ent.
inear Algebra*
‘step, (5) To call attention tothe computational aspect
ier. (6) To
in very die
However, here the word “transversal” will be used
3. Matrices of Zeros and Ones
The subject of SDR’s is frequently treated in the
context of matrices of 0's and I's. The incidence
matrix of the family Q of subsets of E is the mat
A= [ay], ie, jeQ, such that ayy=
if ej, and ay
A matching in a matrix is a subset of its positions
i, ) such that first indices (rows) of members are all
different and second indices (columns) of members are
all different. A transversal (column transversal) of &
matrix is a matching in the matrix which has a member
in each column. The product of the entries of @ trans.
versal, we call a transversal product
each,
ning
ibed as
ing itis JOURNAL OF RESEARCH of the Notional Bureau of Standards—B. Mathematics and Mathematical Physics
edge Vl 718, No 4 Gcsber-Deconber 1067
ongtthe
saning
node is
+E) by .
an Systems of Distinct Representatives and
general
aie | Jack Edmonds
a te Institute for Basic Standards, National Bureau of Standards, Washington, D.C. 20234
eighted | (November 16, 1966)
bored Some purpose ofthis paper are: () To ake seriously the erm, “tera ank," 2) To make an isu
comp | of no "eamancing os and alan” by no “asaging them he fe pce, (To poe The
Suercal are of Grama rl. To uate ani rcleanee of aunbe oor ee
of work depends onthe tout of work
St Sb, aspect where these les bing an nance of la ns
a deserbe an SDR tance of thes on exenel ensures ha wees nin
visio ferent aye than does oul uninoddar they. he ecelng paper, Onley Bence, de
ton Coe Serbs Uterine of ta ead
it pp Key Words: Algorithms, combinatorics, indeterminates, linear algebra, matrids, systems of
ae iinet represen ake
fears
—— 1. Introduction However,
The wellknown concept of term rank [5, 6)! is *
shown here to he a special case of linear-algebra rank.
Thie observation is used to provide a simple linear
“1. Mas, slgebra proof ofthe wellknown SDR theorem. Except
5 for familiar linear algebra, the paper is self-contained.
don with Incidentally to SDIs, an algorithm ie presented for
wes Nw. computing the determinant or the rank of any matrix
ex, er any integral domain. Iisa vatiation of Gaussian otherwise
$22°%., ie. leas) elimination which hae eertan odverae
sbinatoril [tis observed to be an interestingly bad algorithm for
van) computing term rank,
E°CL23| The final part of the paper discusses some simple
th See? mmatroidal aspects of SDR’s.
teri
2. Systems of Distinct Representatives
4-249)
Let Q be any finite family of subsets of a finite
set E. “Different” members of Q may be identical in
content. The number of members of a family H is
denoted by |#|. The union of the members of H is
denoted by U(H). The SDR theorem says that it is
possible to choose a different element from each mem:
ber of if and only if there is no subfamily H of Q
such that |H| > |U(@d)|. The “only if” part is obvious.
A subset of E formed by choosing a different ele-
‘vent from each member of Q is called a system of
distinct representatives of Q, or an SDR of Q. (Some:
vlimes, such as in [1}, it is called a transversal of Q.
“Toperkchictsahern grassy Urey of tn, On
Clearly, Q has an SDR if and only if its incidence
matrix 4’has @ L-valued transversal (or a. L-valued
transversal product.
In the context of 0,1 matrices the SDR theorem says
that a matrix A=layf, ie, je0, of O's and I's has
4 Lvalued transversal if and only if there is no HCO
and KCE such that |] >(K| and such thet 3
matrix [au], ieK, jel, contains all the 1's of submatrix
ay), ie, je.
4, Permanents, Determinants, and Good
Algorithms
The maximum cardinality of @ nonzero-valued
‘matching (matching having no zero entries) in any
matrix 4 is well-known as the term rank 114) of A.
241Clearly 114) is the maximum cardinality of a subset
of columns of 4 which has a nonzero-valued trans-
versal. Thus the term rank of a family of sets is de-
fined to be the maximum cardinality of a subfamily
which has an SDR.
Clearly r(4) equals the maximum order of a square
submatrix of A which has a nonzero-valued transversal.
‘The permanent of a square matrix is defined to be the
sum of all its transversal products. Thus (4), for a
nonnegative matrix 4, equals the maximum order of a
submatrix of A with non-zero permanent. The perma.
nent of a square 0,1 matrix is the number of l-valued
transversals of the matrix,
A good algorithm is not known for computing the
permanent of any square 0,1 matrix (relative to the
order of the matrix as the “size” of input). For an
algorithm to he good we mean that there is a poly-
nomial function fin) which for every n is an upper
bound on the “amount of work” the algorithm does
for any input of “size” n.
