COLUMN RANKS AND THEIR PRESERVERS OF GENERAL BOOLEAN MATRICES

Transcription

1 J. Korean Math. Soc. 32 (995), No. 3, pp COLUMN RANKS AND THEIR PRESERVERS OF GENERAL BOOLEAN MATRICES SEOK-ZUN SONG AND SANG -GU LEE ABSTRACT. We show the extent of the difference between semiring rank and column rank of matrices over finite Boolean algebra. We also obtain the characterizations of linear operators that preserve column ranks of general Boolean matrices. Introduction There is much literature on the study of matrices over a finite Boolean algebra. But many results in Boolean matrix theory are stated only for binary Boolean matrices. This is due in part to a semiring isomorphism between the matrices over the Boolean algebra of subsets of a k element set and the k tuples of binary Boolean matrices. This isomorphism allows many questions concerning matrices over an arbitrary finite Boolean algebra to be answered using the binary Boolean case. However there are interesting results about the general (i.e. nonbinary) Boolean matrices that have not been mentioned and they differ somwhat from the binary case. In many instances, the extension of results to the general case is not immediately obvious and an explicit version of the above mentioned isomorphism was not well known. In [4], Kirkland and Pullman gave a way to derive results in the general Boolean algebra case via the isomorphism from the binary Boolean algebra case by means of a canonical form derived from the isomorphism. In [2], Beasley and Pullman compared semiring rank and column rank of the matrices over several semirings. The difference between semiring rank and column rank motivated Beasley and Song to investigate of the column Received June 0, 994. AMS Classification: Primary 5A03, 5A04. Key words and Phrases: Boolean column rank, linear operator. This research was supported by the Research Fund of the Ministry of Education, Korea in 993, and in part by TGRC-KOSEF

2 532 Seok-Zun Song and Sang -Gu Lee rank preservers of matrices over nonnegative integers [3] and over the binary Boolean algebra [5]. In this paper, we will show the extent of the difference between semiring rank and column rank of matrices over a general Boolean algebra and we also obtain characterizations of the linear operators that preserve column ranks of general Boolean matrices. Let B be the Boolean algebra of subsets of a k element set S k and σ, σ 2,...,σ k denote the singleton subsets of S k. We write + for union and denote intersection by juxtaposition; 0 denotes the null set and the set S k. Under these two operations, B is a commutative, antinegative semiring (that is, only 0 has an additive inverse); all of its elements, except 0 and, are zero-divisors. Let M m,n (B) denote the set of all m n matrices with entries in B. The usual definitions for adding and multiplying matrices apply to Boolean matrices as well. For each m n matrix A over B, thep-th constituent [4] of A, A p,isthe m n(0,)-matrix whose (s, t)-th entry is if and only if a st σ p. Via the constituents, A can be written uniquely as σ p A p, which is called the canonical form of A. It follows from the uniqueness of the decomposition, and the fact that the singletons are mutually orthogonal idempotents, that for all m n matrices A, all n r matrices B and C,and all α B, (a) (AB) p = A p B p, (b)(b +C) p = B p +C p, (c) (α A) p = α p A p for all p k. 2. Boolean rank versus Boolean cloumn rank The Boolean rank, b(a), of a nonzero A M m,n (B) is defined as the least index r such that A = BC for some B M m,r (B) and C M r,n (B). The rank of zero matrix is zero; in the case that B = B ={0,}, we refer to b(a) as the binary Boolean rank, and denote it by b (A). For a binary Boolean matrix A, we have b(a) = b (A) by definition. If V is nonempty subset of M r, (B) that is closed under addition and multiplication by scalars, then V is called a vector space over B. The concepts of "subspace" and of "generating sets" are defined so as to coincide with familar definitions when B is a field. We ll use the notation < F > to denote the subspace generated by the subset F of V. As with fields, a basis for a vector space V is a generating subset of least cardinality. That cardinality is

