3.

5 Special Cases in the Simplex Method 117

The initial tableau of Phase II is thus given as

Basic x1 x2 x3 x4 Solution
1 18
z 0 0 5 0 5

1 3
x1 1 0 5 0 5
x2 0 1 - 35 0 6
5
x4 0 0 1 1 1

Because we are minimizing, x3 must enter the solution. Application of the simplex method will
produce the optimum in one iteration (verify with TORA).

remarks. The removal of the artificial variables and their columns at the end of
Phase I can take place only when they are all nonbasic (as Example 3.4-2 illustrates). If
one or more artificial variables are basic (at zero level) at the end of Phase I, then their
removal requires the following additional steps:

Step 1. Select a zero artificial variable to leave the basic solution and designate its row
as the pivot row. The entering variable can be any nonbasic nonartificial vari-
able with a nonzero (positive or negative) coefficient in the pivot row. Perform
the associated simplex iteration.
Step 2. Remove the column of the (just-leaving) artificial variable from the tableau. If
all the zero artificial variables have been removed, go to Phase II. Otherwise,
go back to Step 1.

The logic behind step 1 is that the feasibility of the remaining basic variables will
not be affected when a zero artificial variable is made nonbasic regardless of whether
the pivot element is positive or negative. Problems 3-47 and 3-48 illustrate this situa-
tion. Problem 3-49 provides an additional detail about Phase I calculations.

3.5 sPeciAL cAses in tHe simPLex metHOd

This section considers four special cases that arise in the use of the simplex method.

1. Degeneracy
2. Alternative optima
3. Unbounded solutions
4. Nonexisting (or infeasible) solutions

The remainder of this section presents a theoretical explanation of these situa-

tions. It also provides an interpretation of what these special results mean in a real-life
problem.
118 Chapter 3 The Simplex Method and Sensitivity Analysis

3.5.1 degeneracy
In the application of the feasibility condition of the simplex method, a tie for the mini-
mum ratio may occur and can be broken arbitrarily. When this happens, at least one basic
variable will be zero in the next iteration, and the new solution is said to be degenerate.
Degeneracy can cause the simplex iterations to cycle indefinitely, thus never terminat-
ing the algorithm. The condition also reveals the possibility of at least one redundant
constraint.
The following example explains the practical and theoretical impacts of degeneracy.

example 3.5-1 (degenerate Optimal solution)

Maximize z = 3x1 + 9x2

subject to
x1 + 4x2 … 8
x1 + 2x2 … 4
x1, x2 Ú 0
Using the slack variables x3 and x4, the solution tableaus are

Iteration Basic x1 x2 x3 x4 Solution

0 z -3 -9 0 0 0
x2 enters x3 1 4 1 0 8
x3 leaves x4 1 2 0 1 4

1 z - 34 0 9
4 0 18
1 1
x1 enters x2 4 1 4 0 2
1
x4 leaves x4 2 0 - 12 1 0
3 3
2 z 0 0 2 2 18
1
(optimum) x2 0 1 2 - 12 2
x1 1 0 -1 2 0

In iteration 0, x3 and x4 tie for the leaving variable, leading to degeneracy in iteration 1
because the basic variable x4 assumes a zero value. The optimum is reached in one additional
iteration.

remarks.
1. What is the practical implication of degeneracy? Look at the graphical solution in
Figure 3.5. Three lines pass through the optimum point 1x1 = 0, x2 = 22. Because this
is a two-dimensional problem, the point is overdetermined, and one of the constraints
is redundant. Redundancy means that an associated constraint can be removed
 3.5 Special Cases in the Simplex Method 119

x2 FiGure 3.5
LP degeneracy in Example 3.5-1
z5
3x
1 1 9x
2
x1 1
Optimal 4x #
x1 12 8 (redun
degenerate 2x dant)
solution 2#
4
x1

