AN ABSTRACT OF THE THESIS OF VISUTDHI UPATISRINGA for the degree of DOCTOR OF PHILOSOPHY July 29, 1975 in Mathematics presented on Title: THE RELATION BETWEEN COMPLEX MATRICES OBTAINED BY COMPOSING SIMILARITY AND CONJUNCTIVITY Abstract approved: Signature redacted for privacy. C. S. Ballantine This dissertation is concerned with the problem of determining, for given two n x n complex matrices A and B, necessary and suf- ficient conditions on them so that A is with B (A is similar to a matrix which is conjunctive with B). In Chapter II, the 2 x 2 complex case is completely solved. In Chapters III and IV, the n x n complex case is investigated and is only partially solved. The n x n complex case includes the following classes of matrices: Hermitian matrices, diagonable matrices, triangular nondiagonal matrices, unitary matrices, involutory matrices, positive stable matrices and accretive matrices. We state three sample results for the n x n complex case. Sample Result 1. An n x n unitary matrix A is with an n x n nonsingular hermitian matrix B if and only if A = A = A-1 and index A = index B. Sample Result 2. An n x n idempotent -natrix A of rank r is Ave, with an n x n complex matrix B if and only if B is conjunctive with 0 [2I 01 e 1r- s a 0 On- r - s for some s < r. Sample Result 3. An n x n real diagonal matrix A with positive distinct eigenvalues is with an n x n complex matrix B if and only if B is a product of three positive definite hermitian matrices. The Relation Between Complex Matrices Obtained by Composing Similarity and Conjunctivity by Visutdhi Upatisringa A THESIS submitted to Oregon State University in partial fulfillment of the requirements for the degree of Doctor of Philosophy June 1976 APPROVED: Signature redacted for privacy. Professor of Mathematics in charge of major Signature redacted for privacy. Ch man of Department of Mathematics Signature redacted for privacy. Dean of Graduate School Date thesis is presented July 29, 1975 Typed by Ilene Anderton and Lyndalu Sikes for Visutdhi Upatisringa ACKNOWLEDGMENT I would like to express my sincere gratitude to Professor C. S. Ballantine for suggesting the problem, for his guidance and encouragement, and for his willingness to sacrifice so much of his time and energy throughout the preparation of this thesis. I would also like to thank Professor Harry E. Goheen for his initial encouragement of my studying mathematics. Finally, I would like to thank my wife, Sally, for her patience and understanding. TABLE OF CONTENTS Chapter I. II. INTRODUCTION AND PRELIMINARIES 2 x 2 COMPLEX MATRICES §2. 1. §2.2. III. Singular Cases Nonsingular Cases n x n COMPLEX MATRICES WITH REAL EIGENVALUES §3. 1. §3. 2. IV. 1±a_se 17 19 35 Diagonable Matrices 64 64 Triangular Matrices 73 n x n COMPLEX MATRICES WITH COMPLEX EIGENVALUES §4. 1. §4. 2. §4. 3. §4. 4. §4. 5. 1 Diagonable Matrices Triangular Matrices Unitary Matrices Positive Stable Matrices Accretive Matrices 77 77 83 93 96 98 BIBLIOGRAPHY 121 PAGE INDEX TO LEMMAS, THEOREMS, ETC. 123 LIST OF TABLES Page Table 2 x 2 singular complex matrices. 61 2 x 2 nonsingular complex matrices. 62 n x n complex matrices. 115 Matrices similar to triangular matrices. 119 Miscellaneous results. 120 THE RELATION BETWEEN COMPLEX MATRICES OBTAINED BY COMPOSING SIMILARITY AND CONJUNCTIVITY I. INTRODUCTION AND PRELIMINARIES In this dissertation we are investigating certain relations between similarity and conjunctivity of n x n matrices with entries in the complex field C. Our problem is to find necessary and sufficient conditions on two n x n complex matrices S and T (or on two classes of such matrices) so that S is similar to a matrix which is conjunctive with T, or S is conjunctive with a matrix which is similar to T. In Chapter II, we deal mainly with 2 x 2 complex matrices. In Chapter III, we restrict ourselves to some n x n complex matrices with real eigenvalues. In Chapter IV, we look into some more general n x n complex matrices with complex eigenvalues. We first recall that two n x n complex matrices S and T are said to be conjunctive if there exists an n x n nonsingular matrix D such that D SD = T (where D denotes the conjugate transpose of D), and S and T are said to be similar if there exists an n x n nonsingular matrix C such that C-1SC = T. We give the following definitions for the relations of our problems. 2 Definition 1. 1 An n x n complex matrix S is said to be CAUL to an n x n complex matrix T if there exist n x n nonsingular matrices C and D such that D* SD = C-1TC. Definition 1. 2 An n x n complex matrix S is said to be -1/kle with an n x n complex matrix T if there exist n x n nonsingular * matrices C and D such that C -1 SC = DTD. We consider a few basic properties of the above relations. Lemma 1. 1 Let S and T be n x n complex matrices. Then S is Ake with T if and only if T is 1t to S. Proof: The proof follows directly from Definitions 1. 1 and 1. 2. Lemma 1. 2 Let S and T be n x n complex matrices and g be a complex number. Then (1) if S is Cdrytto T, then 6S is Gut to ET, and hence (2) if T is Aisle with S, then ET is 404e with eS, Proof: Let S and T be n x n complex matrices and fi be a complex number. Then (1) if S is C64 to T, then there exist nonsingular matrices C and D such that D SD = C - TC. CC -1 So &(D *SD) = TC) and hence D (6S)D = C1- (ET)C. Thus ES is Owl., to ET. (2) follows immediately from Lemma 1. 1. Lemma 1. 3 For every n x n complex matrix S, S is jibe with S. Proof: Trivial. Lemma 1. 4 For every n x n complex matrix T, T is °hut to T. Proof: Trivial. 3 We can easily show that the relations ).,e and eid, are not symmetric. This property is illustrated in the following example. Example 1. 1. Let S = Here S is 10 [ 2 21 and T = r 1 0 01 1' with T, since S is similar to the matrix [ 1 is conjunctive with T. However, T is not 0] 2 which0 with S, since S is not positive definite hermitian. Lemma 1. 5 Let S and T be n x n complex matrices. if S is conjunctive with T, then S is Then with, and eAzi to, T; if S is similar to T, then S is4i with, and ei,ti to, T. Proof: Let S and T be n x n complex matrices. (1) Since S is similar to itself and T is similar to itself, S islose with T, and S is eivg to T, respectively. (2) Since T is conjunctive with itself and S is conjunctive with itself, S is,/Ate with T, and S is 6.4 to T, respectively. We shall always denote the rank of the matrix S by rank S. Lemma 1. 6 If an n x n complex matrix S is .,Aie with an n x n complex matrix T, then rank S = rank T. Proof: If an n x n complex matrix S is 4 with an n x n com- plex matrix T, then there exist n x n nonsingular matrices C and D such that C-1SC = D TD. Since C and D are nonsingular, rank S = rank C-1SC = rank D TD = rank T. 4 Lemma 1. 7 If an n x n complex matrix T is to an n x n complex matrix S, then rank T = rank S. Proof: The proof follows from Lemmas 1. 1 and 1. 6. Lemma 1. 8 Let T be an n x n complex matrix with nonzero determinant and S be an n x n complex matrix. If S is with T, then sgn(det S) = sgn(det T). Proof: Let T be an n x n complex matrix with nonzero deter- minant and S be an n x n complex matrix. If S is,Afe with T, then there exist n x n nonsingular matrices C and D such that C-1SC = D TD. Since det S = det(C- 1SC) = det(D TD) = (det D D)(det T) and det (D D) >0 for every nonsingular matrix D, we see that det S is a positive multiple of det T, i. e. sgn(det S) = sgn(det T). , Lemma 1.9 Let S be an n x n complex matrix with nonzero determinant and T be an n x n complex matrix. If T is Oki to S, then sgn(det T) = sgn(det S). Proof: The proof is essentially the same as that of Lemma 1. 8. We shall follow the notation in [2] for matrix entries and for submatrices. Let S be an n x n complex matrix and let M and N be nonempty subsets of the set {1, 2, .. . , n} and let M' and N' be the subsets complementary to M and N respectively. We denote by S[ M N] the submatrix of S lying in rows whose indices come from M and in columns whose indices come from N. We denote by S.. the entry in the 1 x 1 submatrix S[il j]. We give meanings to three 5 other ways of designating submatrices: S[M S( N] S[ M' I N], and S(M N) = S[ I N' I. = S[MIN'], We abbreviate further for principal submatrices, putting S[ M] = S[M I M] and S(M) S[ ]NT]. Next we say that a nonsingular matrix D defines the D SD and that the order of this conjunctivity is conjunctivity S the order of D. Finally, we let m be the cardinal of M and let Q be an m x m nonsingular matrix. Then by "the [M] subconjunctivity of order n defined by Q" we mean the conjunctivity (of order n) defined by the n x n matrix D satisfying D[M] = Q, D[M1M) 0, DWI Mi I, where I is the identity matrix of order n-m, 0, and D(M) (We shall usually not specify the order of a subconjunctivity when it is clear from context. ) When i < j, the [1, j] subconjunctivity (of order n) defined by r 0l j L1 will be called "the interchanging [1, j] subconjunctivity (of order n). The following fact from [ 2], characterizes the effect of an arbitrary subconjunctivity. Fact 1, 1, ([ 2], Fact 1, I) Let S be an n x n matrix, M be a subset of {I, 2, . , n}, m be the cardinal of M, Q be an m x m nonsingular matrix, and T be the matrix obtained by applying to S the EM] subconjunctivity (of order n) defined by Q. Then 6 T[M] = QS[M] Q T[M1M) =Q*S[MIM) T(MIM] = S(MIM1 Q T(M) S(M) Hence any zero columns of S[M IM) and of T[MIM) correspond, and any zero rows of S(M IM] and of T(MIM] correspond, We state two more facts about triangularizing matrices by conjunctivity, Fact 1. 2. ([2], Fact 1.2) Let A, B, C, D be matrices of respective dimensions p x p, p x q, q x p, q x q, and let A be nonsingular. Then there are uniquely defined matrices C and D1 such J that the conjunctivity of order p + q defined by the matrix (in block form) I [ 0 -1 -AB , I -1 has the following effect: A C B D [ A 0 C1 D1 Fact 1.3. ([2], Fact 1.3) Let S be an n x n matrix and let r be the rank of S. Then S is conjunctive with a lower triangular matrix whose first r diagonal entries are nonzero and whose last n - r columns are zero. 7 Next, let S be an n x n complex matrix. We shall define the "conennS), to be the set of all complex numbers X SX for which X is an n x 1 complex matrix. We list some basic properties of the function Fin the following result. Fact 1.4. ([ 2], Face 3.1). (i) F(zS) = (sgn z) F(S) for every nonzero z C; (ii) ris conjunctively invariant, i, e. , [ID SD) = PIS) for every nonsingular D and every S of the same order as D; (iii) F(S1) FI(S) for every principal submatrix S1 of S; (iv) every diagonal entry of S lies in r(S); (v) r(S) consists of zero, and the numbers T11 for which T is conjunctive with S (in particular, F(S) contains all the eigenvalues of S); (vi) r(S) = r(51) + r(S2) whenever S2 (()denotes the direct sum); (vii) r(S) is a convex cone S = S1 for every S. As in [ 2], we introduce some algebraic definitions about S that will correspond to the various geometric possibilities for the cone r(S). For each n x n complex matrix S and each real 0 (0 will always be real when used in this dissertation) we define a (hermitian n x n) matrix H(O;S) by H(O;S) -iOS + e0 S*). (When S is understood, we shall sometimes write H(0) instead of H(O;S). ) We also use the well-known fact that the only kinds of cones in C are: the zero cone (the origin); half-line (closed half-line 8 with endpoint at zero); line (line through zero); sector (convex sector containing zero with apex at zero); half-plane (containing zero with boundary line through zero); and C itself, An n x n complex matrix S will be called: contrahermitian iff e -i0 S i s nonhermitian for all 0 iff r(S) has nonempty interior; cohermitian iff S is not contrahermitian iff r(S) is a subcone of some line; contradefinite iff H(e) is indefinite for all iff r(S) = C (the whole plane); codefinite iff S is not contradefinite iff r(S) is a subcone of some halfplane; unidefinite iff there is a 0 for which e -i0 S is a nonnegative definite nonzero hermi- tian matrix iff r(S) is a half-line; bidefinite iff there is a 0 for which e -i0 S is an indefinite hermitian matrix iff r(S) is a line; transdefinite iff there is exactly one value of ei0 for which H(0) is nonnegative definite and nonzero 9 iff r(S) is a half-plane; iff S is codefinite but not transdefinite prodefinite nor cohermitian iff r(S) is a sector; iff S. = 0 for all i, j = 1, 2, the zero matrix n iff PS) is the zero cone. In [ 2] the first of each of the above pairs of "iffs" is given as the definition of the underlined word, We next state some properties of the function E defined as in [ 2]. Let denote the set of n x n complex matrices all of whose real eigenvalues are positive. For S E , the function E is defined as 1 tr{[(1-t)I + tS]-1 (S - I) } dt E(S) = 0 (where 77 ,! means "imaginary part" and "tr" means "trace"). Let denote the set of n x n nonsingular complex matrices S for which -1 (if F(S), The following two facts are from [ 2], Fact 1. 