Int. J.

Production Economics 65 (2000) 257}274

The di!erence between the managerial and mathematical

interpretation of sensitivity analysis results in
linear programming
TamaH s Koltai *, TamaH s Terlaky

Technical University of Budapest, Department of Industrial Management and Economics, H-1521 Budapest, Hungary

Delft University of Technology, Faculty of Information Technology and Systems, Department of Technical Mathematics and Informatics,
P.O. Box 5031, 2600 GA Delft, The Netherlands
Received 7 December 1998; accepted 28 March 1999

Abstract
This paper shows that managerial questions are not answered satisfactorily with the mathematical interpretation of
sensitivity analysis when the solution of a linear programing model is degenerate. Most of the commercially available
software packages provide sensitivity results about the optimality of a basis and not about the optimality of the values of
the decision variables. The misunderstanding of the shadow price and the validity range information provided by
a simplex based computer program may lead to wrong decision with considerable "nancial losses and strategic
consequences. The paper classi"es the most important types of sensitivity information, graphically illustrates degeneracy,
and demonstrates its e!ect on sensitivity analysis. A production planning example is provided to show the possibility of
faulty production management decisions when sensitivity results are not understood correctly. Finally the recommendations for the users of linear programing models and for software developers are provided.  2000 Elsevier Science B.V.
All rights reserved.
Keywords: Linear programming; Sensitivity analysis; Production management

1. Introduction
Linear programing (LP) is one of the most extensively used operations research technique in production and operations management [1]. As a
result of the development of computer technology
and the rapid evolution of user friendly LP soft-

* Corresponding author. Tel.: #36-1-463-2456; fax: #36-1463-1606.

E-mail address: koltai@imvt.bme.hu (T. Koltai)

ware, every operation manager can run an LP

software easily and quickly on a laptop computer.
Although solving LP models is now accessible for
everybody, the interpretation of the results requires
a lot of skill. Most of the management science and
OR textbooks pay special attention to sensitivity
analysis, and the problems of degeneracy, but sensitivity analysis under degeneracy is rarely discussed.
Commercially available software do not give
enough information to the user about the existence
and the consequences of these, very common,
`special casesa. In practice, managers very frequently misinterpret the LP results which may lead

0925-5273/00/$ - see front matter  2000 Elsevier Science B.V. All rights reserved.
PII: S 0 9 2 5 - 5 2 7 3 ( 9 9 ) 0 0 0 3 6 - 5

258

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

to erroneous decisions and to important "nancial

and/or strategic disadvantages.
Several papers have addressed this issue. Evans
and Baker draw the attention to the consequences
of the misinterpretation of sensitivity analysis
results in management. They illustrate their point
with a simple example and list some published
cases in which the erroneous interpretation of
sensitivity analysis results is obvious [2]. Aucamp
and Steinberg [3] also warn that shadow price
analysis is incorrect in many textbooks, and that
the shadow price is not equal to `the optimal solution of the dual problem'' when the obtained optimal solution is degenerate. They present some
examples of shadow price calculations by commercial packages. AkguK l [4] re"nes the shadow
price de"nition of Aucamp and Steinberg, and
introduces the negative and positive shadow
prices for the increase and the decrease of the
RHS elements. Greenberg [5] shows that very
frequently practical LP models have a netform
structure, and netform structures are always degenerate. He illustrates sensitivity analysis of
netform type models by one of the Midterm
Energy Market Model of the U.S. Department of
Energy. Gal [6] summarizes most of the critics
concerning sensitivity analysis of LP models and
highlights some important research directions.
Rubin and Wagner [7] illustrates the traps of
the interpretation of LP results by using the industry cost curve model in a tutorial type paper
written for managers and instructors. Jansen et al.
[8] explain the e!ect of degeneracy on sensitivity
analysis by using a transportation model, and
present the shortcomings of the most frequently
used LP packages. They also show how complete,
correct sensitivity analysis can be done. Wendell
[9,10] also pays special attention to the correct
and practically useful calculation of sensitivity information. The biggest problem is not that operations researchers are unaware of the di$culties of
sensitivity analysis. This issue is discussed thoroughly in the scienti"c literature (see for example
[6,10,11]) and a complete, mathematically correct
treatment of sensitivity analysis is presented by
Jansen et al. [8], and by Roos et al. [12]. Practice,
however, shows that the problem is not widely
known among the LP users, and the available

commercial software packages are not helpful in

recognizing the di$culties.
The main objective of this paper is to explain the
di!erence between the managerial questions and
the traditional mathematical interpretation of sensitivity analysis. In the "rst part of the paper basic
de"nitions are introduced, the most important
types of sensitivity information are classi"ed, and
degenerate LP solutions are illustrated graphically.
In the second part a production planning problem is used to demonstrate the consequences of
incorrect interpretations of the provided sensitivity information. Finally, some recommendations are made for both practitioners and software
developers.

2. Basic de5nitions and concepts

Every LP problem can be written in the following standard form:
min +c2x"Ax"b, x*0,,
V

(1)

where A is a given m;n matrix with full row rank

and the column vector b represents the right-hand
side (RHS) terms and the row vector c2 represents
the objective function coe$cients (OFC). Problem
(1) is called the primal problem and a vector x*0
satisfying Ax"b is called a primal feasible solution.
The objective is to determine those values of the
vector x which minimize the objective function. To
every primal problem (1) the following problem is
associated:
max +b2y"A2y)c,.
W

(2)

Problem (2) is called the dual problem and a vector

y satisfying A2y)c is called a dual feasible solution.
For every primal feasible x and dual feasible y it
holds that c2x*b2y and the two respective objective function values are equal if and only if both the
solutions are optimal.
Most computer programs to solve LP's are
based on a version of the simplex method. Modern,
hi-performance packages are furnished with
interior point solvers as well, however, the

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

implemented sensitivity analysis is based always on

the simplex method. The simplex procedure selects
a basis of the matrix A in every step, the selected
basis solution is calculated and the optimality criteria is checked. To de"ne the optimal basis
solution some preparation is needed. Let B be a
set of m indices, and A be the matrix obtained
by taking only those columns of A whose indices
are in B. If A is a nonsingular matrix then by using
the vector x "A\b a basis solution can be de"ned as

(x )
H
x"
H
0

if j3B,
otherwise.

(3)

If in addition x *0 holds then x is called a primal

feasible basic solution. The variables with their
index in B are the basic variables, the others are the
nonbasic variables. Dual variables can be associated
to any basis A as follows:
y"(A\)2c .

