Linear System with Hidden Z-Matrix and Complementary Condition arxiv:1807.04536v1 math.oc 12 Jul 2018 R. Jana a, A. Dutta a,1, A. K. Das b a Jadavpur University, Kolkata, 700 032, India. b Indian Statistical Institute, 203 B. T. Road, Kolkata, 700 108, India. 1 Email: aritradutta001@gmail.com Abstract In this article we study a class of linear system formed by hidden Z-matrix with complementary condition. We extend the results of Fiedler and Pták for the linear system with complementary condition using game theoretic approach. In this context we consider the matrix class based on the concept of principal pivot transform. Examples are presented to demonstrate our results. We show that hidden Z based linear system with complementary condition can be solved by solving a suitable linear program. Keywords: Hidden Z-matrix, linear system, complementary condition, value of a matrix, P 0 -matrix. AMS subject classifications: 90C33, 15A39, 15B99. 1 Introduction The linear system with complementary condition finds a vector in a finite dimensional real vector space that satisfies a particular system of inequalities. The problem is defined as follows: Given A R n n and a vector q R n, we consider the linear system and complementary conditions as q +Az 0, z 0, (1.1) z T (q +Az) = 0. (1.2) More specifically the problem is called linear complementarity problem denoted as LCP(q, A) for which we find a vector z R n satisfying the linear system as well as the complementary condition or show that no such vector exists. The case of LCP(q,A) with q = 0 is known as homogeneous LCP with A for which z in the solution set of LCP(0,A) implies that λz belongs to the solution set of LCP(0,A) for all scaler λ 0. For details see 3. The linear system with complementary condition includes linear programming, linear fractional programming, convex quadratic programming, bimatrix game and a number of applications reported in operations research, multiple objective programming problem, mathematical economics, geometry and engineering. The linear system with complementary condition along with a hidden Z-matrices received wide attention in the literature. A generalization of Z-matrices was addressed by 1 Corresponding author 1
Mangasarian 8 to study the linear complementarity problems solvable as linear program. Pang 11 proposed this class as hidden Z-matrices. A matrix A R n n is called hidden Z-matrix if there exist Z-matrices X, Y R n n and r,s R+ n such that AX = Y, (1.3) r T X +s T Y > 0. (1.4) Fiedler and Pták 5 studied Z-matrix in the context of linear system with complementary condition. In this article we show that a feasible vector satisfying the linear system is a solution of the linear system with complementarity conditions by extending results of Fiedler and Pták 5 in terms of hidden Z-matrix. The paper is organized as follows. Section 2 presents some basic notations and results. In section 3, we show a hidden Z-matrix under some additional conditions to be a P 0 matrix. We settle a result related to singular hidden Z-matrix. We illustrate our result by giving a suitable example of singular hidden Z-matrix. We show that a linear system with hidden Z-matrix and complementary condition can be solved with the help of a linear program. 2 Preliminaries We denote the n dimensional real space by R n. R n + denotes the nonnegative orthant of Rn. We consider vectors and matrices with real entries. Any vector x R n is a column vector unless otherwise specified and x T denotes the row transpose of x. e denotes the vector of all 1. If A is a matrix of order n, α {1,2,,n} and ᾱ {1,2,,n}\α then A αᾱ denotes the submatrix of A consisting of only the rows and columns of A whose indices are in α and ᾱ respectively. A αα is called a principal submatrix of A and det(a αα ) is called a principal minor of A. a ij denotes the ijth element of the matrix A. Given a matrix A R n n and a vector q R n, we define the feasible set FEA(q,A) = {z R n : z 0,q+Az 0} and the solution set of LCP(q,A) by SOL(q,A) = {z FEA(q,A) : z T (q +Az) = 0}. The concept of principal pivot transform plays an important role in this context. Some of the matrix classes are invariant under the principal pivot transform. For details see 2. We state