Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
A STUDY OF PARTIAL ORDERS ON NONNEGATIVE MATRICES AND VON NEUMANN REGULAR RINGS A dissertation presented to the faculty of the College of Arts and Sciences of Ohio University In partial fulfillment of the requirements for the degree Doctor of Philosophy Brian Scott Blackwood August 2008 This dissertation entitled A STUDY OF PARTIAL ORDERS ON NONNEGATIVE MATRICES AND VON NEUMANN REGULAR RINGS by BRIAN SCOTT BLACKWOOD has been approved for the Department of Mathematics and the College of Arts and Sciences by Dinh V. Huynh Professor of Mathematics Benjamin M. Ogles Dean, College of Arts and Sciences BLACKWOOD, BRIAN SCOTT, Ph.D., August 2008, Mathematics A STUDY OF PARTIAL ORDERS ON NONNEGATIVE MATRICES AND VON NEUMANN REGULAR RINGS (84 pp.) Directors of Dissertation: Surender K. Jain and Dinh V. Huynh In this dissertation, we begin by studying nonnegative matrices under the minus partial order. The study of the minus partial order was initiated by Hartwig and Nambooripad independently. We first give the precise structure of a nonnegative matrix dominated by a group-monotone matrix under this partial order. The special cases of stochastic and doubly stochastic matrices are also considered. We also derive the class of nonnegative matrices dominated by a nonnegative idempotent matrix. This improves upon and generalizes previously known results of Bapat, Jain and Snyder. Next, we introduce the direct sum partial order and investigate the relationship among different partial orders, namely, the minus, direct sum, and Loewner partial orders on a von Neumann regular ring. It is proven that the minus and direct sum partial orders are equivalent on a von Neumann regular ring. On the set of positive semidefinite matrices, we show that the direct sum partial order implies the Loewner partial order. Various properties of the direct sum and minus partial orders are presented. We provide answers to two of Hartwig’s questions regarding the minus partial order. One of the main results gives an explicit form of maximal elements in a given subring. Our result generalizes the concept of a shorted operator of electrical circuits, as given by Anderson-Trapp. As an application of the main theorem, the unique shorted operator has been derived. Finally, we consider the parallel sum of two matrices over a regular ring. Previously known results of Mitra-Odell and Hartwig are generalized. We also obtain a result on the harmonic mean of two idempotents. Approved: Dinh V. Huynh Professor of Mathematics Preface The study of nonnegative matrices was initiated by Perron [43] in 1907 and Frobenius [14] in 1912. In 1920, E.H. Moore [40] defined a unique inverse for every square or rectangular matrix but it gained little notice. In 1955, R. Penrose [42] showed that Moore’s generalized inverse is the unique matrix satisfying four equations which is now known as the Moore-Penrose generalized inverse. Subsequently, countless papers have been written on various topics related to generalized inverses. In particular, there has been a great deal of interest in λ − monotone matrices, that is, nonnegative matrices with a nonnegative generalized inverse. The investiga- tion began in 1972, when R. J. Plemmons and R. E. Cline [44] determined when a nonnegative matrix has a nonnegative Moore-Penrose inverse. Soon thereafter, in 1974, A. Berman and R.J. Plemmons [6] provided the solution for when nonnegative matrices possess a nonnegative group-inverse. Investigations into decompositions of nonnegative matrices with nonnegative generalized inverses were made by many authors, notably P. Flor, S. K. Jain and L. Snyder. In addition to the study of nonnegative matrices and generalized inverses for their own sake, there are applications to a variety of other fields including statistics, economics and engineering. Various partial orders on rings and matrices have been studied extensively. In 1934, K. Loewner [32] defined a partial order on the set of positive semidefinite matrices which has been used extensively with shorted operators. The star partial order of M. P. Drazin [12] is on a ring with involution. Motivated by Drazin, R. E. Hartwig [19] and, independently, K.S.S. Nambooripad [41] defined a partial order on a regular ring, namely the minus partial order. Both partial orders have been used on rings and matrices. S. K. Mitra [34] was instrumental in the development of the minus partial order as well as its applications to shorted operators. Chapter 1 provides the basic definitions, notation, lemmas and theorems used in the later chapters. A new lemma is stated and proven as well. In Chapter 2, we begin our study of nonnegative matrices and the minus partial order. Bapat, Jain and Snyder [4] described the structure of nonnegative matrices which are dominated by a nonnegative idempotent matrix under the minus partial order. We improve upon and generalize these results in our investigation into the structure of nonnegative matrices A that are dominated by a given nonnegative λ−monotone matrix B under the minus partial order. We provide necessary and sufficient conditions for a nonnegative matrix A dominated by a nonnegative matrix B that has a nonnegative group inverse. As a special case when B is an idempotent matrix, we provide an explicit description of the class of nonnegative matrices A dominated by a nonnegative idempotent matrix B. Next, we turn our attention to partial orders on a von Neumann regular ring. In Chapter 3, we introduce the direct sum partial order defined as a ≤⊕ b if bR = aR⊕ (b − a)R. The relationship between the minus partial order and the direct sum partial order is investigated. We provide answers to two questions that Hartwig [19] posed. Additionally, we give an explicit description of maximal elements in a subring under the minus partial order. As a special case, we obtain a result similar to the one obtained by Mitra-Puri [38] for the unique shorted operator. In Chapter 4, we consider the parallel sum of two matrices over a regular ring. This generalizes earlier results of Mitra-Odell and Hartwig. We also obtain a result on the harmonic mean of two idempotents. Acknowledgments I would like to acknowledge the invaluable help, patience and encouragement of my advisors, Professor S. K. Jain and Professor Dinh V. Huynh, during the preparation of this dissertation. Next, I would like to thank Dr. Ashish K. Srivastava for the helpful conversations that we had about von Neumann regular rings. Also, special thanks goes to Dr. Pramod Kanwar for his remarks and corrections to my dissertation. I would like to thank my family for their support and understanding. In particular, special thanks goes to my wife, Sara, and my children: Brian, William and Tyler. Contents Abstract 3 Preface 5 Acknowledgments 8 1 Preliminaries 11 1.1 Definitions, Notation, and Conventions . . . . . . . . . . . . . . . . . 11 1.2 Nonnegative Group-Monotone Matrices and the Minus Partial Order 16 1.3 Partial Order on a von Neumann Regular Ring . . . . . . . . . . . . 20 1.4 The Parallel Sum of Two Matrices Over a Regular Ring . . . . . . . . 21 2 Nonnegative Group-Monotone Matrices and the Minus Partial Order 23 2.1 Group-Monotone Matrices . . . . . . . . . . . . . . . . . . . . . . . . 24 2.2 Idempotent Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 9 10 3 Partial Order on a von Neumann Regular Ring 44 3.1 Equivalence of Partial Orderings and Their Properties . . . . . . . . . 46 3.2 A Characterization of the Maximal Elements . . . . . . . . . . . . . . 57 3.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4 The Parallel Sum of Two Matrices Over a Regular Ring 70 4.1 Commutative Regular Ring . . . . . . . . . . . . . . . . . . . . . . . 71 4.2 Regular Ring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 Bibliography 79 Chapter 1 Preliminaries This chapter provides the necessary background and results used throughout the dissertation. The first section lists definitions, notation, and conventions that are used in the dissertation. In the second section, results that are needed for Chapter 2 are given along with a new lemma. In the third section, we state important results that will be needed for Chapter 3. In the final section, known results are presented for use in Chapter 4. 