See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/252003878

Symmetric matrices properties to duality in linear programming problem

Article · April 2011

DOI: 10.1109/ICMSAO.2011.5775555

CITATIONS READS

3 549

4 authors, including:

Ikhsan Romli Taufik Sahroni

Universitas Pelita Bangsa Technical University of Malaysia Malacca
18 PUBLICATIONS 23 CITATIONS 2 PUBLICATIONS 4 CITATIONS

SEE PROFILE SEE PROFILE

Adi Saptari
President University
103 PUBLICATIONS 331 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

A Plug Assembly Line: An Assessment of Workstation Design and Working Posture View project

Automated Real-Time Ergonomic Assessment View project

All content following this page was uploaded by Ikhsan Romli on 08 February 2019.

The user has requested enhancement of the downloaded file.

 Symmetric Matrices Properties to Duality in Linear Programming Problem
Ikhsan Romli1, Taufik2, Adi Saptari3, Muzalna bt. Mohd Jusoh4
1, 2, 3,4
Faculty of Manufacturing Engineering,
Universiti Teknikal Malaysia Melaka,
Durian Tunggal, 76109 Melaka,
MALAYSIA
1
[email protected], [email protected], [email protected],
4
[email protected]

Abstract In LP, the primal and dual problem formulations

Duality is one of the most important topics in solved by the simultaneous solution. The primal-dual
optimization either a theoretical and algorithmic solution will be optimal in the transaction if either the
perspective. Optimization problem usually involved primal or dual problems converge to the same solution.
mathematical model. One of the applications widely For example, if the sales revenue of the producer is
used is Linear Programming. Linear functions maximized by primal formulation, then the purchasing
applications are frequently used in production planning, cost of the customer is minimized by the related dual
networks, scheduling, and other application of a linear formulation [7].
function subject to linear constraints. Extensive number The square matrix termed basis has long played a
of papers related with duality in optimization has been fundamental role in simplex algorithms for solving LP
published. However, studies about the effect of different problem, yet the basis was extended recently to include
characteristics of randomly symmetric matrices in the a deficient case by exploiting primal degeneracy [8].
duality in linear programming have not yet been The symmetric matrices can be classified into five based
discussed widely. This study addresses new findings on different characters of quadratic form i.e. positive
which may contribute to advancement of LP theory and definite, positive semi-definite, negative definite,
practice, particularly on the effects of various number negative semi-definite, and indefinite [9]. Each of the
of variables and different characteristics of matrices in character has its own properties in term of its
LP problems to duality. determinant.
Commonly, there are two methods that are
1. Introduction frequently improved by researcher in LP problem;
A certain group of optimization problems that the Simplex Method discovered by George Dantzig, Interior
objective functions is a linear function of the proposed Point Method by Kamarkar. Recently, the idea of Todd
variables and also the linear equalities or inequalities to [10] introduced that the effectiveness of excellent
be called Linear Programming (LP) [1]. Associated with performances of simplex methods and interior point
every LP problem, original called the primal, there is methods were similar for the practiced applications of
other LP problem called its dual that can be formulated LP. Nevertheless, for particular types of LP problems, it
by slightly altering the form of the primal problem. It is may be that one type of method is usually used in small
able to be applied to obtain the originally program problem properly within admit of measure of
solution and also the remarkably advantageous dimensions of the order 50x100, whereas interior point
knowledge of the optimal solution to the primal should method is more efficient for some large-scale problem.
be given by its variable [2]. This research simulated matrices in duality come
The researchers have worked through the up to dimension of the order 50x50, it used simplex
development of LP duality (primal-dual solution), i.e. method to solve the LP problems. This research focused
the evidence of properties of duality relationships had on estimation in the effect of different characteristics of
been proved utilized by Corley [3], weak and strong symmetric matrices to LP problems in duality. The goal
duality in infinite dimensional [4], LP only using with of this research work is to develop duality theory in LP
Positive Semi-definite Matrices [5]. Theory in LP for various characteristics matrices.
concerns the concept of duality. The dual solutions can
be used to measure the primal sensitivity with respect to 2. Proposed Methodology
changes in the independent term of the restrictions [6]. This paper presents a study on the simulation of
Solving in the area of LP problem normally used optimization linear model for appraising various
Simplex methods in which it could terminate, with characteristics of definite matrices. For the purpose of
returning a feasible solution or non-feasible one. this work, the simplex method will be used in this linear
However, it would not be able to know whether that optimization. The steps of the proposed simulation can
feasible solution is optimal. Therefore, the concept of be summarized as follows:
LP duality is used to prove that it is the optimal
solution.