(4)

If c!A2y*0 then y is a feasible solution for the

dual problem and is called the dual feasible basic
solution. If the basis A is both primal and dual
feasible, then A is an optimal basis, and the corresponding basic solutions x and y are optimal basis
solutions for the primal and the dual problems,
respectively. It might happen, that a basis gives an
optimal primal solution, but the related dual basis
solution is dual infeasible. Such a basis is called the
primal optimal. Analogously, when a basis gives
a dual optimal solution, but the related primal
solution is infeasible, then the basis is called the
dual optimal.
Sometimes the optimal basis is not unique,
more than one basis may yield an optimal solution either for the primal or for the dual problem
or for both. This is called degeneracy and occurs
in almost all practical problems. Formally, a
basis is called primal degenerate when there are
variables with zero value among the basis variables
and it is called dual degenerate when some dual
slack variables s "c !(A2y) , not belonging to
H
H
H
the basis indices B, are zero. In general, if a basis
is either primal, or dual, or from both sides degenerate then we simply say that it is degenerate. In
case of degeneracy many optimal solutions exist

259

that are not basis solutions, e.g. interior point

methods provide the so-called strictly complementary optimal solutions. Strictly complementary optimal solutions are the optimal solutions x, y such
that x#(c!A2y)'0.
Very frequently the basic parameters of an LP
model changes (e.g. cost coe$cients, resource capacities, etc.) and it would be important to know
if any action on behalf of the decision maker is
required as a consequence of these changes.
Sensitivity analysis can help to answer this question if it is applied correctly. The objective of
sensitivity analysis is to analyze the e!ect of
the change of the objective function coe$cients
(OFC) and the e!ect of changes of right-hand
side (RHS) elements on the optimal value of the
objective function, and the validity ranges of
these e!ects. Depending on how this analysis
is performed, three types of sensitivities are calculated:
E Type I sensitivity: Type I sensitivity determines
those values of some model parameters for which
a given optimal basis remains optimal. Sensitivity
analysis of the optimal basis for the OFC elements determines within which range of an OFC
the current optimal basis remains optimal and
what is the rate of change (directional derivative)
of the optimal objective function value when the
OFC changes within this range. In case of the
RHS elements the question is, within which
range an RHS element can change so that the
current optimal basis stays optimal, and what is
the rate of change (shadow price) of the optimal
objective function value within the determined
interval.
Type I sensitivity analysis is implemented in all the
commercial software packages. In case of primal
degeneracy, however, several basis may belong to
the same optimal solution yielding di!erent ranges
and rate of changes for the same parameter to
di!erent optimal basis. In the case of dual degeneracy many primal optimal solutions, and therefore
many di!erent optimal basis may exist resulting in
di!erent intervals and rates of changes. From
a mathematical point of view the provided information is correct, because the question is the
sensitivity of the given optimal basis, but, can be

260

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

misleading for decision makers, if the given information is not interpreted correctly.
E Type II sensitivity: Type II sensitivity determines
those values of some model parameters for which
the positive variables in a given primal and dual
optimal solution can remain positive, and the zero
variables remain zero, i.e. the same activities remain active. More accurately, we have an optimal solution (not necessarily basis solution)
x with its support set supp(x)"+i " x '0,. We
G
are looking for those model parameters, for
which an optimal solution (basis or not basis)
exists with precisely the same support set. Sensitivity analysis of a given optimal solution
for an OFC determines within which range
of the OFC an optimal solution with the
same support set exists and what is the rate of
change (directional derivative) of the optimal
objective function value when the OFC changes
within this range. In case of the RHS elements
the question is, within which range a RHS element can change without the change of the support set of the optimal solution, and what is the
rate of change (shadow price) of the optimal
objective function value within the determined
interval.
Contrary to Type I sensitivity, Type II sensitivity
depends on the produced optimal solution, but not
on which basis } if any } represents the given
optimal solution.
E Type III sensitivity: Type III sensitivity determines those values of some model parameters for
which the rate of change of the optimal objective
function value is the same. Roughly speaking
sensitivity (and range) analysis means the analysis of the e!ects of the change of some problem
data, in particular an objective coe$cient c or
H
right-hand side coe$cient b . Let us assume that
H
either c #c or b #b is the perturbation. It is
H
G
known that the optimal value function is a piecewise linear function of the parameter change (see
for example [9,11]). In performing Type III sensitivity analysis one wants to determine the rate
of the change of the optimal value function and
the intervals within which the optimal value
function changes linearly.

Type III sensitivity information is independent of

the solution obtained, it depends only on the problem data and on which OFC or RHS element is
changing.
As a short recapitulation observe that:
E Type I sensitivity depends on the type of optimal
basis solution produced and on the type of optimal basis found to represent the given optimal
basis solution.
E Type II sensitivity depends on the type of optimal solution (not necessarily a basis solution)
produced. If eventually it is an optimal basis
solution, then Type II sensitivity information is
independent of the basis representation of the
found optimal basis solution.
E Type III sensitivity is independent of the type of
optimal solution produced, and it is also independent of how the optimum is represented. It
depends only on the problem data.
The calculation and importance of the three di!erent sensitivity information depends on the type of
optimum solution produced by the LP solver. Most
of the LP solvers used for small and medium size
problems are based on some versions of the simplex
method and they provide an optimal basis solution.
Other solvers, typically used for (very) large scale
problems are based on interior point methods and
they provide an interior (i.e. strictly complementary) optimal solution. To distinguish among the
three type of sensitivities is necessary because of the
existence of degeneracy. The following cases can be
observed:
E When the optimal solution is neither primal
nor dual degenerate, all the three types of sensitivities are the same, since there is a unique
optimal solution with a unique optimal basis. In
this case the sensitivity analysis output of the
commercially available software packages provide reliable, useful information for decision
making.
E When the optimal solution is only primal degenerate, a unique primal optimal solution exists.
Moreover, several optimal basis belong to the
same, unique primal optimal solution. In this
case Type I and Type II sensitivities may be

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

di!erent since there are di!erent Type I sensitivity information for all the optimal basis. One
important case is when the increase, and the
decrease of an RHS parameter results in di!erent
rate of changes, i.e. the optimal value function at
the current point is not di!erentiable. Due to this
fact the introduction of the right and left side
shadow prices and their respective sensitivities
[3] was needed. Type II and Type III sensitivity
information for the RHS elements are split into
two parts: the left and right side sensitivities. The
left and right linearity intervals of the optimal
value function provide the Type III information.
When the left and right side shadow prices are
identical, then only one interval is given. Type II
and Type III sensitivity information for an OFC
are identical in the case when the solution is only
primal degenerate.
E When the optimal solution is only dual degenerate, several di!erent primal optimal basis and
nonbasis solutions may exist with di!erent support sets, while the dual optimal solution is
unique. In this case Type I and Type II sensitivities at each alternative primal optimal basic solutions are identical, but Type II sensitivities can
be calculated from the nonbasic solutions as
well. Type II sensitivity is interested only in the
optimal solutions belonging to the same support
sets, therefore, Type III sensitivity may be di!erent from the Type II sensitivities of each optimal
solutions, except at the strictly complementary
ones.
E When the optimum is both primal and dual degenerate, then all the three types of sensitivities
may be di!erent. In this case each optimal basis
of each optimal basis solution may have a different Type I sensitivity information. Optimal
solutions with di!erent support sets may have
di!erent Type II sensitivities and can be examined at nonbasis solutions too. As it is known,
Type III sensitivity information is uniquely
determined, it is independent of the optimal
solution obtained. Typically, the intervals provided by Type I and Type II sensitivities are
subintervals of the Type III sensitivity intervals. The rates of changes produced by Type I
and Type II either coincide with Type III information or are useless, their validity (as a sub-

