Ki Hang Kim and Fred W.

Roush
Alabama State University
Montgomery, Alabamu 36101

Submittedby Hans Schneider

ABSTRACT

For a Boolean matrix A, a g-inverse of A is a Boolean matrix G satisfying

AGA = A, and a Vagner inverse is a g-inverse which in addition satisfies GAG = G.
We give algorithms for finding all g-inverses, all Vagner inverses, and all of several
other types of inverses including Moore-Penrose inverses. We give a criterion for a
Boolean matrix to be regular, and criteria for the various types of inverse to exist. We
count the numbers of Boolean matrices having Moore-Penrose and related types of
inverses.

1. INTRODUCTION

Let p = { 0, l} be the Boolean algebra of order two with operations ( + , -)

definedonitasfo~ows:0+0=0~1=1~0=0~0=0and1+0=0+1=1+1=
1. 1= 1. We suppress the dot of a-b and simply write ab. By a Boolean
matrix we mean a matrix whose elements belong to p. Let I$ denote the set
of all such matrices. Then B, is a semigroup under multiplication.
These systems arise in several areas of mathematics. For example, a
binary relation p on an n-set X = {xi,. . . , xn} can be interpreted as a Boolean
matrix A = (aii), where aii = 1 if (xii, xi) Ep, and aii = 0 otherwise. In particu-
lar, the composition of two binary relations on X is represented by the
product of their associated matrices. Dually, a directed graph G, with vertex
set V={v,,..., v~}, determines and is determined by a member MG = (mii) of
B,,, where mii = 1 if there is an arc from vi to vi, and mii = 0 otherwise.
Plemmons [8], Kim [4], P. Rao [ll], P. Rao and B. Rao [12], and 8chein
[14] have previously found algorithms for finding particular g-inverses or
Vagner inverses. Plemmons [9] and P. Rao and B. Rao [12] have studied
some of the other types of inverses which occur in this paper.

*This research was supportedin part by AlabamaState UniversityFaculty Research Grant

R-78-6.

LINEAR ALGEBRA AND ITS APPLICATIONS Z&247-262 (1978) 247

0 ElsevierNorth-Holland,Inc., 1978 0024-3795/78/0022-0247$01.75
248 KI HANG KIM AND FRED W. ROUSH

We give algorithms (i) for finding all g-inverses of a Boolean matrix, (ii)
for finding all minimum-norm g-inverses of a Boolean matrix, (iii) for finding
all least-squares g-inverses of a Boolean matrix, and (iv) for finding all
Vagner inverses of a Boolean matrix. We give a criterion for a Boolean
matrix to be regular, and criteria for Boolean matrices to have minimum-
norm g-inverses, least-squares g-inverses, and Moore-Penrose inverses. We
count the number of Boolean matrices of specified rank and dimension
having one of these last types of inverses. These inverses have applications in
network and switching theory and information theory [3].