5. ... 12X ([2], Fact 3. 2) Let S e , and let X ,X be the eigenvalues of S. For j = I., 2, <,, ) of arg X.. Then E(S) = cr. . j=1 n n let o-. be the prinJ cipal value (1. e., ' 10 (thus E(S) is a particular determination of arg (det S)). Hence e iE (S) = sgn (det S) Fact 1.6. ([2], Fact 3.3). E is conjunctively invariant on E(D SD) = E(S) for every n x n nonsingular D and every a,',i.e., S E . As in [ 2], let 5) denote the set of all n x n positive definite complex (hermitian) matrices. For each positive integer m let denote the set of all matrices S such that S can be written as a product of m matrices from?. We state the following two theorems of [ 2] without proof. Theorem 1. 1. ([2], Theorem 2). Let S be an n x n complex matrix. Then the following four statements are equivalent (to each other): S y--)2 (i. e., S is a product of two positive definite her-. mitian matrices); S is similar to an element of (D2 S is similar to an element of p; s is unitarily similar to a diagonable lower triangular matrix of positive diagonal. (Note: "diagonable" means "similar over the complex field to a diagonal matrix". ) 11 Theorem 1. 2 ([ 2], Theorem 3). Let She an n x n complex matrix. Then the following six statements are equivalent; SE 93(i. e. , S is a product of three positive definite hermitian matrices); S is conjunctive with an element of P3; S is conjunctive with an element of P2; S is conjunctive with a lower triangular matrix of positive diagonal; S is conjunctive with a matrix all of whose leading principal minors are positive; at least one of the following (vi. a) or (vi. b) holds: (vi. a) det S> 0 and S is contradefinite; (vi. b) S 6. and E(S) = 0 and 1 E r(s). An extremely useful result of [ a], which we shall use repeatedly, is the following lemma. Lemma 1.10. ([ 2], Lemma 4.1). (i) Let 13 be real and y be Then there is a conjunctivity having the effect nonnegative. r 1 0, L 2y 1 e r eia L 2(y2 sin2a)k for all real a such that sin a < y and (y - sin a)2 2 + COS a > 0 . 12 > 1, and for all a such that I sin al <y and i. e. , for all «when cos a> 0 when 0 < y <1. (ii) Thus, in particular, whenever 1-Tr > Ial> 161, every 2 matrix of the form = eir3 1 r eia 0 2p e-ja is conjunctive with a suitable matrix of the form io A2 = °-i6] eir3 [ eX 2 (Where p is the same in A2 as in Al), namely, with one for which IP12 + sin2 a = I12 + sin 2 8. (Thus K i0 if 12-Tr > 1a > 161.) Furthermore, " 1 a = Tr and p 0 in Ai, then Ai is contradefinite and is for every 8 conjunctive with a suitable matrix of the (iii) form A2 (namely, with one for which P12 + = lx 2 + sin2 8). 1 In applying Lemma 1. 10, we often use the following notation. Let T be an n x n nonsingular lower triangular matrix and, for j = 1, 2, n, let p- arg T... We then define arg diag T by the J3 equation arg diag T = (cr 1, 0- 2' y 0- n) We shall write (0- 1, 0-2, n) (1-1, T2' n) 13 as an abbreviation for the following conditional: "If T is any (nonsingular lower triangular) matrix whose arg diag is o- 2, ( cr (T1 ' . , T'2 CT n), then there is a matrix whose arg diag is and which is conjunctive with T. n Another useful fact mentioned in [2] is the following lemma. Lemma 1.11. For arbitrary complex numbers a, c, and d, the two matrices [a and o] [ d c o, a -I are conjunctive. Proof: We consider the following three cases: Case 1. c =0 Case 2. c 0 1 [1 0 [ ra 0 0 0 d 0 L1 1 0J = - 0 ----a rc 0I 0 0 _T [ 0 a aE C 0 C a 0 0 1 2 C c a a 0 a 0 0 1 0 --Ed ca 1 T I 1 0 1 Ca-E Case 3. cd d 0, d = 0 ac 1 a c d c-c-1 -Ed c 0 al 0 ca 0 a i 14 We define an n x n complex matrix S to be *-regular (or EP or EPr of [9]), provided SX = 0 implies S*X = 0 for n x 1 complex matrices X. We define an n x n complex matrix S to be *-irregular if it is not *-regular. We state without pi-oof the following known results. Lemma 1.12. ([8], Theorem 1. 1). Every codefinite matrix is *-regular. Lemma 1.13. ([8], Theorem 1.2). An n x n matrix S of rank r is *-regular if and only if S is unitarily similar to the direct sum of a nonsingular r x r matrix and a zero matrix. Lemma 1. 14. ([8], Lemma 6.9). If S is n x n lower triangular nonsingular contrahermitian with all diagonal entries on some line through 0 and not all on one side of 0, then S is contradefinite, in fact, S has a 2 x 2 or 3 x 3 principal submatrix which is contradefinite. Lemma 1. 15. ([8], Lemma 4. 2). Let S be n x n contra- definite nonsingular and t1, such that sgn ( t l't 2'tn) = sgn tn be any complex numbers det S. Then S is conjunctive with a lower triangular matrix T such that diag T =(ti, t2, T , tn) and n] is contradefinite. An n x n complex matrix S is said to be trapezoidal if it is lower triangular and its first r diagonal entries are nonzero (where r = rank S). Thus Fact 1. 3 says that every n x n matrix is 15 conjunctive with a trapezoidal matrix. Lemma 1. 16. ([8], Lemma 3. 10). Let S be *-irregular n x n , of rank r and let t1' t2' tr be any nonzero complex numbers. Then S is conjunctive with a matrix T such that T is trapezoidal, = 1, and Tr+1, r diag T = (t1, t2, , tr, 0, 0, , 0). Finally, we state without proof the following well-known results for reference. Theorem 1.3. ([11], pp. 100-101). Two n x n hermitian matrices are conjunctive if and only if they have the same rank and index. Theorem 1.4. ([10], pp. 84-85). An n x n complex matrix P is positive definite if and only if P = D D for some n x n nonsingular matrix D. An n x n complex matrix U is unitary if U U = I and an n x n complex matrix S is normal if SS = 55. Theorem 1.5. ([7], p. 314, p. 316). If S is an n x n hermitian (or normal) matrix, then there exists a unitary matrix U such that U -1 SU is diagonal. Theorem 1.6. ([11], p. 194). Two n x n normal matrices are unitarily similar if and only if they have the same eigenvalues. 16 Theorem I. 7. ([10], p. 67). (Schur's Triangularization Theorem). If S is an n x n complex matrix with eigenvalues X X 2, , X n, then there exists an n x n unitary matrix U such that U SU = T is a lower triangular matrix with diag T = (X 1,X 2, , X n). Theorem 1.8. ([6], p. 142; [12], p. 357). (Autonne's Lemma). Let S be an m x n complex matrix. Then there exists an m x m unitary matrix V and an n x n unitary matrix W such that v sw A [ 0 o 0' where A is a square diagonal matrix all of whose diagonal entries are positive. 17 2 x 2 COMPLEX MATRICES II. In this chapter we consider only 2 x 2 complex matrices. Here we have determined the necessary and sufficient conditions for two with, or 2 x 2 complex matrices to be 1/4-t& e#10.44, to each other. First, let S Eac loco be an arbitrary 2 x 2 complex matrix. We make the following abbreviations: det S = ad - bc A = k (ad- + We see that I. < I ' - - and we can further verify that A and .T.are invariant under unimodular conjunctivity. The ratios A: I Al are thus invariant under arbitrary conjunctivity, as are the inequalities (I. >0, > 0, A >0, <- , etc. As in 2], we list canonical forms for S under conjunctivity I and give criteria (for deciding which canonical matrix S is conjunctive with) in terms of conjunctivity invariants. the nonsingular cases of S. Here A First, we consider 0, so we can define a real number 3 (mod Tr) and a real nonnegative number y satisfying 18 e2ip 161-. 1 1 - 2y 2 = 16 I -1 Thus 13 (mod Tr) and y are conjunctivity invariants of S. Case 1. S is bidefinite. Here S is conjunctive with il3 r i 01 L0 -i-j e in which either determination may be taken for characterized by the condition: ei13. This case is 0 and S* = -e-2ip S. A Case 2. S is nonsingular but not bidefinite. Here S is conjunctive with ip e 1 2,/ 0 11 where either determination may be taken for eiP if S in contradefinite. When S is codefinite, the determination above is conjunctively invariant and satisfies Re -ip {e (a + d) } >0. Case 2 is characterized by the condition: 6 0 and S* - e-2.43S. Next we consider the singular cases. Case 3. S is singular and contradefinite. Here S is conjunctive with 19 r0 01 12 0-1 Case 3 is characterized by the condition: A = 0 and I> <0. Case 4. S is singular and codefinite and nonzero. Here S must be unidefinite and the trace of S must be nonzero. Thus we can define eip to be the signum of a + d. Then e ip is a conjunctivity invariant of S and S is conjunctive with [e0iP 001 This case is characterized by the condition: S 0 and A = 0 and C.= 0. Case 5. S = 0. §2.1. Singular Cases Theorem 2.1. Let A be the 2 x 2 zero matrix and B be a 2 x 2 complex matrix. Then B is e6/4_, to A if and only if B = 0. Proof: Trivial. Theorem 2. 2. Let B be the 2 x 2 zero matrix and A be a 2 x 2 complex matrix. Then A is A = 0. A-114.& with B if and only if 20 Proof: Trivial. Lemma 2.1. A 2 x 2 complex matrix A is similar to a positive semidefinite hermitian matrix of rank one if and only if det A = 0 and tr A >0. Proof: Suppose a 2 x 2 complex matrix A is similar to a positive semidefinite hermitian matrix H of rank one. Then there exists a 2 x 2 nonsingular matrix C such that C-1AC = H. Since - rank H = 1, det H = 0. Thus det A = det (C 'AC) = det H = 0. Because H is positive semidefinite nonzero, tr H > 0, so tr A = tr (C-1AC) = tr H > 0. Conversely, suppose A is a 2 x 2 complex matrix with detA= 0 andtrA> 0. Then the characteristic equation of A is det (X I - A) = X2 -X (tr A) = O. Thus the eigenvalues of A are trA and 0. Since these eigenvalues are distinct, there exists a nonsingular matrix C such that C-1AC = r trA 0 -0 01 -1 which is clearly a positive semidefinite hermitian matrix of rank one. Since every two 2 x 2 positive semidefinite hermitian matrices of rank one are conjunctive, we have the following theorem. 21 Theorem 2.3. Let B be a 2 x 2 positive semidefinite hermitian matrix of rank one and A be a 2 x 2 complex matrix. Then A is __,.Ziite, with B if and only if det A = 0 and tr A> 0. Proof: The proof follows from Lemma 2.1 and the above remark. A 2 x 2 codefinite complex matrix B of rank one is conjunctive (by case 4 at the beginning of this chapter) with ip e where e 1 [0 0 01 sgn (tr B) and is a conjunctivity invariant of B. Thus we have the following result. Corollary 2. 3.1. Let B be a 2 x 2 codefinite complex matrix of rank one and A be a 2 x 2 complex matrix. Then A is _Abi-t6 with B if and only if det A = 0 and sgn (tr A) = sgn (tr B). Proof: The proof follows from Theorem 2. 3 and the above remark. Theorem 2. 4. Let B be a 2 x 2 contrahermitian matrix of rank one and A be a 2 x 2 complex matrix. Then A is with B if and only if A has rank one. 22 Proof: Let B be a 2 x 2 contrahermitian matrix of rank one and A be a 2 x 2 complex matrix. ("Only if") LetA be ;Me. with B. Since B has rank one, by Lemma 1.6, A has rank one. ("If"). Suppose A has rank one. We consider two cases. Case (i): trA = 0. Here A has single eigenvalue 0 and hence is similar to 0 01 L2 0-1 r which is conjunctive (by Case 3 at the beginning of this Chapter) with B. So A is Case (ii); trA 0. _,1114.E with B. Here A has eigenvalues trA and 0. So A is similar to the matrix rtrA 0 01, L2 which is conjunctive (by Case 3 at the beginning of this Chapter) with 00 E Hence A is 2 0] with B. Lemma 2. 2. If a 2 x 2 complex matrix B is conjunctive with a matrix of rank one and trace zero, then det B = 0 and t.(B) <0. 23 Proof: Suppose a 2 x 2 complex matrix B is conjunctive with a matrix of rank one and trace zero. Thus B is conjunctive with with a2 + bc = 0 and not all of a, b, c are zero. Hence there exists a nonsingular matrix D such that b D BD = [ca -al So det (D*BD) =-a2-bc = 0, and hence det B = 0. Now, BD) = k fa (-a) + a (-a) - bb - cc} ,1/2 {2 aa + bb + cc}. Since 2 aa + bb + cc = 0 forces a = b = c = 0, contradicting our assumption on a, b and c, we must have 2 aa + bb + cc > 0. Thus (D BD) <0. Since the inequality <0 is conjunctively invariant, (B) <0, Theroem 2. 5. Let A be a 2 x 2 complex matrix of rank one and trace zero, and let B be a 2 x 2 complex matrix. Then B is efsIAL, Proof: to A if and only if det B = 0 and .T. (B) <0. Let A be a 2 x 2 complex matrix of rank one and trace zero, and let B be a 2 x 2 complex matrix. ("Only if"). Let B be efra, to A. Then there exist nonsingular matrices C and D such that D BD = -1 CAC. Since 24 -1 * rank (D*BD) = rank (C-1AC) = rank A = 1 and tr (D BD) = tr (C AC) 0, by Lemma 2. 2, det B trA 0 and (B) <0. ("If "). Suppose det B = 0 and .(B) <0. Then B is conjunctive (by Case 3 at the beginning of this Chapter) with the matrix 00 [2 01 which is similar to A. Thus B is to A. Theroem 2, 6. Let A be a 2 x 2 complex matrix of rank one and trace cm". 