261

di!erential) is restricted to the current point

only.
E When the current optimum is a strictly complementary optimal solution, e.g. when the problem was solved by interior point methods, then
Type II and Type III sensitivities are the same.
When the optimal solution at hand is not
a strictly complementary one, Type II and Type
III sensitivities might be di!erent. In this case the
left and right intervals of Type II sensitivities are
possibly subintervals on the Type III sensitivity
intervals.
An important question is, when the di!erence
between Type II and Type III sensitivities is important for the decision maker. If the decision maker
implements an optimal solution, then in many situations, the important information is the sensitivity
of the implemented optimal solution (Type II sensitivity). For example if an optimal production plan,
determined by LP, is already running, then the
important question is how the change of certain
costs, or the change of a machine capacity in#uences the implemented plan. If the question is, how
much an RHS element can be increased with the
same consequences, and independently of the possible change of an optimal solution, then it is
a Type III sensitivity question. For example if
a machine capacity can be increased economically
at the calculated shadow price, important information can be how much the capacity can be increased
economically in total. It is possible that di!erent
production plans (di!erent optimal solutions, especially when optimal basis solutions are implemented) belonging to di!erent capacities increase,
but all capacity extensions are made with the same
marginal bene"ts.
In case of large models, solvers based on interior
point methods (IPMs) are frequently used. IPM
solvers generally provide strictly complementary
optimal solutions. In this case Type I sensitivity
cannot be asked because, in case of degeneracy, the
produced optimal solution is not a basis solution.
When one is interested in obtaining an optimal
basis solution, a basis identixcation procedure
might be applied to produce an optimal basis.
Such procedures are implemented in many commercial software packages. Type II and Type III

262

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

sensitivity information are identical in this case,

because the change of the support set of a strictly
complementary optimal solution is in one to one
correspondence with the linearity intervals of the
optimal value function [12].
The di!erent questions of Type I, Type II and
Type III sensitivity can be made transparent when
one observes that in the case of degeneracy optimal
solutions might exist that are neither the basis
solutions nor strictly the complementary solutions.
In such a case all the three sensitivities might be
di!erent. Type I has no sense when no optimal
basis is at hand. Type II depends on the current
solution while Type III does not. The di!erent type
of solutions of an LP problem are illustrated in
Fig. 1. Let us assume that the set of the feasible
solutions is the unit cube and the objective is to
maximize x , that is,

max (x )

subject to
x


#s
x


"1,


#s

"1,

(5)

x
#s "1,


x , x , x , s , s , s *0.
     

In this case the solution P "(0, 0, 1; 1, 1, 0) is


an optimal basis solution with the basis given
by the variables (s , s , x ). The solution P "
  

(1/2, 1/2, 1; 1/2, 1/2, 0) is a strictly complementary
optimal solution, while the solution P "

(0, 1/2, 1; 1, 1/2, 0) is an optimal solution which
is neither a basis nor a strictly complementary
solution.
As a summary it can be stated that in case of
degeneracy, commercial packages do not provide the sensitivity information useful for the
decision maker (note that in practice almost
all problems are degenerate). They give answer
to a less ambitious question. They provide information about the interval of a parameter
value within which the current optimal basis remains optimal, and at what rate the change of
the parameter varies the optimal function value
in that interval (Type I sensitivity). This answer
is intimately related to the optimal basis obtained by the simplex solver. In case of degeneracy
many di!erent optimal basis exist, thus many
di!erent ranges and rates of changes might be
obtained. To obtain the true Type III sensitivity
information about the change of the value of the
OFC and RHS elements one needs extra e!ort. In
fact one has to solve some subsidiary LPs for determining linearity intervals, and left and right derivatives of the optimal value function. In case of
strictly complementary solutions provided by
IPM solvers Type II and Type III sensitivities coincide for some speci"c solutions. In all other cases
they might be di!erent. The decision maker has to
determine if Type II or Type III sensitivity is important for a certain problem, and LP software
packages should assist to obtain those information
easily.

3. Graphical illustration of the problem of

sensitivity analysis

Fig. 1. Graphical illustration of the di!erent type of LP solutions.

When the LP problem has no more than two

variables, the solution space and all the information concerning the optimum and its sensitivity can be represented in a two-dimensional
space. The following problem will be our prototype

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

problem:
max (12x #10x )


subject to
(C1) x #x )600,


(C2) 2x #x )1000,


(6)
(L1) x
)400,

(L2)
x )500,

x *0, x *0.


The feasible set and the solution of problem (6) can
be seen in Fig. 2. The two constraints (C1) and (C2)
and the upper bounds on x and x (L1) and (L2)


are represented as half spaces. The boundary of
these spaces with the corresponding labels are depicted in the "gure. The intersection of these half
spaces is represented as a shaded area, which contains all the primal feasible solutions. The objective
function (iso-pro"t line) is drawn as a straight
dashed line. The objective function touches the
shaded area at point P , therefore the unique opti
mal solution is at x "400 and x "200.


In order to transform problem (6) into the standard form, indicated by problem (1), slack variables
(denoted by s , i"1,2, 4) are introduced for all
G

Fig. 2. Graphical illustration of the prototype problem.

263

the constraints, and the objective function is

changed to have a minimization problem. The
problem in the standard form is as follows:
min (!12x !10x )


subject to
(C1) x #x #s
"600,



(C2) 2x #x
#s
"1000,



(7)
(L1) x
s
"400,


(L2)
x
#s "500,


x , x , s , s , s , s *0.
     
Problem (7) shows that A is a 4;6 matrix with
rank equal to 4. The values of the slack variables at
P are the following:

s "0; s "0; s "0; s "300.




Since at P there are three nonzero variables

(x , x and s ) and the rank of the matrix A is 4, the
 

optimal solution is degenerate. This can be seen in
Fig. 2. The point P is the intersection of three lines

(the boundary of (C1), (C2) and (L1)). Two lines
would be enough to determine the location of a
point in a two-dimensional space, therefore P is

over determined. Even if we remove any one of
(C2), or (L1), the point P remains the only opti
mal solution. This over determination of the
optimal point is a graphical illustration of primal
degeneracy.
Let us see the consequences of degeneracy on
sensitivity analysis. The shadow prices and the corresponding validity ranges for the optimal solution,
calculated with the help of Fig. 2, are given in
Table 1. The change of an RHS element is represented by a parallel shift of the corresponding line
in Fig. 2.
If the RHS of any of these constraints are decreased, then the left side shadow prices are obtained for each of the constraints respectively
(column three of Table 1). The optimal point P is

at the intersection of constraints (C1), (C2) and (L1).
The decrease of any of the RHS of these constraints
results in the movement of the optimum point, P ,

which consequently changes the objective function
value as well. Since the change of the RHS of (L2)

264

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

Table 1
Shadow prices and validity ranges of the optimal values
Dual
variable

y
!
y
!
y
*
y
*

Current
RHS value

600
1000
400
500

Left side
shadow price

10
2
2
0

Validity range
LL

UL

400
700
100
200

600
1000
400
500

does not a!ect the location of P its shadow price is


zero. (C1) can be moved to P , (C2) and (L1) can be

moved to P with the same shadow price value.