978-1-4577-0005-7/11/$26.00 ©2011 IEEE

2.1 Create the Random Matrices calculation of primal problem as well as dual problem is
Symmetric matrices were randomly generated in need compared both MATLAB and EXCEL Solver as a
five types of matrices, i.e. Positive Definite (PD), validation from the result of LP problem. Hence, the
Positive Semi-Definite (PSD), Negative Definite (ND), previous step is required recalculation by use of the
Negative Semi-definite (NSD), In-definite (ID). EXCEL Solver.
MATLAB software was used to generate random matrix 3. Result and Discussion
LP problem for different characteristics of matrices and The evaluation some simulation result which are
different size of matrices.
illustrated the behavior of randomly symmetric matrices
2.2 Verify the Properties of Symmetric Matrices on simplex method that generated in duality LP
Setting of the simulation design was carried out to problems:
test the symmetric matrices in exact accord with number
of decision variables that followed by five characters set 3.1 Generating Randomly Matrices
matrices differently in LP problem. This step used the The required matrices were generated for the
determinant of the leading principal minor of the sub-
matrix to verify whether it has included into one of five purpose of LP formats. All parameters required consist
types. of the coefficient matrix which represents the
coefficient for constraint, the column vector as the right-
2.3 Solve the Subject in LP Duality hand side, and the row vector as the coefficient for
The matrices are confirmed, seeking for solution objective function. These matrices used to simulate on
of primal problem. The primal problem is need
primal problem as well as dual problem of the LP
converted to find the dual problem in a way that is
coefficient matrix, the objective function of the primal problems.
problem will be done transposing. Meanwhile, once the In particularly, the coefficient matrices were formed in
matrices for dual problem are confirmed, looking for the square matrices that belonging to each one of five
solution of dual problem. A nearly symmetric relation various characters of symmetric matrices that include
between a primal problem and its dual problem can be PD, PSD, ND, NSD, ID. Actually, the size of matrices
seen by considering the following canonical form of
linear inequalities: generated were simulated in this research just a few
orders of matrices (i.e. 5x5, 10x10, 20x20, 30x30,
Primal Problem 50x50). For each of the matrix, 30 samples were
Max c1x1 + c2x2 ............cnxn = f (1) simulated for this study. A result small part of generated
Subject to : a11x1 + a12x2 ....... + a1nxn ≤ b1 symmetric matrices randomly is shown in the Table 1.
a21x1 + a22x2 ....... + a2nxn ≤ b2 (2)
Table 1. The A5x5 of Symmetric Matrices
an1x1 + an2x2 ...... + annxn ≤ bn Type Generate of the Matrices
where x1, x2, ...., xn ≥ 0 (3) 1st 2nd 3rd
It can also be written concisely as: 30
8
8
25
1
2
6
9
9
7
26
4
4
30
4
2
8
7
8
0
21
9
9
21
3
2
9
6
5
4
Max cT x = f (4) PD 1 2 22 4 10 4 2 21 0 4 3 2 23 1 10

Subject to: Ax ≤ b (5) 6

9
9
7
4
10
24
8
8
23
8
8
7
0
0
4
28
7
7
27
9
5
6
4
1
10
24
8
8
25
where A is a matrix, c and b are vectors. Each elements -42 1 2 1 7 -42 4 5 3 5 -42 3 4 3 1

(a, b, and c) is ∈ R. ND
1
2
-47
10
10
-40
4
9
0
8
4
5
-46
4
4
-44
7
7
10
3
3
4
-48
8
8
-44
8
8
1
5
1 4 9 -42 9 3 7 7 -48 6 3 8 8 -46 8
7 0 8 9 -43 5 10 3 6 -48 1 1 5 8 -41

Dual Problem 70
21
21
62
24
14
31
19
31
19
62
17
17
62
16
18
28
26
28
26
57
15
15
69
10
13
24
34
24
34
Min b1y1 + b2y2 ............bnyn = v (6) PSD 24 14 67 28 28 16 18 65 25 25 10 13 58 22 22