the results of two person matrix games in linear system with complementary conditions due to Von Neumann 12 and Kaplansky 7. The results say that x R m,y R n and v R such that m i=1 x i a ij v, j = 1,2,,n, n j=1 y j a ij v, i = 1,2,,m. The strategies (x,y ) are said to be optimal strategies for player I and player II respectively and v is said to be minimax value of game. We write v(a) > 0 to denote the value of the game corresponding to the payoff matrix A. The value of the game v(a) is positive(nonnegative) if a 0 x 0 such that Ax > 0 (Ax 0). Similarly, v(a) is negative(nonpositive) if a 0 y 0 such that y T A < 0 (y T A 0.) The value of the matrix game corresponding to payoff matrix A R n n is preserved in all its PPTs. Now we give the definitions of some matrix classes which will be required in the next section. A matrix A R n n is said to be P (P 0 )-matrix if all its principal minors are positive (nonnegative). almost P (P 0 )-matrix if all its principal minors upto order (n 1) are positive (nonnegative) and det(a) < 0. 2
S-matrix 11 if there exists a vector x > 0 such that Ax > 0 Z-matrix if a ij 0 and K(K 0 )-matrix if it is Z P (P 0 ). N-matrix if all its principal minors are negative. An N-matrix is called an N-matrix of the first category if it contains at least one positive entry otherwise it is called an N-matrix of the second category. Q-matrix if for every q, LCP(q,A) has atleast one solution. Now we give some theorems which will be required for discussion in the next section. Theorem 2.1. 1 Let A R n n be a hidden Z-matrix. Then for any two Z-matrices satisfying AX = Y and r T X +s T Y > 0 for some r,s R n +. Then (i) X is nonsingular, and (ii) an index set α {1,2,,n} such that the matrix Xαα X αᾱ Yᾱα Yᾱᾱ K. Theorem 2.2. 11 Let A R n n hidden Z. Then A P if and only if A S. Theorem 2.3. 5 Let A be a Z-matrix. If v(a) > 0 then A P. Theorem 2.4. 5 Let A be a Z-matrix. If an x > 0 such that Ax 0, then A P 0. 3 Main results We first show that hidden Z-matrices are invariant under principal rearrangement. Theorem 3.1. If A R n n is a hidden Z-matrix then PAP T is a hidden Z-matrix for any permutation matrix P. Proof. Let A be a hidden Z-matrix. Then by the definition of hidden Z-matrix Z-matrices X,Y and two nonnegative vectors r,s such that AX = Y and r T X + s T Y > 0. Now for any permutation matrix P, P 1 PAP T (P T ) 1 X = Y. Thus, (PAP T )(P T ) 1 X = PY. Therefore(PAP T )(P T ) 1 XP T = PYP T. Now letting X 1 = (P T ) 1 XP T and Y 1 = PYP T, we get (PAP T )X 1 = Y 1. It is easy to show that X 1 and Y 1 are Z-matrices. Now for r 1,s 1 R n +,rt 1 X 1+s T 1 Y 1 = r T P T ((P T ) 1 XP T )+s T P 1 (PYP T ). Thereforer T 1 X 1+s T 1 Y 1 = r T XP T + s T YP T = (r T X + s T Y)P T > 0. Thus PAP T satisfies the definition of hidden Z-matrix. Theorem 3.2. Let A R n n hidden Z-matrix. Suppose a vector x > 0 such that Ax 0. Then A P 0. Proof. We prove this result by induction method on n. The result is trivially true for n = 1. Let us consider it is true for all matrices of order < n. Now consider the matrix A R n n hidden Z. Then for some Z-matrices X and Y, AX = Y. Now a vector x > 0 such that Ax 0. This implies YX 1 x 0 since X is nonsingular. Letting x 1 = X 1 x yields Xx 1 = x > 0. Hence Yx 1 0. Then by the Theorem 2.1, an index set α {1,2,,n} Xαα X such that the matrix W = αᾱ K and Wx 1 0. This gives x 1 0. Therefore Yᾱα Yᾱᾱ X is a K-matrix and for x > 0, X 1 x > 0 since X is nonsingular. Therefore x 1 > 0 with Yx 1 0. Then by the Theorem 2.4 of 5, Y P 0. This implies Y K 0. Therefore det 3
A 0. Now it is sufficient to prove that for any β {1,2,,n}, the principal submatrix A β β of A is hidden Z and a y > 0 such that A β βy 0. Now A β β(x β β X ββ X 1 ββ X β β ) = Y β β Y ββ X 1 ββ X β β which implies A β β (X/X Xββ X ββ) = (M/X ββ ), where M = β β. It is easy to show that (M/X ββ ) is a Z-matrix. Since (X/X ββ ) is a K-matrix, therefore A β β uβ is a hidden Z-matrix. Without loss of generality we consider x 1 =. Then Now as X ββ X 1 ββ X ββ X 1 ββ I I X ββ X β β X ββ X β β X ββ X β β uβ X ββ X β β uβ = 0 X/X ββ u β y β = (X/X ββ ) > 0. Again u 0 M/X β ββ y β > 0, A β βy β 0. This implies det A β β 0. Therefore A P 0. Y ββ. Y β β > 0, then (X/X ββ ) > 0. Let us take 0. Therefore (M/X ββ ) 0. So for Remark 3.1. The above result may not hold if the condition x > 0 is dropped to x 0. For 1 0 example if we let A = then it is easy to show that x 0 such that Ax 0. 