2. PRELIMINARIES

In this paper, matrices are denoted by capital letters A, B, C, etc., and

vectors by lower-case letters. Let V,, be the set of all n-tuples of elements of
/3. An element of V,, is called a Boolean vector of dimension n. We will
usually write 0 instead of (0,. . ., 0), since usually it will be clear from the
context what we mean.
A subset W of V, is called a subspace of V,, if (1) 0 E W, and (2)
u + ‘L)E W for all U, 0 E W. The subspace spanned by W is the smallest
subspace of V, which contains W. We denote the subspace spanned by W
by ( W). A vector v is said to be dependent on W if v E ( W>. A set W is
said to be independent if for all v E W, v is not dependent on W \{ v}, where
\ denotes the set-theoretic difference. Let W be a subspace. A subset S of W
is said to be a basis of W if (1) W=(S), and (2) S is an independent set.
Thus the basis of { 0} is 0.
By the row space of A E B,, we mean the subspace of V, spanned by the
rows of A; we denote it by R(A). By b(A) we mean the unique basis of
R (A); we call b(A) a row basis of A. If v E V,,, then by v, we mean the
element of p occurring at the kth position of v. For u and v of the same
dimension we say that u < v if ui = 1 only if vi = I for all i. For each v E b(A),
a vector u with one 1 is called an identificationvector v if and only if: u < w
holds if and only if v < w for w E b(A). Let I(v) denote the set of identifica-
tion vectors of a basis vector v of A. The cardinality of b(A) is called the row
rank of A. The concepts of column space, column basis, and column rank of
A are defined in a similar manner. A matrix A is said to be of rank r if (row
rank of A) = (column rank of A) = r. In general, the row rank and the column
rank of A need not be the same [S, 81. Hereafter, unless otherwise stated, we
consider vectors and matrices over j3.
If A E B,,, then G is said to be a generalized inverse of g-inverse of A if
and only if AGA = A. We call A nonsingular if A has a g-inverse and A has
rank n. Let A r denote the transpose of A. We summarize the definitions of
the various types of inverses in Table 1.
INVERSES OF BOOLEAN MATRICES 249

TABLE 1
DEFINITIONS OF INVERSES

g-inverse AGA=A
Minimum-norm g-inverse AGA=A, (GA)T=GA
Least-squares g-inverse AGA = A, (AG)T= AC
Vagner inverse AGA=A, GAG= G
Moore-Penrose inverse AGA = A, GAG = G,
IAGJT=AG, (GA)*=GA

The second and third inverses were named by P. Rao and B. Rao [12].
The vagner inverse was introduced by Vagner [15] and is of importance in
general semigroup theory. Other authors have called it semi-inverse [8],
Moore-Penrose inverse [4], and reflexive g-inverse [lo, 11, 121.

3. INVERSES AND REGULAR ELEMENTS

Two elements a, b of a semigroup S are said to be C- (CR-) equivalent if

and only if they generate the same principal left (right) ideal in S [2]. The
relation C n 3 is denoted by X, while the composition e ?ll,= % C is called
9. Thus equivalence relations are called Green’s relations on S. (For details,
see [2, Chapter 21 and [5]).
Elements which have g-inverses (or Vagner inverses) are called regular
[2]. The notion of regularity was first introduced by von Neumann [I for
rings. Every % -class consists either entirely of regular elements or entirely of
irregular elements. An element a is said to be an idempotent if a2 = a. A
g-class contains regular elements if and only if it contains an idempotent.
A E B,, is regular if and only if the row space of A forms a distributive lattice
[I71*

4. KNOWN-RESULTS

Only since 1971 has the field of study of g-inverses and Vagner inverses
been investigated systematically. The concept of inverse of matrix was first
introduced by Wedderburn [16] as early as 1934: G is said to be an inverse
of A if AC = GA = Z, where Z is an identity matrix. In 1952, Lute [S] showed
that A possesses a two-sided inverse if and only if A is an orthogonal matrix
in the sense that AA r= I, and that in this case A r is the two-sided inverse.
In 1963, Rutherford [13] showed that if A possesses a one-sided inverse, that
inverse is also a two-sided inverse. Furthermore such an inverse, if it exists, is
unique and is A r. It is easy to see that the permutation matrices are the only
matrices (over p) which have inverses.
250 KI HANG KIM AND FRED W. ROUSH

In 1971, Plemmons [8] gave the following method for finding the Vagner
inverses of a matrix: find idempotent E in the C-class containing A, and
solve the equation XA = E. In 1973, unaware of the earlier work of Plem-
mons [a], Kim [4] showed that if A is nonsingular, then the Vagner inverse of
A is the adjoint A* of A, defined by A* = [Ai J, where Aii is the cofactor of
uji in

