 degrees, and a clarity rating of at least 21 degrees.
The two ingredients to be mixed to produce the new product are Alpha and Beta, which cost $7 per pound
and $9 per pound, respectively. Each pound of Alpha contributes to 3 degrees of brilliance, 8 degrees of
hue, and 7 degrees of clarity whereas each pound of Beta contributes 5 degrees of brilliance, 8 degrees of
hue, and 3 degrees of clarity. Relevant information has been summarized in the following table:

Criterion Alpha Beta Minimum requirement

--------------------------------------------------------------------------------------
Brilliance 3 5 15
Hue 8 8 32
Clarity 7 3 21
--------------------------------------------------------------------------------------
Cost 7 9

Currently, only 1.2 pounds of Beta are available. Let A be the amount (in pounds) of Alpha and B be the
amount (in pounds) of Beta to be mixed to develop the new paint. A linear program for determining the
quantity of each ingredient to be use to minimize the total cost is presented below:

Minimize Z = 7A + 9B
subject to: 3A + 5B  15
8A + 8B  32
7A + 3B  21
B  1.2
A , B  0

(1) Run Solver to solve the LP, summarize the optimal solution as well as the optimal objective function
value, and interpret them. Be sure to show both the Answer Report and the Sensitivity Report.

(2) Based on the Answer Report, how many constraints are nonbinding at optimality? What are the
respective slacks? What do they mean in this problem and where do they come from?

(3) If a local supplier would like to provide additional Beta at a cost of $2 per pound, should the offer be
accepted based on the Sensitivity Report? Why or why not?

{Solution] (1) Both the Answer Report and the Sensitivity Report are displayed below:
 9

It is seen from the Answer Report that the optimal solution is (A*, B*) = (3, 1.2) and the
optimal objective function value is 31.8. In other words, 3 pounds of Alpha and 1.2
pounds of Beta should be mixed to make the new paint so that the total cost is minimized
at $31.8.

(2) According to the Answer Report, the second and the third constraints are nonbinding at
optimality with respective slacks of 1.6 units and 3.6 units. The implication is that the
new paint to be produced contains 1.6 degrees higher than the minimum requirement of
32 degrees of hue and 3.6 degrees higher than the minimum requirement of 21 degrees of
clarity.

Notice that A* = 3 pounds of Alpha and B* = 1.2 pounds of Beta are to be mixed. The
total rating of hue in the new paint is 8A* + 8B* = 8(3) + 8(1.2) = 33.6 degrees and the
minimum requirement is 32 degrees, hence 33.6 - 32 = 1.6 degrees higher than necessary.
Likewise, the total rating of clarity contained in the new paint is 7A* + 3B* = 7(3) +
 10

3(1.2) = 24.6 degrees and the minimum requirement is 21 degrees, hence 24.6 - 21 = 3.6
degrees higher than necessary.

(3) We see from the bottom row of the Sensitivity Report that the shadow price of the Beta is
$-2.6667/pound, which means that the total cost will go down by $2.6667 for each
additional pound of the Beta used so long as the change is within the allowable limit. As
$2.6667 > $2, the offer should be accepted but only up to 0.3 pounds of additional Beta
should be purchased. This is because the allowable increase for the shadow price to
remain $-2.6667 is 0.3.

3. Duality Theory

. Every linear programming model has two forms. The first or original form is called the primal, and the
second form is derived from the primal and is called the dual.

. Since both the primal and the dual are associated with the same problem, their respective solutions are
closely related to each other. Specifically, the optimal solution values in the primal are equal to the shadow
prices in the dual, and the shadow prices in the primal are equal to the optimal solution values in the dual. In
addition, the optimal objective function value for the primal is the same as that for the dual.

. The importance of duality in linear programming lies in the following facts:

(1) The solution to the dual provides a basis for understanding some of the fundamental economic
relationships that exist in the primal. This is particularly useful for managers to make sound business
decisions.

(2) It allows some primal LP problems with more constraints than decision variables to be transformed into
equivalent dual LP problems with less constraints than decision variables, which are much easier to
solve. (In general, the amount of time required to solve an LP is in proportion to the number of
constraints but depends little on the number of decision variables.)

(3) LPs involving many decision variables and only two constraints may be transformed into an equivalent
dual LP involving only two decision variables (and many constraints) that may be solved graphically
when no computer is available.

. Primal/dual transformation - The rules described in the exhibit on the last page of this chapter may be used
to help you with the transformation between the primal and the dual of an LP.

Remark: A decision variable in an LP is said to be unrestricted in sign (UIS) if it is allowed to take on a

positive value, 0, or a negative value.

. Example 4.4: Consider the following linear program, which was formulated for the product mix problem in
Example 2.15. Derive the dual linear program.