1.1 Definitions, Notation, and Conventions Definition 1.1.1. A matrix A = [aij ] is nonnegative if aij ≥ 0 for all i, j which is denoted A ≥ 0. Likewise, A = [aij ] is positive if aij > 0 for all i, j which is 11 12 written A > 0. Definition 1.1.2. An n×n matrix A = [aij ] is said to be row stochastic if aij ≥ 0 and n P aij = 1 for 1 ≤ i ≤ n. An n × n matrix A = [aij ] is said to be column j=1 stochastic if aij ≥ 0 and n P aij = 1 for 1 ≤ j ≤ n. An n × n nonnegative matrix i=1 is said to be doubly stochastic if the matrix is both row and column stochastic. Definition 1.1.3. An n × n real matrix A is symmetric if A = AT , where T denotes the transpose. Likewise, an n × n complex matrix A is hermitian if A = A∗ where A∗ denotes the conjugate transpose of A. Definition 1.1.4. Let A be an m × n matrix. Consider the following equations: AXA = A (1) XAX = X (2) (AX)T = AX (3) (XA)T = XA (4) Ak XA = Ak (1k ) AX = XA where X is an n × m matrix. (5) For a matrix A and a non-empty subset λ of {1, 2, 3, 4, 5}, X is called a λ-inverse of A if X satisfies equations (i) for i ∈ λ. Note that equations 1k and 5 are only valid for square matrices. A {1} − inverse 13 of A will be written as A− or A(1). Also, the {1, 2, 3, 4} − inverse of A is the unique Moore-Penrose inverse and is denoted A†. Definition 1.1.5. Let A be an n × n matrix. The smallest positive integer k such that rank Ak = rank Ak+1 is called the index of a square matrix. Every square matrix A has a unique {1k , 2, 5} − inverse, where k is the index of A called the Drazin inverse which is denoted AD . The system of equations (1), (2) and (5) has a unique solution X, called the group inverse of A, if and only if rank(A) = rank(A2). The group inverse of A is denoted by A# where the index is k = 1. Definition 1.1.6. A matrix A is called λ − monotone if it has a nonnegative λ − inverse. A matrix A is group-monotone if A# exists and is nonnegative. Definition 1.1.7. A matrix J is a direct sum of matrices J1 , ..., Jr, written J = J1 ⊕ · · · ⊕ Jr , if J1 0 · · · 0 . . 0 J . . . . 2 . J = . .. . . . . . . 0 0 · · · 0 Jr Definition 1.1.8. If A, B are m × n matrices then we say that A is dominated by B under the minus partial order, written A ≤− B, if rank B = rank A+rank(B − A). 14 Theorem 1.2.1 will provide us with several equivalent statements. In Chapters 3 and 4, R will be a von Neumann regular ring with identity, unless stated otherwise. Definition 1.1.9. An element a ∈ R is called von Neumann regular if axa = a for some x ∈ R, and x is called a von Neumann inverse of a. A ring R is called von Neumann regular if every element in R is von Neumann regular. For convenience, we will often use the terminology regular ring in place of von Neumann regular ring. Definition 1.1.10. A ring R is unit-regular if for each x ∈ R there is a unit (an invertible element) u ∈ R such that xux = x. Definition 1.1.11. Let S be the set of all regular elements in any ring R. For a, b ∈ S we say that a ≤− b if there exists a {1} − inverse x of a such that ax = bx and xa = xb. This is known as the minus partial order for regular rings. For the ring of matrices over a field, a ≤− b if and only if rank(b − a) = rank(b) − rank(a). Definition 1.1.12. For a, b ∈ R, a ≤⊕ b if bR = aR ⊕ (b − a)R is called the direct sum partial order. Definition 1.1.13. The Loewner partial order on the set of positive semidefinite matrices S is defined as follows: for a, b ∈ S, a ≤L b if b − a ∈ S. 15 Definition 1.1.14. Let R be a regular ring and S be a subset of R. We define a maximal element in C = {x ∈ S : x ≤⊕ a} as an element b 6= a such that b ≤⊕ a and if b ≤⊕ c ≤⊕ a then c = b or c = a. Definition 1.1.15. We say that a, b ∈ R are parallel summable if a (a + b)(1) b is invariant under the choice of {1} − inverse (a + b)(1) . (1) summable then p(a, b) = a (a + b) If a and b are parallel b is called the parallel sum of a and b. Definition 1.1.16. In general, a ring R is called prime if aRb = (0) implies that a = 0 or b = 0. Definition 1.1.17. We say that every principal right ideal of a ring is uniquely generated if any two elements of the ring that generate the same principal right ideal must be right associates. In other words, if for all a, b in the ring R, aR = bR implies a = bu for some unit u ∈ R. Definition 1.1.18. The harmonic mean of two numbers is define the harmonic mean of a and b as 2a(a + b)(1)b. 2ab . a+b For a, b ∈ R, we 16 1.2 Nonnegative Group-Monotone Matrices and the Minus Partial Order We begin by stating some of the known facts and preliminary results that are referred to throughout Chapter 2. Theorem 1.2.1. ([34], Lemma 1.2) Let A and B be m × n matrices. Then the following conditions are equivalent: 1. A ≤− B 2. There exists a {1}-inverse A− of A such that (B −A)A− = 0 and A− (B −A) = 0. 3. Every {1}-inverse of B is a {1}-inverse of A. 4. Every {1}-inverse B − of B satisfies AB − (B − A) = 0 and (B − A)B −A = 0. In other words, the parallel sum of A and B − A is the zero matrix. Theorem 1.2.2. ([13], Theorem 2). If E is a nonnegative idempotent matrix of rank r, then there exists a permutation matrix P such that 17 JD J 0 0 T P EP = CJ CJ D 0 0 0 0 0 0 , 0 0 0 0 where all the diagonal blocks are square; J is a direct sum of matrices xiyiT , xi > 0, yi > 0 and yiT xi = 1, i = 1, 2, . . . , r; and C, D are nonnegative matrices of suitable sizes. In particular, if E is symmetric then J 0 , P EP T = 0 0 where J is a direct sum of matrices xi xTi , xi > 0 and xTi xi = 1, i = 1, 2, . . . , r. Lemma 1.2.1. ([4], Lemma 3) Let A, E be n × n matrices such that E 2 = E, and suppose that A ≤− E . Then A is idempotent and AE = A = EA. The next lemma gives the converse when A is also an idempotent. Lemma 1.2.2. Let A and B be n × n idempotent matrices. A ≤− B if and only if AB = A = BA. Proof. Let A, B be idempotent where AB = A = BA. As A is idempotent, AAA = A2A = AA = A2 = A so A is its own {1}-inverse. Now by Theorem 1.2.1, 18 we will prove the equivalent condition that there exists a {1}-inverse A− of A such that (B − A)A− = 0 and A− (B − A) = 0 to show that A ≤− B. So choose A = A− . Then (B − A)A− = (B − A)A = BA − A2 = BA − A = 0 as BA = A. Also A− (B − A) = A(B − A) = AB − A2 = AB − A = 0 as AB = A. Thus by Theorem 1.2.1, A ≤− B as required. The converse follows by the previous lemma. Theorem 1.2.3. ([24], Theorem 5.2) Let A be a nonnegative matrix and let A(1,2) = p(A) ≥ 0, where p(A) = k P αi Ami , αi 6= 0, mi ≥ 0. Then there exists a permutation i=1 matrix P such that JD J 0 0 T P AP = CJ CJ D 0 0 0 0 0 0 , 0 0 0 0 where C, D are nonnegative matrices of appropriate sizes and J is a direct sum of matrices of the following types (not necessarily both): (I) βxy T , where x and y are positive unit vectors with y T x = 1 and β is a positive root of k P i=1 αi tmi +1 = 1 19 β 12 x1y2T 0 ··· 0 0 . .. 0 T . . 0 β x y . 23 2 3 . . . where xi,yi are positive .. .. .. (II) 0 .. .. . β d−1,d xd−1 ydT . β d1xd y1T 0 ··· ··· 0 unit vectors of the same order with yiT xi = 1; xi and xj , i 6= j are not necessarily of the same order. The numbers β 12, . . . , β d1 are arbitrary positive with d > 1 and d | mi + 1 for some mi such that the product β 12β 23 · · · β d1 is a common root of the following system of at most d equations in t P αi t (mi +1) d = 1, d∈Λ0 P αi t (mi +1−k) d =0 d∈Λk k ∈ {1, 2, . . . , d− 1} where Λk = {d : d | mi + 1− k, d 6= 1} k ∈ {0, 1, . . . , d− 1} with the understanding that if some Λk = ∅ then the corresponding equation is absent. Conversely, suppose we have, for some permutation matrix P , JD J 0 0 T P AP = CJ CJ D 0 0 0 0 0 0 0 0 0 0 where C, D are arbitrary nonnegative matrices of appropriate sizes and J is a direct sum of matrices of the following types (not necessarily both) ′ (I ) βxy T , β > 0 where x,y are positive vectors with y T x = 1. 20 β 12x1 y2T 0 ··· 0 0 . .. 0 T . . 0 β x y . 23 2 3 ′ . . . .. .. .. (II ) 0 .. .. . β d−1,d xd−1 ydT . β d1 xd y1T 0 ··· ··· 0 where β ij ≥ 0, xi and yi are positive vectors with yiT xi = 1. Then A(1,2) ≥ 0 and is equal to some polynomial in A with scalar coefficients. 1.3 Partial Order on a von Neumann Regular Ring The following result of Jain and Prasad [27] will prove to be useful throughout Chapter 3 and, specifically, for providing an equivalent definition of the minus partial order on a regular ring. Theorem 1.3.1. Let R be a ring and let a, b ∈ R such that a+b is a regular element. Then the following are equivalent: 1. aR ⊕ bR = (a + b)R; 2. Ra ⊕ Rb = R(a + b); 3. aR ∩ bR = (0) = Ra ∩ Rb. 21 From Rao-Mitra [45], we have the following characterization of {a(1)} and {a(1,2)}. Lemma 1.3.1. Let R be a ring and let a ∈ R. If x ∈ {a(1)} then {a(1)} = x + (1 − xa)R + R(1 − ax). In addition, {a(1,2)} = {a(1)aa(1)}. 1.4 The Parallel Sum of Two Matrices Over a Regular Ring From Hartwig-Shoaf [22], we have the following theorem. Theorem 1.4.1. Let R be a regular ring and a, b ∈ R. Then the following are equivalent: 1. (a + b)R = aR + bR, R(a + b) = Ra + Rb; 2. The triplets a(a + b)(1)a, b(a + b)(1)b, a(a + b)(1)b, and b(a + b)(1)a are invariant under (·) (1) and a(a + b)(1)b = b(a + b)(1)a. The following theorem of Gillman-Henriksen [15] shows that every commutative regular ring with identity is a unit regular ring. Theorem 1.4.2. For any element a of a commutative regular ring with identity, there exists a unit u such that a2u = a. 22 Marks [33] proved the following result about unit regular rings. Theorem 1.4.3. Let R be a von Neumann regular ring. Then R is unit-regular if and only if every principal right ideal is uniquely generated. From [36], Mitra-Odell proved the following result for matrices over any field. Theorem 1.4.4. Let A and B be matrices of order m × n each, and let there exist a matric C such that {C − } = {A− } + {B − } Then A and B are parallel summable and C = P (A, B). In [17], Hartwig extended the result of Mitra and Odell to matrices over a prime regular ring. Theorem 1.4.5. Suppose R is a prime regular ring such that A, B, C ∈ Rm×n where {C − } = {A− } + {B − }. Then A, B are parallel summable and C = P (A, B). Chapter 2 Nonnegative Group-Monotone Matrices and the Minus Partial Order In this chapter, our focus is on the minus partial order studied by several authors ([19], [34], [36], [41]). We continue our investigation into the structure of nonnegative matrices A that are dominated by a given nonnegative λ−monotone matrix B under the minus partial order. The present chapter studies the case when B is a nonnegative matrix that possesses a nonnegative group inverse. In [4], Bapat, Jain and Snyder considered the case when the matrix B is an idempotent matrix. We provide necessary and sufficient conditions for a nonnegative matrix A dom- 23 24 inated by a nonnegative matrix B that has a nonnegative group inverse (Theorem 2.1.1). In the special case when B is an idempotent matrix, Theorem 2.2.1 provides an explicit description of the class of nonnegative matrices A dominated by a nonnegative idempotent matrix B. For some applications of decompositions of nonnegative matrices, one may refer to the following recent papers of Herrero-Ramirez-Thome [23] and Jain-Tynan [28]. 