Here the summation is taken in the Boolean algebra over all permutations u
of 1,.,., n. Thus ]A] is either 0 or 1.
In 1973, P. Rao and B. Rao [12] studied many types of inverses by
writing regular matrices in the form

where A, is a nonsingular matrix and P,Q are permutation matrices. They

showed that

Pi A,’ 0
[0 OQ
1

is a Vagner inverse of A if Al1 is a Vagner inverse of A,. They developed

criteria for existence of all five types of inverses, which can be shown
equivalent to ours, though we do not use this decomposition. They gave the
following method for finding a g-inverse: first delete dependent columns,
then dependent rows, until the remaining rows and columns are indepen-
dent. Then hunt for columns with a single 1 entry. The row and column of
the 1 entry are deleted and the search continued. Transposing the matrix of
the 1 entries occurring in the search gives a g-inverse, if one exists.
Recently Schein [14] has given a surprising closed formula for the unique
Vagner inverse which is entrywise greater than or equal to all other Vagner
inverses:

where C (complement) denotes the operation of replacing all 0 entries by 1

and all 1 entries by 0. He also gave algorithms for g-inverses and Vagner
inverses, which are related to this result.
INVERSES OF BOOLEAN MATRICES 251

5. ALGORITHMS

Let 4. (Aei) denote the ith row (column) of A. Let p(t)=inf{wE

R (A) : w > t}, where the infimum is taken in the lattice R (A).

1. Algorithm for Computing All g-Znverses of A

(1) Find b(A).

(2) FindZ( u)f or each u E b(A). A g-inverse will exist if and only if Z(v)
is nonempty for each v E b(A).
(3) For each v E b(A), choose a specific identification vector u E Z(u).
(4) For each such chosen U, choose a vector s (to be the image of U)
such that si = 1 only if Ai. < v and such that si = 1 for at least one i such that
Ai. = v.
(5) Choose any vector t with exactly one 1 entry other than the U’S
chosen in step (3), if t is not less than any row vector, and send t to an
arbitrary vector. Otherwise send t to a vector b such that bi =l only if
Ai. < p(t).
(6) Linearly order the set of vectors with only one 1 in such a way that
the mapping i+(&,, Siz,. . . ,a,) is an order isomorphism. Write the vectors s
and b in the order of the vectors u and t.

EXAMPLE1. Let A be

1 1
1 1 1
0 11.
1 0 0

(1) b(A)={@ 1 I),(1 0 O)}.

(2) I((0 1 I))={(0 1 o),(o 0 l)}; Z((1 0 0 ))={(l 0 0)).
(3) There are two ways to choose the U’S.
(4) For u = (1 0 0), s must be (0 0 1). For the other u chosen, s must
be (0 1 0).
(5) For the remaining l-element vector t, which will be one of (0 1 0),
(0 0 I), p(t) will be (0 1 1). The vector b can be either (0 0 0) or (0 1 0).
(6) There are 3 g-inverses of A:
252 KI HANC, KIM AND FRED W. ROUSH

II. Algorithm for Finding All Minimum-Norm g-Inverses of A

(1) Find b(A). A minimum-norm g-inverse will exist if and only if the
sets of l’s in the members of b(A) are disjoint, i.e.. for v,tu E b(A), u# w,
uj = 1 imphes wj =O.
(2) Let E be the matrix with eii = 1 if and only if for some u E b(A),
t$ = vi = 1.
(3) Any solution G of GA = E will be minimum-norm g-inverse of A. To
find such solutions. for each row of E. determine which linear combinations
of rows of A equal that row of E. One such solution can easily be found by
adding all rows of A less than or equal to the row of E in question. For low
dimensions some sort of branch-and-bound method can be used if all
possibildtiesare desired.

EXAMPLE2. Let A be as in Example 1.

(1) See Example 1 (1).

(2) Let

(3) The first row of G must be (0 0 1). The second row of G must be
(0 1 0). The th.’u d row of G must be (0 1 0). Thus G is

L 0
0 01 01

III. For Finding All Least-Squures g-inverses of A

Transpose A, apply the preceding algorithm, and transpose the resulting
inverses.

