Mathematics Magazine

ISSN: 0025-570X (Print) 1930-0980 (Online) Journal homepage: https://maa.tandfonline.com/loi/umma20

Applications of Graph Theory in Linear Algebra

Michael Doob

To cite this article: Michael Doob (1984) Applications of Graph Theory in Linear Algebra,
Mathematics Magazine, 57:2, 67-76, DOI: 10.1080/0025570X.1984.11977080

To link to this article: https://doi.org/10.1080/0025570X.1984.11977080

Published online: 13 Feb 2018.

Submit your article to this journal

Article views: 29

View related articles

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at

https://maa.tandfonline.com/action/journalInformation?journalCode=umma20
 -----A:WCLC)
Applications of Graph Theory in Linear Algebra

Graph-theoretic methods can be used

to prove theorems in linear algebra.

MICHAEL Dooo
The Uniuersitv of Manitoba
Winnipeg, Manitoba, Canada RJT 2N2

Graph theory has existed for many years not only as an area of mathematical study but also as
an intuitive and illustrative tool. The use of graphs in wiring diagrams is a straightforward
representation of the physical elements of an electrical circuit; a street map is also a graph with
the streets as edges, intersections of streets as vertices, and street names as labels of the edges. The
graphs resemble the physical object that they represent in these cases, and so the application (and
sometimes the genesis) of the graph-theoretic ideas is immediate. A flow diagram of a computer
program and a road map with one way streets are examples of graphs which contain the concept
of direction or flow to the edges; these are called directed graphs.
There are applications of graphs and directed graphs in almost all areas of the physical sciences
and mathematics, many of them known for fifty years or more, but very few of these ideas have
percolated down to the undergraduate student. The purpose of this article is to apply graph-theo-
retic ideas to some of the fundamental topics in linear algebra. While there are many such
applications, we shall focus on only two of the most elementary ones, i.e., matrix multiplication
and the theory of determinants. The presentation is usable as a supplement to the usual classroom
lectures; in fact, this paper grew out of notes used in a linear algebra course for sophomores.
The main tool to be used is the directed graph. Intuitively this can be thought of as a set of
points (or vertices) with arrows (or arcs) joining some of the points. A label may be put on an arc.
More formally, a digraph consists of a set of vertices V and a subset of ordered pairs of vertices
called the arcs. A labelling of the digraph is a function from the arcs to the real numbers. A
labelled digraph is usually visualized by considering the vertices as points with arcs as arrows
going from vertex i to vertex} whenever (i, j) belongs to the sets of arcs. The ith vertex of the arc
(i, j) is called its initial vertex, while thejth vertex is called its tenninal vertex. The arc is then
given a label which is the image of that arc under the labelling function. When the initial and
terminal vertices are identical, the arc is called a loop. We shall sometimes say that an arc "goes
out of" its initial vertex and that it "goes into" its terminal vertex. The number of arcs that go out
of a vertex is called its outdegree, and the number of arcs that go into that vertex is called its
indegree.
A digraph G is a subgraph of a digraph H if the vertices and arcs of G are contained in the set
of vertices and set of arcs of H. One type of subgraph of H is a walk: this consists of a sequence of
vertices v0 ,v 1,v 2 , ... ,V11 such that (v,_ 1,u,) is an arc in H for i=l.2, ... ,11. We use
(u 0 , u 1 , v 2 , ... ,t>,) to denote such a walk (which is consistent with the notation for an arc); we then
say that the length of the walk is n. A path is a walk where all the vertices are distinct, and a cycle
has v 1 , P 2 , . .. , P, distinct and u0 = P,. Another type of subgraph is a factor; it has the indegree and
outdegree of each vertex equal to one. A moment's thought reveals that a factor is simply a