Subject to : a11y1 + a12y2 ....... + a1nyn ≥ c1 31

31
19
19
28
28
69
69
69
69
28
28
26
26
25
25
81
81
81
81
24
24
34
34
22
22
74
74
74
74
a21y1 + a22y2 ....... + a2nyn ≥ c2 (7) -238 16 15 26 26 -242 17 8 27 27 -238 9 6 25 25
16 -237 8 23 23 17 -240 8 17 17 9 -243 5 23 23
NSD 15 8 -240 23 23 8 8 -244 16 16 6 5 -247 18 18
an1y1 + an2y2 .......+ annyn ≥ cn 26 23 23 -229 -229 27 17 16 -230 -230 25 23 18 -220 -220
26 23 23 -229 -229 27 17 16 -230 -230 25 23 18 -220 -220
-4 12 -2 1 3 9 -4 7 -16 5 -10 -6 0 9 6
where y1, y2, ...., yn ≥ 0 (8) 12 12 7 11 -13 -4 7 12 3 2 -6 4 -3 7 -3
It can also be written concisely as: ID -2
1
7
11
-6
1
1
-1
7
16
7
-16
12
3
-12
-11
-11
14
-9
-22
0
9
-3
7
11
6
6
0
-4
-3
Min b Ty = v (9) 3 -13 7 16 -7 5 2 -9 -22 -1 6 -3 -4 -3 -15

Subject to: ATy ≥ c (10)

where A is a matrix, c and b are vectors.
3.2 Verifying the Properties of Symmetric Matrices
Note that A in LP general form is Amxn matrix, but it The matrices were verified to make sure the
shall be formed symmetric (Anxn) matrix as in above. properties of matrices were symmetric matrices which
Each elements (a, b, and c) is ∈ R.
generated randomly by MATLAB software. The
2.4 Validate the Primal-Dual Solution verification was conducted by checking the determinant
The result of optimization produces either the
value of objective function or the decision variable. The
of the leading principal minor of the sub-matrix. This
procedure is well-known as Sylvester’s Theorem.
Let αk be the determinant of the leading principal sub-
matrix of order k of A, i.e.

α 1 = a11, α2 = det ,

α3 = det , . . . , αn = det A

Then, the conditional statement fulfilled the criteria of (a)

symmetric matrix A is described in Table 2 as follow:

Table 2. Conditional statement of Symmetric Matrix

Type Relation Condition

PD <=> αk > 0 for all k
ND <=> αk < 0 for odd k & αk > 0 for even k
PSD <=> αn = 0 & all k are ≥0
NSD <=> αn = 0 & odd k are ≤0 and even k are ≥ 0
ID => diagonal elements have different signs or some even k are < 0
(b)
Each result of the generated matrices tested to
verify and ensure that those were included in one of five
types within the symmetric matrices. For example, one
of result generated symmetric matrices will be taken to
verify by adopting the conditional statement above
mentioned to be considered as programming logic in
MATLAB source-code shown as follows.
30 8 1 6 9
8 25 2 9 7
A= 1 2 22 4 10
6 9 4 24 8
9 7 10 8 23 (c)

For the matrix A, there are five principal minors namely

of order 1 until order 5, the following answer of each of
principal minors for detail:

α1 = 30, α2 = det 30 8
= 686,
8 25
30 8 1
α3 = det 8 25 2 = 14979,
1 2 22
30 8 1 6
α4 = det 8 25 2 9
= 298453, (d)
`1 2 22 4
6 9 4 24
30 8 1 6 9
8 25 2 9 7
α5 = det 1 2 22 4 10 = 4421637
6 9 4 24 8
9 7 10 8 23