2.1 Group-Monotone Matrices Theorem 2.1.1. Let A, B be n × n nonnegative matrices such that B # ≥ 0, rank B = r and rank A = s, s ≤ r. permutation matrix P such that JD J 0 0 T P BP = CJ CJ D 0 0 Then A ≤− B if and only if there exists a A11D 0 0 A11 0 0 0 0 T and P AP = 0 0 CA11 CA11D 0 0 0 0 0 0 0 0 , 0 0 0 0 where C, D are nonnegative matrices of appropriate sizes and J is a direct sum of the following types (not necessarily both): (I) βxy T , where x and y are positive unit vectors with y T x = 1, β > 0. 25 β 12x1 y2T 0 ··· 0 0 . .. 0 T . . 0 β x y . 23 2 3 . . . .. .. .. (II) 0 .. .. . β d−1,d xd−1 ydT . β d1 xd y1T 0 ··· ··· 0 with β ij > 0 where xi ,yi are positive unit vectors of the same order with yiT xi = 1; xi and xj , i 6= j are not necessarily of the same order and α1r y1T x1 0 · · · 0 α11 α12 · · · .. .. 0 . . . . . . ... ... . . 0 A11 = . . . . . . . .. . . .. .. .. 0 .. .. 0 αr1 αr2 · · · αrr 0 · · · 0 xr ··· 0 .. . .. . ··· 0 .. .. . . .. . 0 0 yrT where αij ≥ 0 and A11 ≤− J . Proof. Let A, B be n × n nonnegative matrices such that rank B = r and rank A = s, s ≤ r. As B # ≥ 0, by Theorem 1.2.3 [24] there exists a permutation matrix P such that JD J 0 0 T P BP = CJ CJ D 0 0 0 0 0 0 , 0 0 0 0 26 For any {1} − inverse B − of B , P BP T = where J is of type (I) or (II). P BB − BP T = P BP T P B − P T P BP T . It is straightforward that − J −D J 0 0 CJ − CJ − D 0 0 0 0 JD J 0 0 0 0 is a {1} − inverse of CJ CJ D 0 0 0 0 0 0 0 0 0 0 0 0 0 0 because J − is a {1}−inverse of J . Furthermore, we choose a {1}−inverse P B − P T of P BP T as given below: − − J D J 0 0 − T PB P = − CJ − D CJ 0 0 0 0 0 0 . 0 0 0 0 As A ≤− B, by Theorem 1.2.1 [34], AB −B = AB −A = BB −A = A. Partitioning P AP T in conformity with P BP T yields A11 A 21 T P AP = A31 A41 A12 A13 A14 A22 A23 A24 . A32 A33 A34 A42 A43 A44 27 Now P AB − BP T = P AP T P B − P T P BP T A11 A 21 = A31 A41 A12 A13 A14 J − J −D 0 A22 A23 A24 0 A32 A33 A34 CJ − CJ − D 0 0 A42 A43 A44 − A11J J A J − J 21 = A31J − J A41J − J A11 A 21 = A31 A41 0 0 J JD 0 0 0 0 0 0 CJ CJ D 0 0 0 0 0 0 0 0 0 0 0 0 + A13CJ −J A11J − J D + A13CJ −J D 0 0 + A23CJ −J A21J − J D + A23CJ −J D 0 0 + A33CJ −J A31J − J D + A33CJ −J D 0 0 + A43CJ −J A41J − J D + A43CJ −J D 0 0 A12 A13 A14 A22 A23 A24 = P AP T . A32 A33 A34 A42 A43 A44 Thus, Ai3 = 0 and Ai4 = 0 for i = 1, 2, 3, 4. In addition, P BB − AP T = P BP T P B − P T P AP T JD J 0 0 = CJ CJ D 0 0 0 0 J − J −D 0 0 0 0 0 0 CJ − CJ − D 0 0 0 0 0 0 A11 0 0 A21 0 0 A31 0 0 A41 A12 A13 A14 A22 A23 A24 = A32 A33 A34 A42 A43 A44 28 − − − − J J A12 + J J DA22 J J A11 + J J DA21 0 0 CJ J −A + CJ J − DA − − 11 21 CJ J A12 + CJ J DA22 0 0 A11 A 21 = A 31 A41 0 0 0 0 0 0 0 0 A12 A13 A14 A22 A23 A24 = P AP T . A32 A33 A34 A42 A43 A44 Then it follows that A2j = 0 and A4j = 0 for j = 1, 2, 3, 4. Thus P AB − BP T P BB − AP T − − A11J J A11J J D 0 0 = A J − J A J − J D 31 31 0 0 − J J − A12 J J A11 0 0 = CJ J −A − 11 CJ J A12 0 0 0 0 0 0 and 0 0 0 0 0 0 0 0 0 0 0 0 where P AB − BP T = P AP T = P BB − AP T . We will now use the following list (1) of equations to determine P AP T : 29 A11 = A11J − J = J J − A11, A12 = A11J − J D = J J − A12, (1) A31 = A31J − J = CJ J − A11, A32 = A31J − J D = CJ J −A12. From the relations above, A11 = A11J − J = J J −A11, A12 = A11D, A31 = CA11, and A32 = CJ J − A12 = CA11J − J D = CA11D. As a result, A11D A11 0 0 T P AP = CA 11 CA11D 0 0 0 0 0 0 . 0 0 0 0 We claim: A11 ≤− J or equivalently rank(J ) = Rank(A11 ) + rank(J − A11). First note that rank(J ) = rank(B) = rank(P BP T ) and rank(A11) = rank(A) = rank(P AP T ) . Because rank(P BP T ) = rank(P AP T ) + rank(P BP T − P AP T ) it follows that rank(J ) = rank(A11) + rank(J − A11). Thus we obtain A11 ≤− J . Recall that A11 = A11J − J = J J −A11 where J is a direct sum of matrices of type ′ ′ (I) or (II). Now if J is any type (I) summand of J , then J = βxy T where β is a ′ positive scalar and x, y are positive unit vectors. Choose (J )− = β1 xy T . Thus 1 T 1 T ′ − ′ ′ ′ T T T (J ) J = xy (βxy ) = xy = (βxy ) xy = J (J )− β β 30 ′′ Now if J is any type (II) summand of J, then β 12x1 y2T 0 ··· 0 0 . . 0 .. 0 β 23x2y3T . . ′′ . . . .. .. .. J = 0 . .. T . . . β x y d−1,d d−1 d T β d1 xd y1 0 ··· ··· 0 for some positive integer d . Now 0 1 x yT β 12 2 1 ′′ − J = 0 .. . 0 ··· 0 ··· 0 1 x yT β d1 1 d 0 .. . . .. . ··· 0 1 x yT β 23 3 2 .. .. . .. . 1 x yT β d−1,d d d−1 .. . 0 . Then T x1 y1 ′′ ′′ − 0 J J = . .. 0 Thus 0 .. . .. . ··· ··· 0 .. .. . . ′′ − ′′ = J J . .. . 0 0 xdydT 31 T x1 y1 − − J J = JJ = ··· 0 0 .. . .. . .. . 0 ··· 0 .. .. . . .. . 0 0 xr yrT is a block diagonal matrix such that each diagonal block is of rank one where the summands of J are of type (I), (II), or both. Note that the list of equations (1) is valid for any given choice of {1} − inverse of J . We have chosen J # for J − . Now partition A11 in conformity with J − J = J J − . So A11 ′ ′ A11 · · · · · · A1r . . .. .. . . . . = . .. .. .. . . ′ ′ Ar1 · · · · · · Arr ′ ′ ′ From A11J − J = A11 = J J −A11, it follows that xi yiT Aij = Aij = Aij xj yjT . ′ ′ ′ Clearly, each Aij must be of rank 0 or 1. Now Aij = xi yiT Aij xj yjT . Thus, we may ′ write Aij = αij xi yjT . Hence A11 T T T α11 x1y1 α12x1 y2 · · · α1r x1 yr . . α x y T .. .. 21 2 1 = . . .. . . . . . αr1 xr y1T ··· · · · αrr xr yrT 32 α1r y1T x1 0 · · · 0 α11 α12 · · · .. .. 0 . . . . . . ... α . . 21 0 = . . .. .. .. . . . . . . 0 ... . . .. 0 · · · 0 xr αr1 · · · · · · αrr 0 ··· 0 .. . .. . 0 .. .. . . . .. . 0 0 yrT ··· The converse is clear. The proofs of the following corollaries follow immediately from Theorem 2.1.1. Matrices of type (I) and (II) are also the same. Corollary 2.1.1. Let A, B be n × n nonnegative matrices such that B # ≥ 0, B is row (column) stochastic, rank B = r and rank A = s, s ≤ r. Then A ≤− B if and only if there exists a permutation matrix P such that J 0 A11 0 and P AP T = , P BP T = CJ 0 CA11 0 where C is a nonnegative matrix of the appropriate size and J is a direct sum of type (I) and (II) matrices with β ij > 0 where xi,yi are positive unit vectors of the same order with yiT xi = 1; xi and xj , i 6= j are not necessarily of the same order 33 and α1r y1T x1 0 · · · 0 α11 α12 · · · .. .. 0 . . . . . . ... ... . . 0 A11 = . . .. .. . . . . . . 0 ... . . .. .. 0 · · · 0 xr αr1 αr2 · · · αrr 0 ··· 0 .. . .. . ··· 0 .. .. . . .. . 0 0 yrT where αij ≥ 0 and A11 ≤− J . J J D and P AP T = Note that if B was column stochastic, then P BP T = 0 0 A11 A11D above. 0 0 Corollary 2.1.2. Let A, B be n × n nonnegative matrices such that B # ≥ 0, B is doubly stochastic, rank B = r and rank A = s, s ≤ r. Then A ≤− B if and only if there exists a permutation matrix P such that P BP T = [J ] and P AP T = [A11] , where J is a direct sum of type (I) and (II) matrices with β ij > 0 where xi ,yi are positive unit vectors of the same order with yiT xi = 1; xi and xj , i 6= j are not 34 necessarily of the same order and T x1 0 · · · 0 α11 α12 · · · α1r y1 0 · · · 0 . . . . . . . . . 0 .. 0 . . . . . . . . . . . . . . . . A11 = . . . . .. . ... ... . . ... ... . . . 0 . . . 0 T 0 · · · 0 xr αr1 αr2 · · · αrr 0 · · · 0 yr where αij ≥ 0 and A11 ≤− J . The following example demonstrates that when A ≤− B, with B # nonnegative, A# need not be nonnegative. 0 1 Example 2.1.1. Let B = 0 0 0 1 0 0 1 0 0 0 and A = 1 0 0 0 0 0 0 0 1 1 0 1 0 0 where rank(B) = 3 0 0 0 0 0 0 and rank(A) = 2. Obviously both A, B ≥ 0. Now B # 0 0 = 1 0 1 0 0 0 1 0 ≥ 0. We 0 0 0 0 0 0 proceed to show that A is of the form stated in the theorem and A ≤− B. There 35 0 1 exists a permutation matrix P = 0 0 0 0 T P BP = 1 0 1 0 0 0 0 0 such that 0 1 0 0 0 1 1 0 0 1 1 0 1 0 and P AP T = 0 0 0 0 0 0 0 0 1 0 0 0 1 0 where C, D = 0, 0 0 0 0 0 0 0 1 0 J = 0 0 1 is a type (II) matrix 1 0 0 and A11 1 0 0 1 1 0 1 0 0 1 1 0 = 0 1 0 1 0 1 0 1 0 = 1 0 1 . 0 0 0 0 0 1 0 0 0 0 0 1 0 1 0 1 1 0 In addition, J − A11 = 0 0 1 − 1 0 1 = 1 0 0 0 0 0 −1 0 0 −1 0 0. 1 0 0 Therefore, rank A11 + rank (J − A11) = 2 + 1 = 3 = rank J. Thus A ≤− B. Furthermore, rank (A2) = 2 = rank (A) and so A# exists. 36 −1 1 # But A = 0 0 2.2 1 2 0 0 −1 0 0. 