Maximize Z = 4x + 5y
subject to: x + 3y  13
4x + 3y  25
x , y  0

[Solution] Minimize Z = 13u + 25v

subject to: u+ 4v  4
3u + 3v  5
u , v  0

. Example 4.5: Derive the dual of the following LP:

 11

Minimize Z = 3x1 - 4x2 + 5x3

subject to: 2x1 + 3x2 + 3x3  27
x1 - x3  0
4x1 + 2x2 + 4x3  55
x2  4
x1 , x2 , x3  0

[Solution] Maximize Z = 27y1 + 55y3 + 4y4

subject to: 2y1 + y2 + 4y3  3
3y1 + 2y3 + y4  -4
3y1 - y2 + 4y3  5
y1 , y2  0
y3 , y4  0

. Example 4.6: Formulate the dual of the linear program presented below:

Maximize Z = 3x1 + 5x2

subject to: 2x1 + 3x2  8
5x1 + 4x2 = 9
x1 + x2  3
x1 , x2  0

[Solution] Minimize Z = 8y1 + 9y2 + 3y3

subject to: 2y1 + 5y2 + y3  3
3y1 + 4y2 + y3  5
y1 , y3  0
y2 UIS

. Example 4.7: Derive the dual of the following LP:

Minimize Z = 24x1 + 28x2 + 18x3

subject to: 4x1 + x2 + 2x3  3
2x1 + 7x2 + 3x3  7
x1 , x3  0
x2 UIS

[Solution] Maximize Z = 3u + 7v
subject to: 4u + 2v  24
u+ 7v = 28
2u + 3v  18
u , v  0

. Example 4.8: Formulate the dual of the LP model presented below:

Maximize 2x1 + x2
subject to: x1 + x2 = 2
2x1 - x2  3
x1 - x2  1
x1  0
x2 UIS
 12

[Solution] Minimize Z = 2y1 + 3y2 + y3

subject to: y1 + 2y2 + y3  2
y1 - y2 - y3 = 1
y1 UIS
y2  0
y3  0

. Example 4.9: Derive the dual of the following linear program:

Maximize Z = 3x1 + 4x2 - 2x3

subject to: 4x1 - 12x2 + 3x3  12
-2x1 + 3x2 + x3  6
-5x1 + x2 - 6x3  -40
3x1 + 4x2 - 2x3 = 10
x1  0
x2  0
x3 UIS

[Solution] Minimize 12y1 + 6y2 - 40y3 + 10y4

subject to: 4y1 - 2y2 - 5y3 + 3y4  3
-12y1 + 3y2 + y3 + 4y4  4
3y1 + y2 - 6y3 - 2y4 = -2
y1 , y2  0
y3  0
y4 UIS

. Solving special linear programs - A linear programming model involving decision variables that either must
take on negative values or are unrestricted in sign may be reformulated as one where the decision variables
are all nonnegative according to the following rules:

(1) A variable (x) that must take on a negative value - Replace each x with -xp in the LP, where xp is
nonnegative.

(2) A variable (y) that is unrestricted in sign - Replace each y with yp - yn in the LP, where yp and yn are
both nonnegative.

The modified LP is then solved by running Solver in the normal fashion. However, the optimal solution
obtained needs to be converted to find the optimal solution to the original problem.

. Example 4.10: Solve the following linear program by running Solver:

Maximize Z = 12x1 - x2 - 5x3 - 20x4

subject to: -2x1 + x2 + x3 + x4  25
-3x1 + x3 + 4x4  30
x1  0
x2 UIS
x3 , x4  0

[Solution] Replacing x1 with -x1n and replace x2 with x2p - x2n, where x1n, x2p, and x2n are all
nonnegative, the original LP becomes:

Maximize Z = -12x1p - x2p + x2n - 5x3 - 20x4

subject to: 2x1p + x2p - x2n + x3 + x4  25
3x1p + x3 + 4x4  30
x1p , x2p , x2n , x3 , x4  0
 13

Running Solver to solve the new LP, we find that the optimal solution is (x1p *, x2p*, x2n*,
x3*, x4*) = (10, 5, 0, 0, 0) and the optimal objective function value is Z* = -125. It follows that
the optimal solution to the original LP is (x1*, x2*, x3*, x4*) = (-x1p*, x2p* - x2n*, x3*, x4*) =
(-10, 5 - 0, 0, 0) = (-10, 5, 0, 0) with an optimal objective function value of Z* = -125.

. Relationships between primal and dual - As pointed out earlier, the key relationships between the primal
and the dual of a linear program are:

(1) The optimal solution values in the primal are equal to the shadow prices in the dual,

(2) The shadow prices in the primal are equal to the optimal solution values in the dual, and

(3) The optimal objective function value for the primal is the same as that for the dual.

(4) The dual of the dual is the primal.

14