(L2) can be moved to P without a!ecting the

objective function value. The corresponding lower
limits (LL) are given in the fourth column of
Table 1. In case of left side shadow prices the upper
limits (UL) are equal to the current values of the
RHS elements ("fth column of Table 1). If the RHS
of any of these constraints are increased, then the
right side shadow prices are obtained for each of
the constraints, respectively (column six of Table 1).
In case of constraints (C2) and (L2) the increase of
the right-hand side values do not a!ect the location
of the optimum point, because (C1) and either (C2)
or (L1) "xes its place. Therefore the corresponding
right side shadow prices are equal to zero. When
the right hand side of (C1) is increased, then the
optimum point will stay at the intersection of (C1)
and (C2) and the shadow price will be equal to 8.
Since (L2) does not a!ect the location of P its

shadow price is also zero. In case of right side
shadow prices the lower limits (LL) are equal to the
current values of the RHS elements, while the upper
limits (UL) are determined by the geometrical
properties of the solution space. When (C1) is
moved upward the intersection of (C1) and (C2)
P moves upward as well. When P reaches (L2),


then the move of (C1) does not a!ect the location of
P any more, and the shadow price turns into zero.

The RHS value at this point is the UL of the
sensitivity range, and it is equal to 750. The UL of
all the other constrains are equal to in"nity.
Table 2 shows the shadow prices and their validity ranges found by the STORM computer package

Right side
shadow price

8
0
0
0

Validity range
LL

UL

600
1000
400
500

750
R
R
R

Table 2
Shadow prices and validity ranges at the optimal basis B

Dual
variable

y
!
y
!
y
*
y
*

Current
RHS value

600
1000
400
500

Shadow
price

10
0
2
0

Validity range
LL

UL

400
1000
100
200

600
R
400
R

at the optimal basis B "+1, 2, 3, 6,. It can be seen


that at this basis the left side shadow prices and
validity ranges were provided for constraints (C1)
and (L1), and the right side shadow price and validity range was found for the constraint (C2). Table 3
contains the shadow prices and their validity
ranges found at the optimal basis B "+1, 2, 5, 6,.

At this basis the right side shadow prices and validity ranges were provided for constraints (C1) and
(L1), and the left side shadow price and validity
range was found for the constraint (C2). The left
and right side shadow prices for constraint (L1) are
identical, and its correct value and validity range
was found in both optimal basis as seen in the last
rows of Tables 2 and 3.
The reason for the di!erences in Tables 1}3 can
be explained if we look at the mathematical interpretation of degeneracy. Every corner point of the
shaded area of Fig. 2 can be represented by one or
more basis. The corner point which is over determined, i.e. de"ned by the intersection of more than
two lines, represents more than one basis. Depending on which two lines are taken to de"ne this

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

Table 3
Shadow prices and validity ranges at the optimal basis B

Dual
variable

y
!
y
!
y
*
y
*

Current
RHS value

600
1000
400
500

Shadow
price

8
2
0
0

265

Table 4
The increase of the objective function by a unit increment of the
RHS elements

Validity range
LL

UL

600
700
400
200

750
1000
R
R

point, di!erent basis is considered, that is, di!erent

sets of B in (3) may lead to the same basis solution.
This is the case at P , where Table 2 was calculated

with the help of a basis containing columns 1, 2,
4 and 6, and Table 3 was calculated with the help
of a basis containing columns 1, 2, 5 and 6 of
problem (7).
The main problem of RHS sensitivities in the
prototype problem is that in case of a degenerate
primal optimal solution the dual problem has no
unique solution. Di!erent bases belonging to the
same optimal solution provide di!erent shadow
prices and validity ranges. Tables 2 and 3 show that
the results provided by the two optimal bases are
mixtures of the left side, right side and full shadow
prices and validity ranges. The complete Type III
information, similar to Table 1, is not given at any
of the basis. It depends on the computer code at
which basis, among the many optimum ones, the
program stops. Di!erent commercially available
software may report di!erent RHS sensitivities for
the same problem [8]. All these results are correct
mathematically, because they describe the validity
of an optimal basis (Type I sensitivity), but are not
useful for managerial decisions, because they do
not re#ect the validity of the positivity status of the
decision variables at optimality (Type II sensitivity),
and do not characterize the validity range of the
left/right marginal values (Type III sensitivity). The
correct RHS information, which refers to the rate of
change of the optimal objective value, and the
range where these rates are valid are given in
Table 1. The Type II sensitivity of the RHS elements are given in Table 1, in which for
y , y , y , the left and right side sensitivities are
! ! *
Type III information as well for two di!erent

RHS elements

Rate of change of
the objective function

Validity range

C1
C2, L1

10
2

400)*b )600
!
400)*b )600
*
*b "*b
!
*

linearity intervals. For y , the Type II and Type

*
III sensitivity informations are identical.
It can be seen in Table 1 that most of the rightside shadow prices are zero. An interesting question
is, how the optimal objective function value can be
increased by the simultaneous increase of those
RHS elements which have a zero shadow price.
This question is equivalent to the problem of increasing the capacity of bottleneck resources of
production systems. Fig. 2 shows that the RHS of
(C1) can be increased alone, but the RHS of (C2)
and (L1) need to be increased simultaneously. This
information is summarized in Table 4. The optimum value of the objective function increases by 10
if the RHS of (C1) is increased by one unit. This is
true within the interval [400, 600]. When the RHS
of (C2) and (L1) are simultaneously increased by
one unit, the change of the objective function value
is 2 and the validity range is a line segment in
a two-dimensional space, given in the last window
of Table 4.
Since the objective function coe$cient sensitivity
of the primal problem is the same as the RHS
sensitivity of the dual problem, all what was said of
the RHS is valid for the objective function coe$cients as well. Graphically the change of an OFC
can be represented by the change of the slope of the
line of the objective function. In Fig. 2 the optimal
solution of problem (7) is P as long as the objective

function line stays between (L1) and (C1). The corresponding sensitivities are given in Table 5. These
data coincide with the sensitivities provided by the
STORM computer package when the optimum
was calculated at the basis B . The results provided

at the basis B are given in Table 6. The intervals

obtained in this case are the subsets of the correct
sensitivity ranges. The last columns of Tables 5

266

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

Table 5
Objective function coe$cient sensitivities and rate of change at
the optimal basis B