0, and let B be a 2 x 2 complex matrix. Then B is to A if and only if det B = 0 and either (B) <0 or sgn (trB) = sgn (trA). Proof: trace Let A be a 2 x 2 complex matrix of rank one and 0, and let B be a 2 x 2 complex matrix. ("Only if"). Let B be trA CA4/ to A. Since rank A = 1 and 0, B is conjunctive with a matrix of rank one and trace 0. Hence det B = 0 and (by conjunctivity canonical matrix at the begin- ning of this Chapter) B is conjunctive either with [ ei13 0 2 0 where 0] or with eiP [10 0] 0 = sgn (trB), If B is conjunctive with r0 01 L2 0-1 25 then 1)(B) <0; and if B is conjunctive with ip 0 1 01 ' E0 e then sgn (trB) = sgn (trA). ("If"). Suppose det B = 0 and either ts (B) <0 or sgn (trB) = sgn (trA). If (B) <0, then B is conjunctive with the canonical matrix 00 [2 01 which is again conjunctive with rtrA 0 01 L2 Since the eigenvalues of A are trA and 0, and trA rtrA L2 is similar to A. So B is Oktt, 0, 00] to A. If .T.(B) > 0 and sgn (trB) = sgn (trA), then B is conjunctive with the canonical matrix ip 1 [0 e 0 01 ' where eiP = sgn (trB), and hence B is conjunctive with trA I- 0 So B is OPZAL, to A. 0]0 26 §2. 2. Nonsingular Cases We first prove a few lemmas. 0 1 Let B = [ 2.y Lemma 2. 3. I with y > 0, and let D = [a with a, b, c, d EC. Then tr (D BD) = Proof: Let B = a D = [c D*BD 1 0 2.y 1] b = +c + (rd. + 2(aC + bU) y + b-S + with y > 0, and let with a, b, c, {-a1-3 d, E Then C. 0, [1 -az] 2-y + 2 aZ- y -I- bb + cid + 2 ba y y. Hence tr (D BD) = a-at + bE + Cc- +12 + 2 (ac + Let a, b, c, Lemma 2.4. Then (i) 1 1b12 12 + lb 12 + Id 12 > 0; c with ad bc. 0; 12 >0; la 12 1 dE C Id 12 1c12 la + c 12 +lb + d Proof: (i). Suppose 12 > 0. Ia 2 bc 1 r2 2 +1bl +1c1 +idl 2 = 0. 27 Then we must have a=b=c=d= 0, forcing ad = bc, a contradiction. (ii) and (iii) can be similarly argued as in (i). (iv). Suppose + c2 -I- lb + d I2 = O. Then we must have la +cl= lb + di= O. 1 Hence a = -c and b = -d, forcing ad = bc, a contradiction. Let y > 1. Then Lemma 2. 5. rl 0 . 1] is conjunctive with -L [ 1 0, 2y 1 Proof: o 11 1 - [ 2y rl 0] = [ -1 -1 -2y -10] is conjunctive with 01 [ -1 0 [10 -1] -1 = [2y 0, -11 Let y > 1. Then,by Lemma 1.10, r -1 L 2y o.sis conjunc -1 Lemma 2. 6. B= e ] with 1 01 2 1' Let L 2y 1with y > O. Then B is conjunctive with a matrix of zero trace if and only if , 28 Proof: Let B=e 1[3 r 0 1 2y 1] 1 with y > O. We assume without loss of generality that eiP = 1. ("Only if"). Suppose B is conjunctive with a matrix of zero trace. Then there exists a nonsingular complex matrix such that tr (D BD) = 0. Thus,by Lemma 2.3, + dra + 2 (aE + b-a) y = 0. Since by +c a-g: Lemma 2. 4 (i), a + bE + c + d> 0, we must have (aC- + ba) y <0. Since y > 0, it follows that a7 +15a <0. Now, 0= + b17. + = a-a- +b17; + = + bE + y + (1,71. + 2 (a-c- + + + (a-c- + + Ecl ) y + +-aic + -b-d) + (y -1) (aT+b-a+T.c +Ed) + (fa + (a-c- + .1a+c12+1b+d12+ (y -1) (a-C- + Since by Lemma 2.4 (iv), (y - 1) (iE a c 12 + +ac +13d) b + dl2 > 0, we must have +Ecl) <0. Since a-c- + ba +--a.c +Ed <0, it follows that y - 1 > O. Thus y > 1. ("If"). Let y > 1. Then it suffices to show that there exists a 2 x 2 nonsingular matrix 29 ab D = [c such that tr (D BD) = 0. Let 1 D = [ -(y -("Y + j2 - Then D is nonsingular since det D = 1 - (y + iy2 - D4BD = (y 1 eiP / 2 -( y + VN 1) [ - (y +j-y2 -1) 1 ip a. (N+iN [= =e 1 +A2_0 1 - (.y2 - y2 + 1 eip r0 * 0. Thus 1 01 2-y 1 -h, 1 +17-12 )][ 1)2 (Y+42-1)+ where *'s are numbers of no interest to us. Hence tr (D BD) = 0. The selection of the above nonsingular matrix D is motivated by the remark below. Since * DBD = ip r + cc' + 2 a-c- b71; + (fa + 2 ba y it suffices to find a, b, c, d E C with ad bc satisfying 30 aa + cc + 2 ac and bb =0 (1) y=0 dd--+ 2 1371 ( 2) . By Lemma 2. 4 (ii) and (iii), we have a-a- + cc > 0 and bi-; + dc-T> 0. Since N > 1, we must have aT <0 and ba <0. Thus a a_c = cc C <0 ana <u. = Let -- = a and -u = 1 with t > 0 and u > 0. Then c = -at and b = -du. Substituting b and c into equations (1) and (2), we have Ia i2 (t - 2 -y t + 1 ) = 0 = Id 2 - 2 Nu + 1). Thus t= N \i/N 2 - 1 and u= Since bc = (-du) (-at) = (tu) (ad) and ad iN 2 bc, it follows that tu Thus we can take t=u= -y + Lemma 2. 7. 2 -1 and a=d=1. Let 1 0 B = [2y i] -1. 1. 31 with 0 < y < 1 and let Q be any nonsingular 2 x 2 matrix. Then Re {tr (Q*BQ)) > 0. Hence if A is j141.1. with B, then Re {trA} > 0. Proof: Let 1 B = [2y 0 1] with 0 < y <1 and let ab Q = [c d] with ad bc. Then,by Lemma 2.3, tr (Q BQ) = + + c" + d(T1 + 2 (ac + b-c-1) We consider two cases. Case 1. y = 1. Here Re {tr(Q BQ)} =aa + =a + + cc + dd + 2 Re {a.C. + b-a} +c = (a-a: + aE + +d +(a d) + cZ) + (15E + b +6d + da) =la+c12 +1b+d1 2 By Lemma 2.4 (iv), la + c 12 + 1b + d12 > 0. Thus Re {tr (Q BQ)} > 0. + 12 +-a- c + 32 Case 2. 0< <1. Here Re {tr(Q BQ)} = a + b1-3 + c + 2 Re(a-C + ba) y +d >Hal2 +Ib2i +Ic12 +Id12 - 2 yIa-c-+15a1 > I a 12 + ibi12 + 1cl12 +Id!12 - 2 y (211allcl + 21b1 VI) lal2 +1b12 + 1c12 +1d12 >+1b12 1c12 11)71) (la Id 12 Id 12) = (1 - 10(1a 12 + 1b12 Since 1 - rc12 I > 0 and by Lemma 2. 4 (i), ibi2 lc, 1 2 Id 12 > it follows that Re {tr(Q BQ)} > 0. Hence if A is _4426 with B, then there exist nonsingular matrices C and D such that C1 AC = D BD. Thus Re {trA} = Re {tr (C-1AC)} = Re {tr (D*BD)} > 0. Lemma 2. 8. If x2 2 cosa with (sin a) (cos a) 0 2 .2a sin 1.02 33 and if z = x + iy, then sin2 a = Proof: {1 - I z2I + 1z2 - 111. Suppose with (sin a) (cos a) (sin x2 2 cos2a 2 0. a) a) sin a - 1 Then x2 - (cos2a) y2 = sin2 a cos2a. Thus (1-cos 2 a)x2 - (cos2a) y2 = (1 - cos a) a) 2 cosa So cos 4a - (1 + x2 + y2) cos2a + x = 0 . Hence 2 cos a - (1 + x2 + y2) ± 1(1 + x2 + y2)2 - 4 x 2 2 < 1. We claim that cos2a- 1+x2+y2 il+x 2 +y 2) 2 -4x2 2 For if 2 22 2 2 + y 2 +J(1 + x + y ) - 4 x 1 + x cos a2 2 5_ 1, . 34 then J (1 + x2 + y2)2 - 4x 2 < 1 - x2 - y2. Thus we would have x 2 +y 2 < 1 and (1 + x2 + y2)2 - 4 x2 < (1 - x2 - y2)2 . So 1 + 2 (x2 y2) + + (x2 + y2)2 - 4 x2 < 1-2(x2 Thus y2 < 0 which forces y = 0 and x2 < 1. cos2a - 1+ x2 +/(1 x2)2 - 4 x 2 + 2 + y2) + (x2 +y2)2. So -1 Thus we would have sin2a = 0, contradicting our assumption that (sin a) (cos a) Therefore 0. sinn a = 1 - cos 2 a = = If z 1 1 - (1 + x2 + y2) -/(1 + x2 +y2)2 - 4 x2 2 {1 - (x2 ±y 2) + j[(1 + x)2 +y 2] [ (1 - x)2 +y 2]} x + iy, then I z2 iz 12 x2 2 . 35 Since 1 + z = (1 + x) + iy 12 +z and = (1 + x)2 +y 2 1 - z = (1 - x) - iy, and 11 - z 12 = (1 - x)2 +y2 Thus J 4 y1 2 [0 x)2 y1 2 )02 /I +z jiz2 12 11 z 12 112 z 12 Therefore sin2a {1- = 1-2 z 2 + I z2 11} Let a c C\{0}. Then Lemma 2.9. Re la + a-11 >0 if and only if Re {a} > 0. Proof: Let a cC\{0}. ("Only if"). Re {a + - 1) Since 1 + Suppose Re {a+ a- 1 } > 0. Then = Re la + lai2 1> 0, } - Re {a + 2 } = (1 + lal 1T) Re {a} >0. lal it follows that Re {a} > 0. lal2 ("lf"). Suppose Re {a} > 0. Then Re {a-1} = Re { 2 }= lal 12 lal Re {a} - 1 lal 2 Re {a} >0. 36 Thus Re {a+ a-1} = Re {a} + Re {a-1} Let A be the 2 x 2 identity matrix and B be a Theorem 2. 7. 2 x 2 complex matrix. Then B is B >0 C014- to A if and only if p. Proof: Let A be the 2 x 2 identity matrix and B be a 2 x 2 complex matrix. Let B be (1Only if"). C11/24, 2 x 2 matrices C and D such that to A. Then there exist nonsingular D* BD = C-1 AC= C-1 IC= by Theorem 1. 4, B is positive definite hermitian, ("If"). Let B 1. e. , I. Thus, B (/). Then,by Theorem 1. 4, there exists a nonsingular matrix D such that B = D* D. Thus (D-1)*B (D-1) = I = A and so B is conjunctive with A. Therefore,by Lemma 1. 5, B is Oka, to A. Corollary 2. 7. 1. Let A be a 2 x 2 nonzero scalar complex matrix and B be a 2 x 2 complex matrix. Then the following three statements are equivalent. B is CA/J,Z, to Ai * AB E B is conjunctive with A. Proof: Let A be a 2 x 2 nonzero scalar complex matrix, i. e. , A = aI for a 0, and let B be a 2 x 2 complex matrix. 37 )4 (ii). Suppose B is CA/4 to A. Then,by Lemma 1. 2, elltd, to ;1A = I. Thus,by Theorem 2. 7, B a1 2r1 (Th a-1B P. Hence A B=aB= lal2 (a-1B) ra I=y_/. . (iii). Suppose A*13 Then,by Theorem 1. 4, there exists a nonsingular matrix D such that D (A B)D = I. Thus (a D) B (a D) (i). B is aI = A. So B is conjunctive with A. Suppose B is conjunctive with A. Chili Then, by Lemma 1. 5, to A. Let B be a 2 x 2 positive definite hermitian Theorem 2. 8. matrix and A be a 2 x 2 complex matrix. Then A is B if and only if A ..1frte with E9. . Proof: Let B be a 2 x 2 positive definite hermitian matrix and A be a 2 x 2 complex matrix. ("Only if"). Let A be ..1411-6 with B. Then there exist non- singular matrices C and D such that C-1AC = D BD. Since B is positive definite, by Theorem 1. 4, there exists a nonsingular matrix E such that B = EE. Thus 38 A = C (D E E D) C1 = C (C C = ) (D E E D) C1 (CC*) NEDC-1)* (EDC-1)] = (CC) Q). Since C and Q EDC-1 are nonsingular, CC and Q Q are positive definite. Hence A92. ("If"). Suppose A 6m2. Then, by Theorem 1. 1, A is similar to a positive definite hermitian matrix and hence to B. Thus by Lemma 1. 5 A is __,1/ne with B. Corollary 2. 8. 1. Let B be a 2 x 2 unidefinite nonsingular complex matrix and A be a 2 x 2 complex matrix. Then A is with B if and only if (trB ) A Let B be a 2 x 2 unidefinite nonsingular complex Proof: matrix, i. e. , B = EP for some P ET= 1. P2. Then E Pand some complex with sgn (trB) and hence * hence (trB ) B, is positive definite. = sgn (trB*). Thus &3, and Let A be a 2 x 2 complex matrix. Then A is 4kZ, with B if an only if (by Lemma 1. 2) (t.rB ) A is with (trB ) B if and only if (by Theorem 2. 8) .,1416 (tr13*) A Ep2. 39 Corollary 2. 8. 2. Let B be a 2 x 2 unidefinite nonsingular complex matrix and A be a 2 x 2 nonscalar complex matrix. Then A is ..,44.i.e. with B if and only if (i) sgn (det A) = sgn (det B), (ii) (trA) (trB*) > 0 and (iii) (trA)2 det A Proof: >4. Let B be a 2 x 2 unidefinite nonsingular complex matrix. Then B is conjunctive with ip e 0 10 11 for some real p. Thus e-iPB is conjunctive with and hence is positive definite. Suppose A is a 2 x 2 nonscalar complex matrix. Then e ("Only if"). Let A be PA is also nonscalar. with B. Since det B 0, by Lemma 1.8, sgn (det A) = sgn (det B) and (i) holds. By Lemma 1. 2, e'iPA is _Id< with e-iPB. Thus by Theorem 2.8 e PA e-iPA EP2. Since is nonscalar, tr(e-IPA) >0 and [tr(e-iPA)] 2 >4 det (jiPA) >0. Since e iP = sgn (trB) and = sgn (trB*), 0 <tr(e-iPA) = e- iPtrA (sgn trB ) (trA). Thus (trB ) (trA) > 0 and (ii) holds. Also [tr(e-iPA)]2 = e-2113(trA)2 and det (e-IPA) = e-2iPdet A. Thus 40 e-2i13(trA)2 > 4 e -2ip det A > o. So (trA)2> 4 det A and (iii) holds. ("If"). Suppose A and B satisfy the conditions (i) sgri (det B) sgn (det A), (ii) (trA) (trB*) > 0 and (iii) 2 det A >4. Since e-iPB is positive definite, det (e-1113) > 0 and from (i) det (e PA) > 0. Also from (ii) we have 0 <(trA) (sgn tr B*) = (trA) (e-i13) = tr (e-iPA). From (iii) we have e-2ip (trA)2 det (e-ipA)e-2ip det A r tr (e-ipA)12 thus [tr (e-iPA)12 > 4 det Theorem 2. 8, e-iPA is (e-iPA) > 0. ..,,Ifteze, (trA)2 det A 4 Hence e-iPA 7.--D 2. Solby with e-iPB and by Lemma 1. 2, A is 441,e, with B. Theorem 2. 9. Let B be a 2 x 2 indefinite hermitian matrix and A be a 2 x 2 complex matrix. Then A is only if det A <0 and (trA)2 > 0. with B if and 41 Proof: Let B be a 2 x 2 indefinite hermitian matrix and A be a 2 x 2 complex matrix. with B. Then there exist 2 x 2 * -1 nonsingular matrices C and D such that C AC = DBD. Since B ("Only if"). Let A be is hermitian, D BD, and hence C-IAC, is also hermitian. Thus trA = tr (C-IAC) is real and so (trA)2 > 0. Since B is 2 x 2 indefinite hermitian, the eigenvalues of B are of opposite signs. Thus det B <0 and by Lemma 1. 8, det A <0. ("IVO. Suppose det A <0 and (trA)2 > 0. Then the eigenvalues of A are real and of opposite signs. Thus A is similar to the matrix with a and b real and ab <0. Since [a0 b01 is indefinite hermitian, it is conjunctive with B. Thus A is with B. Corollary 2.9. 1. Let B be a 2 x 2 bidefinite complex matrix and A be a 2 x 2 complex matrix. Then A is only if sgn (det A) = sgn (det B) and (det B*) .A5L-Ate (trA) with B if and < 0. Proof: Let B be a 2 x 2 bidefinite complex matrix. Then B is conjunctive (by Case 1 at the beginning of this Chapter) with 42 i i(3 e 0 [0 -i1 for some real p. Thus ie -ip B is conjunctive with -1 [0 0 11 Hence ie-ir3B is indefinite hermitian and det (ie-i13B) <0. Also sgn (det B) = e2, so sgn (det B*) = e. Let A be a 2 x 2 complex matrix. ("Only if"). Let A be ,..,416 with B. Since det B 0, by Lemma 1.8, sgn (det A) = sgn (det B). By Lemma 1. 2, ie-iPA is ..Witte, with ie-iPB. Since ie-iPB is indefinite hermitian, by Theorem 2. 