3 Column ranks and their preservers 533 the dimension,dim(v),ofv. Since B is canonically identified with the subsemiring {0, } of B, a binary Boolean matrix can be considered as a matrix over both B and B. The Boolean column rank, c(a), ofa M m,n (B)is the dimension of the space < A > generated by the columns of A. In the binary Boolean algebra, we denote it by c (A) for A M m,n (B ). (2.) It is known [2] that for all m n matrices A over B, 0 b(a) c(a) n. (2.2) For any p q matrix A over B, the Boolean rank of A 0 is b(a) 0 0 and its Boolean column rank is c(a). (2.3) The Boolean rank of a matrix is the maximum of the binary Boolean ranks of its constituents. ([4]) Let us start with the relationship between c(a) and c (A) for a binary Boolean matrix A when considered in B and B. LEMMA 2.. For any binary Boolean matrix A, we have c(a) = c (A). Proof..SinceB can be considered as a subsemiring of B,wehavec(A) c (A). Conversely, if c(a) = r, then there exists a basis B ={x,x 2,...,x r } of the column space of A such that each x i is a linear combination of the columns of A over B. Then the p-th constituents (x ) p,(x 2 ) p,...,(x r ) p can generate all columns of A p over B. Hence c (A p ) r. But A = A p for all p, soc (A) r =c(a). Let µ(b,m,n)be the largest integer γ such that for all A M m,n (B), b(a) = c(a) if b(a) γ. (2.4) It follows from definition of µ and (2.2) that if c(a) >b(a)for some p q matrix A, then µ(b, m, n) <b(a)for all m p and n q. Beasley and Pullman determined the value of µ on B in [2] as follows; LEMMA 2.2. ([2]) if min(m,n) = µ(b, m, n) = 3 if m 3 and n = 3 2 otherwise. Now we determine the value µ for a nonbinary Boolean algebra B.

4 534 Seok-Zun Song and Sang -Gu Lee LEMMA 2.3. If B is any Boolean algebra and m 3 and n > 3, Proof. Let µ(b, m, n) 2. [ 0 ] 0 A = Then c(a) = c (A) by Lemma 2.. Since the four columns of A consitute a basis for the column space of A over B, c (A) = 4. Thus c(a) = 4. But b(a) min{m, n} =3.The result follows from the property (2.4). LEMMA 2.4. If c(a) = r and σ p A p is the canonical form of A M m,n (B) then max{c (A p ) p k} r. Proof.. Assume that c(a)=r. Then there exists some basis {x,...,x r }for the column space of A. Then the pth constituents (x ) p,...,(x r ) p can generate all columns of A p over B. Therefore c (A p ) r for all p. We remark that the inequality in Lemma 2.4 may be strict for r > as shown in Example 2. below. EXAMPLE 2.. Let A = [ σ σ ] 0 be a matrix over the nonbinary Boolean algebra of the k element set S k, where σ is a singleton subset of [ S k.thenc(a)=3 ] by the proof [ of Theorem] below. But c (A ) = c ( ) = 2andc 0 (A p )=c ( ) = 0 2forallp=2,3,...,k. LEMMA 2.5. If B is any Boolean algebra and m and n, then µ(b, m, n). Proof..Ifc(A)=thenb(A)= by the property (2.). If b(a) =, then A can be factored as xa t where a t = [a,...,a n ] M,n (B) and x M m, (B). Thus the column space of A can be generated by one column vector x. Soc(A) =b(a). On the other hand b(a) c(a) by (2.).