It appears these results of all principal minors are

greater than zero (non-negative). Thus, it is clearly
answered that the matrix A has been verified as a PD
symmetric matrix.
The following is one of 5 types of symmetric (e)
matrices 50x50 depicted in surface plots.
Figure 1. Surfaces plot of Symmetric Matrix. (a) PD,
(b) PSD, (c) ND, (d) NSD, (e) ID
 3.4 Converting the Matrices to Dual Problem
In Figure 1, PD matrix has a surface with the diagonal The primal problem is required conversion to find
elements are higher than the other elements, beside that the dual problem in a way that is the coefficient matrix,
all of the elements are linearly independent because its the column vector, and the row vector of the primal
determinant value equal to positive or greater than zero. problem will be done transposing. Especially, since the
This character called as non-singular matrix (full-rank coefficient matrix has symmetric-shaped so it will be
and invertible). Likewise it occurs in PSD that random congruent although having taken transposition. On the
number of the diagonal elements always higher, other hand, the column vector will be interchange the
however it has linearly dependent its 2 rows and 2 row and column turning into the row vector, and vice
columns, such that determinant of this matrix always versa. It can be shown as follows.
equal to zero or as singular matrix (non-invertible).
Meanwhile, ND and NSD matrices have diagonal 30 8 1 6 9 66
elements are negative numbers. These matrices are 8 25 2 9 7 24
T
distinguishable that NSD is contained dependent linear A = 1 2 22 4 10 , c = 30 ,
in 2 rows and 2 columns, so it always singular. ID has 6 9 4 24 8 68
different sign or multiple diverse in the diagonal 9 7 10 8 23 53
elements and also in the rest elements. Hence some
these matrices have determinant value that equal to zero bT = 41 60 75 58 55
and the rest matrices are non-singular.