9, [tr (ie-iPA)]2 > 0. Thus 0 < [tr (ie-iPA)]2 So e (ie-iP)2 (trA)2 =(trA)2 -2ip (trA) 2 < 0 and hence (det B ) (trA)2 < 0. ("If"). Suppose sgn (det A) = sgn (det B) and (det B ) (trA)2 < 0. Since sgn (det B*) = e-21.13, e-2iP (trA)2 < 0. Thus 0 < [tr (ie-i1A)]2. Since det (ie-iPB) <0, det (ie by Theroem 2.9, ie-iPA is (trA)2 PA) <0. Sol with ie-iPB. By Lemma 1.2, A is All< with B. Theorem 2.10. Let B=[ 1 02y 1] 43 with y > 1 and A be a 2 x 2 nonscalar complex matrix. Then A is _A/lye with B if and only if det A > 0. Proof: Let 0, 2y 1' 1 B=[ with y > 1 and A be a 2 x 2 nonscalar complex matrix. ("Only if"). Suppose A is }1/146 with B. Since det B = 1, by Lemma 1. 8, det A > 0. ("If"). Suppose det A > 0. For convenience we denote 5 = + \Met A and consider the following two cases. Case (i): (trA)2 = 4 det A. Here the eigenvalues of A are equal. Since A is nonscalar, A is similar to 01 r1 2-y 1-1 or to 0, 1 -5[2y (for the same y as above). 1 Now 0 1.1 with (5-kI) 8[2y 1] is conjunctive 1 and -5[2y 0 11 (6L 4- 2y 1 01)16-kI) r 1 0 1 1 L 2y 1 -I 0 I is conjunctive with (821) -5[ 2y 1] 1 (5 -1/2 I) =-[ 1 0, 2y 11 44 which is again conjunctive (by Lemma 2.5) with ,1 0, L 2.y IJ Hence A is Alivie with B. 4 det A. Here the eigenvalues of A are distinct. Case (ii): (trA)2 We denote one eigenvalue of A by 5 a; then the other is 6a-1, and Thus A is similar to the matrix a-1. a _1 5 Fa L2p a _I with arbitrary p. Let a= tea with a real and t > 0. Then 0 -1 =6 [2p aa 5 2 Lt 0 2 0 1- 6 Lteia 2p rte t2 2p is conjunctive with t -l-ia e 6-1/2 t-le-ial 01 t _0 2 0 t2] Leia 0 2p Let p = iy 2 - sin2a. Then by Lemma 1, 10 there is a conjunctivity having the effect [eia 24 2 Hence A is --d/116 sin2a with B. e-ia 45 Corollary 2.10.1. Let B be a 2 x 2 nonsingular contradefinite matrix and A be a 2 x 2 nonscalar complex matrix. Then A is with B if and only if sgn (det A) = sgn (det B). Proof: Let B be a 2 x 2 nonsingular contradefinite matrix and A be a 2 x 2 nonscalar complex matrix. Then B is conjunctive with e ip [ 1 o, 2y 1 with y > 1 and p real. The proof follows routinely from Theorem 2.10 by showingthat eA is .)1/46 ,with e -ipB if and only if . det (e -'1A) > 0. So by Lemma 1. 2 A is }446 with B if and only if sgn (det A) = sgn (det B). Theorem 2. 11. Let B [ 1 0 2y 1 , with 0 < y < 1 and let A be a 2 x 2 nonscalar complex matrix. Then A is ...,41,e, with B if and only if (i) det A > 0, (ii) Re {trA} > 0 and (iii) 2y Proof: 2 >1+ (trA)2 - 1 4 det A Let B=[ with 0 < (trA)2 4 det A 1 0 2y 1] < 1 and let A be a 2 x 2 nonscalar complex matrix. 46 ("Only if"). Suppose A is }44.4e, with B. Then there exist non- singular matrices C and D such that C-1AC = D*BD. Since det B = 1, by Lemma 1.8, det A > 0 and (i) holds. Since 0 r 1 lj 2y B with 0 <y < 1, it follows from Lemma 2.7 that Re {trA} >0 and (ii) holds. Finally, to show (iii), by Schur's Triangularization Theorem, there exists a unitary matrix U such that * * U-1 (C-1AC) U = U (D BD) U = T is a lower triangular matrix whose diagonal entries are the eigenvalues of A. Thus we denote the matrix T as ia 5 2q 0 t- 1 e-ial with t > 0, 5 > 0, a real. Further, we may assume q > 0 here. Since 1)(T) {62 e2ia = 1/2 = + 62 e-Zia - 4 62q21 62 {cos 2a - 2q2} =6 2 {1 - 2 sin2a - 2 q 2 } and det T = o2, 1 2Y2 - (3) Idet 131 - 1 - 2 sina - 2 q2. 47 Thus y 2 = sina + q 2 > sin2a. Also, from (ii), we have . 5 Re{teia = Re {trT} = Re {trA} > 0. + Thus 5(t + t-1) cos a > 0 and hence cos a > 0. Let z = x + iy = k (teia +t -1 e-ia ) - trA - ) cos a and y = (t + t1 Then x then x = If sin a (t + (1) and y = 0. cosa sin a. t') If sin a = 0, Thus 1+ (trA)2 4 det A = 1+ (t = 1+ (t - t-1)2 +t1)2 (trA)2 4 det A 1 (t + t-1)2 -1 4 4 (t + t-1)2 4 4 - 0 < 2 y2 0, then 2 x2 2 1/2 (t - 1/2 .2a sin (t + t-1)2 2 cosa .2a (t - t -12 ) sin 2 4 cosa 4 sin 2a Thus by Lemma 2.8 .2a =2 {1 + sin z2 - (trA)2 -1 4 detA z2 (trA)2 4 det A =1 . 48 Since sin a < y 2 2y 2 (trA)2 4 det A >1+ (trA)2 4 det A 1 which is the desired result. ("If"). Suppose A satisfies the following conditions: (i) det A > 0, (ii) Re {trA} > 0 and (iii) 2 y2 > 1 + (trA)2 4 det A (trA)2 -1 4 det A For convenience, we denote trA by c and \idet A by 6, and consider the following two cases. Case 1. c2 = 4 62. Here the eigenvalues of A are equal. Since det A = 62 > 0 and A is nonscalar, A is similar to 1 0 ± 6 [2y 1] with y > 0. Since Re {trA} > 0 and 6 > 0, ± 5 is in fact, + 6 (61) 6 [2y 1 -1 . A1 = 6 0 z 1 ] 1 Thus 6[ 2y 0] (6-1/21) 1 Case 2. c2 ba 1 4 62. = [1 2y 0] 0 1.] 01 -.1 2q a 2.y i s conjunctive with So A is Akvie, with B. Here A has distinct eigenvalues, say 6aand Then A is similar to [a [1 with arbitrary q > 0. Since - Re {trA} > 0, Re {a+ a1 } = 1 Re {trAi} = 1 5 Re {trA} > 0. 0] 1 49 Thus by Lemma 2.9 Re {a} > 0. Here we consider two subcases. Subcase 2a. Im {a} = 0. Here a> 0 and a cr-1 and let q = y in Al. Then Al is conjuctive with -- -1-2 ,5 a2 a01 0 [a0 k 2 L0 2-y 1 with B. So A is Subcase 2b. Im {a} A 0. Let a= teia with t > 0 and a real. x+ z = 1/2(a+ ;1) - Let trA 25 Since Re {a} > 0, it follows that cos a > 0. Here sin a = t1Im {a} 0. Also x = 1/2 (t + t-1) cos a and y = k (t - t-1) sin a. Thus x 2 2 .2a sin 2 cosa - (t t-1)2 cos2a 2 4 cosa (t t-1)2 in s2 2 4 sina So by Lemma 2.8 sin2a = {1 + lz 2 (trA)2 -1 4 det A =k + < 1/2 { 2 N2} = -ii _ (trA)2 4 det A -1 . 50 sin a and hence L[t-1/2 8 -t = Now' Al is conjunctive with < Y. [ eia 0 3- t2 0 with q I Al 5 2[ 2q 01 -ia e 42 - sin2a > 0, and by Lemma 1.10, [eia 0 2 4/y 2 [2 - sin2a e-ia Therefore A is 104 Corollary 2.11.1. 1 is conjunctive with .y 10 with B. Let B be a 2 x 2 nonsingular codefinite contrahermitian matrix and A be a 2 x 2 nonscalar complex matrix. Then A is ..x/Ait with B if and only if (i) sgn (det A) = sgn (det B), (ii) (iii) trB Re { (trA) IdetBIdet B (B) Idet BI Proof: tr B 5- (trA)2 4 det A } (trA)2 4 det A > 0 , and -1 Let B be a 2 x 2 nonsingular codefinite contrahermitian matrix and A be a 2 x 2 nonscalar complex matrix. Then B is conjunctive (by Case 2 at the beginning of this Chapter) with 51 with 0 <-y < 1 and p real, so e . B is conjunctive with r1 12y 01 1J Also e-if3A is nonscalar. ("Only if"). Suppose A is JAZ with B. Since det B Lemma 1. 8, sgn (det A) = sgn (det B) and (i) holds. e A ts 0, by By Lemma 1. 2, By Theorem 2.11, with e-iPB. Re {tr(e-iPB)} > 0 and Re {tr(e-iPA)} > 0. Now trB sgn (e -ip tr B) = sgn eip ("pr. v. " means 2 = sgn (pr. v. = So e-*= sgn e-41 det B ) "principal value". (trB)2 ) det B sgn( (trB)2 det B 1 sgn (trB) = sgn 1 tr B = sgnIdet---r (trBB1 sgn ( ( (trB)2det B (trB) (trB Idet B I trB det B (trB)2 det B (tr B) 2 det B ) ) 52 Since Re {e-iP trA} = Re {tr(e-i(3 A)} > 0 and multiple of trB( trB \ det B ldet Re ((trA) e-i13 is a positive trB trB idet BI det B >0 and (ii) holds. Also by Theorem 2.11, we have -) r tr (eif? A)12 4 det (e -i(3 4 det (e PA) A) (trA)2 - 1 4 det A = (e-43A)12 (trA)2 4 det A i. e. N2 < (trA)2 4 det .T.(B)_ 2 .y2 1-2 (trA)2 - 1 4 det A Hence (trA)2 4 det A (trA)2 4detA Net B1 -1 which is the desired result (iii). ("If"). Suppose A and B satisfy the following conditions: sgn (det A) = sgn (det B), Re {(trA) t.(B) Idet BI ( trB trB idet BI det B (trA)2 4 det A ) } > 0, and (trA)2 4 det A -1 53 Since eB is conjunctive with [12y with 0 <y < 1, det (e > 0. 13B) 01 Thus from (i) we have det (e-ii3A) > 0. From (ii) we see that Re {tr(e-43A)} = Re {e-ii3trA}, which is a positive multiple of Re {(trA) ( trB* pet BI From (iii) and the fact that So I - 2 y2 < 22 > dce t( BB) + (trA)2 4 det A - 1 - 2 y2, we have (trA)2 4 det A 1 + 2 e-2ip (trA) -1 >1+ rtr (e-ipA)12 -1 = 1 . I (trA)2 4 det A 1 trB det B } > 0 + 4 edet A (trA)2 4 det A - 2 ip (trA)2 4 e- 2ip det A Hence we have (3) 2y2 4 det (e 'A) Therefore by Theorem 2.11 e-iI3A is Lemma 1. 2, A is with B. rtr (e-iPA)12 4 det (e-iPA) with e-iPB. By 54 Theorem 2.12. Let A be a 2 x 2 nonsingular complex matrix with distinct eigenvalues a and [3, and B be a 2 x 2 complex matrix. Then B is 0444 to A if and only if B is conjunctive with for some c > 0. Proof: This is a special case of Theorem 4. 4, proved in Chapter IV. Corollary 2. 12. 1. Let A=[ i _i] 00 and B be a 2 x 2 complex matrix. Then the following three state- ments are equivalent. B is C44,6 to A; B is conjunctive with [ic -101 for some c > 0; det B > 0 and B is either bidefinite or contradefinite. Proof: Let i 0 A = [0 _i] and B be 2 x 2 complex. 55 (ii): Replace aby i and p by -i in Theorem 2.12. Suppose B is conjunctive with [ic -1o] for some c > 0. Then det B > 0. If c = 0, then B is bidefinite. (iii): If c > 0, then by Lemma 1.14 B is contradefinite. (i): Suppose det B > 0 and B is either bidefinite or contradefinite. Also we have det A = 1 and trA = 0. If B is bidefinite, then by Corollary 2.9.1 B is &Ate to A. If B is contradefinite, then by Corollary 2.10.1 B is C/1144 to A. Theorem 2.13. Let A be a 2 x 2 nonscalar matrix with det A = 1 and trA = 2, and let B be a 2 x 2 complex matrix. Then B is ern.' to A if and only if (i) det B 0, (ii) det B> (B), and (iii) either (D (B) <- det B or Re {trB} > 0. Proof: Let A be a 2 x 2 nonscalar matrix with det A = 1 and trA = 2, and let B be a 2 x 2 complex matrix. ("Only if"). Suppose B is end to A. Then there exist nonsingular * matrices C and D such that DBD = C-1AC. Since det A = 1, by Lemma 1. 9, det B > 0 and (i) holds. Since A is nonscalar with both eigenvalues 1 and by Schurts Theorem, there is a unitary * -1 matrix U such that U (D BD) U = U-1 (CAC) U= 56 for some y A 0 and we may assume y > 0. Thus from the intro- 10) - 1 &action of this chapter, we see that det B y2 <1. Hence (B) <det B and (ii) holds. To show (iii), we consider two cases. 2 1(2 '3-)B <- 1 y >1. Here we have 1 - 2 y <- 1. Thus det and hence cD(B) <- det B. Case 1. Case 2. 0 <y < 1. Here we have by Lemma 2. 7, Re {tr B} > 0. ("If"). Suppose B satisfies the conditions: (i) det B > 0, (ii) and (iii) either 1. (B) <- det B or Re {tr B} > 0. Then det B > B is conjunctive with eiprl 2' 01 L with y >0 or else Bis conjunctive with e iB, i I. Since det B > 0,2iI3 e= land e iB = ' with ± [o o ,.] 0, 0 for we would have 1. Now B cannot be conjunctive (B) =-det B and Re {trB} = 0, contradicting (iii). We claim that y > 0. For if y = 0, we would have 1)(B) = det B, contradicting (ii). Hence B is conjunctive with ± [1 2y 0] 1 with y > O. 57 2 (B) 0 < y < 1. Here det B - 1 - 2 y > - 1, and det B >0 1 and by Case 2 at the by (i), so Re {trB} > 0 by (ii). Since e Case 1: beginning of this chapter satisfies Re tr B} > 0, we must have {e = 1 here, and hence B is conjunctive with rl 0, 2y Case 2: 1J y > 1. Here 0 [12y 11 is conjunctive (by Lemma 2. 5) with rl. L 2y 0, 1J Thus in either case B is conjunctive with rl 12y 0, 1J which is a nonscalar matrix of trace = 2 and det = 1. Since A is nonscalar with single eigenvalue 1, the matrix rl L 0, 2-y with y > 0 is similar to A. Thus B is elkft to A. Corollary 2.13.1. Let A be a 2 x 2 nonsingular nondiagonable complex matrix and B be a 2 x 2 complex matrix. Then B is eAti 58 to A if and only if (i) sgn (det B) = sgn (det A), det B (iii) > (1)(B), and either IT. (B) < - det B1 or Re {(tr A*) (trB)} > 0. Proof: Let A be a 2 x 2 nonsingular nondiagonable complex matrix. For convenience let 8 = ildet Aland 6= sgn (trA). Then both eigenvalues of A are 68. Thus by Schurls Theorem there is a unitary matrix U such that U-1AU = 88 [1 2y for some y 0, 11 0 and we may assume y > 0. Let B be a 2 x 2 complex matrix. ("Only if"). Suppose B is eikkt to A. Since det A 1. 9, 6 0, by Lemma sgn(det B) = sgn (det A) and (i) holds. By Lemma 1. 2, B is Cif4.1 to 8-4,A. However, 8A is unitarily similar to [ 1 0 2y 11 -1_ and so 8 8 B is ei1444 to r1 L 2,y Thus by Theorem 2.13 we have 0, 11. 59 8-2 Idet BI B = Idet (6 5-2 det = E B) = det (6 -1_E B) -_6 B) = 62 1 > B) = 6-2sT.(B), i.e., Idet B > c(B) and (ii) holds; also we have either 62(1.(B) = (6 E B) <- det (5 6 B) = - det (6 -1E B) -1_ -8-2 det B I i.e., (D(B) <- or 0 <Re {tr (5-1F B)} = Re det B I , =5 -1 6 16-1Z- , tr B} Re {(sgn tr A*) (trB)} -1 ItrAl i. e. , Re {(trA*) (trB)}, Re {(trA ) (trB)} > 0 and (iii) holds. ("If"). Suppose A and B satisfy the following conditions: (i) sgn (det B) (i1) det B I > sgn (det A), (B), and either <1.(B) <- [(Diet B j or Re {(trA ) (trB)} > 0. Since 6 - 1_ E A is unitarily similar to -- with y > 0, det (5 1E A) = 1. Thus by (i) we have (1) det (5 -1_E B) > 0. From (ii) we have (2) det (5 6 B) = Idet (5& B)1= 5 -2 bet B I > 5-2 .T.(B) = 'I. (5 -1_ E B). 60 From (iii) we have (3) either 1)(5-1-g B) = or Re {tr(5-1-E B)} 5-2(B) bet 131 = - det (5-1E B) <- tr B} = Re {5-1 = Re {5-1 (sgn tr A*) (trB)} > 0. Thus by Theorem 2.13 5 -1 _ E B is which is unitarily similar to 5 - 51 to 11 0, 1-2,/ 11 6 A. Hence 5 -1 6 B is e A and so by Lemma 1. 2, B is to 0,14,6 to A. We summarize the main results of this Chapter in Tables 1 and 2. Table 1. 2 x 2 singular complex matrices. A B A is 41,withB Reference zero matrix complex 4=4 B=0 Theorem 2.1 complex zero matrix <==> A=0 Theorem 2. 2 complex codefinite of rank one <=> det A = 0 and sgn (trA) = sgn (trB) Corollary 2. 