‘The tansversals of any square matrix partition
uniquely into two classes such that any two trans-
versals which differ by just two positions (of each)
are in different classes. (The proof is left to the reader.)
The determinant of a Square matrix with a prescribed
1-1 correspondence between rows and columns,
ive., with a prescribed transversal, is defined to be the
sum of transversal products over whichever of these
classes contains the prescribed transversal minus the
sum of transversal products over the other clas
A very remarkable formal property of determinants,
of say integers, is that there exists a good algorithm
for computing them—a version of Gaussian elimination
which we will describe in section 7, It is also remark-
able that there exists a good algorithm, the same one
as above, for computing the linear-algebra rank of
any matrix of say integers. The largest number of
digits of an entry as well as the dimensions of the
matrix must of course he figured somehow into the
measure of “size” of possible inputs to the algorithm.
For example, the “size” of an input may be taken to
be the maximum of these numbers, or may be taken
to be a vector consisting of these several numbers.
[An algorithm which is good in the sense used here
is not necessarily very good from a practical viewpoint.
However, the good-versus-not-good dichotomy is
useful. It is easily formalized (say, relative to a Turing
machine, or relative to a typical digital computer with
an unlimited supply of tape), and usually it is easily
recognized informally. Within limitations it does have
practical value, and it does admit refinements to “how
00d” and “how bad.” The classes of problems which
are respectively known and not known to have good
algorithms are very interesting theoretically.
Good algorithms, not Gaussian elimination, are well-
known for computing the term rank of any 0,1 matrix.
For instance see a remark in section 8 of this paper
together with [3]. We shall see that Gaussian elimina-
tion is also an algorithm for computing term rank.
1...
5. Rank
Aron pint of this paper is the following obser 40) of
vatton Tae in (i]uses theseme iden ina deaperet {gre
‘Tusonem 1. The term ronk of 0.1 mainte A,
the same. as the linearalaebra rank ofthe matng @
tbvained by replacing the 1's in A by distinct nde
terminates (over any integral domain). here 0
Several “numbers” are called indeterminates if , Th
oe mE a cle Indeterminate | Treo
egnyeRae a eeepc Ra
OOF OF THEOREM 1. The linear-algebra rank 2
imate equals the maximum order of «minor wig S278
nonzero determinant. (A minor of a matrix is a sub. 7° % 4
matrix with a. prescribed 1-1 correspondence
tween its rows and columns.) The determinant of a
minor is a linear combination, with +1’s and —1’s ’
Coefficients, of alts transversal. products. Thus
Where the entries are. zeros. and distinct indeterm |
hates, a determinant is nonzero if and only if one or | TeTa"
more of its transversal products is nonzero, End of | “&0rith
proof. PRoo
in Aas
6. A Linear-Algebra Proof of the
SDR Theorem
Suppose a matrix 4 of zeros and distinct indetermi. simply
nates has no transversal of nonzeros..Then, by Theorem Each of
1h its rank is less than the number of columns, so the | mentary
columns ae linearly dependent. Let 4; bea submatrix the dete
Consisting of a minimal dependent subset Hof the the row
Columns of 4. Let ds be a submatrix consisting of ¢ rows. an
maximal independent subset K of rows of dy, Since termine:
row rank equals column rank, we have [?f](K[1. factor 0
fA has a column of all zeros then that column alous | of detex
isa dependent set H, and K is empty. Otherwise,there anes th
fa column vector + such that 4,2 ie all zeros and such by e, an
that the components of are nonzero polynomials in another
the entries of determi
If any row, a, of 4s, not in de, contained one or By it
mote indeterminates, then ax would be a polynomial minents
function of indeterminetes, equal to zero, bit at
identically zero. Tlence, the rows of dy, not in 4s,
Gontain only zeros. End of proof
7. ABad Algorithm for Term Rank at 10,
‘Theorem 1 also suggests using Gaussian elimination ‘hi mi
as an algorithm for computing term rank. With he "!*
Skeuse of trying to give this approach every posible a4,
Gdvantage,. wer desenbe an ‘improved. version of (“230
Gaussian elimination (possibly due to Gauss). ev
"The rank of any matrix 4°, or the determinent of
any matrix 4° (with mutually corresponding index
fet), with entries from an integral domain, ean be
Computed as follows. On th
Inch steph beining wih k= tan «ye eine
matrix 4*=(at] from 4*-!=[aty, by’ fist choosing
ma cor cll by and some olan, ell HO,
neither of them chosen on previous steps and such th
aslo” 0. The algorithm stops when this i= 90 because
longer possible. and aby
2427
& obser.
ne
rit A's
ct inde.
nates if
intogral *
80 the
ubmateix,
H of the
se, there
and sich
vials in
done or
}ynomial
Dut not
vt in As,
nk
imination
With the
possible
‘The matrix 4*is determined as follows. Each row i(t),
1