1 2 Note that A / P 0 but A is a hidden Z-matrix. Corollary 3.1. Let A R n n hidden Z-matrix and a vector x > 0 such that Ax 0. Then every Schur complement (1, p. 75) in A hidden Z P 0. Theorem 3.3. Let A be a singular hidden Z-matrix. Then v(a) 0. Proof. Let us consider A be a singular hidden Z-matrix but v(a) > 0. We show that an x > 0 such that A x > 0. By definition of value positivity an x R+ n such that Ax > 0. Let x = x+ǫe > 0, where ǫ > 0. Then A x = A(x+ǫe) = Ax+ǫAe. If Ae 0, it is enough to choose any ǫ > 0. If not. Let a = min i (Ax) i > 0 and b = max i (Ae) i. Now choose ǫ such that a > ǫb. This implies ǫ < a/b. Now for 0 < ǫ < a/b, we can get x = x+ǫe such that x > 0 and A x > 0. Now A is a hidden Z-matrix with v(a) > 0. We say that an x > 0 such that A x > 0. Again as A is hidden Z-matrix then for some Z-matrices X and Y, AX = Y. Since X is nonsingular by the Theorem 2.1, then YX 1 x > 0. Let x 1 = X 1 x. Then Y x 1 > 0 and X x 1 > 0. Then by the Theorem 2.1, an index set α {1,2,,n} Xαα X such that the matrix W = αᾱ K and W x 1 > 0. Let x 2 = W x 1 > 0. Then Yᾱα Yᾱᾱ x 1 = W 1 x 2 > 0 since W 1 0. Hence for any x 1 0, X x 1 > 0 and Y x 1 > 0. Therefore v(x) > 0 and v(y) > 0. Now as X and Y are Z-matrices then by the Theorem 2.3, X and Y are P-matrices. Thus we have dety > 0 and detx 1 > 0. Therefore deta > 0. It contradicts the fact that A is singular matrix. Example 3.1. Let A = Z-matrix with X = 1 1 0 1 1 0 0 0 1 2 1 0 1 1 0 0 1 3. Note that A is singular matrix. Now A is hidden and Y = 1 0 0 1 0 0 0 1 3. Take r = 1.6 4 2 and s = 4
4 0 0.1, then it is easy to show r T X +s T Y > 0. Then by Theorem 3.4, v(a) 0. Now we show that for a hidden Z-matrix with v(a) > 0, LCP(q,A) has a unique solution. Theorem 3.4. Let A be a hidden Z-matrix and v(a) > 0. Then the linear system with complementary condition given by 1.1 and 1.2 has unique solution. Proof. Let A R n n with v(a) > 0. Then x R n + such that Ax > 0. In view of Theorem 3.3 x > 0 such that A x > 0. Then by Theorem 2.2, the linear system with complementary condition given by 1.1 and 1.2 has unique solution. Corollary 3.2. Let A be a hidden Z-matrix and v(a) > 0. Then the linear system with complementary condition given by 1.1 and 1.2 has a convex and connected solution set. Remark 3.2. For a hidden Z-matrix A with v(a) > 0, LCP(q,A) is processable by Lemke s algorithm and criss cross method 4. Consider the following example. 1 2 0 Example 3.2. Let A = 0 1 0. For X = 1 0 1 3 0 r = 8 and s = 0 0 1 For x = 1 4 5 solution for each q R n. 1 2 0 0 1 0 1 2 1 and Y = 1 0 0 0 1 0 2 0 1, we obtain r T X + s T Y > 0. Hence A is a hidden Z-matrix., v(a) > 0. Therefore by Theorem 3.4, the system LCP(q,A) has a unique Here we propose a method to find whether a hidden Z-matrix A is P-matrix or P 0 - matrix. Algorithm: Step I: Choose ǫ, δ > 0. Consider the following linear programming problem, minimize s subject to Ax se 0 x δe s ǫ (3.1) If solution of the linear programming problem exists then by Theorem 3.4, A P, else go to Step II. Step II: Choose ǫ = 0, δ > 0 and consider the linear programming problem (3.1). If the solution of the linear programming problem exists then by Theorem 3.2, A P 0, else decision is inconclusive. Note that all 2 2 P-matrices are hidden Z but in general there are P-matrices which are not hidden Z 11. Now we show the condition under which a P-matrix is a hidden Z-matrix. 5
Definition 3.1. A matrix A R n n is said to be type D 9 if real numbers {α i } n i=1 with α n > α n 1 > > α 1, such that { α i if i j; a ij = α j if i > j. It is easy to show that positive type D matrices are P-matrices. Theorem 3.5. Suppose A is positive type D matrix. Then A hidden Z. Proof. Suppose A is positive type D matrix. Then A is nonsingular and A 1 is Z-matrix as shown in 9 which in turn implies A 1 is hidden Z-matrix. Now A is a PPT of A 1 and PPT of a hidden Z-matrix is hidden Z 1. Therefore A hidden Z. Theorem 3.6. Let A R n n hidden Z. Then A almost P-matrix if and only if A 1 N-matrix of second category. Proof. LetA