without changing the solution space. Thus, in Figure 3.5, x1 + 4x2 … 8 is redundant
but x1 + 2x2 … 4 is not. The mere knowledge that some resources are superfluous can
be important during the implementation phase of the solution. The information may
also lead to discovering irregularities in the modeling phase of the solution. Unfor-
tunately, there are no efficient computational techniques for identifying redundant
constraints.
2. From the theoretical standpoint, degeneracy can lead to cycling. In simplex iterations
1 and 2, the objective value does not improve (z = 18), and it is thus possible for the
simplex method to enter a repetitive sequence of iterations, never improving the objective
value and never satisfying the optimality condition (see Problem 3-54). Cycling may not
be a common occurrence, but there have been reports of it being encountered in practice.8
Though algorithms have been developed for eliminating cycling, their use can lead to dras-
tic slowdown in computations and hence they should not be implemented unless there is
evidence that cycling is actually taking place.9
3. Although an LP model may not start with redundant constraints (in the direct sense shown
in Figure 3.5), computer roundoff error may actually create degeneracy-like conditions
during the course of solving a real-life LP. In such cases, the iterations will “stall” at a solu-
tion point, thus mimicking cycling. Commercial codes attempt to alleviate the problem by
periodically perturbing the values of the basic variables (see Section 3.7 for more details
about how commercial codes are developed).

3.5.2 Alternative Optima

An LP problem may have an infinite number of alternative optima when the objective
function is parallel to a nonredundant binding constraint (i.e., a constraint that is satis-
fied as an equation at the optimal solution). The next example demonstrates the practical
significance of such solutions.

8
See T. C. Kotiah and D. I. Steinberg, “Letter to the Editor-On the Possibility of Cycling with the Simplex
Method,” Operations Research, Vol. 26, No. 2, pp. 374–376, 1978.
9
See R. Bland, “New Finite Pivoting Rules for the Simplex Method,” Mathematics of Operations Research,
Vol. 2, No. 2, pp. 103–107, 1977.
120 Chapter 3 The Simplex Method and Sensitivity Analysis

example 3.5-2 (infinite number of solutions)

Maximize z = 2x1 + 4x2

subject to
x1 + 2x2 … 5
x1 + x2 … 4
x1, x2 Ú 0
Figure 3.6 demonstrates how alternative optima can arise in the LP model when the objec-
tive function is parallel to a binding constraint. Any point on the line segment BC represents an
alternative optimum with the same objective value z = 10.
The iterations of the model are given by the following tableaus.

Iteration Basic x1 x2 x3 x4 Solution

0 z -2 -4 0 0 0
x2 enters x3 1 2 1 0 5
x3 leaves x4 1 1 0 1 4
1 (optimum) z 0 0 2 0 10
x1 enters x2 1 1 1 0 5
2 2 2
x4 leaves x4 1 0 - 12 1 3
2 2

2 z 0 0 2 0 10
(alternative optimum) x2 0 1 1 -1 1
x1 1 0 -1 2 3

Iteration 1 gives the optimum solution x1 = 0, x2 = 52, and z = 10 (point B in Figure 3.6).
The existence of alternative can be detected in the optimal tableau by examining the z-equation

x2 FiGure 3.6
LP alternative optima in Example 3.5-2

z5
2x
1 14
x1

x2
1
x2
#
4

x1
B 12
x2 #
5 Optimal basic solutions

A D
x1
 3.5 Special Cases in the Simplex Method 121

coefficients of the nonbasic variables. The zero coefficient of nonbasic x1 indicates that x1 can be
made basic, altering the values of the basic variables without changing the value of z. Iteration 2
does just that, using x1 and x4 as the entering and leaving variables, respectively. The new solution
point occurs at C1x1 = 3, x2 = 1, z = 102. (TORA’s Iterations option allows determining one
alternative optimum.)
The simplex method deals with corner point optima only—namely points B and C in the
present example. Mathematically, we can determine all the points 1x1, x2 2 on the line segment
BC as a nonnegative weighted average of points B 1x1 = 0, x2 = 52 2 and C1x1 = 3, x2 = 12—
that is,

xn 1 = a102 + 11 - a2132 = 3 - 3a
f, 0 … a … 1
xn 2 = a 1 52 2 + 11 - a2112 = 1 + 32 a

remarks. In practice, alternative optima are useful because we can choose from many solutions
without experiencing deterioration in the objective value. For instance, in the present example,
the solution at B shows that activity 2 only is at a positive level. At C, both activities are at a posi-
tive level. If the example represents a product-mix situation, it may be advantageous to market
two products instead of one.