IV. Algorithm for Finding All V’ner Inverses of A

(1) Find b(A).

(2) FindZ( v)f or each v E b(A). A Vagner inverse will exist if and only
if Z(v) is nonempty for each v E b(A).
(3) Choose a specific u E Z(v) for each v E b(A). We will write u(v) for
u.
INVERSES OF BOOLEAh’ MATRICES 2%

(4) For each chosen U, choose a vector s such that si = 1 only if A,* d fj
and si = 1 for at least one i such that A,. = o. We will sometimes write S(O)
for s.
(5) If it is not true that o’> u implies s(u’)= s(u). redefine s(u) to he
sO(z))+~z~O~O(z), where s”(u) is the s(z)) obtained in step (4).
(6) For a vector t which is not one of the chosen identification vectors,
if t is not less than any row basis vector, send t to an arbitrary sum of the
vectors s(u) or 0. Otherwise let p(t) be the infimum of the row vectors which
are greater than or equal to t, taken in the row space of A regarded as a
lattice. Choose a set P of row basis vectors whose sum is less than or equal to
p(t), and let t be sent to b=ZzErs(z), or if l? is empty, to 0.
(7) Write the vectors s and b in the order of the t’s and u’s.

EXAMPLE 3. Let A be

i 0111 1 01.
1 1
(1) b(A)= ((1 1 01,(0 1 1)).
(2) I((1 1 O))={(l 0 0)); I((0 1 l))={(O 0 1)).
(3) (Not needed.)
(4) For (1 1 0), s is (1 0 0). For (0 1 l), s is (0 0 1).
(5) (Not needed.)
(6) For t=(O 1 0), p(t)=O. So b-0.
(7) Thus the Vagner inverse of A is

BEMARK 1. If it is only desired to find a single Vagner inverse, this

procedure can be considerably simplified, We omit step (3), take s to be the
vector si = 1 if and only if 4. < 0, and send all identification vectors to this s
and all other vectors containing only one 1 to zero.

V. Moore-Penrose Inverses of Matrices

REMARK2. As far as Moore-Penrose inverses are concerned, if a Boolean

matrix A has a Moore-Penrose inverse, then A r is its unique Moore-Penrose
inverse [lo].
254 KI HANG KIM AND FRED W. ROUSH

6. ANALYSIS OF EFFICIENCY

The number of operations required in these algorithms can be estimated.

Take for instance the Vagner inverse algorithm. The number of different
Vagner inverses of a matrix may be extremely large, so we will consider only
the estimated number of operations per Vagner inverse.
Let A be a Boolean matrix. First, for every ordered pair of distinct rows
A,,,A,., we determine whether Ai. < Ai.. This requires n3 - n2 comparisons.
This data should be stored as a matrix E. By intersecting E with its
transpose, requiring i intersections of entries, we can determine which
( >
rows are equal to which other rows. Next we find a set of rows which are
different from each other, but such that each row is equal to some row in the
set. We include a row in this set if and only if it is not equal to a prior row.
Using the intersection of E and ET this requires about operations
(; >
(which could be addition of entries in the intersection matrix). Let this set be
called S,.
We next test each member of S, to see whether it is zero. This will
require no more than n2 comparisons. Let the set S, consist of the nonzero
members of S,. Next we test each row of S, to see if it is a linear
combination of other rows. Specifically we determine if a member 4. of S, is
the sum of all rows. Ai. of S, which are less than Ai,. From the matrix E, we
knbw which rows these are. Computing their sums will take about kin2
additions, where k, is the average number of rows less than or equal to a row
of A. Comparing these sums with the rows Ai. will take about n2 compari-
sons. By this method a row basis S, for A can be determined with about
n3 + (k, + 2)n2 - n operations.
After this we determine the sets of identification vectors for the elements
of the row basis. For each row basis vector compute the sum of all row basis
vectors which are not < the given row basis vector. The set of such vectors
can be determined from the matrix E. This computation requires about
(n - k&n additions, where r is the row rank of the matrix. Next compare
these vectors with the row basis vectors to determine all entries where the
row basis vectors are 1 but the sum vectors are zero. This gives the sets of
possible identification vectors for the row basis vectors. The steps so far have
required at most 2n3 +3n2 - n operations. And by adding a little to the
comparison routine (at most n2 additional steps) we can find out whether the
matrix is regular.
The fourth step will not require a significant number of operations. The
fifth step (where it is simplest to omit the comparisons and go ahead and
add) requires about k,m additions.
For the sixth step, it will usually be the case that the vectors t will be less
than some row basis vector. We find those row basis vectors which are
INVERSES OF BOOLEAN MATRICES 2.55