Objective
function
coe$cient

Current
value

c

c


12
10

Validity range
LL

UL

10
0

R
12

Rates of
changes

400
200

Table 6
Objective function coe$cient sensitivities and rate of change at
the optimal basis B

Objective
function
coe$cient

Current
value

c

c


12
10

Validity range
LL

UL

10
6

20
12

Rates of
changes

400
200

and 6 show the rate of changes of the optimum

value function. The identical rate of changes of the
respective coe$cients in both optimal basis B and

B indicate that the optimal solution is not a dual

degenerate. This is also clear from Fig. 2 since the
optimal solution is unique.
Fig. 3 illustrates a slight modi"cation of the
sample problem. A new constraint ((C3): x !2x


)200) is added to the problem and the objective
function is also modi"ed (min[!12x !0x ]). In


this case the optimal objective function coincides
with constraint (L1), and all the points in the interval [P , P ] are optimal. Consequently all the basis
 
at P and the basis at P are optimal and the


optimal solution is both primal and dual degenerate, and we expect di!erent Type I, Type II and
Type III sensitivities. Let us consider now the
shadow price and sensitivity range of the RHS of
constraint (L1). It can be seen that as long as (L1)
increases or decreases the shaded area the shadow
price is equal to 12. This is true between points
P and P and corresponds to the RHS values of

(L1) in the interval [0, 440], which is the Type III
sensitivity information for the RHS of (L1). If,

Fig. 3. Graphical illustration of the modi"ed prototype

problem.

however, the problem is solved by a computer code

of the simplex method, then depending on the basis
found by the program, the following intervals can
be obtained: [100, 400], [200, 440], [400, 440].
There are four di!erent Type I sensitivities. In this
modi"ed example the left and right-hand side
shadow prices are equal, and there are two di!erent
Type II sensitivities for the two optimal basis solutions at P and P , which are [100, 440] and


[200, 440] respectively.
Finally let us take a strictly complementary
solution of the modi"ed problem found by an
IPM solver. On the P }P line segment the
 
point x "400, and x "150 can be such an


optimum. For this solution Type I sensitivity has
no sense because this is not a basis solution,
and Type II and Type III sensitivities coincide,
because this is a strictly complementary optimal
solution. On the interval [0, 440] of the RHS of
(L1) we always have solutions where both x and

x are positive while (L1) is the only active con
straint.
As a conclusion it can be said that the sensitivity
results based on an optimal basis characterize correctly the optimality of that optimal basis. The
graphical representation, however, shows that

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

several results are either incomplete or irrelevant

from the point of view of the information required
by a decision maker.

4. Illustration of the problems of degeneracy in

a production planning model
Production planning is one of the most important application areas of linear programing. Either
directly, as individual planning models, or indirectly, as part of a complex production planning and
control system, linear programs are run and their
results used for production management decisions.
In most of the cases sensitivity analysis is more
important than the optimum itself. On one hand,
the exact result of an LP model can be implemented
very rarely in practice, but small modi"cations of
the optimal solutions can still give acceptable
results. Thus it is necessary to be aware of multiple
optimal solutions and how they behave under perturbations. Sensitivity analysis provides information about the possible modi"cations and their
e!ect on the optimal objective function value and
on the optimal solutions. On the other hand, in
practice the conditions (available capacities, demand, etc.) may change but sensitivity analysis can
help to react appropriately to the new situation.
Most of the production planning models have the
following structure:
* ,
min (p x #h I )
GR GR
GR GR
R G
subject to
x #I
!I "D , i"1,2, N, t"1,2, ,
GR
GR\
GR
GR
,
m x )K , t"1,2, ,
G GR
R
G
,
I )IN< ,
GR
R
G
x , I *0,
GR GR

t"1,2, ,

i"1,2, N, t"1,2, ,

(8)

267

where the following notations are used:

N

x
GR
I
GR
D
GR
p
GR
h
GR

number of di!erent products,

number of planning periods,
produced quantity of product i in period t,
inventory of product i at the end of period t,
demand of product i in period t,
production unit cost of product i in period t,
inventory holding cost of product i in
period t,
m
resource consumption of product i,
G
K
available production capacity in period t,
R
IN< available inventory capacity in period t.
R
The objective of problem (8) is to determine the
optimal production quantity and the inventory
level in all the L planning periods. This plan minimizes the production and inventory holding costs
and, at the same time, production and inventory
capacities are respected. Problem (8) can be extended by considering several stages of the production, distinguishing regular and over time
productions, incorporating subcontractors, etc.
Special conditions, production requirements and
initial inventory levels can be formulated with
additional constraints as well. Such models are
widely discussed in the literature (see for example
[1,13,14]). In almost all cases, however, the solution
is degenerate. The problems and traps of sensitivity
analysis of this type of models will be illustrated
with a small size model determined by the data of
Table 7.
The production quantity of two products (P

and P ) in two production periods ( and )



should be determined. The demand for P is zero in

the "rst and 200 units in the second period. The
demand for P is 100 units in both periods. The

production cost is the same ($10 per unit) for both
products in , and $25 per unit for P , and $20 per


unit for P in . The inventory holding cost is the


same in all periods for all products ($5 per unit).
There is the capacity to produce 300 units in the
"rst period, and 200 units in the second period. The
inventory can not exceed 200 units in any of the
two periods.
Substituting these data in (8) a production planning model with eight variables, four equalities and
four inequalities are obtained. Completing this
model with four slack variables the following LP

268

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

Table 7
Data of the production planning model
Model parameters

Demand
Prod
(units/period) (D ) Prod
GR
Production cost
Prod
($/units) (p )
Prod
GR
Inventory cost
Prod
($/units) (h )
Prod
GR
Production capacity
(unit/period) (K )
R

Period 1
( )


Period 2
( )


1
2

0
100

200
100

1
2

10
10

25
20

1
2

5
5

5
5

300

200

200

200

Inventory capacity
(unit/period) (IN< )
R

scheduled for as well. The planned 200 units of


P and 100 units of P are exactly equal to the


production capacity of the "rst period. There is also
enough inventory capacity to store the 200 units of
P until the second period. Since there is no more

free production capacity, the second period demand of P can not be produced in , although it


would be advantageous "nancially. The optimal
solution therefore is the following:
x "200; x "100; x "0; x "100;




I "200; I "0;
I "0; I "0;




s "0;
s "100;
s "0;
s "200.




There are six nonzero values and the rank of the
matrix of problem (9) is 8. Therefore the solution is
primal degenerate, and since there is no alternative
optimum, the solution is not dual degenerate. In

model in the standard form is obtained:

min (10x #10x #25x #20x #5I #5I #5I #5I )








subject to
x


"0,
"100,

x
#I
!I
"200,



x
#I
!I
"100,



x #x
#s
"300,



x #x
#s
"200,



I #I
#s
"200,



I #I
#s "200,



x

!I


!I


(9)

x , x , x , x , I , I , I , I *0.
       