VOL. 57, NO. 2. MARCH 1984 67

(vertex) disjoint union of cycles. The weight of a subgraph, W( G), is the product of the labels of
the arcs in that subgraph. Weights of walks, paths, cycles, and factors are similarly defined.
Matrix multiplication and the Konig digraph
In this section we examine a particular directed graph associated with an m X 11 matrix A. The
multiplication of two matrices is equivalent to "glueing" their digraphs together to form a single
graph. An entry in the product matrix is then related to the weights of certain paths in the new
graph. Most standard proofs about matrix multiplication involve the manipulation of subscripts
and/or the interchanging of summations. These techniques, while valid, tend to obscure the
underlying ideas; this is avoided by the graph-theoretic approach.
Let A = (a, 1 ) be a matrix with m rows and n columns. The Konig digraph, G (A), is a labelled
digraph with m+ 11 vertices-m of these vertices correspond to the rows and 11 correspond to the
columns. The arc from vertex i to vertex) has a,, as its label. When drawing the graph we shall
omit arcs with a zero label; this will clarify the relationships between several matrix and graphic
operations. As a further convention, we shall draw the row vertices on the left and the column
vertices on the right. The top vertex will correspond to the first row or column, the one below it to
the second, etc. An illustration of a matrix A and its Konig digraph G( A) is shown in FIGURE 1.
D. Konig [5] used this digraph in his well-known and fundamental book in 1936; he referred to it
tangentially even earlier [4] in 1916.

2 ())
-]'

FIGl'RI l. A 2 X 3 matrix A and it~ Konig digraph G(A).

If A and Bare both m X 11 matrices, then G(.4) and G(B) are essentially the same digraph with
different labels. The matrix sum A + B is defined, and the digraph G (A + B) obviously is
obtained by adding the weights on the corresponding arcs of G(A) and G(B). The usual additive
properties of matrices carry over directly to the digraphs. The proofs of these familiar properties,
e.g., the associative, commutative, and distributive laws, are identical to the usual ones, but rather
than focusing on a particular row and column entry in a matrix, a label of a particular arc is used
to determine the validity of these properties. The multiplication of a matrix by a scalar is also easy
to visualize. The digraph G(rA) is obtained from G(A) by multiplying every arc label by r. The
usual properties of scalar multiplication are also easily verified.
Given an m X 11 matrix A and an n X r matrix B, the concatenation graph G (A)* G (B) is
defined by taking the digraphs G(A) and G(B) and identifying then column vertices of G(A)
with the n row vertices of G( B). An example of the concatenation of two graphs is given in
FIGURE 2.

fiGl 'RI. , The concatenation of 1'1\0 Konig digraph~.

68 MATHEMATICS MAGAZINE
 THEOREM 1. The (i, j) entry of the matrix product AB is equal to the sum of the weights of the
paths in G (A)* G ( B)from the ith row vertex of G (A) to the jth column L•ertex of G (B).
Proof Each path of length two from vertex i to vertex j passes through a unique vertex k. By
definition, the weight of the path is the product of the labels of its arcs, which is au,bk;· Summing
over all possible k gives the desired result.
Note that the exclusion of edges of weight zero from the digraph according to our convention
still leaves the theorem valid. We see, for example, that for the digraph in FIGURE 2, the (1,2)
entry of the product AB is -2 since there is only one path between the appropriate vertices.
Several properties of matrix multiplication are immediate consequences of Theorem 1. To show
matrix multiplication is associative, assume that matrices A, B, and C are of appropriate orders,
and consider the graph G(A)•G(B)•G(C). Then the (i.J) entry of (AB)C and A(BC) are
clearly both equal to the sum of the weights of the paths of length three from the i th row vertex of
G (A) to the j th column vertex of G (C). It is an easy exercise to show that matrix multiplication
distributes over addition by looking at a particular path (i. j, k) in the Konig digraphs of
G(A)•G(B), G(B)•G(C), and G(A)•G(B+ C).
If Pis a permutation matrix, then G(P) consists of a set of arcs with no common vertices. In
other words, the outdegree of each row vertex and the indegree of each column vertex is equal to
one, and all labels of arcs are 1. The product of permutation matrices is a permutation matrix. Look
at the concatenation digraph! Likewise, the digraph of a diagonal matrix makes the following facts
trivial. The product of diagonal matrices is a diagonal matrix. Each diagonal entry of the product is
the product of the corresponding diagonal entries in the factors.
A matrix is upper (lower) triangular if a,J = 0 whenever i > j (i <j). Thus G(A) is the Konig
digraph of an upper (lower) triangular matrix if and only if all the arcs in the drawing are
horizontal or downward (upward). The product of upper (lower) triangular matrices is upper (lower)
triangular. The digraph makes it trivial!
The Konig digraph of the transpose of A is obtained from G (A) by reversing the direction of
all of the arcs. If we wish to adhere to the convention of having row vertices on the left and
column vertices on the right, we must then reflect the new graph with respect to a line through the
column vertices (see FIGURE 3). The properties of the reversals and reflection make the following
Theorem obvious.
THEOREM 2. Let AT be the transpose of the matrix A; then
(1) (Ar)r =A,
(2) (A+B)T=AT+Br,
{3) (cA)T =cAr, and
{4) (AB)T = BTAT.