0 0 0 0 0 0 Idempotent Matrices As a special case of the previous theorem, we provide necessary and sufficient conditions that give the structure of a nonnegative matrix A satisfying A ≤− B where B is a nonnegative idempotent. The condition obtained in this situation is simpler to verify. Theorem 2.2.1. Let A, B be n × n nonnegative matrices such that B 2 = B, rank B = r and rank A = s . Then A ≤− B if and only if there exists a permutation matrix P such that JD J 0 0 T P BP = CJ CJ D 0 0 A11D 0 0 A11 0 0 0 0 and P AP T = CA 0 0 11 CA11D 0 0 0 0 0 0 0 0 , 0 0 0 0 where C, D are nonnegative matrices of appropriate sizes and J is a direct sum of matrices xi yiT where xi , yi are positive unit vectors of the same order with yiT xi = 1; 37 xi and xj , i 6= j are not necessarily of the same order and A11 T x1 0 · · · 0 y1 0 · · · 0 . . 0 . . . . . . .. 0 . . . . . . .. E = . . .. . . . . . . 0 .. . . . . . . 0 0 · · · 0 xr 0 · · · 0 yrT where E is a nonnegative idempotent r × r matrix. Proof. As B ≥ 0 and B is idempotent, B is its own group inverse. By assumption A ≤− B and so by Theorem 2.1.1 JD J 0 0 T P BP = CJ CJ D 0 0 A11D 0 0 A11 0 0 0 0 T and P AP = 0 0 CA11 CA11D 0 0 0 0 0 0 0 0 , 0 0 0 0 where C, D are nonnegative matrices of appropriate sizes and J is a direct sum of matrices xiyiT where xi , yi are positive unit vectors of the same order with yiT xi = 1; xi and xj , i 6= j are not necessarily of the same order and T x1 0 · · · 0 α11 α12 · · · α1r y1 0 · · · 0 . . . . . . . . . 0 .. . . .. .. .. .. . . .. .. 0 . A11 = . . . . .. . ... ... . . ... ... . . . 0 . . . 0 0 · · · 0 xr αr1 αr2 · · · αrr 0 · · · 0 yrT 38 As B is idempotent, it follows from Lemma 1.2.1 that A is idempotent and hence 2 T P AP Tis idempotent. Thus A11 = A11 because P AP is idempotent.Furthermore, T x1 0 · · · 0 y1 0 · · · 0 0 . . . . . . ... 0 . . . . . . ... has a left inverse and has a right since . . .. . . . . . . 0 .. . . . . . . 0 0 · · · 0 xr 0 · · · 0 yrT inverse, α11 α12 · · · α1r . .. .. .. . . E= . .. .. . . . . αr1 αr2 · · · αrr is also an idempotent matrix because yiT xi = 1. To prove the converse, we first show that A11 ≤− J , i.e. rank(J ) = rank(A11) + rank(J − A11). We have by assumption that rank B = r and rank A = s. follows that rank P BP T = r and rank P AP T = s. It Now as the rank P BP T is completely determined by J and rank P AP T is completely determined by A11, rank J = r and rank A11 = s. Now J − A11 = 39 T 0 x1 0 · · · 0 y1T 0 · · · 0 x1 y1 0 · · · . . . . . . . . . 0 0 0 . . . . . . . . . . . . . . . . . . − E = . . . .. .. .. .. . . . . . . 0 .. . . . . . . 0 . . 0 T T 0 · · · 0 xr yr 0 · · · 0 xr 0 · · · 0 yr T y1 x1 0 · · · 0 0 . . . . . . ... 0 . . . ... ... 0 .. . 0 0 · · · 0 xr .. . .. . ··· T y1 0 x1 0 · · · 0 .. .. .. .. .. . . . . . 0 0 E − . . . . .. .. .. 0 . 0 .. .. 0 0 · · · 0 xr 0 yrT ··· 0 T y1 0 · · · 0 x1 0 · · · 0 0 . . . . . . ... 0 . . . . . . ... . [I − E] = . . .. . . . . . . 0 .. . . . . . . 0 0 · · · 0 yrT 0 · · · 0 xr ··· 0 .. . .. . ··· 0 .. .. . . .. . 0 0 yrT T y1 0 · · · 0 x1 0 · · · 0 0 . . . . . . ... 0 . . . . . . ... ≤ rank ([I − E]). [I − E] Thus rank . . . . . . . . .. .. 0 .. .. 0 . . 0 · · · 0 yrT 0 · · · 0 xr But as Y T X [I − E] Y T X = [I − E] , 40 T y1 x1 0 · · · 0 0 . . . . . . ... 0 [I − E] rank . . .. . . . . . . 0 .. 0 · · · 0 xr 0 ··· 0 .. . .. . ··· 0 .. .. . . = rank ([I − E]) . .. . 0 0 yrT Since an idempotent matrix is diagonalizable, there exists an invertible matrix U such that 1 0 · · · · · · . 0 . . . . . . −1 . U EU = 0 .. . . 1 0 0 0 0 ··· ··· ··· 0 .. . .. . .. . 0 where there are exactly s 1’s ones along the diagonal as rank E = s. Now U −1 [I − E] U = I − U −1 EU. Hence the rank(J − A11) = r − s. Thus rank(J ) = r = s + (r − s) = rank(A11) + rank(J − A11) . This yields A11 ≤− J and hence A ≤− B as required. Although we do not specifically state corollaries for the row, column and doubly stochastic idempotent cases, they are analogous to the corollaries given previously 41 for the group-monotone case. However, we do present the symmetric idempotent case which follows from the previous theorem. Corollary 2.2.1. Let A, B be n × n nonnegative matrices such that B 2 = B, B is symmetric, rank B = r and rank A = s . Then A ≤− B if and only if there exists a permutation matrix P such that J 0 A11 0 and P AP T = , P BP T = 0 0 0 0 where J is a direct sum of matrices xi xTi where xi is a positive unit vector with xTi xi = 1; xi and xj , i 6= j are not necessarily of the same order and A11 T x1 0 · · · 0 x1 0 · · · 0 . . . . . . 0 . . . . .. . . . . .. 0 E = . . . ... ... . ... ... . 0 . 0 0 · · · 0 xTr 0 · · · 0 xr where E is a nonnegative idempotent r × r matrix. We present the following example to demonstrate the structure described in Theorem 2.2.1. In this case, B is a doubly stochastic idempotent matrix. 42 1 0 0 0 1 1 2 2 Example 2.2.1. Let B = 0 1 1 2 2 0 0 0 0 1 0 0 and A = 0 0 1 0 0 0 0 0 0 0 . Clearly A ≤− B. 0 0 0 0 0 1 Following the notation in the theorem P = I. We may express T 0 0 x1 0 0 y1T 0 0 x1 y1 = 0 x 0 0 yT 0 T B= 0 x y 0 2 2 2 2 T T 0 0 x3 y3 0 0 x3 0 0 y3 1 0 = 0 0 1 0 0 1 0 0 1 0 2 0 0 and clearly A = 0 1 0 2 0 0 0 0 1 0 0 1 0 0 0 1 0 2 0 1 1 0 1 0 2 0 0 0 1 0 1 0 1 0 0 0 1 0 0 0 0 1 1 0 where E = 0 0 0 is an 1 0 0 0 1 0 0 1 idempotent. Recall that according to Lemma 1.2.1, A ≤− B where B is an idempotent matrix implies that A is an idempotent matrix. In Corollary 2.2.1, we characterized 43 matrices dominated by a symmetric idempotent matrix B. A natural question is whether or not A is also a symmetric idempotent matrix. In the following example we show that this need not be the case. 0 1 0 1 0 0 Example 2.2.2. Let B = 0 1 0 and A = 0 1 0 . 0 1 0 0 0 1 symmetric idempotent matrix. Clearly, B is a Now A ≤− B as rank(B − − A) = 2 = rank(B) 0 1 0 rank(A) and, furthermore, A is idempotent. But A = 0 1 0 6= 0 1 0 AT and thus A is not symmetric. 0 0 0 1 1 1 = 0 0 0 Chapter 3 Partial Order on a von Neumann Regular Ring Various partial orders on an abstract ring or on the ring of matrices over the real and complex numbers have been introduced by several authors either as an abstract study of questions in algebra, or for the study of problems in engineering and statistics (See, e.g. [1], [3], [12], [26], [32] and [38]). Also of interest are partial orders on semigroups which are studied by several authors (See, e.g. [19], [39], and [41]). In this chapter we study the well-known minus partial order on a von Neumann regular ring which is simply a generalization of a partial order on the set of idempotents in a ring introduced by Kaplansky. Recall that for any two elements a, b in a von Neumann regular ring R, we say a ≤− b (and read it as a is less than or equal 44 45 to b under the minus partial order) if there exists an x ∈ R such that ax = bx and xa = xb where axa = a. Furthermore, we define the partial order ≤⊕ by saying that a ≤⊕ b if bR = aR ⊕ (b − a)R, and call it the direct sum partial order. The Loewner partial order on the set of positive semidefinite matrices S is defined by saying that for a, b ∈ S, a ≤L b if b − a ∈ S. The direct sum partial order is shown to be equivalent to the minus partial order on a von Neumann regular ring. We also demonstrate that the minus partial order on the subset of positive semidefinite matrices in the matrix ring over the field of complex numbers implies the Loewner partial order. In addition, two questions posed by Hartwig about the minus partial order are answered. One of the main results of this chapter gives an explicit description of maximal elements in a subring under minus partial order (Theorem 3.2.1). As a special case, we obtain a result identical to the one obtained by Mitra-Puri [38] for the unique shorted operator; which, in turn, is equivalent to the formula of Anderson-Trapp ([3] Theorem 1) for computing the shorted operator of a shorted electrical circuit (Theorem 3.3.1). 46 3.1 Equivalence of Partial Orderings and Their Properties We now investigate properties of the direct sum partial order and its relation to the minus partial order. Let R be a regular ring. Recall a ≤⊕ b if and only if bR = aR ⊕ (b − a)R. By Theorem 1.3.1 this is equivalent to Rb = Ra ⊕ R(b − a). It is straightforward to see that ≤⊕ is a partial order. Next, we show that the minus partial order is equivalent to the direct sum partial order on a regular ring. Theorem 3.1.1. Let R be a regular ring and a, b ∈ R. Then the following are equivalent: 1. a ≤⊕ b; 2. a ≤− b; 3. {b(1)} ⊆ {a(1)}. Proof. 1 =⇒ 2 : As a ≤⊕ b, bR = aR ⊕ (b − a)R. Hence, a ∈ bR and thus a = bx for some x ∈ R. any g ∈ {b(1)}, bgb = b. It follows that aR ⊆ bR. As R is a regular ring, for Thus bga = bg(bx) = (bgb)x = bx = a. Now aga = 47 bga − (b − a)ga = a − (b − a)ga. Thus a − aga = (b − a)ga. But aR ∩ (b − a)R = (0) and a − aga = (b − a)ga ∈ aR ∩ (b − a)R. Hence a − aga = 0 and (b − a)ga = 0. Therefore aga = a = bga and hence {b(1)} ⊆ {a(1)}. Indeed, this demonstrates that 1 =⇒ 3. Now choose x = gag. Then axa = a(gag)a = aga = a and x ∈ {a(1)}. Now bx = (bga)g = ag as bga = a. Furthermore, ax = agag = ag as aga = a. Thus ax = bx. Now bg(b − a) = bgb − bga = (b − a) and (b − a)g(b − a) = bg(b − a) − ag(b− a) = (b−a)−ag(b−a). Hence ag(b−a) = (b−a)−(b−a)g(b−a) ∈ aR∩(b−a)R = (0). Thus (b − a) = (b − a)g(b − a) and ag(b − a) = 0. It follows that agb = aga = a. Now xb = (gag)b = g(agb) = ga and xa = gaga = ga. Therefore xb = xa. Thus ax = bx and xa = xb for some x ∈ {a(1)} and it follows that a ≤− b. 2 =⇒ 3 : This is well-known. It is proven here for completeness. As a ≤− b, there exists some x ∈ {a(1)} such that ax = bx and xa = xb. It follows that a = axa = bxa = axb and for any y ∈ {b(1)}, aya = (axb)y(bxa) = ax(byb)xa = axbxa = (axb)xa = axa = a. Thus {b(1)} ⊆ {a(1)}. 