This small size problem can be solved with any LP
software, but the optimal solution can be found
easily by simple reasoning as well.
The data show that there is a considerable di!erence between the production costs in and in .


It would be cheaper to produce all the products in
. The products demanded in are produced


"rst. If there is free capacity, products demanded in
can be produced in as well. After producing


100 units of P , there is free production capacity,

therefore the production of P demanded in is



this case Type I and Type II sensitivities are the

relevant information.
Solving the model with the STORM computer
program, the same solution is gained. Apart from
the optimal values of the production quantities and
inventory levels, several important questions can be
asked. First we will try to answer these questions
logically, then, we will see what answer is provided
by the sensitivity analysis results of the STORM
computer program. The RHS sensitivities at two
optimal basis (B "+1, 2, 4, 6, 7, 10, 11, 12, and


T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

B "+1, 2, 3, 4, 5, 10, 11, 12,) are given in Tables


8 and 9. The objective function coe$cient sensitivities at B and B are provided in Tables 10 and 11.


(1) A new customer wants to buy P in the xrst

period. How much will be the marginal cost of this
extra production?
The production cost of P in is $10. If one


extra unit should be produced in , then the pro
duction of another unit, which was originally
produced in , but demanded in , has to be


produced in because of production capacity

limitations. The shift of the production of this one
unit from to will increase production cost


from $10 to $25, and at the same time eliminates the
$5 inventory cost. The total cost of the shift is
therefore $10 for every unit (25!10!5). The
result is the sum of the production cost of the new
product ($10) and the cost of shift ($10), which
yields a $20 increase of the objective function for
every unit of new P produced for . There is 100


units free capacity in to reschedule P . There

fore, the $20 shadow price is valid as long as the
new demand for P in the "rst period is less than

100 units.
When the program "nds the optimal solution at
the basis B the sensitivity information is incorrect.

The $15 shadow price, shown in the "rst row of
Table 8, is not the deduced value. The validity
range shows that we are at a break point of the
piecewise linear optimal value function. $15 is just
a sub-di!erential, the true right and left derivatives
might be di!erent from $15. When the optimal
solution is found at the basis B then the sensitivity

information coincides with the deduced values ("rst
row of Table 9).
(2) The demand for P in the second period may

decrease or increase. How will this change the optimal value of the total cost?
Producing one unit less from P for will result


in the savings of $10 production cost, and in the
savings of $5 inventory holding cost because all
P is produced in . Since there will be free capa

city in , one unit of P produced for can be



shifted to , saving by this way another $5

(20!10!5). This is altogether $20 per unit. Since
100 units of P can be substituted by 100 units of

P , this $20 left side shadow price is true as long as

the demand decreases from 200 units to 100 units.

269

Table 8
RHS sensitivity of the production planning model at the optimal
basis B

Dual
variable

y
"
y
"
y
"
y
"
y
!.
y
!.
y
',4
y
',4

Current
RHS value

0
100
200
100
300
200
200
200

Shadow
price

15
15
20
20
!5
0
0
0

Validity range
LL

UL

0
100
100
0
300
100
200
0

0
100
200
200
300
R
R
R

Table 9
RHS sensitivity of the production planning model at the optimal
basis B

Dual
variable

y
"
y
"
y
"
y
"
y
!.
y
!.
y
',4
y
',4

Current
RHS value

0
100
200
100
300
200
200
200

Shadow
price

20
20
25
20
!10
0
0
0

Validity range
LL

UL

0
100
200
0
200
100
200
0

100
200
300
200
300
R
R
R

When the demand for P increases by one unit in


, this extra quantity should be produced in

because in there is no free production and


inventory capacity. This will result in a $25 increase
of the objective function for every unit (right side
shadow price). Production can be increased up to
100 units as a consequence of the 100 units free
production capacity in .

When the program "nds the optimum at the
basis B then the left side shadow price and validity

range is provided (third row of Table 8). When the
optimum is found at the basis B then the right side

shadow price is obtained (third row of Table 9).
Which one is found, however, cannot be in#uenced
by the user.

270

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

(3) How much does the optimal total cost change if

production capacity is increased by one unit in the
xrst period?
Since the 200 units of P is produced for , it


should be kept in the warehouse. Inventory constraints indicate that there is no more space to
store, therefore the demand of P in the second

period cannot be produced earlier, although, "nancially it would be advantageous. Therefore, no matter how much the production capacity is increased
in , it will not in#uence the objective function;

the shadow price is zero.
The zero shadow price cannot be found at any of
the two basis, as the "fth rows of Tables 8 and
9 show.
(4) How much does the optimal total cost change if
the production capacity in the xrst period decreases
by one unit?
Since production capacity is fully utilized in ,

the lost capacity will decrease the production of P .

If P is produced in production cost increases


by $15 (from $10 to $25) but inventory cost disappears ($5). The objective function therefore increases by $10 per unit. No more than 100 units of
production can be shifted to because of capacity

limitations, therefore the $10 is valid when production capacity does not decrease below 200 units.
When the program "nds the optimum at the
B basis then the provided shadow price is wrong

("fth row of Table 8). The validity range indicates
that this is true just in the very near neighborhood
of the current capacity, but from a practical point of
view, this information is irrelevant. When the optimum is found at the B basis then the answer is

right ("fth row of Table 9). The negative sign indicates that the decrease of capacity increases the
objective function.
(5) How much does the optimal total cost change if
inventory capacity in the xrst period decreases by one
unit?
When the inventory capacity in decreases,

then production of P in should be decreased,


because there is not enough inventory space to
store those units of P which are produced for .


Shifting the production of one unit of P to will


increase the costs by $10 (25!10!5"10).
No matter which basis is found, this information
is not provided at any of the basis (seventh row of

Tables 8 and 9). The zero right side shadow price

correctly indicates that there is no use increasing
the inventory capacity because production capacity
will impede any improvement of the production
plan. There is no information, however, about the
e!ect of the capacity decrease (left side shadow
price).
(6) The production cost of P in the second period

decreases. Will this require the change of the optimal
production plan?
At the current production cost it would be better
to move the production of P to , but the pro

duction capacity is fully utilized. When production
cost of P decreases in , the possible bene"t by


producing P in decreases as well. When pro

duction cost drops to $15, the production cost in
will be equal to the production plus inventory

cost in , therefore shifting the production to is


not advantageous any more. Since the production
was not moved to because of the capacity con
straints, this $15 is just a symbolic value. This value
indicates that if we could change the plan it would
be advantageous to do it as long as the production
cost is higher than $15. But the correct answer to
the question is that no matter how much the production cost of P in decreases, the production


plan will stay optimal.
When the program "nds the optimal solution at
the basis B then we may conclude that the produc
tion plan should be changed when the cost decreases below $15, because the fourth row of Table
10 indicates a $15 lower limit for the validity of the
optimal production plan. When the B basis is

found, then the right sensitivity information is
given (fourth row of Table 11).
(7) The inventory holding cost of P in the xrst

period increases. Will this change the optimal production plan?
Since only the demand of P in is scheduled


for production in , there is no inventory of P in


the optimal production plan. It means, that no
matter how much the inventory holding cost of
P increases it will not in#uence the optimal plan. It

is true, however, that, reaching $10 has a symbolic
importance. Above this level it will not be worth to
move the production of all the P to even if it


were possible, because the high production cost in
the second period will still be better than the low