greater than or equal to the t’s, requiring r(n - r) comparisons. For such a t
we next determine the set Ii of all row basis vectors less than or equal to
every row basis vector which is greater than or equal to t. This is done by
intersecting at most r columns in the matrix E. This requires at most
f(n - r)n steps for all the vectors t. Now r is chosen to be any subset of I,.
The sum required in computing z will be at most r(n - r)n separate addi-
tions. In general, to obtain a new Vagner inverse we need only perform one
of the operations under step (6), and many times step (5) need not be
performed. Step (7) requires at most n2 comparisons.
Thus our algorithm requires 2n3+4n2 - n initial steps and then k,n2 +
1.5n2+0.5n3 (or usually fewer) additional operations per Vagner inverse.
To find a single Vagner inverse we can choose si = 1 if and only if ri Q T,
so that step (5) will be automatic. Send all the t vectors to zero. In this way a
special Vagner inverse can be found with about 2n3 + 6n2 - n operations.
By comparison, Rao’s algorithm may require something like n3 steps to
find a row basis, n3 steps to find a column basis, and 2n2 steps of hunting for
columns with a single one entry, to obtain a g-inverse. Then somewhat more
steps may be needed (about n3) to obtain a Vagner inverse.

7. THEOREMS.REGARDING g-INVERSES AND VAGNER INVERSES

The following lemma is a key step in proving theorems concerning

g-inverses and Vagner inverses.

LEMMA 1. If A E B,, is regular, fm every v E b(A) there exists at kast

une uEl(v).

Proof. The proof that we give is based on Zaretskii’s criterion for

regularity [17]: a Boolean matrix is regular if and only if its row space is a
distributive lattice.
A vector u containing only one 1 is an identification vector for z, if and
only if u < v and

where S is the set of basis vectors which are not greater than or equal to v.
Thus there exists at least one identification vector for v if and only if

vg c w.
WES
256 K1 HANG KIM AND FRED W. ROUSH

slppse on tne contrary

Multiply both sic& .by v, in tne row space regarded as a lattice. That is

v=
(2 1WES
20 v.

By ciisrributimty,

Each vector WL)is properly less r&n 0, since each w was not greater than or
equal ro ~1.Thus v IS a sum of pop&y smaller vectors. Thus v is not a basis
vector, wnich is a conuadiction. Thrs proves the lemma. q

r~I’HEOM!M 2. lhe Jirst algtitiwn, fbr all allowable &&es, yi4dd.s all
g- imwrs~sof A.

Proof: First we will show that matrices G constructed in this way are
always g-mvtxses of A. ‘kt e7 be a row basis vector ior A. Then VG is the
sum of s(v), of Y(Q) for aU other basrs vectors v, whose identikicdtion vectors
iue b O, and or b(t) for all t< u. By the definition of s, sjv)A=u and
Y( SJA = v,. By the dehmtion of an identification vector, U, < v for each v.
And by deiimtion of b, b (t)A < y(t) and y(t) < v, for each t < v. Therefore

uGA==v+ c p(t)+ &,=v.