Notice how the reversal of the arcs makes the reversal of the order of the matrices A and B in
Equation (4) intuitively clear.

FIGURE 3

VOL. 57, NO. 2, MARCH 1984 69

 ,___ _ _ _ _,0 1
Io >()1 1o o.o 1 1o

2o >()2 2o ~o 2 2o-------l>~O 2

'X'
i - 10

~z:
"A
io- - - - - - " ' O i

J J i+1o- - - - - - - - ' ' > ( ) i +1

m - 1 o------0>0 m - 1 l1l- l 0------~o m- 1 m-lo------~m -1

m o--------;-,o m m o--------,>0 m Ill o------~0 m

(a) (b) (c)

FIGURE 4. Konig digraphs of the elementary row operations on an m X n matrix: (a) adds a multiple of the ith row to
the jth row: (b) interchanges the ith and jth rows: (c) multiplies the ith row by )\.

The ( i, j) element of the product matrix AB is completely determined by the sub graph of
G(A)• G(B) consisting of the ith row vertex of G(A), thejth column vertex of G(B), and all
paths of length two joining them. A partition of the column vertices of G (A) ( = the row vertices
of G( B)) produces a partition of these paths. The validity of block multiplication is now easy to
understand since each block arises from a partition of the rows and columns of the matrix.
Properties of matrices associated with elementary row operations are easily seen using Konig
digraphs, since multiplying by a matrix on the left is nothing more than concatenation on the left
by a Konig digraph. For example, if we wish to multiply the i th row of the m X n matrix A by 'A
and add it to thejth row, we simply concatenate the digraph shown in FIGURE 4(a) with G(A). It
is an easy exercise to construct the Konig digraphs that represent the interchange of two rows of a
matrix or the multiplication of a row by 'A. See FIGURE 4(b), (c). The digraphs in FIGURE 4 are
called the Konig digraphs of the elementary row operations. A glance at these quickly shows that
the Konig digraphs of an elementary row operation and its inverse are identical except for at most
one label; in one case 'A is replaced by -'A and in the other it is replaced by 'A - 1 .
The Coates graph and the determinant
In this section we look at a different digraph associated with a square matrix. The determinant
of the matrix can be computed from the weights of the cycles in this graph. This allows a relatively
straightforward computation of the determinant of a matrix of arbitrary order, especially if there
are many zero entries. Usually determinants are defined by making an excursion into the theory of
permutations, a subject which by its nature is deeper than the determinant concept itself and
necessitates a relatively difficult digression. An alternative approach is to define the determinant
inductively using the cofactors, but like many inductive approaches, the results of the proofs are
often believed but not really understood. The graph-theoretic approach avoids both of these
pitfalls. The proof, for example, that the determinant of a matrix A and the determinant of AT are
equal tends to become lost in notation when using either the permutation or inductive approach.
With the graph-theoretic approach this result is proven in one sentence!
Recall that a factor F of a digraph H is a subgraph containing all the vertices of H in which
each vertex has both indegree and outdegree equal to one. In other words, it consists of a
collection of disjoint cycles that go through each vertex of H. The number of cycles in the factor F
is denoted n (F). If the digraph His labelled, then W( F) denotes the weight of the factor.
Given a square matrix A = ( a; 1 ) of order n, the Coates digraph D(A) is a labelled digraph with
n vertices where the arc from vertex i to vertex j has a , 1 as a label. In FIGURE 5, the Coates