5 Column ranks and their preservers 535 THEOREM 2.. For a nonbinary Boolean algebra B, { 2 if 2 = n m µ(b, m, n) = otherwise Proof. By Lemma 2.5,µ(B,m,n) for all positive integers m and n. Let σ B be a singleton subset of S k. Consider the matrix A = Then the column space of A is V = { σ σ σ x + y + z + w 0 σ σ σ. 0 Let be any subset of V generating V. Let If a, then a= x+y+w y+w a= [ σ } z,w B, and x, y < σ >. ]. for some w B, x, y < σ >. Now = y + w and y = 0orσ,sothatw=or σ (which is the complement of σ ).Butthenσ =x+y+w=or σ, a contradiction, since σ. Hence a. Let If b, then b= [ x+y+zσ y+z b =. ] for some z B, x, y < σ >.

6 536 Seok-Zun Song and Sang -Gu Lee Since = y + z and y = 0orσ,wehavez=or σ. But then = x + y + zσ = zσ ( σ ) or σ, a contradiction, since σ. Hence b. Let c = [ σ 0 Then c would imply that 0 = y + z + w for some y, z and w, one of which is nonzero, which is impossible. Hence {a, b, c}. Therefore c(a) = 3. But b(a) 2 by definition. Thus µ(b,m,n) whenm 2 and n 3, by property (2.4). Hence µ(b, m, n) = form 2andn 3 by Lemma 2.5. Evidently µ(b, m, ) = forallm.if = m n, then the fact that < x, x 2,...,x n >=< n i= x i >implies that µ(b,, n) =. If 2 = n m, thenc(a)=2 whenever b(a) = 2 by property (2.). Thus µ(b, m, 2) = 2form 2 by Lemma 2.5. ]. 3. Linear operators that preserve cloumn rank of the nonbinary Boolean matrices In this section, we obtain the characterizations of the linear operators that preserve Boolean column rank of the nonbinary Boolean matrices. A linear operator T on M m,n (B) is said to preserve Boolean column rank if c(t (A)) = c(a) for all A M m,n (B). Itpreserves Boolean column rank r if c(t (A)) = r whenever c(a)=r. For the terms Boolean rank preserver and Boolean rank r preserver, they are defined similarly. If T is a linear operator on M m,n (B), for each p k define its p-th constituent, T p,byt p (B)=(T(B)) p for every B M m,n (B ). By the linearity of T,wehaveT(A)= σ p T p (A p )for any matrix A M m,n (B). Since M n,n (B) is a semiring, we can consider the invertible members of its multiplicative monoid. The permutation matrices (obtained by permuting the columns of I n, the identity matrix) are all invertible. Since is the only invertiblemember of the multiplicativemonoidof B, the permutation matrices are the only invertible members of M n,n (B). LEMMA 3.. The Boolean column rank of a Boolean matrix is unchanged by pre- or post-multiplication by an invertible matrix.

7 Column ranks and their preservers 537 Proof. This follows from the fact that an invertible matrix is just a permutation matrix. LEMMA 3.2. Suppose T is a linear operator on M m,n B). If T preserves Boolean column rank r, then each constituent T p preserves Boolean column rank r on M m,n (B ). Proof. Assume that A M m,n (B ) is a binary Boolean matrix with c (A)= r. By Lemma 2., we have c(a) = r and c(σ p A) = r for each p =,...,k. Since T preserves Boolean column rank r, c(t (σ p A))=r.But c(t(σ p A)) = c(σ p T (A)) = c(σ p σ i T i (A i )) = c(σ p T p (A)) for each p. Therefore c(σ p T p (A)) = r for each p =,..., k, and hence c (T p (A)) = r. LEMMA 3.3. Suppose T is a linear operator on the m n matrices over B. If each constituent T p preserves binary Boolean rank r, thent preserves Boolean rank r. Proof. Let b(a) = r for A M m,n (B). Then there exists some p such that b (A p ) = r and b (A q ) r for q k by property (2.3). Thus b (T p (A p )) = r and b (T q (A q )) r for q k. Since b(t (A)) = max{b (T p (A p )) p k} by property (2.3), T preserves Boolean rank r. Now we need the following definitions of linear operators on the m n matrices over B. For any fixed pair of invertible m m and n n Boolean matrices U and V, the operator A UAV is called a congruence operator. Let σ denote the complement of σ for each σ in B. For p k,we define the p-th rotation operator, R (p),onthen nmatrices over B by R (p) (A) = σ p A t p + σ p A, where A t p is the transpose matrix of A p. We see that R (p) has the effect of transposing A p while leaving the remaining constituents unchanged. Each rotation operator is linear on the n n matrices over B and their product is the transposition operator, R : A A t. i