‘l’hereiore tar any sum vl of basis vectors in H (A),

Now let G be any g-inverse of A. Let t be a vector less than or equal to a

basrs vector v. We have tG < vG. But since v = v,A for some vector v2,
INVERSES OF BOOLEAN MATRICES 257

(vG)A = v,AGA = v,A = o. ‘lhxe~me (C)A < v. Thererore

Therefore (C)j = 1 only if Ai. < p(t).

This implies also that for arry idermrrcatron vector U, we nave (uG)A < U,
and (uG)~ = 1 only if Ai, < v. Suppose that ror all ioentmcauon vectors
uEI(v), (uG)~=I oufy if Ai* < v. Then (&)A is a sum of vec‘cors properly
less than v. II_& t he a vector with ody one I wiiicn is % ‘u ‘buurnot an
identification vector for 0. Then there exists w E b(A) such tylat t d w aad
v < w. Otherwise t would be an identrfication vector tor v. Thererore p t t) wd
be properly less than v. Therefore (C)A will be properly less than v.
Therefore

(vG)A= c (&)A+ 2 (tG)A

u<v t<o

will ‘be a sum of vectors properly iess than v. Blur (v G)A = r). T’ms contradicts
the assumption that v is a basks vector. Thererore ror some mentnricauon
vector tl of v, there will be some i such that (uG), = I and Air- O. This
completes the proof of the thtxxem.

Proof. The mrmber of ways to map the set ot identiucauon vectors ior r)
subject to si = 1 only if Ai. < v and si = 1 for at Ieast one i sucfl mat A~, = u,
for some identification vector, is given by
25s KI HANG KIM AND FRED W. ROUSH

The number of ways to map the non-identification vectors is given by the

second factor. n

THEOREM 5. The fourth algorithm, forall allowable choices, yields all

Vagner inverses of A.

Proof. First we will show that all matrices G constructed by this

algorithm are Vagner inverses of A.
First we note that the definition of s(v) implies s (v)A = v. Also, for t < v,

tGA= xs(z)A= &< p(t)=v.

r r

Thus for a basis vector v,

vGA=s(v)A+ 2 xs(z)A=v.
t<o r

Thus as before, for any vector va, vsAGA = v,A and AGA = A. The image of
G is generated by the vector s(v). We have s(v)A= v and

vG=+)f tFoFs(z,.

Each z in the second part of the sum is less than or equal to p(t) and so less
than or equal to v. Therefore by step (S), s(z) < s(v). Therefore vG=s(v). So
s(v)AG = s(v), and AG is the identity on the image of G. Therefore
GAG = G.
Next suppose G is a Vagner inverse of A. Since G will also be a g-inverse
of A, there must be an identification vector u mapped to vectors s(v) as
described in step (4) By the construction, s(v)A = v. Also VGA = v, since v is
in the image of A. Therefore s(v) = uG and vG are vectors in the image of G
which have the same image under A. But GAG = G implies A is l-l on the
image of G. Therefore s(v) = vG. Therefore v < v’, s(v) < s(v’).
Let L be the subspace spanned by the vectors s(v). We have s (v)A =
uGA = v, so L maps onto the space or row vectors of A, under A. Also L is
contained in the image of G. Since A is l-l on the image of G, L must be the
complete image of G, and A must be l-l on L.
Lettbeavectorwithexactlyonel,andletM={wER(A):w>t}.For
w E M, wGA= w, so tGA < w. Since tGA is in the row space of A,
INVERSES OF BOOLEAh’ MATRICES 259

K7A < p(t). Suppose

tGA= &.
r

Since

and A is l-l on sums of s(z)‘s, tG must be

If it4 is empty, G Et (image of G) = L. This proves the theorem.

REMARK3. The number of Vagner inverses of A is equal to the product

of the number of idempotents in the ~-class of A and the number of
idempotents in the %-class of A [5].

8. THEOREMS REGARDING OTHER INVERSES

THEOREM 6. A has a minimum norm g-inverse if and only if no two

row basis vectors of A have a 1 in the same column. If an inverse exists, all
inverses are given by the second algtithm.