70 MATHEMATICS MAGAZINE
 all {/I, aD 1 1

A- (I ~I

a., 1
u~~

a,:
lJ ~-\

aJI
2/---~3
"
()
20, Q
1

()2 30 2<._ J 3

~~.
u2:!
0 2 ~3 lf 3\)
a~'

FIGl'Rf 5. The Coates digraph D( A) of a matrix A of order n = 3. and the six factor; of D( A).

digraph D( A) of a 3 X 3 matrix A is shown. The digraph D( A) for n = 3 has six factors; the two
having no loops are the cycles (1,2,3,1) and (1,3,2,1), the three with one loop are (1,1) (2.3,2).
(2,2) (1,3,1) and (3,3) (1,2,1); finally. (1,1) (2,2) (3,3) consists of three loops.
Given a square matrix A of order n, D( A) its Coates digraph, and F the set of all factors of
D(A), then the determinant of A is defined

detA=(-1r L: (-1)" 1 F 1W(F). (1)

FEF
This formulation of the determinant first appeared in 1959 [1], although some of the relationships
between graphs and determinants were known much earlier [4]. Except for the work of Cvetkovic
[2], it has received attention only in a few scattered research papers. In the example of the 3 x 3
matrix (FIGURE 5), the factors having no loops contribute -a 12 a 23 a, 1 and -a 13 a 32 a 21 to the
sum, those having one loop contribute aua 23 a 32 , a 13 a 22 a 31 , and a 12 a 21 a, 1 to the sum, and the
final factor adds - au a 22 a 33 . Since (- 1) 3 = - 1, we observe that equation (1) yields the standard
result. It is straightforward to verify the validity of formula (1) for the determinant for n = 1,
n = 2, and n = 4. For n = 4, we note that the set of all factors can be partitioned into 9 factors
with no loops, 8 factors with one loop, 6 factors with two loops, and one factor with 4 loops.
We now prove that formulation (1) of the determinant is identical with the usual permutation
definition.
THEOREM 3. The graphic and permutation definitions of det A are equimlent. i.e., if det A is
defined hy equation (1 ), then
detA = L sgn(o)al.o(!)a2 of2) . • . an.o{n)•
nES"

where S, is the set of all permutations of the integers {1. 2.... , n }.

Proof. For any permutation o, the set of arcs {(i, o(i))li = 1, ... ,n} forms a factor F.
Conversely, each factor F also corresponds to a permutation o. Hence to complete the proof we
need only show that the two definitions produce values with the same sign. i.e., that
( -1 )" ( -1)" 1 Fi = ~gn( o ). The vertices in a cycle of the factor F correspond precisely to the
integers in a cycle in the decomposition of the corresponding permutation o. In addition. a cyclic
permutation is even if and only if the corresponding cycle in the digraph has an odd numher of