3. 1 complex contrahermitian <=) rank A = 1 Theorem 2. 4 of rank one rank = 1 complex det B = 0 and trace = 0 rank = 1 trace i 0 complex < (B) < 0 det B = 0 and either T (B) < 0 or sgn (trA) = sgn (trB) Theorem 2. 5 Theorem 2. 6 Table 2. 2 x 2 nonsingular complex matrices. A is A Corollar:- 2. 7. 1 nonzero scalar complex complex unidefinite (trB*) A nonscalar unidefinite (i) .(=> A*BE bide finite nonscalar contradefinite nonscalar contrahermitian <=) E Corollary 2. R.1 Corollary 2. 8. 2 r(i) sgn (det A) = sgn (det B) Corollary 2. 9.1 A 2 (det B*) (trA) < 0 Corollary 2.10.1 sgn (det A) = sgn (det B) { (i) sgn (det A) = sgn (det B) trB (ii) Re {(trA)(Idet BI code finite (m) Idet(... B) complex p2 sgn (det A) = sgn (det B) (ii) (trA) (trB ) > 0 (trA)2 >4 (i")det complex Reference with B (=> BreP[ci: (trA)2 4 det A + Corollary 2. 11. 1 tr B det BP> (trA)2 4 det A-1 for some p > 0 Theorem 2.12 Table 2. Continued. A B [io nondiagonable A is AC., with B complex det B > 0 and B is either bidefinite or contradefinite complex (i) <=> sgn (det B) = sgn (det A) B I> t.(B) either I, (B) < - Idet B or Re {(trA*) (trB)} > 0 Symbols: f If-go! denotes "is similar to"; 4V-11 denotes "is conjunctive with". Reference Corollary 2. 12. 1 CorollarN,- 2.13.1 64 n x n COMPLEX MATRICES WITH REAL EIGENVALUES In this chapter we consider only certain classes of n x n complex matrices whose eigenvalues are real. These matrices which will be discussed are: hermitian matrices (since hermitianness is invariant under conjunctivity), diagonal matrices (or diagonable matrices, since diagonability is invariant under similarity), and triangular nondiagonal matrices. Although a few specific results have been proved in Chapter II for 2 x 2 matrices, we restate these results here for completeness. § 3.1. Diagonable Matrices We restate Theorems 2. 7 and 2. 8 for the n x n complex matrices Theorem 3.1. Let A be the n x n identity matrix and B be an n x n complex matrix. Then B is a/n/L to A if and only if B Proof: See Theorem 2. 7. Theorem 3. 2. Let B be an n x n positive definite hermitian matrix and A be an n x n complex matrix. Then A is ..,1-#6 with B if and only if A E p2. 65 Proof: See Theorem 2. 8. Theorem 3. 3. Let B be an n x n hermitian matrix and A be an n x n complex matrix. Then A is .)Me,, with B if and only if A is similar to a hermitian matrix of same rank and signature as B. Proof: Trivial. Theorem 3. 4. Let B be an n x n hermitian matrix and A be an n x n normal matrix. Then A is _ill/6 with B if and only if A is conjunctive with B. Proof: Let B be an n x n hermitian matrix and A be an n x n normal matrix. ("Only if"). Suppose A is _Ike. with B. Then there exist non- singular matrices C and D such that C-1 AC = D* BD. Since B is hermitian, D*BD,an.d hence C-1AC, is also hermitian and is there- fore normal. Hence A and C-1AC are (by Theorem 1.6) unitarily similar. So A is conjunctive with D BD, hence with B. ("If"). Suppose A is conjunctive with B. Then by Lemma 1. 5 A is ..V./kLZ.,with B. Theorem 3. 5. Let A be an n x n real diagonal matrix with positive distinct eigenvalues and B be an n x n complex matrix. Then B is ejn4, to A if and only if B EP 3. 66 Let A be an n x n real diagonal matrix with positive Proof: distinct eigenvalues and B be an n x n complex matrix. ("Only if"). ei,d, to A. Then there exist nonsingular Suppose B is matrices C and D such that D* BD = C-1AC. Thus by Schur's Triangularization Theorem there exists a unitary matrix U such that * * U (D BD) U = U-1(C-1AC)U = T is a lower triangular matrix whose diagonal entries are the eigenvalues of A. Since A has positive eigenvalues, B is conjunctive with a lower triangular matrix of positive diagonal. Therefore, by Theorem 1. 2, B E 9)3. ("If"). Suppose B 9-)3. Then, by Theorem 1. 2, there exists a nonsingular matrix P such that P*BP is a lower triangular matrix of positive diagonal. Thus there exists a nonsingular positive diagonal * * matrix Q such that Q (P BP)Q= (PQ) B (PQ) is a lower triangular matrix whose diagonal entries are the same as those of A. Since the diagonal entries of A are distinct, (PR) B(PQ) is similar to A. Hence B is en"- Theorem 3. 6. to A. Let A be an n x n nonsingular complex matrix similar to a real diagonal matrix and B be an n x n nonsingular cohermitian matrix. Then B is entL to A if and only if B is hermitian and sig B Proof: sig A. Let A be an n x n nonsingular complex matrix similar to a real diagonal matrix and B be an n x n nonsingular cohermitian 67 matrix. ("Only if"). Suppose B is Cliut to A. Then there exist nonsingular matrices C and D such that D*BD = C-IAC. Since B is nosingular cohermitian, B = EH, where H is nonsingular hermitian and E is a nonzero complex number. Hence there exists a unitary * * -1 -1 U HI is a diagonal matrix U such that U (D 8 HD)U U(CAC) matrix, and this diagonal matrix must be real because all the eigenvalues of A are real. Thus HI is hermitian and nonsingular, E must be real and so B must be hermitian. Thus, by Theorem 1. 3, index (B) = index (H1) = index (A) and hence sig B ("If"). Suppose B is hermitian and sig B sig A. sig A. Since there exists a nonsingular matrix C such that C-1AC is a nonsingular real diagonal matrix, C-1AC is nonsingular hermitian and sig (C 'AC)= sig B. Thus, by Theorem 1. 3, B is conjunctive with C-IAC and hence B is entd., to A. We next prove a Lemma. Lemma 3.1. If S is n x n complex lower triangular nondiagonal with all diagonal entries real and two of them of opposite signs, then S is contradefinite. Proof: Let S be n x n complex lower triangular nondiagonal with all diagonal entries real and two of them of opposite signs. First, suppose S is nonsingular. Since S is lower triangular nondiagonal, S is contrahermitian and has all diagonal entries on the real line through 0 and not all on one side of 0. Hence, by Lemma 1.14, S is contradefinite. Next, suppose S is singular of 68 * rank r and codefinite. Then, by Lemma 1.12, S is -regular. So, by Lemma 1.13, S is unitarily conjunctive with the direct sum of a nonsingular r x r matrix S1 and a zero matrix. Since S is lower triangular nondiagonal, S is contrahermitian. Thus the direct sum, and hence S1' is also contrahermitian. Since S has real diagonal entries with two of them of opposite signs, the direct sum, and hence S1' has real eigenvalues with two of them of opposite signs. Thus by Schur's Theorem S1 is unitarily conjunctive with a lower triangular nonsingular matrix S2 with real diagonal entries and two of them of opposite signs. S2 is also contrahermitia.n, and hence nondiagonal. Thus S2 is contradefinite (as in the first case). Therefore S1' and hence 5, is also contradefinite, giving a contradiction. Theorem 3. 7. Let A be an n x n nonsingular real diagonal matrix with distinct eigenvalues and B be an n x n nonsingular nonhermitian codefinite matrix. Then B is if either A (.." and B Proof: or else A E- eind, to A if and only rTh3 and B e-t3- . Let A be an n x n nonsingular real diagonal matrix with distinct eigenvalues and B be an n x n nonsingular nonhermitian codefinite matrix. ("Only if"). Suppose B is ell4Z to A. Then there exist nonsingular matrices C and D such that D BD = C-1AC. Thus by Schur's 69 Theorem there exists a unitary matrix U such that U (I) BLI\U = U-1(C-1AC)U = T is a lower triangular matrix whose diagonal entries are the same as those of A. Since B is nonhermitian and T is conjunctive with B, T is also nonhermitian, and hence nondiagonal. First we claim that A Epu e., A is positive definite or -A is positive definite. For if A q-Dti-P, then A has real diagonal entries with two of them of opposite signs, so T also has real diagonal entries with two of them of opposite signs. Thus, by Lemma 3.1, T is contradefinite. Since B is conjunctive with T, B is also contradefinite, contradicting the fact that B is codefinite. Hence we must have A cpu _p. Therefore either A EPand, by Theorem 3.5, B ER3, or A E _p , i. e., -A Ep; so, by Theorem 3. 5, -B EP3 and hence B ("If"). Suppose A EPand B E (P3. to A. Suppose A E_p and Then,by Theorem 3. 5, B is B -y. Then - A ePand -B eP3. Thus, by Theorem 3. 5, -B is 6/14Z to -A. Therefore, by Lemma 1. 2, B is 0/41.4 to A. Corollary 3. 7.1. Let A be an n x n nonsingular complex matrix whose eigenvalues are real and distinct, and let B be an n x n nonsingular contrahermitian codefinite matrix. Then B is 6,4 to A if and only if either A EP2 and B EP3 or else A E - (P2 and B e -CP3. 70 Proof: Let A be an n x n nonsingular complex matrix whose eic-,envalues are real and distinct and let B be an n x n nonsingular contrahermitian codefinite matrix. Since A is similar to an n x n nonsingular real diagonal matrix with distinct eigenvalues, by to A if and only if either A is similar Theorem 3. 7,B is to an element of sand B or else A is similar to an element of _9D and B e -F3, if and only if (by Theorem 1.1) either A eP 2 and B or else A E 2 and B e - Although an idempotent matrix is not necessarily diagonal, it is similar to a diagonal matrix. Theorem 3. 8. Let A be an n x n idempotent matrix (A = A) of rank r, 0 < r < n. Let B be an n x n complex matrix. Then B is &//42 to A if and only if B is conjunctive with ,0 21 01 0 e I r-s EDI On- r- s for some s < r. Proof: Let A be an n x n idempotent matrix of rank r, 0 < r < n. Let B be an n x n complex matrix. ("Only if"). Suppose B is efriz to A. Then there exist non- singular matrices C and D such that D BD = C-1AC. Since A is idempotent of rank r, A is similar to the matrix 71 [ Ir 0] 0 0 Thus by Schur's Theorem there exists a unitary matrix U such that .1, U (D BD)U = U-1(C-1AC)U = 1T L 01 V where T is arxr lower triangular matrix all of whose diagonal entries = 1 and V is an (n - r) x r matrix. Since rank T = r and rank A = r, the lower right block W must be On-r. Since [T V TO , LV 01 0 0 n-r is idempotent, rT rT r LV 0-1 LV L T2 01 VT 0 -I Therefore T2 = T so T = I (and V is arbitrary), 01 i.e.,1(C-1AC)U = [IrV is conjunctive with P 0 0 0 Now, n-r Ir V 01 [Ir V Oi po 0 0E 0 n- r [Ir 01 LE *VP 0 j for every r x r unitary matrix P and every (n - r) x (n - r) nonsingular matrix E. Let s = rank V. Then we can here choose E nonsingular so that 1/2 E V is in row-echelon form and then choose P as a permutation matrix so that E VP = 2Q1 o [2Is oi 72 where Q is some s x(r - s) matrix. Thus we may assume V itself = [I r 2Q [2Is x s(r-sj °(n-r-s) So 0 V 0 0 9 0 n- r- s. Now, 0 21 -21 0 I -2Q-2Q Is r-s ..- , I s s 0 0 0 0 0 0 0 0 0 0 -2Is -Q I 0 0 -2Q-' I 0 Is I 0 0 , which is again conjunctive 0 * I r- S (1 +QQ r-+Q s [021 s Q) 0 21s +Q Q) 0 0 0 0 0 0 0 0 * r- S 0 r-0 0 (I -4(I5+QQ ) 0 0 0 0 s r-] 0 ) Is ) 0 conjunctive with 0 r-s 2Q 0 -1/2(I5+QQ*)-1 Is -12-(Is+QQ 0 r-s s 0 with I 02Q -4(Is+QQ 0 0 2Q2Q * 0 0 0 r-s n-r 2Is 1i s 01 0 s 0 I r -s 00 0 0 01 I r-s +Q*Q 73 Since B is conjunctive with [0 01 e r-s e 0 21s r V 0 B is conjunctive with On-ri n-r - S. [0 ("If"). Suppose B is conjunctive with 21s 01 r-s @0 n- r- s, ED 0 0 with 0 < s < r < n. Since I 0 r-S 0 215 (3 01 e I r-s is conjunctive with 0 0 Is _2Is Is B is conjunctive with 0 0 0 ,_2Is s 0 0 - 0 e 0 n- r - s. 0 Thus there exists an n x n nonsingular matrix D such that Is D BD = 0 2Is 0 0 I 0 r-s 0 @0 n-r-s. Hence D BD is an idempotent Os matrix of rank r. Since A is also an n x n idempotent matrix of rank r, D BD is similar to A. Thus B is CALL to A. §3. 2. Triangular Matrices We weaken the hypotheses of Theorems 3. 5, 3. 6 and 3. 7 by allowing the matrix A to have repeated eigenvalues. It is not clear whether necessary and sufficient conditions can be found so that the 74 matrix B is CAZ to A. However, we have the following extension of Theorems 3. 5, 3. 6 and 3. 7, by considering a class of matrices similar to A. For X. ER (R denotes the real field), we define 1 minus the number of negative X ..1 When A is similar to diag (X 1,X 2, sig A = sig (X1, X 2, . n. , . , X n). Let A1 Theorem 3.9. i = 1, 2, . Let 2' aAbe the set of all n x n matrices Similar to triangular matrices of diag =A. Let B be an n x n complex matrix. Then there exists A E which is ."Ime with B if and only if B Proof: The proof is a slight modification of the proof of Theorem 3. 5. Theorem 3.10. Let A (x. 1, X 2, for i 1, 2, . . . , n. Let . . , X n) withX.ERVO} be the set of all n x n nonsingular matrices similar to triangular matrices of diag = A Let B be an n x n nonsingular cohermitian matrix. Then there exists A e which is with B if and only if B is hermitian and A is diagonable and sig B = sig A . Proof: The proof is a slight modification of the proof of Theorem 3. 6. 75 Theorem 3. 11. 