3.5.3 unbounded solution

In some LP models, the solution space is unbounded in at least one variable—meaning
that variables may be increased indefinitely without violating any of the constraints.
The associated objective value may also be unbounded in this case.
An unbounded solution space may signal that the model is poorly constructed.
The most likely irregularity in such models is that some key constraints have not been
accounted for. Another possibility is that estimates of the constraint coefficients may
not be accurate.

example 3.5-3 (unbounded Objective Value)

Maximize z = 2x1 + x2
subject to
x1 - x2 … 10
2x1 … 40
x1, x2 Ú 0
Starting Iteration

Basic x1 x2 x3 x4 Solution

z -2 -1 0 0 0

x3 1 -1 1 0 10
x4 2 0 0 1 40
122 Chapter 3 The Simplex Method and Sensitivity Analysis

x2 FiGure 3.7
LP unbounded solution in Example 3.5-3

Unbounded
objective
Unbounded value
solution

z5
space

2x 1
1
2x1 # 40

x2
10
#
x2
2
x1

x1

In the starting tableau, both x1 and x2 have negative z-equation coefficients—meaning that
an increase in their values will increase the objective value. Although x1 should be the enter-
ing variable (it has the most negative z-coefficient), we note that all the constraint coefficients
under x2 are … 0—meaning that x2 can be increased indefinitely without violating any of the
constraints (compare with the graphical interpretation of the minimum ratio in Figure 3.4). The
result is that z can be increased indefinitely. Figure 3.7 shows the unbounded solution space and
also that x2 and z can be increased indefinitely.

remarks. Had x1 been selected as the entering variable in the starting iteration (per the opti-
mality condition), a later iteration would eventually have produced an entering variable with the
same properties as x2. See Problem 3-58.

3.5.4 infeasible solution

LP models with inconsistent constraints have no feasible solution. This situation does
not occur if all the constraints are of the type … with nonnegative right-hand sides
because the slacks provide an obvious feasible solution. For other types of constraints,
penalized artificial variables are used to start the solution. If at least one artificial vari-
able is positive in the optimum iteration, then the LP has no feasible solution. From the
practical standpoint, an infeasible space points to the possibility that the model is not
formulated correctly.

example 3.5-4 (infeasible solution space)

Consider the following LP:
Maximize z = 3x1 + 2x2
 3.6 Sensitivity Analysis 123

subject to
2x1 + x2 … 2
3x1 + 4x2 Ú 12
x1, x2 Ú 0
Using the penalty M = 100 for the artificial variable R, the following tableau provide the
simplex iterations of the model.

Iteration Basic x1 x2 x4 x3 R Solution

0 z - 303 - 402 100 0 0 - 1200

x2 enters x3 2 1 0 1 0 2
x3 leaves R 3 4 -1 0 1 12
1 z 501 0 100 402 0 - 396
(pseudo-optimum) x2 2 1 0 1 0 2
R -5 0 -1 -4 1 4

Optimum iteration 1 shows that the artificial variable R is positive ( =4)—meaning that the
LP is infeasible. Figure 3.8 depicts the infeasible solution space. By allowing the artificial vari-
able to be positive, the simplex method has in essence reversed the direction of the inequality
from 3x1 + 4x2 Ú 12 to 3x1 + 4x2 … 12 (can you explain how?). The result is what we may call
a pseudo-optimal solution.

3.6 sensitiVity AnALysis

In LP, the parameters (input data) of the model can change within certain limits with-
out causing changes in the optimum. This is referred to as sensitivity analysis and will
be the subject matter of this section. Later, Chapter 4 will study post-optimal analysis,

x2 FiGure 3.8
Infeasible solution of Example 3.5-4

3x
1 1
4x
2 $
12
Psuedo-optimal
z5

solution
3x 1
2x 1

1
1 x2

2x 2
#2

0 x1