VOL. 57, NO. 2, MARCH 1984 71

vertices. Let e(11) and o(11) denote the number of even and odd cycles in the permutation 11 (or,
equivalently, the odd and even cycles in the factor). Then
sgn( 11) = ( - 1) o( a l, n (F) = e ( 11) + o ( 11),
and n and e( 11) have the same parity. Hence
sgn( 11 ) = ( _ 1) o( a)= ( _
1)"(F)- e(a) = ( -1} e(a)( _ 1 ) n( F)= ( _
1)" ( _ 1)"(F).
Using definition (1), here is the proof that the determinant of a matrix and the determinant of its
transpose are equal. The Coates digraph D(Ar) is obtained from D(A) by reversing the
orientation of each arc; this leaves the factors, the weight of each factor, and the number of cycles
in each factor unchanged, and hence the determinant is unchanged.
In the last section, we considered matrices that correspond to the elementary row operations.
The matrix that interchanges two rows will have Coates digraph of a single factor, i.e., n- 2 loops
and one cycle of length two. Since all weights are equal to one, the determinant is ( -l)n( -1)n-l
= -1. Adding a multiple of one row to another also corresponds to a Coates digraph with only
one factor, and the determinant is 1 in that case. Finally, multiplying a row by A corresponds to a
diagonal matrix, and this matrix has a Coates digraph with only loops, hence its determinant is A.
Consider an upper triangular matrix as a further example. As can be seen in FIGURE 6, if the
vertices of its Coates digraph are placed horizontally in increasing order, then all arcs are loops or
go from left to right, and hence the only possible cycles are loops. But this means that the only
(nonzero) factor is (1,1) (2,2) (3,3) .. . (n, n), hence
detA =(-1)"(-l)"aua 22 ••• a,11,.
We have proved the determinant of an upper triangular matrix is the product of its diagonal
elements.
A slightly harder combinatorial result can be obtained if we let C( n, r) denote the number of
combinations of n things taken r at a time, and let A be a cyclic tridiagonal matrix, i.e., the
nonzero entries of A satisfy a"= d, a; i+ 1 = c, or a;" 1 , =b. The reader can prove that
[n/2]
detA= L C(n-r,r)brcrdn- 2 r.
r=O

Now let us look at the effect of elementary row operations on the Coates digraph and
consequently on the determinant.
THEOREM 4. Let A be a square matrix of order n, 1 ~ i, j ~ n, and let
B be obtained from A by multiplying the ith row by A,
C be obtained from A by interchanging the ith andjth rows,
E be obtained from A by adding thejth row to the ith row.
Then det B = Adet A,
det C = - det A, and
det E= det A.

2 3 11- l 11

FIGURE 6. The Coates graph of an upper triangular matrix.

72 MATHEMATICS MAGAZINE
 h a

FIGURE 7. lbe change in a factor under row interchange.

Proof. D(B) is obtained from D(A) by multiplying the label of each arc going out of the ith
vertex by A. Thus Fis a factor of D(A) if and only if Fis a factor of D( B). Since the outdegree of
each vertex in F is one, the weights of all arcs of Fin D( A) but one are the same as the weights of
F in D(B), the exception being that single arc going out of the ith vertex which has been
multiplied by A. Thus the weight of Fin D( B) is A times that in D( A), and, summing over all of
the factors, det B = Adet A.
The Coates graphs D(A) and D(C) can be related as follows: we wish to take the digraph
D(A) and move some of the arcs. Since a;k = c1A and a 1k = c;k• D(C) is obtained from D(A) by
taking each arc going out of the ith vertex and moving it so that it goes out of thejth vertex, and
conversely, moving the arcs going out of the jth vertex so that they can then go out of the ith
vertex (the terminal end of the arc remains fixed). The net change of these movements of arcs is to
interchange a;k and a 1k for every k, which, of course, is just what is desired. Now consider a
particular factor F, and suppose that the labels of the arcs going out of vertex i and vertexj are a
and b. How do these row interchanges affect F? It remains a factor with W( F) unchanged! As
can be seen in FIGURE 7, if both the ith and thejth vertices are in the same cycle to begin with,
then the movement of the arcs splits the cycle into two new ones with the same weight (left to
right in FIGURE 7). On the other hand, if the i th and j th vertices are in different cycles, then the
movement of the arcs causes them to combine into one cycle with the same weight (right to left in
FIGURE 7). In either case, W(F) is unchanged and n(F) is increased or decreased by one. Thus
the sign of each summand changes, and det C = - det A.
Notice that if A has two identical rows, then interchanging them leaves A unchanged but
reverses the sign of det A, and so det A = 0. Now let A' be obtained from A by replacing the ith
row of A by thejth row of A (so that det A'= 0). Consider a factor F of D(E); it is also a factor
of D(A) and D(A'), and all the weights of the arcs are identical except for the arc in F going out
of the ith vertex. The weight of this arc is a;k + a 1k in E, is a;k in A, and is a 1k in A'. Thus the
weight of Fin D(E) is the sum of the weight of Fin D(A) and the weight of Fin D(A').
Summing over all factors F, we get det E = det A + det A' = det A.
If one already knows how to reduce a matrix to reduced row echelon form, then it is easy to see
from Theorem 4 that det EA = det E · det A for an elementary matrix E and consequently that
det AB= det A· det B holds in general. In the next section we shall prove this result with
graph-theoretic tools.
Sometimes the determinant is defined abstractly as a function from the set of square matrices
of order n to the real numbers which, viewed as an n-variable function of the columns, is
alternating, multilinear, and takes the identity matrix to 1. Since det AT= det A, we can use rows
instead of columns, of course. Suppose F is a factor of a directed graph with n vertices. Then, by
the argument in Theorem 4, the mapping that takes a square matrix M into ( -1)"W(F) is a
linear function in each row of M. Hence, summing over all rows, we see from (1) that the
determinant is indeed an alternating multilinear form on the rows of M, and that the determinant
of the identity matrix is 1. Since it is easily seen that such a form is unique (see [6] p. 191, for
example), we now see that the graphic definition is equivalent to the abstract definition.