8 538 Seok-Zun Song and Sang -Gu Lee EXAMPLE 3.. Let A = [ 0 0 ] 0 σ σ 0 be a matrix over B. Then c(a) = 3 by Example 2. and property (2.2). But R () (A) = A t, the transpose matrix of A, has Boolean column rank 2. Consider B = A 0 n 3,n 3 for n 3. By property (2.2), the rotation operator does not preserve Boolean column rank 3 on M n,n (B). LEMMA 3.4. ([4]) If T is a linear operator on the m n matrices (m, n ) over a general Boolean algebra B, then the followings are equivalent. () T preserves Boolean ranks and 2. (2) T is in the group of operators generated by the congruence (if m = n, also the rotation ) operators. THEOREM 3.. Suppose T is a linear operator on M m,n (B) for m 3 and n. Then the following are equivalent. () T preserves Boolean column rank. (2) T preserves Boolean column ranks, 2 and 3. (3) T is a congruence operator. Proof. Obviously () implies (2). Assume that T preserves Boolean column ranks, 2 and 3. Then each constituent T p preserves binary Boolean column ranks, 2 and 3 by Lemma 3.2. For A M m,n (B), Lemma 2.2 implies that b (A) = c (A) for b (A) 2. Thus T p preserves binary Boolean ranks and 2. Then T preserves Boolean ranks and 2 by Lemma 3.3. So T is in the group of operators generated by the congruence ( if m = n, also the rotation ) operators. But the rotation operator does not preserve Boolean column rank 3 by Example 3.. Hence T is a congruence operator since T preserves Boolean column rank 3. That is, (2) implies (3). Now, assume that T is a congruence operator of the form T (A)=UAV, where U and V are invertiblem m and n n Boolean matrices respectively. Then T preserves Boolean column rank by Lemma 3.. Hence (3) implies (). If m 2, then the linear operators that preserve column rank on M m,n (B) are the same as the Boolean rank-preservers, which were characterized in [4].

9 Column ranks and their preservers 539 Thus we have characterizations of the linear operators that preserve the Boolean column rank of general Boolean matrices. References. L. B. Beasley and N. J. Pullman, Boolean rank-preserving operators and Boolean rank- spaces, Linear Algebra Appl 59 (984), L. B. Beasley and N. J. Pullman, Semiring rank versus column rank, Linear Algebra Appl 0 (988), L. B. Beasley and S. Z. Song, A comparison of nonnegative real ranks and their preservers, Linear and Multilinear Algebra 3 (992), S. Kirkland and N. J. Pullman, Linear operators preserving invariants of nonbinary matrices, Linear and Multilinear Algebra 33 (992), S. Z. Song, Linear operators that preserves Boolean column ranks, Proc. Amer. Math. Soc. 9 (993), S. Z. Song, Linear operators that preserve column rank of fuzzy matrices, Fuzzy Sets and Systems 62 (994), S. Z. Song, S. G. Hwang and S. J. Kim, Linear operators that preserve maximal column rank of Boolean matrices, Linear and Multilinear Algebra 36 (994), S. G. Hwang, S. J. Kim and S. Z. Sone, Linear operators that preserve spanning column ranks of nonnegative matrices, J.KoreanMath.Soc.3 (994), Seok-Zun Song Department of Mathematics Cheju National University Cheju , Korea Sang-Gu Lee Department of Mathematics Sung Kyun Kwan University Suwon , Korea