1, 2, . i . , n. Let A= (x x 2, . . . ,x n) with X E eVO for Let a be the set of all n x n nonsingular matrices similar to triangular matrices of diag = A . Let B be an n x n nonsingular contrahermitian codefinite matrix. Then there exists with B if and only if either sig which is A B e?3 or else sig A= -n and B c n and rTh3 . Proof: The proof is a slight modification of the proof of Theorem 3.7. An n x n complex matrix S is called nilpotent if A = 0 for some positive integer m. Theorem 3.12. Let A be an n x n nilpotent matrix of rank n- 1 and let B be an n x n complex matrix. Then B is e/14.Z to A if and only if B is conjunctive with a strictly lower-triangular matrix of rank n Proof: 1. Let A be an n x n nilpotent matrix of rank n - 1 and let B be an n x n complex matrix. ("Only if"). Suppose B is 6,11./Z to A. Then there exist nonsingular matrices C and D such that D BD = C-1AC. Furthermore, by Schur's Theorem there exists an n x n unitary matrix U such that * U (D BD)U = U-1(C-1AC)U = T is a lower triangular matrix. Since A is nilpotent of rank n - 1 and T is similar to A, T is strictly 76 lower triangular of rank n - 1. ("If "). Suppose B is conjunctive with a strictly lower triangular matrix T of rank n - 1. Since A is nilpotent of rank n - 1, A and T have the same invariant factors. Thus T is similar to A and hence B is eAkt to A. 77 IV. n x n COMPLEX MATRICES WITH COMPLEX EIGENVALUES In this chapter we again consider only certain classes of n x n complex matrices whose eigenvalues are complex numbers. These matrices include diagonable matrices, triangular nondiagonal matrices, unitary matrices, positive stable matrices and accretive matrices. §4. 1. Diagonable Matrices Theorem 4.1. Let A be an n x n nonzero scalar complex matrix and B be an n x n complex matrix. Then the following three statements are equivalent. to A; B is * c,Th A B e7-'; B is conjunctive with A. Proof: See Corollary 2. 7.1. Theorem 4. 2. Let B be an n x n unidefinite nonsingular complex matrix and A be an n x n complex matrix. Then A is with B if and only if (trB ) A F2. Proof: See Corollary 2. 8.1. 78 Let A be an n x n nonsingular complex matrix Theorem 4. 3. all of whose eigenvalues are distinct and B be an n x n nonsingular contradefinite complex matrix. Then B is efrkt to A if and only if sgn (det B) = sgn (det A). Proof: Let A be an n x n nonsingular complex matrix all of whose eigenvalues are distinct and B be an n x n nonsingular contradefinite complex matrix. ("Only if"). Suppose B is CM," to A. Since A and B are nonsingular, by Lemma 1. 8 sgn (det B) = sgn (det A).. ("If"). Suppose sgn (det B) = sgn (det A). We denote the eigenvalues of A by t1, t2, . t and hence tn. Thus det A = t1 t2 . , . sgn (detB) = sgn (t1 t2 . . . tn). Hence by Lemma 1.15 B is conjunctive with a lower triangular matrix T such that diag T = (t1, t2, . . , tn). Since t1, t2, . similar to A. So B is . , tn are distinct, T is to A. Theorem 4. 4. Let A be an n x n complex matrix similar to a diagonal matrix with exactly two eigenvalues a and p, with rank (A - p1) = r, 0 < r < n. Let B be an n x n complex matrix. Then B is efr1.4 to A if and only if B is conjunctive with S a 0 j=1 Y R 9 al r-s e p in- r - s 79 for some s < min {r, n-r} and for some positive y l' Y2' Let A be an n x n complex matrix similar to a Proof: diagonal matrix with exactly two eigenvalues a and p, with rank Let B be an n x n complex haatrix. (A - pI) = r, 0 < r < n. ("Only if"). Suppose B is CM/4, to A. Then there exist non* singular matrices C and ID such that DBD = C-1AC. Since A is similar to a diagonal matrix with exactly two eigenvalues aand p, with rank (A - pI) = r, so is C-1 AC and hence D* BD. Thus by Schur' s Triangularization Theorem there exists a unitary matrix U1 such that * U1 0 (D BD)U = [air V * with rank V = a for pin_r 1 some s < min {r, n-r}. Since V is an (n-r) x r matrix with rank s, by Autonnels Lemma (Theorem 1. 8) there exist arxr unitary matrix Q and an (n-r) x (n-r) unitary matrix P such that 00i P VQ = , where As is an s x s diagonal matrix with each diagonal entry Nj > 0, j = Then U2 1, 2, , s. Q Now let U2 = [0 is unitary. Thus 0 [otI * U2 U1 . (D BD)U1U2 = 0 PO V pin_r Q 0 0 P 0 pi. 80 0 P VQ Pin- r 1 Since P*VQ , = o 0 o 0 (DU1U2)* B(DU1U2) = 0 S o pin-r- s s_ ,e, 0 0 0 0 0 I 0 0 Is 0 0 Is let U3 = s- * * Then U3(DU1 U2) B(DU1U2)U3 = = r-s 0 0 0 In-r- s aIs As 0 0 0 pis o 0 o 0 aI r- s 0 0 0 0 pin-r- s .., , e aI r- s 1 El) pin-r-s. 81 1 0 0 2 Since As = , aI conjunctivity by a permutation o conjunctive with _As i3Is Therefore B is conjunctive with ( s \j10 [a y. 0 ) C) aI r-s e pIn-r-s r3 for some s < min {r, n-r} and some positive y1,1 v' 2' ("If"). Suppose B is conjunctive with B1= jc:?1 [ a 01 ) Yi P '1r -s e for some s < min {r, n-r} and some positive Let C= ( J=1 (s) Then C-1= ( j=1 )(:) in-2s Yr 112, 82 Thus -1 B1C o Clin-2s} (c)' j=1 e ocar-spIn-r_ y 0 1 0 j=1 1 0 1 al r-s 0 fl In- r- s 0 ( j=1[ y. P- a ( s [a 0 )133, j=1 0 p al r-s 0 13 In- r- s , which is similar to air 0 p In-r , which in turn is similar to A. Hence B is e/14'L to A. An n x n complex matrix S is called involutory if S2 = I. Corollary 4. 4.1. Let A be an n x n involutory matrix with rank (A + I) = r, 0 < r < n. Let B be an n x n complex matrix. Then B is e/P1.4/ to A if and only if B is conjunctive with 83 r-s (-In-r-s) for some s < min {r, n-r} and for some positive yl, y2, Proof: Replace a by "1" and p by "-1" in Theorem 4.4. §4. 2. Triangular Matrices In this section we consider those matrices which are not conjunctive with diagonal matrices. Theorem 4. 5. Let A be an n x n singular complex matrix all of whose nonzero eigenvalues are distinct and such that A has only linear elementary divisors at 0, and let B be an n x n *-irregular complex matrix. Then B is ei444 to A if and only if rank B = rank A. Let A be an n x n singular complex matrix all of Proof: whose nonzero eigenvalues are distinct and such that A has only linear elementary divisors at 0, and let B be an n x n *-irregular complex matrix. ("Only if"). Suppose B is 0/11./i to A. Then by Lemma 1. 7 rank B = rank A. ("If"). Suppose rank B = rank A. Let r = rank A with 0 < r < n, and t1, t2, . . , tr be the nonzero eigenvalues of A. Since B is 84 *-irregular of rank r, by Lemma 1.16 B is conjunctive with a lower with diag T = (t t2 triangular matrix [TV 0] 0 . . . t r) and for some (n-r) x r matrix V. Since A has only linear elementary divisors at 0, the matrix [VT 0 ] 0 . similar to A. Hence B is e/41,1 to A. For the case where B is an n x n singular *-regular contradefinite matrix, by Lemma 1.131B is *-regular if and only if B is conjunctive with the direct sum of a nonsingular matrix B1 and a zero matrix, and the conjunctivity class of any such B1 is uniquely determined by B. Thus n (B) = sgn (det B1) is defined and conjunctively invariant for every *-regular B (if we define n(0) Thus n = 1). (B) is the signum of the product of nonzero eigenvalues of B. We state without proof the following result. Fact 4.1. Let A be an n x n complex matrix of rank r (0 < r < n) such that its nonzero eigenvalues X 1,X 2, . are distinct and it has only linear elementary divisors at 0. Let B be an n x n *-regular contradefinite matrix. Then B is e/kut to A if and only if rank B = r and 7-1 ( B) = sgn (X 1X2...X r). An n x n complex matrix A is said to be nonderogatory if and only if the minimal polynomial of A is equal to its characteristic polynomial. It is also known that A is nonderogatory if and only if 85 the Jordan form for A has just one block for each characteristic root of A. We are unable to find (useful) necessary and sufficient conditions for an n x n complex matrix B to be CMAt to an arbitrary given nonderogatory matrix A. However, we have solved the prob- lem for a nonderogatory matrix A with only one characteristic root (see Theorem 4. 6). We first introduce some preliminary lemmas. Lemma 4.1. Let S be a 3 x 3 nondiagonal lower triangular matrix with diag S = (1,1,1). Then S is conjunctive with a 3 x 3 lower triangular matrix T with diag T = diag S and such that the first subdiagonal of T has no zero entries. Proof: Let S be a 3 x 3 nondiagonal lower triangular matrix with diag S = (1,1,1). We consider the following three cases. Case 1. -1 S is of the form w _0 0 O- 1 0 0 , w 0. 1_ Here we apply to S the [2, 3] - subconjunctivity defined by the matrix 0 *- 1 1 -5 , 4 3 4] and get thereby a matrix T 1= 4 5 w 3 -w 0- 1 0 0 1 We further apply to T1 the [1, 2]-subconjunctivity defined by the 86 [1 o and get thereby a matrix T matrix w1 4 0 6 1 0 --5 w2 where w1 = 4 5 0 and w2 Case 2. S is of the form = 5 0 w o. 0 1 0. g , Here we apply to S the [1,2] -subconjunctivity defined by the matrix 0 3 0 1 4 g -34] and 4 [3 0' 0 get thereby a matrix T2 = 1 g We then apply . 1 [4 5 to T2 the [ 2, 3] -subconjunctivity defined by the matrix 1 and get thereby a matrix T = g2 = 5 g g1 0 0 1 0 g2 1 where g1 = 3 g 0 and 0. ^1 Case 3. S is of the form 0 0 0 1 0 x 0 1 ,x 0. Here we apply to S the interchanging [2,3] -subconjunctivity and get thereby a matrix 87 'l x 0 0 1 0 0 1 , which is covered by Case 1. Lemma 4. 2. Let 000 1 00 10 1 with xy 0. Then S is conjunctive, by a vw 01 [ 2, 3, 41 -subconjunctivity, with a lower triangular matrix T of the 1 00 0 .x1 1 0 0 with x'y'z' form ul V y` 1 W 0. 0 1 Proof: Let 1 000 1 00 10 with xy 0. vw 01 We consider the following three cases. Case 1. w 0, u = 0. Here we apply to S the [ 2, 3] -subconjunctivity defined by the matrix 88 1 ry x' 11 and get thereby a matrix T = LI' 000 1 0 0 [ III yi 1 0 V1 W' z' with 1 ._., x'y' z' 0, since x' = xy , y' = y and z' = w. 0, u Case 2. w 0. Here we apply to S the [2, 3] -subconjunctivity defined by the matrix 0 11 y x' 00 0 0 0 1 0 z 1 1 and get thereby a matrix T = UI V1 0, since x1 = - with xlyizt , W1 w y = y and zr = w. Case 3. w = 0. Here we apply to S the [3, 4] - subconjunctivity defined by the matrix 1 4 3 [-3 4] and get thereby a matrix 00 00 y1 0 1 1 T= 5 * 3 5 4 0 3 3 Y 0 1 Case 1 if u= 3v 4 which is covered by Case 2ifu 4 89 Lemma 4. 3. Let S be an n x n nondiagonal lower triangular matrix with diag S = (1,1, ... ,1). Then S is conjunctive with a lower triangular matrix T with diag T = (1,1, ... ,1) and such that the first subdiagonal of T has no zero entries. Proof: We use induction on n, and we state the induction assertion, Pn, as a sublemma. Sublemma 4. 3. 1. (= induction assertion, Pn). Whenever S is n x n lower triangular nondiagonal of diag S = (1,1, ... , 1), then S is conjunctive with a lower triangular matrix T of diag T = (1,1, ... , 1) such that the first subdiagonal of T contains no zero entries. Proof of Sublemma 4.3.1. and P3 P1 is vacuous' P2 is trivially true, is true by Lemma 4.1. So we assume n> 4 and Pn-1 is true. Let S be n x n lower triangular nondiagonal of diag S = (1,1, ... , 1). Since s= S[nin) _S(n) 01 1 we consider the following two cases. Case 1. S(n) is (n-1)x (n-1) nondiagonal. From our induction hypothesis, S(n) is conjunctive with a lower triangular matrix T1 whose diagonal = (1,1, ... ,1) such that the first subdiagonal of T1 90 contains no zero entries. Thus S is conjunctive with a matrix [Ti 01 T2 = V 1 for some 1 x (n-1) matrix V. Now T2 [n-3, n-2, n-1, n] is lower triangular nondiagonal of diag (1,1,1,1). Then we apply to T2 the [n-2, n-1, n] -subconjunctivity defined in Lemma 4. 2 and get thereby a matrix T such that T[ n-3, n-2, n-1, n] is the matrix 1 0 0 0 xl 1 0 0 with xty'z' 1 VI wtZ 0. Thus we have a 0 1 triangular matrix T of diag T = (1,1, ... ,1). Except possibly for the last three entries, the first subdiagonal of T is the same as that of T2, so all entries on the first subdiagonal of T are nonzero. Case 2. S(n) is diagonal. Here we have S[nin) 1 0. We consider two subcases. Subcase 2a. Sn, n-1 =x 0. Here S[n-2, n-1, n] is of the form 1 0 01 010 91 with x We apply to S the [ n-2, n-1, n] -subconjunctivity defined 0. in Lemma 4.1 and get thereby a matrix T3 such that T3[ n- 2, n-1, n] = Since T3(n) 1 0 0 yr 1 0 x' 1 with xlyt 0. is nondiagonal, T3 is a matrix covered by Case 1. Subcase 2b. S[nin-1, n) 0 and Sn, n-1 = 0. Suppose Sn, r = x 0 for some r such that 1 < r < n -2, and furthermore such that this is the last nonzero off-diagonal entry in the nth row, i.e. , such that S[ n Ir+1, , n-11 = 0. We then apply to S the interchanging r+1, n] -subconjunctivity and get thereby a matrix T4 such that is lower triangular nondiagonal. Thus the matrix T4 is again covered by Case 1. T4(n) Theorem 4. 6. 0 and n> 2 and let A be an n x n Let X. nonderogatory complex matrix with X as its only eigenvalue (i. e. , A - X I is nilpotent of rank n-1). Let B be an n x n complex matrix. Then the following three statements are equivalent. (i). B i s eiL to A; (ii) B is conjunctive with a lower triangular nondiagonal matrix with diag = (X (iii). B E ,X, ,X ); xcp 3 but not Exp . 