Proof. Proposition 1.1 of P. Rao ill] states that G is a minimum-norm

g-inverse of A if and only if GA is a symmetric idempotent with the same
row space as A.
For any idempotent E there exists a permutation matrix P such that
PEPT is a block triangular matrix in which the main diagonal blocks are
either all l’s or all 0’s. In fact,
260 KI HANG KIM AND FILE3 W. ROUSH

If E is symmetric, so will be PAP”. Thus all the blocks below the main
diagonal wiU be zero as shown. This implies that any two row basis vectors
for PEPT have nonoverlapping sets of 1’s. The same must be true of E. If A
has the same row space as E, the sarle will be true of a row basis for A.
If A has this property, the matrix E such that ei i = 1 if and onIy if uJ = 1
and 4 = 1 for some basis vector o of A wiIl be a symmetric idempotent with
the same row space as A.
E is unique if E’ is another symmetric idempotent with the same row
space as A; E’ by symmetry must also have the same column space as A.
Therefore E, E’ lie in the same 3c-class [5]. But no X-class in any semigroup
can contain more than one idempotent [2]. This proves the theorem. I

PROl?oSrrION7. G is a least-squares inverse of A if and only if GT is a

minimum-norm inverse of A T.

Proof. This follows from the defining formulas AGA = A, (AG)?‘= A@.
Ip

THEOREM 8. The number of matrices of rank k and dimension n which

have a minimum-rwrm g-inverse is

[S(n,k)+(k+l)S(n,k+l)]~ (-1)‘(!)(2’-i)“.
i=l

Here the S (n, k) are the Stirling numbers of the seed kind.

Proof. Choosing a row space amounts to dividing the set of n positions

of vectors in V, into either k disjoint nonempty sets or k+ 1 disjoint
nonempty sets, with one set selected as those columns which are always
zero. This can be done in S (n, k) + (k + 1)s (n, k + 1) ways.
A matrix with such a row space wiII have a minimum-norm g-inverse. By
Brandon, Hardy, Kim, and Markowsky [l] there are

$(-l)i(i)(2k-i)n
i=l

matrices with a particular choice of such row space. (This uses the fact that
the space generated by k vectors with disjoint sets of l’s has Zk elements.) n
INVERSES OF BOOLEAN MATRICES 261

THEOHEM 9. A has a Moore-Penrose inverse if u& only if any two TOWS

of A either are equal or have disjoint sets of 1’s.