3 =⇒ 1 : Given that {b(1)} ⊆ {a(1)}, ab(1)a = a for any b(1) ∈ {b(1)}. By Lemma 1.3.1, {b(1)} = g + (1 − gb)R + R(1 − bg) for g ∈ {b(1)}. For each x ∈ {b(1)} there exists some r1, r2 ∈ R such that x = g + (1 − gb)r1 + r2 (1 − bg). Multiplying on the left and right by a yields axa = a [g + (1 − gb)r1 + r2 (1 − bg)] a. Hence a = axa = a [g + (1 − gb)r1 + r2 (1 − bg)] a = aga + a(1 − gb)r1 a + ar2(1 − bg)a = 48 a + a(1 − gb)r1a + ar2 (1 − bg)a. Thus a(1 − gb)r1 a + ar2 (1 − bg)a = 0. As a(1 − gb)r1a + ar2(1 − bg)a = 0 holds for all r1 and r2 , we can take, in particular, r2 = 0 which gives a(1−gb)r1 a = 0 for all r1 and hence a(1−gb)Ra = (0). Similarly, by taking r1 = 0, we conclude aR(1 − bg)a = (0). As r1 , r2 were arbitrary, it follows that a(1 − gb)Ra = (0) = aR(1 − bg)a. Now (a(1 − gb)R)2 = (a(1 − gb)R) (a(1 − gb)R) = (a(1 − gb)Ra) ((1 − gb)R) = (0) ((1 − gb)R) = (0). Similarly (R(1 − bg)a)2 = (0). ring, it has no nonzero nilpotent left or right ideal. Since R is a regular Thus, a(1 − gb)R = (0) and R(1 − bg)a = (0). As 1 ∈ R, a(1−gb) = 0 and (1−bg)a = 0. Therefore, bga = a = agb. Now for any t1, t2 ∈ R, at1 = (bga)t1 = b(gat1) ∈ bR and (b − a)t2 = bt2 − at2 = bt2 − (bga) t2 = b(t2 − gat2) ∈ bR. Hence, aR + (b − a)R ⊆ bR . Thus aR + (b − a)R = bR. Now we want to show that aR ∩ (b − a)R = (0). For some u, v ∈ R, suppose au = (b − a)v ∈ aR ∩ (b − a)R. Then au = agau = ag(b − a)v = agbv − agav = av − av = 0 as a = agb. Thus aR ∩ (b − a)R = (0) and so bR = aR ⊕ (b − a)R. Hence, a ≤⊕ b as required. Remark 3.1.1. It is stated in [35] that Hartwig-Luh have demonstrated that, when R is a regular ring, 2 is equivalent to 3 with the additional hypothesis that a ∈ bRb. We also note that proving directly 2 =⇒ 1 requires a brief argument. 49 The Corollary that follows shows, in particular, that the minus partial order defined on the set of idempotents is the same as the partial order defined by Kaplansky on idempotents (See e.g. Lam [29], page 323). Corollary 3.1.1. Let R be a regular ring and a, b ∈ R such that b = b2. Then the following are equivalent: 1. a ≤− b; 2. a = a2 = ab = ba. Proof. 1 =⇒ 2 : If a ≤⊕ b then bR = aR ⊕ (b − a)R and Rb = Ra ⊕ R(b − a). It follows that aR ⊆ bR and Ra ⊆ Rb. Hence a = bx and a = yb for some x, y ∈ R. So ab = (yb)b = yb2 = yb = a as b is idempotent. Similarly, ba = b(bx) = b2x = bx = a as b is idempotent. Now a − a2 = a(1 − a) ∈ aR and a − a2 = ba − a2 = (b − a)a ∈ (b − a)R. But aR ∩ (b − a)R = (0). Therefore a − a2 = 0 and a = a2. Hence, a = a2 = ab = ba. 2 =⇒ 1 : Conversely, for any x ∈ R, as b = a + (b − a), it follows that bx = [a + (b − a)] x = ax + (b − a)x . So bR ⊆ aR + (b − a)R. Now for any r1 ∈ R, ar1 = (ba)r1 = b(ar1 ) ∈ bR. Also, for any r2 ∈ R, (b−a)r2 = br2 −ar2 = br2 −bar2 = b(r2 − ar2) ∈ bR. As right ideals are closed under addition, aR + (b − a)R ⊆ bR. Hence bR = aR + (b − a)R. 50 Now we want to show that aR ∩ (b − a)R = (0). Suppose au = (b − a)v ∈ aR ∩ (b − a)R for u, v ∈ R. Multiplying on the left by a yields a2u = abv − a2v. By assumption, a = a2 = ab, so au = a2u = abv − a2v = av − av = 0. Thus aR ∩ (b − a)R = (0) and a ≤⊕ b. The proposition that follows shows that, under a certain condition, in a subring S of a regular ring R, the minus partial order on R implies the direct sum partial order on S. Proposition 3.1.1. Let S be a subring of a regular ring R such that bR ∩ S = bS for some a, b ∈ S. Then a ≤− b on R implies a ≤⊕ b on S. In the case where b is an idempotent, bR ∩ S = bS if and only if a ≤⊕ b on S. Proof. Suppose that a, b ∈ S and bR∩S = bS. By Theorem 3.1.1, a ≤− b on R if and only if bR = aR ⊕ (b − a)R. We have, bS ⊂ aS ⊕ (b − a)S ⊆ aR ∩ S ⊕ (b − a)R ∩ S ⊆ (aR ⊕ (b − a)R) ∩ S = bR ∩ S = bS. Thus aS ⊕ (b − a)S = bS and so a ≤⊕ b. For the second part of the proposition, we just need to show the “if” part. Let a, b ∈ S and suppose a ≤⊕ b on S. Clearly, bS ⊆ bR ∩ S. For any x ∈ R, if bx ∈ S, then bx = bbx = b(bx) ∈ bS . Thus bR ∩ S = bS. Next, we show that the condition bR ∩ S = bS is not necessary in the first part of the proposition. 51 Example 3.1.1. Let R be the ring of 3 ×3 matrices and S be the subring consisting 0 1 0 0 1 0 and a = 0 0 0. of upper triangular matrices. Let b = 0 0 1 0 0 0 0 0 0 Clearly bR ∩ S 6= bS, a ≤− b on R and a ≤⊕ b on S as well. Corollary 3.1.2. Let S be a subring of a regular ring R such that a, b ∈ S and a ≤− b on R where b is idempotent. Then a ≤⊕ b on S. Proof. We always have bS ⊆ aS ⊕ (b − a)S. We claim that for any s1 , s2 ∈ S there exists s3 such that as1 + (b − a)s2 = bs3. Now as1 + (b − a)s2 = bs3 = bbs3 = b [as1 + (b − a)s2] ∈ bS. Thus aS ⊕ (b − a)S ⊆ bS and hence a ≤⊕ b on S. Remark 3.1.2. One can choose the subring S in the above proposition to be any right (left) non-singular ring as such a ring is always embeddable in a right (left) maximal quotient ring which is a regular ring (Lam [30], page 376, 13.36). In [19], Hartwig posed the following two questions, among others: (1) If R is a regular ring and aR ∩ cR = (0) = Ra ∩ Rc, does there exist a(1) such that a(1)c = 0 = ca(1)? (2) Does a ≤− c, b ≤− c, aR ∩ cR = (0) = Ra ∩ Rc imply a + b ≤− c? Below, we answer Question 1 in the affirmative and Question 2 in the negative by providing a counterexample. We do not know whether or not someone has 52 answered these questions, as we could not find this in the literature. In any case, we believe that the answers we have given would be of interest to the reader. It is first necessary to prove a lemma. Note that, in a regular ring, if we let b = a + c then, a ≤− a + c if and only if a ≤⊕ a + c if and only if (a + c)R = aR ⊕ cR. Lemma 3.1.1. Let R be a regular ring and let a, b, c ∈ R with b = a + c. Then the following statements are equivalent: 1. a ≤− b; 2. aR ∩ cR = (0) = Ra ∩ Rc. Proof. It follows immediately from Theorem 3.1.1 that a ≤− b implies aR ∩ cR = (0) = Ra ∩ Rc. Conversely, for any g ∈ {b(1)}, bgb = b. Now a + c = b = bgb = bg(a + c) = bga + bgc and consequently a − bga = bgc − c ∈ Ra ∩ Rc. But by assumption Ra∩Rc = (0). Thus a−bga = 0 and bgc−c = 0. It follows that a = bga and c = bgc. So a = bga = (a + c) ga = aga + cga which gives us a − aga = cga ∈ aR ∩ cR = (0). Thus a − aga = 0 and a = aga for any g ∈ {b(1)}. Hence, {b(1)} ⊆ {a(1)} and by the previous theorem, a ≤− b. Proposition 3.1.2. (Hartwig Question 1) If R is a regular ring and aR ∩ cR = (0) = Ra ∩ Rc, for some nonzero elements a, c ∈ R, then there exists a nonzero a(1) such that a(1)c = 0 = ca(1). 53 Proof. Let b = a+c. By the previous lemma, a ≤− b. Then, by the definition of the minus partial order, for some a(1), aa(1) = ba(1) and a(1)a = a(1)b. Now substituting b = a + c yields aa(1) = (a + c)a(1) and a(1)a = a(1)(a + c). Thus aa(1) = aa(1) + ca(1) and a(1)a = a(1)a + a(1)c. It follows that ca(1) = 0 = a(1)c as required. Example 3.1.2. (Hartwig 2) Question 0 0 Let a = 0 0 0 1 0 0 0 0 0 0 , b = 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 and c = 0 0 0 0 0 0 0 0 and b ≤− c. We show that a + b − c. Now, F F F F 0 0 0 0 , aR = 0 0 0 0 0 0 0 0 0 0 0 0 F F F F , bR = 0 0 0 0 0 0 0 0 0 1 1 0 0 1 . Then a ≤− c 0 0 0 0 0 0 54 0 0 Ra = 0 0 0 0 Rb = 0 0 So aR ∩ bR = (0) = Ra ∩ Rb. 0 0 Next a + b = 0 0 0 F 0 0 F 0 and 0 F 0 0 F 0 0 0 F 0 0 F . 0 0 F 0 0 F 0 1 0 0 0 0 0 1 , and a + b = 0 0 0 0 0 0 0 0 0 1 0 0 0 1 − 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 1 = 0 0 0 0 0 0 c because rank(c) − rank(a + b) = 2 − 2 = 0 and rank(c − (a + b)) = rank 0 0 0 0 0 0 1 0 0 0 = 1. 0 0 0 0 0 0 The following result is contained in (Lemma 1, [39]) where the author proves the equivalence of 11 statements. However, for the sake of completeness, a direct argument is provided. 55 Lemma 3.1.2. Suppose R is a regular ring and a, b ∈ R such that {a(1)}∩{b(1)} = 6 ∅. Then the following are equivalent: 1. aR ⊂ bR and Ra ⊂ Rb; 2. a ≤⊕ b. Proof. Suppose aR ⊂ bR and Ra ⊂ Rb. It follows that a = rb = bs for some r, s ∈ R. We claim that ab(1)a is invariant under any choice of b(1). Let x, y ∈ {b(1)} be arbitrary. Now axa = (rb)x(bs) = r(bxb)s = rbs as bxb = b. Similarly, aya = (rb)y(bs) = r(byb)s = rbs as byb = b. Thus axa = aya for every x, y ∈ {b(1)}. Hence ab(1)a is invariant under any choice of b(1). By assumption, {a(1)} ∩{b(1)} = 6 ∅ so there exists some g ∈ {a(1)} ∩ {b(1)}. Therefore ab(1)a = aga = a for all b(1). Hence {b(1)} ⊆ {a(1)} and by Theorem 3.1.1, a ≤⊕ b. Conversely, if a ≤⊕ b, then aR ⊂ bR and Ra ⊂ Rb follow by definition. We now demonstrate an important relationship between {2}-inverses and {1,2}inverses under the direct sum partial order. Lemma 3.1.3. Let a ∈ R where R is a regular ring. equivalent: 1. b is a {2}-inverse of a; 2. There exists a {1,2}-inverse c of a such that b ≤⊕ c. Then the following are 56 Proof. Suppose b is a {2}-inverse of a for a ∈ R. a(1)(a − aba)a(1) and c = b + u. For any fixed a(1), define u = Then aca = aba + aua = aba + aa(1)aa(1)a − aa(1)abaa(1)a = aba+a−aba = a and cac = (b+u)a(b+u) = bab+bau+uab+uau = b + ba(a(1)aa(1) − a(1)abaa(1)) + (a(1)aa(1) − a(1)abaa(1))ab+ (a(1)aa(1) − a(1)abaa(1))a(a(1)aa(1) − a(1)abaa(1)) = b + baa(1) − baa(1) + a(1)ab − a(1)ab + a(1)aa(1) − a(1)abaa(1) − a(1)abaa(1) + a(1)abaa(1) = b + a(1)(a − aba)a(1) = b + u = c. This shows that c is a {1,2}-inverse of a. Now we want to show that b ≤⊕ c. In other words, we will prove that bR⊕uR = cR. Observe that cab = [b + a(1)(a − aba)a(1)]ab = bab + a(1)(ab − abab) = bab. Therefore b ∈ cR. As c = b + u, it is clear that cR ⊆ bR + uR. and b ∈ cR, uR ⊆ cR. It follows that cR = bR + uR. As u = c − b Now we want to show that bR ∩ uR = (0). Let bp = uq ∈ bR ∩ uR for some p, q ∈ R. Multiplying ba on both sides yields bp = babp = bauq = ba[a(1)(a − aba)a(1)]q = (ba − baba)a(1)q = (ba − ba)a(1)q = 0. Therefore bR ∩ uR = 0. Thus bR ⊕ uR = cR and we have demonstrated that b ≤⊕ c. Conversely, suppose that there exists a {1,2}-inverse c of a such that b ≤⊕ c. As c is a {2}-inverse of a, cac = c and thus a ∈ {c(1)}. By assumption b ≤⊕ c and it follows from Theorem 3.1.1 that {c(1)} ⊆ {b(1)}. Thus a ∈ {c(1)} ⊆ {b(1)} and it follows that bab = b. Hence b is a {2}-inverse of a. 57 Lemma 3.1.4. Suppose R is a regular ring. Let y be a {2}-inverse and z be a {1,2}-inverse of an element α in the subring fRe such that y ≤⊕ z. Then eyf ≤⊕ ezf. Proof. Let α = fxe ∈ fRe. Since y ≤⊕ z, yR ⊆ zR and Ry ⊆ Rz. Thus, y = rz = zs for some r, s ∈ R. It is straightforward to verify that zαy = y = yαz. This gives (ezf)x(eyf) = (ezf)x(e(zs)f) = ez(fxe)zsf = ezsf = eyf. Similarly (eyf)x(ezf) = eyf. Thus (eyf)R ⊆ (ezf)R and R(eyf) ⊆ R(ezf). As α = fxe is a common {1}-inverse of y and z, it follows that (eyf)x(eyf) = eyf and (ezf)x(ezf) = ezf and so x is a common {1}-inverse of eyf and ezf. By Lemma 3.1.2, eyf ≤⊕ ezf . 3.2 A Characterization of the Maximal Elements Let R be a regular ring and S be a subset of R. We define a maximal element in C = {x ∈ S : x ≤⊕ a} as an element b 6= a such that b ≤⊕ a and if b ≤⊕ c ≤⊕ a then c = b or c = a. For fixed elements a, b, c ∈ R, we give a complete description of the maximal elements in the subring S = eRf, where e and f are idempotents given by eR = aR ∩ cR and Rf = Ra ∩ Rb. Here, C = {s ∈ eRf : s ≤⊕ a}. Before we prove our main result, we give two key lemmas. We assume through- 58 out that a ∈ / S. Lemma 3.2.1. Let R be a regular ring. Then d ∈ C is a maximal element in C ′ ′ ′ if and only if for any d ≤⊕ a such that dR ⊆ d R ⊆ eR, Rd ⊆ Rd ⊆ Rf, we have ′ d=d. ′ ′ Proof. Let d be a maximal element in C. If d is any element in R such that d ≤⊕ a ′ ′ ′ ′ ′ and dR ⊆ d R ⊆ eR, Rd ⊆ Rd ⊆ Rf, then clearly d ∈ eRf. As d ≤⊕ a, d ∈ C. ′ ′ Then {a(1)} ⊆ {d(1) } ∩ {(d )(1)}. Hence, d ≤⊕ d by Lemma 3.1.2. Then by the ′ maximality of d in C, d = d . The converse is obvious. Lemma 3.2.2. C = {euf : u is {2}-inverse of fa(1) e}. Proof. Let s = etf ∈ C for some t ∈ R. Then s ≤⊕ a. {a(1)} ⊆ {s(1)}. By Theorem 3.1.1, Therefore, we have (etf)a(1)(etf) = (etf). In other words, (etf)(fa(1)e)(etf) = (etf), proving that s = etf is a {2}-inverse of fa(1)e. This shows that s = euf for some {2}-inverse u of fa(1)e. Conversely, consider any u ∈ (fa(1)e)(2) and let x = euf. that x ≤⊕ a. u ∈ (fa(1)e)(2). x = euf ∈ C. We want to show Now xa(1)x = (euf) a(1) (euf) = eu fa(1) e uf = euf = x as Hence {a(1)} ⊆ {x(1)}. By Theorem 3.1.1, x ≤⊕ a and so 59 Theorem 3.2.1. max C = {evf : v is a {1,2}-inverse of fa(1)e}. Proof. Suppose x = euf ∈ C where u = fa(1)e (2) . By Lemma 3.1.3, there is a {1,2}-inverse v ∈ eRf of fa(1)e such that euf ≤⊕ evf. Note that evf ≤⊕ a. Thus max C ⊆ {evf : v is a {1,2}-inverse of fa(1)e} unless evf = a, for every choice of v, but this cannot happen because by hypothesis a ∈ / S. Now suppose evf, ev ′f ∈ C such that v, v ′ are {1,2}-inverses of fa(1)e and evf ≤⊕ ev ′f. Therefore ev ′fR = evfR⊕(ev ′f −evf)R. Now we want to show that ev ′fR = evfR. As evf, ev ′f ∈ C, evf ≤⊕ a and ev ′f ≤⊕ a. (1) and {a(1)} ⊆ {(ev ′f) }. Thus {a(1)} ⊆ {(evf)(1)} So let a(1) be a common {1}-inverse of evf and ev ′f. By assumption evfR ⊆ ev ′fR. As shown in Lemma 3.1.4, (ev ′f) a(1) (evf) = evf and (ev ′f) a(1) (ev ′f) = (ev ′f). Now (ev ′f) R = ev ′fa(1)R = ev ′fa(1) eR = ev ′(fa(1)evfa(1)e)R ⊆ ev ′fa(1)evfR = evfR ⊆ ev ′fR. Thus ev ′fR = evfR. Similarly we can show that Rev ′f = Revf. As Rev ′ f = Revf, we claim that ev ′f = evf. Let ev ′f = revf for some r ∈ R. Now evf = ev ′fa(1)evf = (revf)a(1)evf = r(evf) = ev ′f. Hence max C = {evf : v is a {1,2}-inverse of fa(1)e}. Thus evf = ev ′f. 60 We now provide an example to illustrate the previous theorem. Example 3.2.1. Note that we are choosing f to be of rank two. Soany maximal element will have, at most, rank two. 1 1 2 4 0 1 1 0 2 0 0 0 0 0 0 1 0 0 0 0 0 . 0 1 1 0 Suppose a = 0 0 1 0 0 0 1 1 2 2 Choose e = 1 1 0 2 2 0 0 0 0 0 0 1 0 0 . 0 1 0 0 0 1 0 0 and f = 0 1 Then one choice for a(1) is a(1) = 1 1 1 0 0 0 2 8 8 0 1 1 1 0 1 0 0 4 4 and fa(1)e = . 0 0 0 0 0 1 0 0 0 1 0 0 0 1 For our choice of a {1, 2}-inverse of fa(1)e, we first choose its Moore-Penrose inverse and later its {1, 2, 5}-inverse, known as the group inverse, as both arealso {1, 2}-inverses. Let v1be the Moore 16 45 4 45 (1) Penrose inverse of fa e. Then v1 = 4 45 0 − 32 45 8 45 8 45 0 8 4 0 0 9 9 2 1 0 0 and ev1f = 9 9 2 1 0 0 9 9 0 1 0 0 0 0 0 0 . 0 0 0 1 Now ev1f ≤ a because rank(a − ev1f) = 2 = 4 − 2 = rank (a) − rank(ev1f). Thus 61 ev1f ∈max C. Wenow find another element of max C. 8 2 2 9 9 9 16 4 4 9 9 9 v2 = 0 0 0 0 0 0 0 0 . 0 1 2 1 3 3 2 1 3 3 Then ev2f = 2 1 3 3 0 0 Thegroup-inverse v2 of fa(1)e is 0 0 0 0 . 0 0 0 1 Now ev2f ≤− a because rank(a − ev2f) = 2 = 4 − 2 = rank (a) − rank(ev2f). Thus ev2f ∈max C. 3.3 Applications In this section, as an application of our main theorem on maximal elements, we derive the unique shorted operator aS of Anderson-Trapp [3] that was also studied by Mitra-Puri (See Theorem 2.1, [38]). Throughout this section R will denote the ring of n × n matrices over the field of complex numbers, C. For any matrix or vector u, u∗ will denote the conjugate transpose of u. In this section S will denote the set of positive semidefinite matrices. Recall that any element a ∈ R has a unique Moore-Penrose inverse, denoted a†. For w ∈ S and b ∈ R, x is the unique w -weighted Moore-Penrose inverse of b if x is {1, 2} -inverse of b and satisfies (3)w (wbx)∗ = wbx and (4)w (wxb)∗ = wxb. Recall, the Loewner order, ≤L , on the set S of positive semidefinite matrices in 62 R is defined as follows: for a, b ∈ S, a ≤L b if b − a ∈ S. Suppose a ∈ S and c ∈ R. As in the previous section, eR = aR∩cR, e = e2, and choose f = e∗. Clearly, f ∈ Ra because a is hermitian. Let CL = {s ∈ eRf ∩ S : s ≤L a} = {s ∈ eSf : s ≤L a}. Under this terminology, the set C in the previous section will become, C = {s ∈ eSf : s ≤⊕ a}. We assume that rank(e) 6= rank(a), equivalently, a ∈ / eSf as shown in the remark below. Remark 3.3.1. rank(e) = rank(a) if and only if a ∈ eSf. Proof. Suppose rank(e) = rank(a). So eR = aR as eR ⊆ aR. Then a = ex for some x ∈ R and by taking conjugates, a = x∗ e∗, i.e., a ∈ Re∗ . Hence, a ∈ eRe∗. As a ∈ S, a ∈ S ∩ eRe∗ = eSe∗. For if exe∗ ∈ S then exe∗ = e (exe∗) e∗ ∈ eSe∗ and so S ∩ eRe∗ ⊆ eSe∗. The reverse inclusion is obvious. Conversely, suppose a ∈ eSf. As eR = aR ∩ cR, we have e = ax and so rank(e) ≤ rank(a). As a ∈ eSf, a = ese∗ for some s ∈ S. Therefore rank(a) ≤ rank(e). Hence, rank(e) = rank(a). Lemma 3.3.1. Suppose a, b ∈ S. If a ≤⊕ b then a ≤L b. Proof. Suppose a ≤⊕ b. Equivalently, (b − a) ≤⊕ b and by Theorem 3.1.1 we know that {b(1)} ⊆ {(b − a)(1)}. Thus, b† is a {1}-inverse of (b − a). From [31], as b is 63 positive semidefinite, b† is positive semidefinite. Thus b−a = (b − a) b† (b − a) ≥L 0. Hence (b − a) ∈ S and a ≤L b. Theorem 3.3.1. Let a ∈ S and let fa† be the a-weighted Moore-Penrose inverse of f. Then max C = max CL = {afa† f}. Proof. By Theorem 3.2.1, max C = {evf : v is a {1,2}-inverse of fa(1) e}. By assumption, e ∈ aR and so e = ax for some x ∈ R. By taking conjugates, e∗ = x∗a as a ∈ S. In addition, as f ∈ Ra, f = ya for some y ∈ R. This yields that fa(1)e = yaa(1)ax = yax and thus fa(1)e is independent of the choice of a(1). We may then choose the Moore-Penrose inverse a† for a(1). Next, we want to show that a {1, 2} − inverse of fa† e is also unique. Note that fa† e = e∗ a†e is positive semidefinite, as the Moore-Penrose inverse of a positive semidefinite element is positive semidefinite [31]. As a ∈ S, we can write a = zz ∗ for some z ∈ R. Now fR = yaR = yaa†aR = fa†aR = fa† R = fzz ∗ R = fzR = (fz) (fz)∗ R = fzz ∗ f ∗ R = fa† eR. Similarly Re = Rfa† e. It follows that f = fa† ep and e = qfa†e for some p, q ∈ R. Consider an element evf ∈ max C. Then evf = qfa† evfa†ep = qfa†ep, showing that evf is independent of the choice of {1,2}-inverse v of fa† e. Thus max C is a singleton set consisting of the element † † † e fa†e f. Since a ∈ S, a† ∈ S and hence e e∗a† e f = e fa† e f ∈ S. Next, we proceed to show that max C = {afa†f} also. Recall that afa† f is 64 hermitian and so afa† f = afa†f ∗ = f ∗ fa† ∗ idempotent, we get afa† f = a(fa†f)(fa† f) = fa† f We now prove that afa† f ≤⊕ a. a∗ = ∗ fa† f ∗ a. Since fa† f is an a(fa† f) and thus afa† f ∈ S. Let a(1) be an arbitrary {1}-inverse of a. Then afa†f a(1) afa†f = (afa†)(ya)a(1) afa† f = (afa† y)aa(1)a(fa† f) = afa† yafa†f = afa† ffa† f = afa† f. Hence {a(1)} ⊆ { afa† f (1) }. Consequently, by Theorem 