3.3 Solving the Subject in Primal Problem 3.5 Solving the Subject in Dual Problem
Every one of LP problems always comprises of a The formula of Dual Problem is:
symmetric matrix, and two vectors that will be used to Min b Ty = v
simulate on both primal problem and dual problem of Subject to: ATy ≥ c
the LP problems. In primal problem, the simulations of where y1, y2, ...., yn ≥ 0
LP are performed using MATLAB software. For So that,
example of simulation result is provided as follows.
30 8 1 6 9 41
8 25 2 9 7 60 Min 41 60 75 58 55 *
A= 1 2 22 4 10 , b= 75 ,
6 9 4 24 8 58
9 7 10 8 23 55
30 8 1 6 9 66
cT = 66 24 30 68 53 8 25 2 9 7 24
The formula of Primal Problem is: Subject to 1 2 22 4 10 * ≥ 30 , where ≥0
6 9 4 24 8 68
Max cT x = f 9 7 10 8 23 53
Subject to: Ax ≤ b
where x1, x2, ...., xn ≥ 0 The optimal value of the objective function (v) is
So that, 268.9187, and shown that it has converged in an optimal
solution.
The result of the decision variables value is
Max 66 24 30 68 53 * y1 = 1.544933, y2 = 0, y3 = 0.586596, y4 = 2.112695,
y5 = 0.709916.
In this research, the steps in above are also
conducted in MATLAB Software as many thirty times
30 8 1 6 9 41
8 25 2 9 7 60
repetitively with produce the matrices and vector that
Subject to 1 2 22 4 10 * ≤ 75 , where ≥0 have the different matrices elements randomly.
6 9 4 24 8 58
9 7 10 8 23 55 3.6 Validating the Primal-Dual Solution
The optimal value of the objective function (f) is The result of optimization produces either the
268.9187, and shown that it has converged in an optimal value of objective function or the decision variable. The
solution. calculation of primal problem as well as dual problem is
The result of the decision variables value is need compared both MATLAB and EXCEL Solver as a
x1 = 0.88725, x2 = 0, x3 = 2.992081, x4 = 1.638394, validation from the result of LP problem. Hence, the
x5 = 0.173338. previous step is required recalculation by use of the
In this research, the above procedures are Excel Solver. The results of calculation on Excel solver
conducted in each of five its characteristics as many are shown that both objective function values in primal
thirty (30) times with different set of variables i.e. 5, 10, solutions are 268.9187 as viewed in Figure 2 and also
20, 30 and 50 variables. These are based on the matrices dual solution as shown in Figure 3. It seen that those
and vector that have randomly generated the different values are similar precisely among primal and dual
matrices elements. solution and then the Excel solver have found the
optimal solution from the two of them.
3.7 Sensitivity Analysis 2) There exist differences between the primal solution
To determine the sensitivity of the optimal and the dual solution or duality gap in ND, NSD,
solution, sensitivity analyses were done by changing the and some other ID matrices.
data value. This includes by changing either Right-Hand 3) For symmetric matrices, the study found out that
Side (RHS) or Objective Function Coefficient (OFC). It almost all of those were non-singular of PD, ND, ID
can be appeared in the Table 3 - 6 that in the Constraints matrices. Others were singular i.e. PSD, NSD
section, if Final Value = Constraint RHS, then slack matrices.
should be zero, and because the resource is in a little 4) The experimental results by simulation appear that
supply, it should have a non-zero Shadow Price. On the the optimal solution of LP problems for both primal
other hand, slack must be positive, and Shadow Price and dual for feasible problems, MATLAB or
must be zero, if Final Value ≠ Constraint RHS. The EXCEL Solver provided the same solution.
Allowable Increase as well as the Decrease will
illustrate Unlimited (1E+30) depending on which 5. Acknowledgement
direction is not significant (where slack/surplus keeps The authors would like to thank The Ministry of
growing). Higher Education Malaysia for the financial support
under Fundamental Research Grant Scheme
3.8 Comparison of the Whole Experiment Result No.FRGS/2010/FKP/TK03/6-F00104, Faculty of
The overall simulation results are compared Manufacturing Engineering, Universiti Teknikal
among primal and dual solution in different setting of Malaysia Melaka (UTeM) for support and providing
Problems and those are given in Table 7. And then, the facilities.
result of primal and dual solution is evaluated either the
theory of duality in problem or the theory of matrix in References
linear algebra. [1] R a v i n d r a n , A., Ragsdell, K.M.,
The Fundamental Theorem of Duality said that for any Reklaitis, G . V . , Engineering Optimization
couple of dual problem, one and only one of the Methods and Applications, 2nd ed.., New Jersey,
following situations are possible; John Wiley and Sons, 2006.
i) Both problem have solutions: therefore, they have [2] Bazarra, M. S., J arvi s, J .J ., Li near
optimal solutions and the optimal values of the P r o gr a m mi n g and Network Flows, Canada, John
objective functions coincide; Wiley and Sons, 1977.
ii) One of the problems has a solution, the other does [3] Corley, H.W., Duality Theory for the Matrix LP
not: therefore, the former problem has finite Problem. Journal of Mathematical Analysis and
optimization; Applications, Vol. 104, No. 1, (1983), pp. 47-52.
iii) Neither of the two problems has a solution. [4] Romeijn, H.E., Smith, R.L., and Bean, J.C. Duality
in Infinite Dimensional Linear Programming.
Strong Duality. If solution couple (x, y) of the two Mathematical Programming, Vol. 53, (1992), pp. 79-97.
problems has the feature that f(x) = v(y), then x is [5] Lasserre, J.B. Linear Programming with Positive
optimal solution of primal LP problem and y is the best Semi-definite Matrices. Proceeding of the 34th IEEE
solution of dual LP problem one. Conference on Decision & Control, Vol. 2, (1995), pp.
1127-1132.
Weak Duality. If the primal LP problem does not have [6] Balbas, A., Jimenez, P., and Heras, A. Duality
finite optimization, then dual LP problem does not have Theory and slackness Conditions in Multiobjective
any admissible solutions, and vice versa. Linear Programming. An International Journal
Computers & Mathematics with Applications, Vol. 37,
(1999), pp. 101-109.
It can be seen that the relations of the [7] Scherpereel, C.M. Alignment: the duality of
computerized experimental result showing PD, PSD, decision problems. Management Decision, Vol. 44, No.
and some ID have equal in values between the primal 9, (2006), pp. 1258-1276.
objective value and the dual objective value lead to [8] Pan, P.-Q. a Revised Dual Projective Pivot
achieve optimal solution. Meanwhile, ND, NSD, and Algorithm for Linear Programming. SIAM Journal
the rest ID are different values between the primal Optimization, Vol. 16, No. 1, (2005), pp. 49-68.
objective value and the dual objective value or duality [9] Laub, A.J. Matrix Analysis for Scientists and
gap, in these conditions arising three (3) possibilities Engineers. The United States of America, Society for
any one of optimal, unbounded and infeasible. Industrial and Applied Mathematics, 2005.
[10] Todd, M.J. The Many Facets of linear
4. Conclusions programming. Mathematical Programming, Vol. 91,
1) The primal solution and the dual solution for LP (2002), pp. 417-436.
problem with PD, PSD, and some ID matrices obtain
optimal solution.
APPENDIX

Figure 2. The Maximizing Primal Model in Excel Solver

Figure 3. The Minimizing Dual Model in Excel Solver

 Table 3. Sensitivity Report in Adjustable Cells of Primal Problem

Table 4. Sensitivity Report in Constraints of Primal Problem

Table 5. Sensitivity Report in Adjustable Cells of Dual Problem

Table 6. Sensitivity Report in Constraints of Dual Problem

Table 7. The Comparison of Simulation Result for Different Setting of LP Problems

View publication stats