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

Table 10
The sensitivity of the objective function coe$cients of the production planning model at the optimum basis B

Objective
function
coe$cient

Current
value

p

p

p

p

h

h

h

h


10
10
25
20
5
5
5
5

Validity range
LL

UL

!15
5
20
15
!20
0
!20
!20

15
15
R
25
10
10
R
R

Table 11
The sensitivity of the objective function coe$cients of the production planning model in the optimal basis B

Objective
function
coe$cient

Current
value

p

p

p

p

h

h

h

h


10
10
25
20
5
5
5
5

Validity range
LL

UL

!R
5
20
!5
!R
0
!25
!20

15
R
R
25
10
R
R
R

production cost in plus the increased inventory


cost.
When the program "nds the optimal solution at
basis B then we may conclude that the production

plan should be changed when the cost increases
above $10, because the sixth row of Table 10 indicates a $10 upper limit for the validity of the optimal production plan. When basis B is found, then

the right sensitivity information is given (sixth row
of Table 11).
(8) What is the bottleneck of this production system, and how can the bottleneck capacity be increased?
The bottleneck of this production system is constituted by both the production and inventory

271

capacities in . Both should be increased if de

mand in increases or production has to be

rescheduled from to . If both capacities


are increased by one unit, then one unit of production of P can be shifted from to . This will



result in a $5 (20!10!5) decrease of the optimal
total cost. This simultaneous increase of two RHS
elements cannot be analyzed with the help of
Tables 8 and 9.
In the following, some parameters of problem (9)
will be modi"ed to get a more general case. Let the
production cost of P in be $5. As a consequence


of the decrease of production cost of P in it is


even more advantages to produce all P in . The


savings when the production of one unit of P is

shifted from to is $10 (20!5!5), which is


equal to the saving when one unit of P is shifted

from to (25!5!10). Since in the original


problem 200 units of P were produced in advance,

and there is no production capacity in there are

alternative optima, because every unit of P pro
duced in can be substituted by P . Obviously,


then the substituted P should be produced in .


Let us relax the production capacity constraints as
well, saying that there is capacity to produce 500
units in both periods. This is more than the total
demand for both products, therefore there is
enough capacity to produce all P and P in .



Obviously the existing inventory capacity does not
allow to produce more units in , although it

would be advantageous "nancially. The two alternative optimal basis solutions in this case are the
following:
Optimum 1:
x "200;

I "200;

s "200;


x "100; x "0;


I "0;
I "0;


s "400;
s "0;



Optimum 2:
x "100;

I "100;

s "200;


x "200; x "100; x "0;




I "100; I "0;
I "0;



s "400;
s "0;
s "200.




x "100;

I "0;

s "200.


Optimum 1 is the same as the optimum of the

original problem, except, that the values of the

272

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

slack variables s and s are increased as a conse


quence of the increase in the production capacities
in each period. The value of s is not zero any more.

Despite this change still there are seven nonzero
values among the variables, and the rank of the
matrix of problem (9) is 8. Therefore Optimum 1 is
a primal degenerate optimal solution. This also
means that more than one optimal basis belong to
Optimum 1. In Optimum 2 the production of 100
units of P in is substituted by 100 units of P .



The 100 units substituted quantity of P has to be

produced in . In this solution there are 8 non
zero variables. Therefore this optimum is not a primal degenerate. Since the LP problem has two
alternative optimal basis solutions, it is a dual degenerate.
Let us analyze the change in the capacity of the
inventory (RHS of the inventory capacity constraint) in . Since inventory capacity is the limit
ing constraint, it does not allow the shifting of the
production to . When this 200 units capacity

increases by one unit, then one unit of P or P can


be shifted to , and the gain is $10 in case of both

products. Therefore, the right side shadow price of
the inventory capacity in is $10 per unit. In case

of Optimum 1, 100 units of P and in case of

Optimum 2, 100 units of P can be shifted to .


Therefore the capacity upper bound is 300 units
(200#100). When the capacity of the inventory
decreases, then there is no capacity to store the
products produced in and used only in ,


although it would be advantageous "nancially.
This "nancial advantage is $10 in case of each
product, which is the left side shadow price. In case
of Optimum 1, 200 units of P can be shifted to ,


which results in a zero (200!200) lower bound of

the capacity decrease. In case of Optimum 2 there

are two possibilities. When P is shifted from to


there is no change of production plan, since 100

units of P are produced anyway in . There is no


need to set up machines for new products, just more
should be produced from P . In this case 100 units

of P can be shifted to which results in a 100


units lower bound for the validity range of the
shadow price belonging to Optimum 2. Finally,
100 units of the P can also be shifted to but


this is a di!erent production plan, because originally machines were not set up for producing
P in .


Table 12 summarizes the provided sensitivity
results for the RHS of inventory constraint in .

The modi"ed version of problem (9) is dual degenerate, therefore there are two optima. Optimum 1 is primal degenerate, and the sensitivity
ranges found at the two optimal basis B and

B are listed in the third column of Table 12.

These are the left and right side sensitivity ranges.
Since the left and right side shadow price is the
same ($10 per unit) the sensitivity of Optimum 1
is the union of the left and right sensitivity ranges
(column 4). Optimum 2 is not primal degenerate,
therefore the sensitivity of the optimal basis is
equal to the sensitivity of the optimal solution at
B , as it is indicated in column three and four.

Finally the optimal value function changes linearly
with a $10 per unit gradient within the [0, 300]
interval embracing two optimal solutions and three
optimal basis, as it is indicated in column "ve. It
should be noted again that the STORM software
package provides one of the three intervals given in
column three, depending on the optimal basis
found "rst.

Table 12
The three di!erent types of sensitivity ranges of the shadow price of the RHS of the inventory constraint of the "rst period (y "10)
',4
Solution

Basis

Sensitivity range
of the basis

Sensitivity range
of the optimal
solution

Optimum 1

B "+1, 2, 4, 5, 6, 9, 10, 12,


B "+1, 2, 3, 4, 5, 9, 10, 12,

B "+1, 2, 3, 5, 6, 9, 10 ,12,


200)b )300
',4
0)b )200
',4
100)b )300
',4

0)b )300
',4

Optimum 2

100)b )300
',4

Sensitivity range
of the objective
function

0)b )300
',4

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

Finally if this modi"ed problem is solved by an

IPM solver we may get the following optimum:
IPM Optimum:
x "150; x "150; x "50; x "50;




I "150; I "50;
I "0;
I "0;




s "200;
s "400;
s "0;
s "200.