3.1.1, afa† f ≤⊕ a which gives afa† f ∈ C. Furthermore, by Lemma 3.3.1, afa†f ≤⊕ a gives afa† f ≤L a and hence afa†f ∈ CL . Finally, we show that for every d ∈ CL , d ≤L afa† f. As d ∈ S ⊆ Rf, write d = uf for some u ∈ R. Then dfa† f = uffa† f = uf = d = (fa† f)∗ d fa† f as d is hermitian. Now consider afa† f − d = fa† f ∗ ∗ ∗ a fa† f − fa† f d fa† f = fa† f (a − d) fa† f , which is positive semidefinite and thus afa†f − d ∈ S. Hence d ≤L afa† f. Thus afa† f is the unique maximal element in CL provided afa† f 6= a. We have shown above that afa† f ∈ CL and thus afa† f ∈ eSf. But by assumption a ∈ / eSf . So afa†f 6= a. Therefore, afa†f is unique maximal element in CL and it also belongs to C as we have already proven that afa†f ≤⊕ a. † Now, because e fa†e f is the unique maximal element in C and afa† f ∈ C, † † † afa† f ≤⊕ e fa†e f . By Lemma 3.3.1, afa†f ≤L e fa†e f as e fa† e f ∈ CL . It has been shown above that for every element d ∈ CL , d ≤L afa† f and thus afa† f † = e fa† e f. Hence, max C = max CL = {afa† f} as desired. The following examples demonstrate the result proven in the previous theorem, 65 † i.e. afa† f = e fa†e f and so max C = max CL = {afa† f}. Furthermore, max C agrees with the formula given by Anderson-Trapp for computing the shorted operator aS when we are given the impedance matrix a. The Anderson-Trapp formula states that if a is the n×n impedance matrix then the shorted operator of a withrespect to the k-dimensional subspace S (shorting † a11 − a12a22a21 0 , where a is partitioned as a = n − k ports) is given by aS = 0 0 a11 a12 such that a11 is a k × k matrix. We show that the maximal element a21 a22 afa† f that we obtain is permutation equivalent to aS , i.e. P afa†f P T = aS for some permutation matrix P. 1 2 1 2 0 0 0 0 Example 3.3.1. Let e = 1 0 1 2 2 0 0 0 1 0 a= 1 0 0 1 0 1 0 0 . 0 1 0 0 0 1 1 2 0 0 0 and then f = e∗ = 1 0 2 1 0 1 2 0 † Then one may check that fa = f = 1 2 0 1 2 0 0 0 0 0 0 . Suppose 0 12 0 0 0 1 1 2 0 0 0 0 . 0 12 0 0 0 1 66 1 0 † So afa f = 1 0 1 4 0 1 4 0 0 1 4 0 1 0 0 0 0 . We now show that af †f = e fa† e † f. Now, a† = a 0 1 0 0 0 1 0 1 0 0 0 1 0 0 † † and fa e = 1 0 4 0 1 0 0 0 1 0 0 1 0 † Hence e fa† e f = 1 0 verify that afa† f ≤⊕ a. 1 0 0 0 . 1 0 0 1 1 0 † † Thus e fa e f = 1 0 0 1 0 0 0 0 = af † f as proven in the theorem. a 0 1 0 0 0 1 0 1 0 0 0 0 . 0 1 0 0 0 1 We may This follows from rank(a) − rank(afa†f) = 3 − 2 = 1 = rank(a − afa†f). We know then afa† f ≤L a. Thus max C = max CL = {afa†f}. We now compute operator as given by Anderson-Trapp. We partition the shorted 1 0 a as follows: a = 1 0 0 1 1 0 0 1 0 0 0 0 . 0 1 67 1 0 1 0 1 † 0 1 0 − 0 1 0 0 0 0 0 = Then aS = 1 0 1 1 0 0 0 0 0 1 0 1 0 0 . 0 1 0 0 0 0 1 0 P = 0 0 0 0 0 0 0 1 , P af † fP T = aS . a 0 1 0 1 0 0 1 0 Example 3.3.2. Let e = 0 0 1 0 a= 0 0 Now for 1 0 0 1 1 0 0 0 and then f = e∗ = 0 0 0 0 0 0 0 0 1 1 0 0 0 2 2 0 0 1 0 0 . Then f † = a 0 0 0 2 2 0 2 2 0 0 0 0 1 0 0 0 . So af † f = a 0 0 0 0 0 0 0 0 0 0 0 0 . Suppose 0 0 0 0 0 0 0 0 0 0 0 0 . Now a† = 0 0 0 0 0 0 68 1 0 0 0 1 1 0 0 0 4 4 1 1 1 0 0 † † and fa e = 4 4 0 0 1 1 0 8 8 0 18 18 0 0 1 0 † † Hence e fa e f = 0 0 0 0 0 0 . 0 0 0 0 1 0 † † Thus e fa e f = 0 0 0 0 0 0 0 0 . 0 0 0 0 0 0 0 0 0 0 0 0 = af † f. Now af † f ≤⊕ a as rank(a)−rank(af † f) = a a a 0 0 0 0 0 0 3 − 1 = 2 = rank(a − afa† f). We know then afa† f ≤L a. Thus max C = max CL = {afa† f}. We now compute operator as given by Anderson-Trapp. We partition the shorted 1 0 a as follows: a = 0 0 0 0 0 1 0 0 . 0 2 2 0 2 2 69 Then aS † 0 1 0 0 1 − 0 0 2 2 0 0 0 0 0 2 2 = 0 0 0 0 0 0 0 0 0 0 0 0 1 0 = 0 0 0 0 0 0 0 0 0 0 0 . 0 0 0 0 0 0 In this case, the permutation matrix is just the identity matrix and afa† f = aS . Chapter 4 The Parallel Sum of Two Matrices Over a Regular Ring In this chapter, our focus turns to the parallel sum. The concept of the parallel sum comes from electrical engineering. Two resistors may be wired in series or in parallel. If two resistors R1 and R2 are wired in series then the total resistance is r1 + r2 . If two resistors R1 and R2 are wired in parallel then the total resistance is r1 r2 r1 +r2 unless r1 = r2 = 0 where the parallel sum would then be 0. In [2], Anderson and Duffin generalized the scalar case to that of positive semidefinite matrices. The motivation to solve this problem arises from the computation of the parallel sum of two impedance matrices. They expressed the parallel sum as p(A, B) = A(A+B)†B where A, B are positive semidefinite matrices and (A + B)† is the Moore-Penrose 70 71 inverse of (A + B). In [37] Mitra-Puri generalized the concept of the parallel sum to arbitrary matrices over the complex numbers and introduced the more general definition of parallel summability which is where A(A + B)(1)B is invariant under the choice of {1} − inverse (A + B) (1) and p(A, B) = A(A + B)(1)B is the parallel sum of A and B. In [36], Mitra-Odell introduced the problem that is the consideration of this chapter: Let A, B be matrices of order m×n and let there exist a matrix C such that {C (1)} = {A(1)} + {B (1)}. Then A and B are parallel summable and C = P (A, B). Hartwig [17] later generalized the result of Mitra-Odell to matrices over a prime regular ring with identity. His proof closely followed that of Mitra-Odell. The goal of this chapter is to come closer to solving the problem in matrices over a commutative regular ring and, more generally, in matrices over a regular ring. 4.1 Commutative Regular Ring We first consider the case of a commutative regular ring. The following lemma is straightforward and well known. It is proven here for completeness. Lemma 4.1.1. Suppose R is a commutative ring and e, f ∈ R are idempotent. Then Re + Rf = R(e + f(1 − e)). 72 Proof. Suppose x ∈ R(e + f − fe) then x = u(e + f − fe) for some u ∈ R. So x = ue + uf − ufe = (ue − ufe) + uf = (u − uf)e + uf ∈ Re + Rf. Conversely, suppose x ∈ Re. Then x = ve = ve + vef − veef = vee + vef − vefe = ve(e + f − fe) ∈ R(e + f − fe). Thus Re ⊆ R(e + f(1 − e)) Similarly, Rf ⊆ R(e + f(1 − e)). Thus Re + Rf = R(e + f(1 − e)). Recall that every principal right ideal of a ring is uniquely generated if any two elements of the ring that generate the same principal right ideal must be right associates (if for all a, b in the ring R, aR = bR implies a = bu for some unit u ∈ R). Theorem 4.1.1. Suppose R is a commutative regular ring such that a, b, c ∈ R where {c− } = {a− } + {b− }. Then c = abu where u is a unit. Proof. By Lemma 1.3.1 and the commutativity of R we have that c− + R(1 − cc− ) = a− +R(1−aa− )+b− +R(1−bb− ). As {c− } = {a− }+{b− }, it follows that R(1−cc− ) = R(1 − aa− ) + R(1 − bb− ) = R [1 − aa− + (1 − bb− )aa− ] = R (1 − bb− aa− ) = R(1 − aba−b− ) by the previous lemma and commutativity. As R(1−cc− ) = R(1−aba− b− ), it follows that Rc = Rab. Now given that R is a commutative regular ring, we know by Theorem 1.4.2 that it is unit-regular. As R is unit regular and Rc = Rab, then c = abu where u is a unit by Theorem 1.4.3. Theorem 4.1.2. Suppose R is a commutative regular ring such that a, b, c ∈ R where {c− } = {a− } + {b− }. Then, c = ab(a + b)(1,2). 73 Proof. Suppose that {c− } = {a− } + {b− }. Multiplying both sides by c2 yields c2 (a− + b− ) = c2 c− = c. Again, multiplying both sides by a2b2 yields c2 a2b2 (a− + b− ) = ca2b2 . Thus c2 ab(a + b) = ca2b2 . Multiplying both sides by (a− b− ) gives us c2 ab(a−b− )(a + b) = ca2 b2(a− b− ). We want to show that caa− = c and cbb− = c. From Lemma 1.3.1 and {c− } = {a− } + {b− }, we have c− + (1 − c− c)R + R(1 − cc− ) = a− + (1 − a− a)R + R(1 − aa− ) + b− + (1 − b− b)R + R(1 − bb− ). Multiplying on the left and right by c yields c = ca− c + c(1 − a− a)Rc + cR(1 − aa− )c + cb− c + c(1 − b− b)Rc + cR(1 − bb− )c. As c = c (a− + b− ) = ca− c + cb− c, we have 0 = c(1 − a− a)Rc + cR(1 − aa− )c + c(1 − b− b)Rc + cR(1 − bb− )c. Now, cR(1 − aa−)c = 0, so (1 − aa− )cR(1 − aa− )c = 0 and as R has no nonzero nilpotent left or right ideals, (1 − aa− )c = 0. Thus c = caa− and similarly c = cbb− . Therefore c2 (a + b) = cab. Hence, c2 (a + b) (a− b− ) = cab(a−b− ) = c. So a {1} − inverse of c is (a + b)a−b− and so c− = a− b− (a + b). As R is commutative, we may choose c = ab(a + b)(1,2) as c = cc− c = ab(a + b)(1,2) [a− b− (a + b)] ab(a + b)(1,2) = ab(a + b)(1,2). We are unable to show that c = p(a, b), that is, invariance under {1} − inverses of (a + b). But we have shown that it is invariant under {1, 2} − inverses. This problem remains open. 74 4.2 Regular Ring We now present our results for regular rings and matrices over a regular ring. The proof of the following theorem can be given by suitably modifying the proof of Mitra-Odell [36] and Hartwig [17] but it is given here for completeness. Theorem 4.2.1. Suppose R is a regular ring such that a, b, c ∈ R where {c− } = {a− } + {b− }, bR ⊆ cR and Ra ⊆ Rc. Then a and b are parallel summable. Proof. By assumption, {c− } = {a− } + {b− }. As, c− + (1 − c− c) R + R(1 − cc− ) = a− + (1 − a− a) R + R(1 − aa− ) + b− + (1 − b− b) R + R(1 − bb− ), it follows that (1 − c− c) R + R(1 − cc− ) = (1 − a− a) R + R(1 − aa− ) + (1 − b− b) R + R(1 − bb− ). Now aR ∩ bR = eR where e is idempotent and R(1 − aa− ) + R(1 − bb− ) = R(1 − e). Similarly, Ra ∩ Rb = Rf where f is idempotent and (1 − a− a) R + (1 − b− b) R = (1 − f)R. Now, choose e = b(1 − m− m)b− , f = b− (1 − nn− )b where m = (1 − aa−)b and n = b(1 − a− a). That eR ⊆ bR and eR ⊆ aR is clear. Conversely, let y = ax1 = bx2 ∈ aR ∩ bR. So y ∈ eR. Similarly, Ra ∩ Rb = Rf. Therefore (1 − c− c) R + R(1 − cc− ) = (1 − f)R + R(1 − e). So c(1 − f)Rc + cR(1 − e)c = 0 implies that c(1 − f)Rc + c(0)(1 − e)c = c(1 − f)Rc = 0. c(1 − f)Rc(1 − f) = 0. So Hence c(1 − f) = 0 as R has no nonzero nilpotent left or right ideals. Hence, c = cf. Similarly, c(1 − f)(0)c + cR(1 − e)c = 0 and so 75 cR(1 − e)c = 0. Thus (1 − e)cR(1 − e)c = 0 and as R has no nonzero nilpotent left or right ideals, (1 − e) c = 0. So c = ec. So c = cf = ec. By assumption, bR ⊆ cR and Ra ⊆ Rc. So cR ⊆ eR = aR ∩ bR ⊆ bR ⊆ cR and it follows that eR = cR. Also, Rc ⊆ Rf = Ra ∩ Rb ⊆ Ra ⊆ Rc. Thus Rf = Rc. Thus, aR ∩ bR = eR = cR and Ra ∩ Rb = Rf = Rc. As R is regular, cR = eR implies that e = cc= for some {1} − inverse c= , perhaps different from those used previously. Likewise, f = c≡ for some other {1} − inverse c≡ . As c= = a= + b= for some {1} − inverses a= , b= and c = cf ∈ Rb, b − bm− m = b(1 − m− m)b− b = eb = cc= b = c (a= + b= ) b = ca= b + cb= b = ca= b + c = ca= b + ca= a = ca= (a + b). Also, we have bm− (1 − aa−) (a + b) = bm− m. Combining these yields [ca= + bm− (1 − aa−)] (a + b) = b(1 − m− m) + bm− m = b. Hence Rb ⊆ R(a + b). Clearly, (a + b) [a= c + (1 − aa−) n− b] = b and thus bR ⊆ (a + b) R. Similarly Ra ⊆ R(a + b) and aR ⊆ (a + b) R. Hence aR + bR = (a + b)R and Ra + Rb = R(a + b). By Theorem 1.4.1, a(a+b)−b and b(a+b)− a are invariant under choice of (a+b)(1) and a(a + b)− b = b(a + b)− a. Thus, p(a, b) = p(b, a) are equal and well-defined. It is still an open problem to show that c = p(a, b) under the conditions that bR ⊆ cR and Ra ⊆ Rc. It is known that if R is regular then so is the set Rm×n of m × n matrices over R in the sense that for C ∈ Rm×n , CXC = C always has a solution X = C − ∈ Rn×m 76 [17]. For convenience, let Γ = Rn×m . In the previous theorem, R is a regular ring. However, in the following analogous theorem, the result is stated in terms of m × n matrices over a regular ring R. Theorem 4.2.2. Let Rm×n be the set of m × n matrices over R , where R is is a regular ring. Suppose that A, B, C ∈ Rm×n where {C − } = {A− } + {B − }, BΓ ⊆ CΓ and ΓA ⊆ ΓC. Then A and B are parallel summable. Theorem 4.2.3. Suppose R is a regular ring such that e1, e2 are idempotent, 2 is invertible in R and p(e1 , e2) = p(e2 , e1). Then the harmonic mean of two idempotents is idempotent. Proof. Let e21 = e1 and e22 = e2. Now as p(e1, e2) = p(e2 , e1), we may set p = e1 (e1 + e2)(1) e2 = e2 (e1 + e2)(1) e1 where we have invariance by the definition of the parallel sum. We want to show that pR = e1R ∩ e2 R (similarly Re1 ∩ Re2 = Rp). Now pR ⊆ e1R ∩ e2R by definition of p = e1 (e1 + e2)(1) e2 = e2 (e1 + e2)(1) e1 . We need to show that e1 R ∩ e2R ⊆ pR. Let x ∈ e1R ∩ e2R. Note that (e1 + e2 ) (e1 + e2 )(1) (e1 + e2 ) = (e1 + e2), so (e1 + e2 ) (e1 + e2 )(1) (e1 + e2 ) x = (e1 + e2) x and 2 (e1 + e2 ) (e1 + e2 )(1) x = 2x. As 2 is invertible, x = (e1 + e2) (e1 + e2 )(1) x = e1 (e1 + e2) (1) x + e2 (e1 + e2) (1) x. Set x = e1z1 = e2 z2 for z1, z2 ∈ R. Then x = e1 (e1 + e2 )(1) e2z2 +e2 (e1 + e2 )(1) e1z1 = pz2 +pz1 = pz1 +pz2 = p(z1 +z2) ∈ pR. Hence pR = e1R ∩ e2R. 77 2 2 h (1) i Now we will show that 2p = p. So, 2p = e1 (e1 + e2) e2 2p h i = e1 (e1 + e2 )(1) e2 (e1 + e2) p = e1 (e1 + e2 )(1) e2e1 p + e1 (e1 + e2 )(1) e2e2 p = e1 (e1 + e2)(1) e2p + e1 (e1 + e2 )(1) e2p = e1 (e1 + e2)(1) p + e2 (e1 + e2)(1) p = (e1 + e2) (e1 + e2)(1) p = p. Now, for the harmonic mean, set p = 2p = 2e1 (e1 + e2)(1) e2. So p ′ ′ 2p = p as required. ′ 2 = 4p2 = 78 OPEN QUESTIONS Question 1 Characterize nonnegative matrices A dominated by a Drazin-monotone matrix B under the minus partial order. Question 2 Given a regular ring R such that a, b, c ∈ R where {c− } = {a− } + {b− }, is it the case that a and b are parallel summable and c = p(a, b)? Bibliography [1] W. N. Anderson Jr., Shorted operators, SIAM J. Appl. Math. 20 (1971), 522525. [2] W. N. Anderson, R. J. Duffin, Series and parallel addition of matrices, J. Math Anal. Appl. 26 (1969), 576-594. [3] W. N. Anderson Jr., G. E. Trapp, Shorted operators II, SIAM J. Appl. Math. 28 (1975), 60-71. [4] R. B. Bapat, S. K. Jain, L. E. Snyder, Nonnegative idempotent matrices and minus partial order, Linear Algebra Appl. 261 (1997) 143–154. [5] A. Ben-Israel, T. N. E. Greville, Generalized Inverses: Theory and Applications, second ed., Springer-Verlag, Berlin, 2002. [6] A. Berman and R. J. Plemmons, Matrix Group Monotonicity, Proc. of the Amer. Math. Soc. 46 (3) (1974) 355-359. 79 80 [7] A. Berman and R. J. Plemmons, Nonnegative Matrices in the Mathematical Sciences, first ed., SIAM, Philadelphia, 1994. [8] P. B. Bhattacharya, S. K. Jain, and S. R. Nagpaul, Basic Abstract Algebra. New York: Cambridge Univ. Press, 1986. [9] B. Blackwood, S. K. Jain, Nonnegative group-monotone matrices and the minus partial order, Linear Algebra and its Applications, accepted for publication. [10] B. Blackwood, S. K. Jain, The parallel sum of two matrices over a regular ring, to be submitted. [11] B. Blackwood, S. K. Jain, K. M. Prasad and A. Srivastava, Shorted operators relative to a partial order on a regular ring, submitted to Communications in Algebra. [12] M. P. Drazin, Natural structures on semigroups with involution, Bull. Amer. Math. Soc. 84 (1978), 139-141. [13] P. Flor, On groups of nonnegative matrices, Compositio. Math. 21 (1969) 376– 382. [14] G. Frobenius, Über Matrizen aus nicht negativen Elementen, S.-B. Preuss. Akad. Wiss. 23 (1912), 456-477. 81 [15] L. Gillman, M. Henriksen, Some remarks about elementary divisor rings, Trans. Amer. Math. Soc. 82 (1956) 362–365. [16] K. R. Goodearl, von Neumann Regular Rings, second ed., Krieger Publishing Company, Malabar, Florida, 1991. [17] R. E. Hartwig, A remark on the characterization of the parallel sum of two matrices, Linear and Multilinear Algebra 22 (1987), no. 2, 193–197. [18] R. E. Hartwig, Block generalized inverses. Arch. Rational Mech. Anal. 61 (1976), no. 3, 197–251. [19] R. E. Hartwig, How to partially order regular elements, Math. Japon. 25 (1980) 1-13. [20] R. E. Hartwig, 1-2 inverses and the invariance of BA+ C, Linear Algebra Appl. 11 (1975), 271-275. [21] R. E. Hartwig, J. Luh, On finite regular rings, Pacific J. Math. 69 (1977), no. 1, 73–95. [22] R. E. Hartwig, J. M. Shoaf, Invariance, group inverses and parallel sums, Rev. Roumaine Math. Pures Appl. 25 (1980), no. 1, 33–42. [23] A. Herrero, A. Ramirez, N. Thome, An algorithm to check the nonnegativity of singular systems, Appl. Math. Comput. 189 (2007), no. 1, 355–365. 82 [24] S.K. Jain, Linear systems having nonnegative best approximate solutions - a survey, in: Algebra and its Applications, Lecture Notes in Pure and Applied Mathematics 91, Dekker, 1984, pp. 99–132. [25] S.K. Jain, E.K. Kwak, V.K. Goel, Decomposition of nonnegative groupmonotone matrices, Trans. Amer. Math. Soc. 257 (2) (1980) 371–385. [26] S. K. Jain, S. K. Mitra and H. J. Werner, Extensions of G-based matrix partial orders, SIAM J. Matrix Anal. Appl. 17 (1996), 834–850. [27] S. K. Jain and K. M. Prasad, Right-Left Symmetry of aR ⊕ bR = (a + b)R in Regular Rings, J. Pure and Applied Algebra, 133 (1998) 141-142. [28] S.K. Jain, J. Tynan, Nonnegative matrices A with AA# ≥ 0, Linear Algebra Appl. 379 (2004) 381-394. [29] T. Y. Lam, A First Course in Noncommutative Rings, second ed., SpringerVerlag, 2001. [30] T. Y. Lam, Lectures on Modules and Rings, first ed., Springer-Verlag, 1999. [31] T. O. Lewis and T. G. Newman, Pseudoinverses of positive semidefinite matrices, SIAM J. Appl. Math., 16 (1968) 701-703. [32] K. Loewner, Über monotone matrixfunktionen, Math. Zeitschrift 38 (1934), 177–216. 83 [33] G. Marks, A criterion for unit-regularity, Acta Math. Hungar. 111 (2006), no. 4, 311–312. [34] S.K. Mitra, Matrix partial orders through generalized inverses: Unified theory, Linear Algebra Appl. 148 (1991) 237–263. [35] S. K. Mitra, The minus partial order and shorted matrix, Linear Algebra Appl. 81 (1986), 207-236. [36] S.K. Mitra, P.L. Odell, On parallel summability of matrices, Linear Algebra Appl. 74 (1986) 239–255. [37] S. K. Mitra and M. L. Puri, On parallel sum and difference of matrices, J. Math. Anal. Appl. 44 (1973),. 92-97. [38] S. K. Mitra and M. L. Puri, Shorted operators and generalized inverses of matrices, Linear Algebra Appl. 25 (1979), 45–56. [39] H. A. Mitsch, A Natural Partial Order for Semigroups, Proc. Amer. Soc.97 (3) (1986), 384–388. [40] E. H. Moore, On the Reciprocal of the General Algebraic Matrix, Bulletin of the American Mathematical Society 26 (1920) 394-395. [41] K. S. S. Nambooripad, The Natural Partial Order on a Regular Semigroup, Proc. Edinburgh Math. Soc. 23 (1980) 249-260. 84 [42] R. Penrose, A Generalized Inverse for Matrices, Proceedings of the Cambridge Philosophical Society 51 (1955) 406-413. [43] O. Perron, Zur Theorie der Matrizen, Math. Ann. 64 (1907) 248-263. [44] R. J. Plemmons and R. E. Cline, The Generalized Inverse of a Nonnegative Matrix, Proc. of the Amer. Math. Soc. 31 (1) (1972) 46-50. [45] C.R. Rao, S.K. Mitra, Generalized Inverse of Matrices and Its Applications, first ed., Wiley, New York, 1971.