92 0 and n> 2 and let A be an n x n nonderogatory Proof. Let X Let B be an n x n complex matrix with X as its only eigenvalue. complex matrix. 4 (ii). Suppose B is C.A4 to A. Then there exist nonsingular matrices C and D such that D BD = C-1 AC. Since A is non- derogatory with X as its only eigenvalue, by Schur's Theorem there exists a unitary matrix U such that * -1 -1 U (D BD)U = U(CAC) U=T is a lower triangular nondiagonal matrix with diag T = (X , X , ,X). Hence B is conjunctive with a lower triangular nondiagonal matrix with diag = (X , X , (i). , X ). Suppose B is conjunctive with a lower triangular non- diagonal matrix with diag = (X , X , , ). Then X -1B is conjunctive with a lower triangular nondiagonal matrix with diag = (1,1, ... , 1). Thus,by Lemma 4.31X 1B is conjunctive with a lower triangular matrix whose diag = (1,1, ... , 1) and whose first subdiagonal has no zero entries. Hence B is conjunctive with a lower triangular matrix T with diag T = (X ,X,. ) and the first subdiagonal of T having no zero entries. So T is nonderogatory with only one nonzero eigenvalue X. Since T and A have the same invariant factors, T is similar to A. Therefore B is ein,6 to A. 93 The following statements are clearly equivalent. (ii)<=j(iii). B is conjunctive with a lower triangular nondiagonal matrix T with diag T X (X -1B is conjunctive with a lower triangular nondiagonal matrix V with diag V = (1, 1, X , X ). ,X, -1 BE 3 (". B EXP 3 but X . . , 1). -1B OD. but B xp . (a)<=>(b). Trivial. (b)44(c). By Theorem 1.2 and the fact that V is nondiagonal. (c)<4(d). Trivial. §4.3. Unitary Matrices In this section we restrict ourselves to a special class of unitary matrices. It is not clear at present that we can completely solve the problem for unitary matrices. We first state a wellknown result. Theorem 4. 7. (Polar Factorization Theorem, [10], pp. 74-75; [12], p. 357) Let S be an n x n nonsingular complex matrix. Then there exist unique n x n positive definite hermitian matrices P and Q and n x n unitary matrices U and V such that S = PU = VQ. Furthermore, we have U = V. 94 An application of Polar Factorization Theorem is the following theorem which was pointed out by M. -D. Choi. (see e. g. [5], Theorem 7 (iii)). Let A be an n x n nonsingular complex matrix. Theorem 4. 8. Then there exists a unitary matrix B such that B is both CAZ to, and AlAte with, A. Proof: Let A be an n x n nonsingular complex matrix. Then by Theorem 4. 7 there exists a positive definite hermitian matrix P and a unitary matrix B such that A = PB. Since P is positive definite, by Theorem 1. 4 there exists a nonsingular matrix D such that P = DD . Thus A = PB DD B = D(D*BD)D 1. and hence B is eAira to A. So A = (C - * - (C 1)B = Now let D* (C-1)*C-1BCC-1. = C-1 . So D BD = D lAD -1* Then P = (C)(C-11_). Thus C-1BC C AC and so B is _14444 with A. An n x n complex matrix S is said to be a symmetry if S* = S -1 = S. Theorem 4.9. Let A be an n x n unitary matrix and B be an n x n nonsingular hermitian matrix. Then the following three statements are equivalent. A is At#16 with B; A is a symmetry and index A = index B; (iii) A is conjunctive with B. Proof: Let A be an n x n unitary matrix and B be an n x n nonsingular hermitian matrix. (ii). Suppose A is _JAC with B. matrices C and D such that C-1 AC Then there exist nonsingular D* BD. Since B is hermitian, D BD,and hence C-1 AC, is also hermitian. Thus l C AC _1 (C AC) * * = C A (C -1* . So A(CC ) = (CC )A ) . Since is positive definite and A and A are unitary, by CC Theorem 4. 7 we have A = A. Together with the unitary property of A, this gives us A-1 = A* = A. So A is a * symmetry. Also index A = index (C-1AC) = index (D BD) = index B. Suppose A is a symmetry and index A = index B. Since A-1 = A = A, A is hermitian and nonsingular, So A is conjunctive with B. Suppose A is conjunctive with B. Then by Lemma 1. 5 A is ..,,,hfrze, with B. We state without proof a dual of Theorem 4.9. Theorem 4.10. Let B be an n x n unitary matrix and A be an n x n involutory matrix. Then the following three statements are equivalent. 96 B is CAZ to A; B isa symmetry andindex B = index A; B is similar to A. §4. 4. Positive Stable Matrices We denote by0, the set of all n x n complex matrices S such that every eigenvalue of S is in ORHP (the open right-half plane). An element of / is called a positive stable matrix. We also denote by ,a , the set of all n x n complex matrices S such that S + S is positive definite. An element of is called an accretive matrix. A well-known result of Lyapunov may be stated in the following form. Theorem 4.11. (Lyapunov, [3], [4], [14]). cz_pg=0P A simple application of the above theorem can be stated in the following result, Theorem 4.12. Theorem 4.12. Let A be an n x n positive stable matrix. Then there exists an n x n accretive matrix B such that B is both to, and with, A. Proof: We use Theorem 4.11 and the proof is the same as that of Theorem 4. 8. 97 In view of Theorems 4. 8 and 4.12, we have the following generalization. Theorem 4.13. Let a= . there exists Then for Ae to, and likte with, A. B EB such that B is both Proof: See the proof of Theorem 4. 8. eft" to, and However, it is false to say that if F is both xrjne, with an n x n complex matrix E, then F = PE for some n x n positive definite matrix P. We illustrate this with an example. Example 4.1. F- (,r2 [1 1 1 -11 1 [10 Let E = [01 0, 1 0 0] and F 1 [ -1 11) 1 1 1 Then would be unitarily similar to (= unitarily conjunctive with) E. So by Lemma 1. 5 F is both /11-1 to, and _APlie with, E. Now suppose there is a positive definite matrix P = [ y y] such that F = PE. Then 1 F l 2l = x7 Z] [10 00] 0 which is impossible. [y 0] Finally, we state without proof a result of Stein-Pfeffer which may be expressed in the following form. 98 Theorem 4. 14. (Stein-Pfeffer, [13]). Let A be an n x n stable matrix. Then an n x n hermitian matrix H is the hermitian part of a matrix B Ott to A if and only if index (H) > the maximum number of Jordan blocks at any one eigenvalue of A. §4. 5. Accretive Matrices In this section we consider a special class of matrices called accretive matrices, which are defined in §4. 4. We first state without proof a lemma from [3]. Lemma 4.4. ([3], Lemma 5). If B is an n x n accretive matrix, then B is conjunctive with a diagonal matrix diag (e1P1 e432, , e n) with Tr 2 t 2 > P> - > 2 In [ 3], the authors gave a necessary and sufficient condition for the existence of a lower triangular matrix conjunctive with an accretive matrix. However, the proof was omitted in [ 3]. We supply here a proof by Ballantine. Theorem 4.15. ([3], Theorem 5). Tr z > 13 > > . . Let . >p >- Tr Then there exists a lower triangular n x n complex matrix T conjunctive with diag (e iP1 , e i P-2 , , eii311), and having diagonal 99 entries t1 e t 1, 2e ia 2 , tneian (in some order) with t1, t, , tn positive and >a Z> a 2 1 > >a Pk al + a2 + n > - Tr 2 if and only if (3.1 + P2 + ak for 0 < k < n, with equality holding for k c {0, n}. Proof: The "only if" part follows from results proved in [3] plus Facts 1. 5 and 1. 6. For the "if" part, we will rely mainly on Facts 1. 5 and 1. 6, and Lemma 1.10 and use induction on n. We will state the induction assertion, Pn, as a lemma. Lemma 4. 5. (= induction assertion, Pn). Whenever >1 and and for all 7T 2 ?' 132 Pn > 12 >a >a Pl + P2 + k E {0, 1, 2, P1 + P2 + > >a P-->- > - + a2 -E. Tr 2 ak ... , n} and Pn = +a2 + an' then there is a lower triangular matrix T conjunctive with 100 diag (el, e.432, and having diagonal entries , eial eia2 e ian in this order. Proof of Lemma 4. 5. P1 is trivially true, so suppose n> 2 Let pi., i3', pn, a 1, az, ..., a n be as given and Pn-1 is true. in the hypothesis of Pn and let D = diag Note that {i {1, 2, a1 > 1 n-1 2 ... , n-1} : (ei131, el-132, + a)n > (p 2 n-1 , + p ), e''). so the set (pi + pi+1) } is nonempty. Let j be the ai > minimum of this set. Case 1. j = 1. Here a D[1, 2] = so we have a conjunctivity (see Lemma 1.10) 1 e1 0 eicel L2p with p = isin 2(131 - 132 - Let 81 = 2 P2) . si31- n 2 ( a1 - 0 e1(P1 + P2 - al) Pi2 "2 ) , (which is real because (2 al- pi - p2) = 2 (pi - ai)_> 0). 2 = p1 + p2 - al , and 6. = p. for i > 3. Then (applying the [1, 21 -subconjunctivity indicated by the above conjunctivity to D gives us that) D is conjunctive with a lower triangular matrix B 16 i8) i61 en and such whose diagonal entries are (in order) e, e 101 that B(1) is diagonal. Also,-> 6 2 + P2 - al 52 = 6 + 2 52 + 53 > 6> >62 > 2 a2 a2 63 a3 + (P1 + P2 5k = > + al + (al a2 +. Pk) ak) = a2 + a3 +... + ak for all k e {2,3, n} and 62 + 63 +... +6n = a2 + a3 +... + an. , Thus, by Pn_1, B(1) is conjunctive with a lower triangular matrix T1 of diagonal (e1a2 , elan) and hence B itself is conjunctive e1a3 with T = reial G 01 T1.1 which satisfies the conclusion of Pn. Case 2: j > 1 and a. < 13. . J Again we have (j < n and) arg diag D[i, i+11 = (pi, R D (ay Pi 4- Pj+1 ) a.) and hence B, a lower triangular matrix whose arg diag is (61, 62, . with 63. = a 3., 6i41 Pj aj ' P P j --I- 1 and 6. = 13. for i {j, j+1}, and the only (possible) nonzero ,6) 102 off-diagonal entry of 13 is permuting -1- 1, 2, Now . B+1 apply the cyclicly- j] -subconjunctivity to B and get a lower , triangular matrix e c= of arg diag = y2, 0 LG B(j) = (a), 5 8 ,/.8 . 2' ) with C(1) = B(j) diagonal. Also Nj T y 3+. 2 + P2 +" = + a1 Y2 + Y3 Yk Pk-1 +... + ak-1 for k < j a2 Pj-1) = (P1 +P2 + 3+2 (3* "." 2, T :3 Then2 Cr-2> 3° >> T n) > . +. . A = (al, a2, ... , aj , n > -11-- and -1.--r> T 2 2 + .yk> Tz Y 2 + Y3 + Yn = (P1 + P2 + . +k +... a1+a2-F... 3 > Tn> - >T> E .= 2 and) {2, 3... , }. and aj + Pn) A +oz.+ for k > j , an). T3 +... + Tk for k N2 + aj) Pk) - ai +a2 +... + (T j+1 + Pk) (P11 + P2 Let (Pj , .. +an 103 = T2 + T3 Tn. Tht.s we can apply Pn-1 to C(1) and get it conjunctive with a lower triangular matrix T2 of arg diag = (T2, T3, Tri) = ( ae A a., an). Applying to C the corresponding (1) -subconjunctivity, we get T= e G1 0 T2_, By successively applying the [1, 2], [2,3], ..., j] subconjunctivities given by Lemma 1.11 (for the [1, 2], ... , [ j-1, j] principal submatrices, respectively), we get a matrix conjunctive with T which satisfies the conclusion of Pn. Case 3. (j > 1 and ) a. > J J Here, by the minimality of j, we have a.j-1 < 2 j-1 +p.), j SO a,j <(pJ-1 + 13.). Thus 3 arg diag D j-1, j] = (pi_i, 13i) ---> + pi - ai, ai) and hence B, a lower triangular matrix of arg diag = (Sr 82, . Pj+1' (Pr Pj-1 Pj aj' Pn) and such that the only nonzero off-diagonal entry ofB is J, J- 1 he 104 rest is pretty much as in Case 2, n(J) C= B ia Arg diag C = eyl . e3 LG by the cyclic-permuting 0 n] -subconjunctivity. [j, j+1, 'Z' ..., yn) =(61, ..., on, ai) with , C(n) = B(j) diagonal. Also Y1 + Y2 + + Yk = P1 +132 + >a1 +a2 + Y2 + +... +ak for k < j-1 , Pj-2) + (Pj-1 + P. Yk = 431 +P2 + 1 +a2 +... +a. j-1 Y1 + Y + + Yk = (P1 + f rk=j-1, and o>a (P j-1 (13j-i + Pk+1) = (31 + P2 + Pk+1) cri + a2 + aj) C\r. + + Pj aj) aj + ak+i for n> k > j (and equality holds for k = n-1). Thus let (T1, I 2' Tn-1) = (al, aj' an). 105 Then y + y + yk +. for k E + T 2 +. T1 . n-1) with eqqality for k = n-1. Also Tr T 2 1r.. 2 > n-1 > > and 2 1n-l> Yj-2 = Pj-2> 13j-1> 15j-1 =01. Thus, by P1, C(n) Tr Tr /'2 2 >- = (aj , because 13j) = 6j-1 = Yj-1 + [3.) - a.> a.> 13.> f3. +1 =y B(j) is conjunctive with T2, lower triangular A with arg diag = (T1, ) = (a j an). Apply to C the corresponding (n) -subconjunctivity gives us [ T2 T= 0 ia. which is again conjunctive with, by G1 repeated application of Lemma 1.11, a lower triangular matrix satisfying the conclusion of P. We invite the readers to compare Theorem 4.12 and the following result, Theorem 4.16. Theorem 4.16. Let B be an n x n accretive matrix and A be an n x n complex matrix. If A is with B, then A is stable. Proof: Let B be an n x n accretive matrix and A be an n x n complex matrix. Suppose A is With B. Then there exist 106 nonsingular matrices C and D such that C-1AC= D* BD. Since B is accretive, B + B is positive definite. * * * Now * D BD + (D BD) = D BD +D B D=D (B + B )D is also positive definite, so D BD is also accretive. Since every accretive matrix is a stable matrix, C-1AC is stable and hence so is A. Let B = diag (e Theorem 4.17. 2 >p >-.2 >p >p 2 with eigenvalues tle positive and jr- > 2 Then a1 ia1ia t2 e > Let A be an n x n complex matrix 2 tn e + a2 + . +13k > 131 +132 + ian with t . , t2' t n > a> - Tr. Let A be 14,4 with B. . 