VOL. 57, NO. 2, MARCH 1984 73

 We now consider expansion of the determinant by cofactors. Given a square matrix A, let D;1
be defined as the determinant of the matrix obtained by deleting the ith row andjth column from
A. The (i,j) cofactor of A, denoted A;1 , is then defined by

A;1 = ( -lf+J D;1 .

Now consider the set F of all factors of D(A) and a particular vertex i. Each factor contains a
unique arc going out of the ith vertex. Let~ be the set of factors containing the arc (i, j). Then
clearly F is partitioned by~.}= 1,2, ... ,n. The (i, j) cofactor can now be expressed in terms of
~-
THEOREM 5. Let A = (a ,1 ) be a square matrix of order n, 1 ~j ~ n, and let ~ be the set of factors
of D(A) containing the arc (i, }). Then

a;;Aij=(-lr L (-lr(F)W(F). (2)

FEFj
Proof First, suppose that i = j, and let A' be the square matrix of order n - 1 obtained by
deleting the ith row and column from A. By definition, any Fin F; will contain the loop (i, i).
Thus the factors in F, are precisely the factors of D(A') plus the loop (i, i). Since the weight of a
factor in F; is the product of a;; and the weight of a factor in D(A'), the result is clear except,
possibly, for the sign. The factor in F; has one more cycle than the corresponding factor in D(A'),
so an extra multiple of -1 is introduced inside the summation. However the order of D(A) is one
more than that of D(A'), and so one fewer multiple of -1 appears outside the summation; thus
the theorem is valid when i = j.
To complete the proof of the theorem, we must see the effect of interchanging two adjacent
columns of A on the left and right side of the equation (2). Let A' be obtained from A by
interchanging columns rand r + 1. We certainly have a;,= a; r+ 1 and
1
A 1r =(-l)i+'D.tr = -((-1)'+'+ D'r r+ 1 ) ·

Thus interchanging two adjacent columns causes the value of the left side of the equation (2) to
change only in sign. For the right side, we proceed as in the proof of Theorem 4 and rearrange the
arcs of D(A) to form D(A'). Since we are interchanging columns, we move the terminal end
rather than the initial end of the arcs. Thus each arc terminating at the rth vertex is moved
so it terminates at the r + 1st vertex and vice-versa In a manner analogous to that illus-
trated in FIGURE 7, each factor Fin F, becomes a factor F' in F,+ 1 with W(F) = W(F') and
In( F)- n(F')I = 1. Summing over all Fin F, we get

(-1)" L: (-lr<F>w(F)= -(<-Ir L: (-l)"<F'>w(r)).