R n n hiddenz almost P. ThenA 1 hiddenz N. SupposeA 1 N- matrix of first category. Then by 10, A 1 Q. Therefore by Theorem 3.4 it contradicts that A 1 N. This implies A 1 N-matrix of second category. let B = A 1 R n n hiddenz N-matrix of second category. As B is an N-matrix, all of its proper principal minors are negative and det(b) < 0. So det(b 1 ) < 0 i.e. det(a) < 0. Let B 1 = (m ij ). Then m ij = ( 1) i+j detb(j:i) detb, where B(j : i) is the submatrix of B deleting the jth row and ith column 6. Letting j = i yields m ii = ( 1) i+i detb(i:i) detb. This implies that m ii > 0. Let α {1,2,,n} and ᾱ = {1,2,,n} \ α. Then det(b 1 ) αα = detbᾱᾱ detb > 0. Therefore B 1 = A is an almost P-matrix. Now we introduce a linear program to solve hidden Z-matrix based linear system with complementary condition using KKT condition. Theorem 3.7. Let A hidden Z such that AX = Y and r T X +s T Y > 0 where X,Y are Z-matrix and r,s R+ n. Then the linear system with complementary conditions denoted by LCP(q,A) can be written as minimize (r +A T s) T z 1 +q T z 2 subject to A T s+r A T z 2 0, Az 1 +q 0, z 1,z 2 0 Proof. For LCP(q,A) with A hidden Z, we consider another LCP( q,ā) where Ā = 0 A T p and q = with p = r+a A 0 q T z1 s. Let z = bethe solution of z LCP( q,ā). 2 AgainLCP( q,ā) canbewrittenasequivalent quadraticprogrammingproblem. NotethatĀ isaskewsymmetricmatrix. HenceweobtainthesolutionofLCP( q,ā)bysolvingequivalent quadratic programming problem. Again equivalent quadratic programming problem can be rewritten as minimize (r +A T s) T z 1 +q T z 2 subject to A T s+r A T z 2 0, Az 1 +q 0, z 1,z 2 0. Now following the Lemma 1 of Mangasarian 8, z 1 is the solution of LCP(q,A). 6
4 Conclusion In this article, we study the linear system with complementary condition. The class of hidden Z-matrix and some additional conditions provide solution for this system. To prove our result we apply the concept of principal pivot transform and game theoretic approach. We establish certain characterization of hidden Z-matrix to identify the properties of P 0 - matrix. The main advantage of a linear system with hidden Z-matrix and complementary condition is that the system allows to find a solution by solving a suitable linear program instead of solving a complementary system. Acknowledgement The author R. Jana is thankful to the Department of Science and Technology, Govt. of India, INSPIRE Fellowship Scheme for financial support. References 1 RW Cottle, JS Pang, and RE Stone. The linear complementarity problem. 1992. AP, New York. 2 AK Das. Properties of some matrix classes based on principal pivot transform. Annals of Operations Research, 243(1-2):375 382, 2016. 3 AK Das, R Jana, and Deepmala. On generalized positive subdefinite matrices and interior point algorithm. In Frontiers in Optimization: Theory and Applicationseds: Samarjit Kar et al., pages 3 16. Springer, 2016. 4 AK Das, R Jana, and Deepmala. Finiteness of criss-cross method in complementarity problem. In Mathematics and Computing, eds: Debasis Giri, Ram N. Mohapatra, Heinrich Begehr, Mohammad S. Obaidat, pages 170 180. Springer, 2017. 5 Miroslav Fiedler and Vlastimil Pták. On matrices with non-positive off-diagonal elements and positive principal minors. Czechoslovak Mathematical Journal, 12(3):382 400, 1962. 6 Gregery A Johnson. Inverse N 0 -matrices. Linear algebra and its applications, 64:215 222, 1985. 7 Irving Kaplansky. A contribution to von neumann s theory of games. Annals of Mathematics, pages 474 479, 1945. 8 Olvi L Mangasarian. Linear complementarity problems solvable by a single linear program. Mathematical Programming, 10(1):263 270, 1976. 9 Thomas L Markham. Nonnegative matrices whose inverses are M-matrices. Proceedings of the American Mathematical Society, 36(2):326 330, 1972. 10 Czeslaw Olech, T Parthasarathy, and G Ravindran. Almost N-matrices and linear complementarity. Linear Algebra and its Applications, 145:107 125, 1991. 7
11 Jong-Shi Pang. Hidden Z-matrices with positive principal minors. Technical report, WISCONSIN UNIV MADISON MATHEMATICS RESEARCH CENTER, 1977. 12 John Von Neumann. A certain zero-sum two-person game equivalent to the optimal assignment problem. Contributions to the Theory of Games, 2:5 12, 1953. 8