2 eiPn) with e I, . . + ak for all k {1, 2, . and equality holds for k = n. Let B = diag (e .431 , Proof: Tr 2 >p>p >. . . > 1 with eigenvalues pn ia 1 t1 e t2e 1(32 , , e ipn ) with Let A be an n x n complex matrix > - Z. ia , e 2, ... , tne ian with t1, t2, ..., tn >Let A be ilfra6 with B. 2 > an Then there exist nonsingular matrices C and D such that positive and IT- > a1 2 &2 a2 > n} 107 C -1 AC = D * BD. Then by Schur's Triangularization Theorem there exists a unitary matrix U such that * U-1 (C-1AC) U = U (D BD) U = T is a lower triangular matrix with ia ia diag T = ( t le 1, ian ten). ... , 2, t2e Since T is conjunctive with B, by Theorem 4.15 + for all k E Pk P2 + {1, 2, al + a2 + ak ... , n} and equality holds for k = n. We could extend the following result, Theorem 4.18, in the obvious way to the case where A is not assumed stable. The increased intricacy needed to word this extension would however tend to obscure the content, so we satisfy ourselves with the statement below. Theorem 4.18. 2 > > 13 1 p2 >. . . >p Let B = diag (e 11. ia positive and 2> a1 > a2 ia 1, > .. e Pi131 2 , e n) with Let A be an n x n complex matrix with > - -2. distinct eigenvalues tle , t2e . > 2' n ... , > - 3-r-. ix n with . t tne Then A is t 2' with B 108 k E{i., . , . + ak for all + a2 + +pk > if and only if pi +132 + n} with equality holding for k = n. Proof: The "only if" part follows from Theorem 4.17. ("If"). 2 i132 Suppose B = diag (e > pn > - >p >p 1 1_2_> e , ia with distinct eigenvalues tle tl' t , t positive and > 2, t2e a1 ) with > a2 , tne ian with > a> - > + ak for all k E {1, 2, ...n} and + a2 + +13k> (31 + [32 + iP n ia 1, 2 e , Suppose A is an n x n complex matrix . 2 , equality holds for k = n, then by Theorem 4.15 there exists a lower triangular n x n complex matrix T conjunctive with B and having al diagonal entries tie ia1 eigenvalues tie ia 1 t2e , 2e ian ia2 tne ia 2 ian , A is similar to T and so tne , . Since A has distinct . A is Ike" with B. ip Theorem 4.19. >P1 > P2 ?- with t Pn t' tn 2' Let B = diag (e >- 2 . , e LetA = (tie, t2e positive and Tr > 1> a2 , , e ipn ) with ia ia2 . > a >- tea) , 2 109 be the set of all n x n matrices similar to lower triangular Let matrices of diag = A . Then there exists A E OA which is with B if and only if pi + p2 + . al Pek a2 + for all k E {1, 2, ... n} with equality holding for k = n. Proof: The proof is a slight modification of the proof of Theorem 4.18. Theorem 4.20. Let r3 . r+1' 3r+2' , (3n t (i) (32, 2 , and B be conjunctive with pr by 13 and by pl. Let A have eigenvalues 1 ia2 (tie > (3 >p, > - (Denote (3 . n- r 2 , t2e . . , te t2" tn positive. 4. (kf) for all k E ) with 2 > a.i > a2 > . . . > a > -2 and Let A be .,141..e.. with B. Then a2 + {1, 2, ..., n) and and equality holds for k = n; ia. A has only linear elementary divisors at t.e if a. p or = p.17 at each eigenvalue of sgn ei(3, A has no more than n - r 110 elementary divisors, or at each eigenvalue of sgn e ipt , A has no more than r elementary divisors. Proof: Let > 2 p> p' > - 2 and B be conjunctive with ia ia2 eipr 9 with Tr > Let A have eigenvalues (tle i eip In-r . > a2 > . . . A be xidne with B. > a> - Tr 2 and t t $ , ...,tne t2 e Let tn positive. 2' n) Then follows immediately from Theorem 4.17. Let a = p. Since p > that % = a2 = tl' t2' .. t2 = p. = then rp + (j-r) p,' < > a1 + a2 > . . . > a. > . . > an it follows j > r, We also have j < r.ifFor, contradicting (i). Since . . . tn are positive, it suffices to assume = 1. Hence diag T[1, 2, ..., j] = (e (3, e ei43), 13, (same matrix T conjunctive with B as in Theorem 4.17). We claim that T [1, 2, ..., j] = diag (eiP, suffices to show that T[p,q] = diag (eiP, eiP) , ern). It 1<p<q<j For if T [p,q] is nondiagonal lower triangular, say . 111 eip 0 -2y 11 with y >0, then by Lemma 1.10 there is a conjunctivity having the effect [1 e16 0 eip eip 2y L 2,4 2 1 for all real 6 >0 such that 0 I sin 6 _ sin26 . and < 42 sin2 + cos 6 >0. Thus, by Fact 1. 4, ITT) contains ei(P+6) Since ei(P+6) is outside the r(B) contradicting sector r (B) for all sufficiently small 5 > 0, r(T) the fact that T is conjunctive with B. Thus T [p,q] must be a Now, since A is similar to T, A has only linear i« divisors at t.e if ai = 13. Similarly, if a -= P , then diagonal matrix. e Jlementary J a. = a. = . . . = . . = pi and r < j < n as al. > a2 > ... > a > pt. Thus an by the same argument we can show that T [ j, j+1, J . , n] = diag (eiPl, and eiPI , . I e143 ) , J if a. = pi. J (iii). To show that A has no more than n - r elementary divisors at each eigenvalue of sgn eiP, it suffices to show that T has the same property. Since - ieiPB* = Or = Or -i(eiO3-(31) - e-ig3-131)) 9 [ 2 sin (p-p1)] In-r In-r 112 and sin (13- p' . > 0, - . B- B * is a matrix of rank = n - r. r and T has m elementary divisors at some Now suppose m> n eigenvalue teia of T of arg = a, with 3> a> and t > 0. For the sake of convenience, let X = teia. (We denote nsp (M) as null space of the matrix M). So the number of elementary divisors of T at X = dim (nsp (T - X I) = dim (nsp (T ) ) = m. * be an orthonormal basis for nsp (T --RI) and m Let xl' x2' xl, x2, ..., xm, xm+i, ..., xn be an orthonormal basis for Cnx1 . Let U = [x1 x2 ... xn] = [U1 U2] , where U1 = [x1 x2 . X] m is n x m. We note that U is unitary, UU 1 [U,U2] = U2 U1 U2 Thus U T U U1 U1 [Ul* - U1 U2 Im U2 U2 0 T [Ui U2] U2* U, U2 [T 'U1 T U2] U, [T U, U1 T U2] I=i In-m 113 T1 X ww10 Im S* 0 X T1 Im Hence U TU =S T1 where G S[m+1, m+2, . , n I 1, 2, ..., m] and T1 = S[m+1, m+2, w". Thus ie-iPS - ieil3S* 2 sin (B-a) Im GieI -ie = . 1 Since 0 < - a< Tr, sin (13- > 0. So the rank of ie-iPS must be at least m. Since m > n-r and the rank of ie-if3B is n-r, .ip S - ie iR ieiPS* iei1313* cannot be conjunctive with ie PB-ie PB * Hence S cannot be conjunctive with B. For if S were conjunctive with B, then there would exist a nonsingular matrix D such that D BD = S, and thus we would have . -ipS - ie ipS . . = (D BD) ie = (D *BD) i.eie iB (D BD) (DBD) * 114 =D (ie - ieipB*) D hence would be conjunctive with le -ip B - ieipB*. Therefore T can. not be conjunctive with B, this is a contradiction. Hence A has no more than n-r elementary divisors at each eigenvalue of sgn eir3 Similarly, we can show that A has no more than r elementary divisors at each eigenvalue of sgn e'. We summarize our main results in Chapters III and IV in the following Tables 3, 4 and 5. . Table 3. n x n complex matrices. B is CAtri to A A nonzero scalar complex complex unide finite > A B E7-1 (trB ) A Reference B`eA Theorem 4.1 2 Theorem 4. 2 no Theorem 3.5 Real diagonal, positive distinct complex BE nonsingular, nonsingular cohermitian B =Band nonsingular {(A, B), (-A, - B))CP x ciD 3 ch. roots similar to real Theorem 3.6 sig B = sig A diagonal matrix Nonsingular, real diagonal, distinct ch. roots nonsingular, nonsingular contrahermitian nonsingular, nonsingular contradefinite nonzero ch. roots are distinct, has only linear elem. divisors at 0. Theorem 3.7 codefinite with ch. roots real and distinct distinct ch. roots 3 contrahermitian -{(A, B), (-A, -B)} C p2 x p3 Corollary 3. 7.1 codefinite -irregular sgn (det A) = sgn (det B) <=> rank B = rank A Theorem 4. 3 Theorem 4. 5 Table 3. Continued. A B nonzero ch. roots ,X are X 1,X 2, -regular contradefinite distinct, has only linear elem. divisors B is efriZ to A Reference rank B r and (B) = sgn (X IX 2...X r) n Fact 4.1 at 0 idempotent of rank r, complex 0< r< n 0 a 9 13 ED I On-r-s, Theorem 3. 8 S for s < r [a s is-is air ol [Is In-r complex ED a I 1 13, 0< r < n y r-s ED 13 In -r-s 13_1 Theorem 4. 4 s < min {r,n-r} Yy r 2' " Ys Positive nilpotent rank = n-1 complex Brestrictly lower triang. matrix of rank n-1 nonderogatory, with single ch. root X 0, n > 2 complex B?dlower-triang. nondiagonal Theorem 4. 6 matrix with diag= (X , , ...X ) B EX p 3 but B N.p Theorem 3.12 Table 3. Continued. B is CAut, to A A Reference normal hermitian A'eB Theorem 3. 4 complex hermitian A oehermitian matrix of same rank and signature as B Theorem 3. 3 unitary nonsingular < hermitian involutory > A is a symmetry and Theorem 4. 9 index A = index B <=fr A'ZIB unitary B is a symmetry and Theorem 4.10 index B = index A <=4 13/A accretive complex A is stable with distinct ch.ia lai roots(tle )..., t1, t2, tea), , tn positive, >a>a>...>a >- 2 7r Tr diag. (e IT >@_>B > 2- rir2 Theorem 4.16 ip i@ , ... , e >a, >--Tr 2 n) #=? Pi+1324-- -hk?-0i+a2+. for all k E {1, 2, +Cek ... , n} and equality holds for k=n Theorem 4.18 Table Continued. B is A stable c er e e ip' in-r, Tr 2 C/frj,1 to A Referen.c,, A has only linear elem. divisors at teia if a = p or = pi i(1) (2) At each ch. root of A has no more than n-r elem. divisors, or at each ch. root of sgn eip, sgn e', A has more thanr elem. divisors. no Theorem 4. 20 Table 4. Matrices similar to triangular matrices.(Let (x' , 2 n n x n nonsingular matrices similar to triangular matrices of diag = A a,_ complex X.>0, i--=-1, 2, ... , n 1 X 1 i --, \{0}, 1, 2, <=> B e cp3 nonsingular cohermitian ... , n ,( nonsingular x. e GR\{0}, i = 1, 2, B is CM/ to some A E B .(=;, contrahermitian ... , n , 1j31ii3n code finite be the set of all ) and . ) Reference Theorem 3. 9 B is hermitian, A is diagonable, and sig B = sig A Theorem 3.10 either sig A = n and BEP 3 or else sigh= -n and B Theorem 3.11 ia. X = t.e i t i=1, 2, ... Tr 2 >c11>a2. , t.>0, 1 diag(e , ... e ), ,n, <=> Pi+P)2+. +Pikai+az+. for all k e {1, 2, Tr >an>-2 Tr >3 >. 2 F1-1-2 Tr >f3 >-n 2 ... ,n} and equality holds for k = n. Theorem 4. 19 Table 5. Miscellaneous results. A Conditions Reference nonsingular There is a unitary matrix B such that B is both CA.-a, to, and 4111/C with, A Theorem 4. 8 stable There is an accretive matrix B such that B is both CALL to, and _14/2C, with, A Theorem 4. 12 stable Hermitian matrix H is the hermitian part of a Theorem 4.14 matrix B Cm,L to A < > index H > max. no. of Jordan blocks at any one eigenvalue of A 121 BIBLIOGRAPHY C. S. Ballantine, Triangularizing Matrices by Congruence, Linear Algebra and Its Applications, 1(1968), 261-280. C. S. Ballantine, Products of Positive Definite Matrices. Linear Algebra and its Applications, 3(1970), 79-114. IV, C. S. Ballantine and C. R. Johnson, Accretive Matrix Products, Linear and Multilinear Algebra (to appear). S. Barnett and C. Storey, Analysis and Synthesis of Stability Matrices, J. of Differential Equations, 3(1967), 414-422. M. -D. Choi, Adjunction and Inversion of Invertible HilbertSpace Operators, Indiana University Mathematics Journal, Vol. 23, 5(1973), 413-419. A. W. Gillies, The Physical Content of Autonne's Lemma, Siam J. Applied Mathematics, Vol. 19, 1(1970), 142-143. K. Hoffman and R. Kunze, Linear Algebra, 2nd ed. Prentice Hall, 1971. D. G. Hook, Effects of Conjunctivity on the Inertia of Complex Matrices, Ph. D. Thesis, Oregon State University, 1974. I. J. Katz and M. H. Pearl, On EPr and Normal EPr Matrices, Journal of Research of the National Bureau of Standards (1966), 47-77. M. Marcus and H. Minc, A Survey of Matrix Theory and Matrix Inequalities, Allyn and Bacon, Boston, 1964. S. Perlis, Theory of Matrices, Addison-Wesley, Reading, Mass. H.d n 2 1958. Schneider and G. P. Barker, Matrices and Linear Algebra ed. , Holt, Rinehart and Winston, Inc., New York, 1973. P. Stein and A. Pfeffer, On the Ranges of Two Functions of Positive Definite Matrices II, ICC Bulletin, 1967, vol. 6, 81-86. 122 14. 0. Taussky, A. Remark on a Theorem of Lyapunov, J. Math. Anal. and Appl. 2(1961), 105-107. APPENDIX Page Index to Lemmas, Theorems, Corollaries, Facts, Table 1, p. 2, p. Definitions, Examples and Tables Theorem 1. 1, p. 10 Theorem 4. 11, p.96 4. 12, p.96 1. 2, p.11 61 62 3, p. 115 4, p. 119 5, p. 120 Fact 1. 1, p. 1. 2, p. 1. 3, 1. 4, 1. 5, 1. 6, Fact 4. 1, Lemma 1. 5 6 6 p. p. 7 p. 9 Theorem p. 10 p. 84 1, p. 1. 2, p. 1. 3, 1. 4, 1. 5, 1. 6, 2 2 p. 2 p. 2 p. 3 p. 1. 7, p. 3 1. 8, p. 1. 9, p. 123 p. 11 p. 13 p. 14 p. 14 p. 14 p. 14 I. 16, p. 15 Lemma 3. 1, p.67 Lemma 4. 1, p.85 4. 2, p. 87 4. 3, p. 89 4. 4, p.98 4. 5, p.99 1. 7, p. 15 p. 15 p. 16 1. 8, p.16 2. 2. 2. 2. 2. 2. 2. p. 19 p. 19 p. 21 p. 21 Theorem 3. 1, 2, 3, 4, 5, 6, 7, 1, 3. 3. 3. 3. 3. 3. 2, 3, 4, 5, 6, 7, 3. 8, 3. 9 , p.23 p. 24 p.36 p.64 p.64 p.65 p.65 p. 65 p. 66 p.68 p.70 p. 74 3. 10, p. 74 3. 11, p. 75 3. 12, p. 75 2. 2, p. 22 2. 3, p. 26 2. 4, p. 26 2. 5, p. 27 2. 9, p. 35 1. 5, 1. 6, p.37 p.40 2. 10, p.42 2. 11, p.45 2. 12, p.54 2. 13, p.55 Lemma 2. 1, p. 20 2. 6, p. 27 2. 7, p. 30 2. 8, p.32 p.15 p.15 2. 8, 2. 9, 4 4 4 1. 10, 1. 11, 1. 12, 1. 13, 1. 14, 1. 15, 1. 3, 1. 4, Theorem 4. 4. 4. 4. 4. 4. 4. 1, 2, 3, 4, 5, 6, 7, 4 . 8, p. 77 p. 77 p. 78 p. 78 p. 83 p.91 p.93 p. 94 p.94 4. 10, p.95 4. 4 9, 4. 13, p. 9 7 4. 14, p.98 4. 15, p.98 4. 16, p. 105 4. 17, p. 106 4. 18, p.107 4. 19, p. 108 4. 20, p. 109 Sublemma 4.3. 1, p. 89 Corollary 2. 3. 1, p. 21 Corollary 2. 7. 1, p. 36 2. 8. 1, p.38 2.8. 2, p.39 2.9. 1, p. 41 2. 10. 1, p.45 2. 11. 1, p.50 2. 12. 1, p.54 2. 13. 1, p.57 Corollary 3. 7. 1, p.69 Corollary 4. 4. 1, p. 82 Definition 1. 1, p. 2 1. 2, p. 2 Example 1. 1, p.3 Example 4. 1, p. 9 7