FEF, F'EF,+I

Thus the right side of equation (2) also changes sign when two adjacent columns are interchanged.
Hence equation (2) is valid for A if and only if it is valid for A'. In other words, if the theorem is
valid for a particular i and j, then it is also valid for i and j + 1.
Now suppose i > j; we then successively interchange columns} andj + l,j + 1 andj + 2,j + 2
andj + 3, etc., until we have finally interchanged columns i - 1 and i. By the preceding argument,
all of the resulting matrices will simultaneously satisfy or not satisfy equation (2) of the theorem.
Since the final matrix is just the case where i = j, we see that the theorem is valid for all of the
matrices, and, in particular, for A. The argument fori <j is symmetric.
COROLLARY (Expansion by cofactors). Let A be a square matrix of order n, 1 ~ i, j ~ n; then
n n
det A= L a;kAik = L akJAkJ·
k=l k=l

Proof As was noted previously, F is partitioned by Fk> k = 1,2, ... ,n. Thus

74 MATHEMATICS MAGAZINE
 I) I)

detA=(-1)" L (-l)"<Flw(F)= L (-l)" L 1

(-1)" FlW(F) L l a"'Aik·
FEF k ~

The Coates digraph has application in other areas of linear algebra including powers of
matrices and spectral theory, although these applications generally focus on the paths in the
digraph rather than on the cycles. For the sake of brevity, these applications must be omitted, hut
even the results presented here establish the natural connection between the Coates digraph and
linear algebra.
Another view of the determinant
We now wish to return to the determinant of the product of two matrices. Our purpose is to
give a graph-theoretic proof, not only as an end in itself, but also as a means of revealing some of
the underlying combinatorial structure. For a square matrix of order n, we must first observe how
a factor in the Coates digraph translates to the Konig digraph. Such a factor is simply a
vertex-disjoint set of n arcs, and so, disregarding the signature for the moment, the summands of
the determinant (Equation (1 )) are precisely the weights of the sub graphs consisting of n
vertex-disjoint arcs.
Now consider the concatenation of the Korug digraphs of two square matrices A and B of
order n. How does this relate to the product of det A and det B? If F1 is the set of factors of G (A)
and F2 is the set of factors of G (B), then

detA·detB=( L (-1)" 1 W(F1 ))( L (-1)"'W(f~))

F, EF, F:EF,

= L ( -1) 111
+
11
'W(FI)W(F2).
F1 EF1
F1 EF 1

In other words. each summand corresponds to the weight of a subgraph in G(A)* G(B)
consisting of 11 vertex-disjoint paths of length 2. Disregarding the sign for the moment, we see that
these subgraphs in fact correspond precisely to the summands of det A · dct B. What about the
summands of det AB? Each arc in AB is obtained from the paths of length 2 in G (A)* G (B).
Further, each arc of length 2 in G( A)* G (B) will contribute to some element in AB and hence to
det AB. Again, expanding the products of sums we see that the summand~ of det AB are the
weights of sub graphs of G (A)* G( B) consisting of n paths of length 2 where the initial and
terminal (but not necessarily the middle) vertices are distinct. The non-distinctness of the middle
vertex accounts for the larger number of term~ ( 11 !n") in det AB as compared with the number
( n ! ) 2 in det A · det B.
To complete the proof we return to the postponed consideration of the signs of the terms. We
shall show that each of the summands in det A · det B also appears in dct AB with the same sign
and that the remaining terms of det AB sum to zero. To do this we must first see a new
relationship between the parity of a permutation a and the Konig digraph. Let the Konig digraph
of a permutation a be the digraph of the corresponding permutation matrix, so the Konig digraph
of a simply consists of the set of arcs (i, tT(i)).
THEOREM 6. The paritv of a permutation a and the paritr of the number of arc intersections in the
Konig digraph are the same.
Proof Suppose a( i) = i for i = 1, 2, ... ,n. Then the number of intersections in the Konig
digraph of a is 0 and a has even parity. If a is an arbitrary permutation. we may proceed by
uncrossing the arcs one pair at a time, while observing that the parity of both the permutation and
the number of arc intersections change with each uncrossing.
Indeed, if the arcs (i, a(i)) and (j, a(j)) intersect, we may say without loss of generality that
i <j and a(i) > a(j). Define a' by a'(i) = a(j), a'(j) = a(i ), and a'( k) = a(k) for all other k.
How does the number of arc intersections of a' compare with that of a'l Any knot equal to i or j