This is a strictly complementary optimal solution.
Type II sensitivity in this case determines the range
of a parameter within which the support sets of
some appropriate optimal solutions are identical.
The support set in this case represents a production
plan in which every product is produced in every
period and there is inventory of both products in
the "rst period. Because this is a strictly complementary optimal solution, Type II and Type III
sensitivities coincide.
In this small example it was possible to "nd out
the right answers to the questions and check to
what extent the sensitivity results at the di!erent
optimal basis are correct. In complex problems,
with several thousand constraints and variables, it
is very di$cult to know how sensitivity analysis,
provided by current programs, should be interpreted (see for example [5,15]). Note that there is
consistent mathematical theory to answer all the
above questions correctly, typically one smaller LP
has to be solved to get the answer to each question
separately (see [8,12,16]). This is, however, not implemented yet in the existing commercial LP packages, which is the shortcoming of the software and
not the shortcoming of the theory. There is a lot of
possibility to improve the sensitivity output of LP
software.
E When these software are designed for managerial
use the sensitivity information concerning the
decision variables and/or the objective function
and not sensitivity information of the optimal
basis should be presented. This requires some
extra computation, but would certainly be a considerable help to users in the management area.
Table 1 is an example of this type of sensitivity
information.
E Bottleneck analysis is one of the most important
area in production management. When several
resources constitute the bottleneck of the system

273

(the solution is degenerate) the information

about the set of bottleneck resources and about
the multiple change of the RHS elements is an
essential information. Table 4, and the analysis
of question 8 in Section 4 illustrates this problem.
Note that this question can be treated by introducing a new variable that represents the
simultaneous change of the RHS elements and
sensitivity on this new variable is to be obtained. An easy and automatic performance of
this analysis by the computer packages would be
important.
E Finally the existence of alternative solutions
should be expressed more explicitly. The problem of primal degeneracy is that, although the
primal solution is unique, there are alternative
dual optima. Turning this around, when a dual
optimal solution is degenerate, then the primal
optimum is not unique. In practical terms it
means that several optimal production plans
exist. For operations managers it would be useful
to generate information about the multitude of
the optimal plans, and decision on the selected
one can be based on intangible criteria. Here we
stress again, consistent mathematical theory
exists for determining the dimension of the optimal sets when one uses IPMs. To provide this
information would not increase the computational cost at all.

5. Conclusions
The main objective of the paper is to show that
implementation of sensitivity analysis in commercial packages, and managerial interpretation of sensitivity analysis of linear programming models are
di!erent. The sensitivity information given by the
simplex based commercial packages tell the user in
what range the data can vary to keep the obtained
optimal basis optimal, and how the current optimal
basis solution changes as a function of the problem
data. When an optimal solution of an LP model is
degenerate then there are several optimal basis
providing the same optimal value, and possibly
all optimal basis provide di!erent sensitivity results. These results are mathematically correct, but
their information content is either incomplete or

274

T. Koltai, T. Terlaky / Int. J. Production Economics 65 (2000) 257}274

irrelevant from the management decision point of

view. Management wants to know either the sensitivity information concerning activities in an optimal solution (Type II sensitivity), or the sensitivity
information concerning the objective function
(Type III sensitivity).
The situation is a little di!erent in case of solvers
based on the IPM, because these solvers provide
strictly complementary solutions. These optimal
solutions are not basis solutions. Therefore Type I
sensitivity has no relevance. However using e$cient
optimal basis identi"cation techniques, an optimal
basis can be produced when needed. On the other
hand, Type II and Type III information coincide in
this case. But to distinguish among Type II and
Type III sensitivity is important for decision making purposes, specially when a non-strictly complementary optimal solution is implemented.
Both the graphical solution of the small LP
model and the logical solution of the production
planning model have illustrated the existence of the
three type of sensitivities.
Users should be careful when sensitivity results
of an LP package are used for management decisions.
Almost all practical size problems are degenerate, and
the sensitivity information depends on the basis
found by the computer program. Di!erent software
may give di!erent results to the same model. Sometimes the goodness of the sensitivity output can be
checked by simple logic, but in most of the cases
there is no direct way of evaluating the results.
Linear programming will probably stay one of
the most popular operations research tool used in
practice. The development of computer technology
brought nearer this tool to inexperienced users. The
interpretation of the sensitivity output of the currently available software packages is di$cult and
contains several traps because they provide only
Type I sensitivity information. Software producers
have a lot of possibility to help avoid the presented
problems and to serve better the users in the management area.

References
[1] L.A. Johnson, D.C. Montgomery, Operations research in
production planning, in: Scheduling and Inventory Control, Wiley, New York, 1974.
[2] J.R. Evans, N.R. Baker, Degeneracy and the (mis)interpretation of sensitivity analysis in linear programming,
Decision Science 13 (1982) 348}354.
[3] D.C. Aucamp, D.I. Steinberg, The computation of shadow
prices in linear programming, Journal of the Operational
Research Society 33 (1982) 557}565.
[4] M. AkguK l, A note on shadow prices in linear programming, Journal of the Operational Research Society 35
(1984) 425}431.
[5] H.J. Greenberg, An analysis of degeneracy, Naval Research Logistics Quarterly 33 (1986) 635}655.
[6] T. Gal, Shadow prices and sensitivity analysis in linear
programming under degeneracy, OR Spektrum 8 (1986)
59}71.
[7] D.S. Rubin, H.M. Wagner, Shadow prices: Tips and
traps for managers and instructors, Interfaces 20 (1990)
150}157.
[8] B. Jansen, J.J. de Jong, C. Roos, T. Terlaky, Sensitivity
analysis in linear programming: just be careful! European
Journal of Operational Research 101 (1997) 15}28.
[9] R.E. Wendell, The tolerance approach to sensitivity analysis in linear programming, Management Science 31 (1985)
564}578.
[10] R.E. Wendell, Sensitivity analysis revisited and extended,
Decision Sciences 23 (1992) 1127}1142.
[11] T. Gal, Postoptimal analysis, in: Parametric Programming and Related Topics, McGraw-Hill, New York,
1986.
[12] C. Roos, T. Terlaky, J.-Ph. Vial, Theory and Algorithms
for Linear Optimization } An Interior Point Approach,
Wiley, Chichester, UK, 1997.
[13] S.C. Graves, A.H.G. Rinnooy Kan, P.H. Zipkin, Logistics
of Production and Inventory, North-Holland, Amsterdam, 1993.
[14] S. Nahmias, Production and Operations Analysis, Irwin,
Burr Ridge, 1993.
[15] A. Farkas, T. Koltai, A. Szendrovits, Linear programming
optimization of a network for an aluminum plant: A case
study, International Journal of Production Economics 32
(1993) 155}168.
[16] A.B. Berkelaar, C. Roos, T. Terlaky, The optimal set and
optimal partition approach, in: H.J. Greenberg, T. Gal
(Eds.), Advances in Sensitivity Analysis and Parametric
Programming, Kluwer Academic Publishers, Dordrecht,
The Netherlands, 1997.