PTOOf: *: Ry Theorem 7 the sets of I’s in any two row basis vectors are
disjoint, and by Theorem ‘7 applied to the transpose, the sets of l’s in any
two column basis vectors are disjoint. Suppose some row A,. is a nontrivial
sum invobing two row vectors A,,, A,, which are different basis vectors.
There will be some i, i such that (Aa*)i-l, (Ab.)i=O, (A,.)j=O, (Ab.)j=l.
Therefore there will be vectors v, w in the cdumn basis such that v, = 1,
u, = 0, w, = 0, w, = 1. IIowever a91column vectors w.hich are 1 in either the a
row or the b row will have a 1 in the c row. Thus v, w do not have disjoint
sets of l’s, a contradiction.
3: C. R. Rao [lo] has proved that if A has a Moore-Penrose inverse G, G
must be AT. We will show that AT is a Moore-Penrose inverse of A if A
satisfies the condition of this theorem.
The (i, /)-entry of XX TX is 1 if and only if there exist p, q such that
~~~=(X~)~~=x~~==l, or qp=xqp = xqi= 1. Therefore the i row and the q row
are not disjoint, so they must be equal, so xii = 1. Conversely if xii = 1, taking
p = i, y = i gives (XXTX)ii = 1. Therefore XX’X= X.
Transposing this equation, X ‘XX r = XT. The matrices XX and X TX will
always be symmetric. This proves the theorem. II

THEOREM 10. The number of matrices dimension n and rank k huving a

Moore- Penrose irrverse is

k![S(n,k)+(k+l)S(n,k+l)]“,

where the S (n, k) are the Stirling numbers of the second kind.

Proof. Choosing a two-space such that different basis vectors have

disjoint sets of l’s amounts to dividing the set of n positions of row vectors
into k disjoint, nonempty sets, or into k + I disjoint, nonempty sets, with one
of the sets singled out as the set of columns which are identically zero. This
can be done in S( n, k) + (k + l)S( n, k + 1) ways. Likewise the column space
can be chosen in S( n, k) + (k + 1) S( n, k + 1) ways. Let a row basis be written
xi,...,~ and a column basis be written yi,...,y’. Then the matrix A must be
a sum of certain matrices A, such that (AcJii = 1 if and only if ( xJj = 1,
( y,), = 1. Moreover, since no nontrivial sums of basis vectors can be row or
column vectors, the set of a, b such that Aab is a summand is a permutation
on a set of k elements. Thus given the row and column spaces, A can be
chosen in k! ways. It can be seen that all of these choices satisfy the
condition of Theorem 8. This completes the proof. n
262 KI HANG KIM AND FRED W. ROUSH

REFERENCES

1 Ft. L. Brandon, D. W. Hardy, K. H. Kim, and G. Markowsky, Cardinality of

D-classes in B,,, Semigroup Forum 4 (1972), 341344.
2 A. H. Clifford and G. B. Preston, The algebraic theory of semigroups, Math.
Surveys, Vol. 1, No. 7, Amer. Math. Sot., Providence, RI., 1961.
3 J. Denes, Transformations and transformation semigroups, Seminar Report,
University of Wisconsin, Madison, Wisconsin, 1976.
4 Ki Hang Kim, A Moore-Penrose inverse for Boolean relation matrices, in Proc.
Second Australian Conf. on Combinatorial Math., Melbourne, Australia, August
1973, pp. 18-28. MR 50#6955.
5- Combinatorial properties of binary semigroups, Period Math. Hungar. 5
(1974j, 3-46.
6 R. D. Lute, A note on Boolean matrix theory, Pmt. Amer. Math. Sot. 3 (1952),
382-388.
7 J. von Neumann, On regular rings, Proc. Nut. Acud. Sci., U.S.A. 22 (1936),
296300.
8 R. J. Plemmons, Generalized inverses of Boolean relation matrices, SIAM I.
A&. Math. 20 (1971), 4266433.
9- Regular nonnegative matrices, Centre des Recherches Math., Universite
de Montreal, technical report, Aug. 1972.
10 C. R. Rao, Generalized inverse of Boolean valued matrices, lecture delivered at
the Conf. On Combinatorial Math., Univ. of Delhi, Delhi, India, 1972.
11 P. S. S. N. V. Prasada Rao, Generalized inverse of special types of matrices,
Dissertation, Indian Statistical Institute, Calcutta, India, 1974.
12 - and K. P. S. Bhaskara Rao, On generalized inverse of Boolean matrices,
technical report, Indian Statistical Institute, Calcutta, India, 1973.
13 D. E. Rutherford, Inverses of Boolean matrices, Proc. CZasgow Math. Assoc. 6
(x%3), 49-53.
14 B. M. Schein, Regular elements of the semigroup of all binary relations,
Semigroup Forum 13 (1976), 95-102.
15 V. V. Vagner, Generalized groups, Dokl. Akad. Nauk SSSR (N.S.) 84 (1952),
1119-1122. MR 14#12.
16 J. H. M. Wedderbum, Boolean linear associative algebra, Ann. of Math. 35-l
(1934), 185194.
17 K. A. Zaretskii, Semigroup of binary relations, Math. Sbornik 61 (1963), 291-365.

Received 16 March 1977; revised 3 October 1977