VOL. 57, NO. 2, MARCH 1984 75

must satisfy 1 ::::; k < i or i < k <j, or j < k::::; n. Similarly, 1 ::::; a(k) < a(j), a(j) < a(k) < a(i), or
a(i) < a(k)::::; n. Thus there are nine possible configurations fork and a(k). For eight of them,
the number of arc intersections is unchanged by uncrossing the arcs (i, a(i)) and (j, a(j)). In the
configuration where i < k <j and a(j) < a(k) < a(i) the number of intersections drops by two.
Thus the total number of arc intersections has been decreased by twice the number of arcs of the
form {(k, a( k))ii < k <j, a(j) < a(k) < a(i)} plus one, the extra intersection coming from the
arcs ( i, a (i)) and (j, a (j)) themselves. Hence the transposition ( ij) that takes a to a' also causes a
change in t!".e parity of the number of arc intersections. Eventually, we get to the identity
permutation, and hence the result is established.
COROLLARY. The permutation a and the number of pairs in the set {( i, j) 11 : : ; i < j ::::; n, 1 ::::; a (j)
< a(i)::::; n} have the same parity.
The set defined in the Corollary is called the set of inversions of the permutation a.
Note that in the proof of Theorem 6, each uncrossing represents a transposition, and the
sequence of uncrossings that are used to obtain the identity permutation yields the product of
transpositions that equals the original permutation. Thus the proof also shows that any permuta-
tion is the product of transpositions, and the graph yields an intuitive representation of a as the
product of transpositions.
Now let us apply these results to the products of determinants. One summand of det A · det B
corresponds to n vertex-disjoint arcs of length two. The sign of this term is the product of the
signs of the terms in det A and in det B, each of which can be obtained from the number of arc
intersections in its particular case. How does this compare with the same term as it appears in
det AB? The sign of this term comes from the parity of the number of intersections of the paths of
length two. A pair of paths will intersect if the first arcs of the paths intersect in G(A) or if the
second arcs intersect in G(B). But notice that if both the first arcs and second arcs intersect, then
they yield a pair of non-intersecting arcs as far as det AB is concerned. Thus the number of
intersecting arcs in det AB and the product of those in det A and det B have the same parity,
which is what was desired.
Finally, let us take care of those extra terms that were in det AB but did not appear in
detA ·detB. Let (i,k,/) and (j,k,m) be two paths of length two in G(A)*G(B). How can
these paths appear in det AB? There are two ways: they come from terms of the form
(a;kbk 1 )(a1kbkm)x or terms of the form (a,kbkm)(a 1kbk 1 )x where x represents the remaining
terms in each of the products. Hence the permutations that give rise to these paths can be paired
so that they differ only by an interchange of I and m. Thus the same product appears with
opposite sign, and they sum to zero.
It is interesting to note that if one considers more than two paths passing through the vertex k,
then the fact that the alternating group contains exactly half of the members of the symmetric
group can yield the same result.
The author would particularly welcome comments from readers. Further material can be supplied for those with
'uch interest.
This paper results in part from joint work of the author and his friend and colleague Dragos Cvetkovic of the
University of Belgrade, whose textbook [31 presents linear algebra from a graph-theoretic viewpoint.

References
[l 1 C. L Coates. Flow graph solutions of linear algebraic equations. IREE Trans. Circuit Theory. CT-6 (1959)
no-un.
[ 21 D. M. Cvetkovic. The determinant concept defined by means of graph theory. Mat. Vesnik, 12 (1975)
333-336.
[ 3] _ _ , Kombinatorina Tcorija Matrica, Haunca Rnjiga, Belgrade. 1n0.
[4 ] D. Konig, Ubcr Graphen und ihre Anwendung auf Determinantentheoric and Mengenlehrc, Math. Ann., 77
(1916) 453-465.
[5 ] _ _ , Thcorie dcr end lichen und uncndlichen Graphen, Akadcm. Verlagsges., Leipzig, 1936.
[6 1 S. Lang. Linear Algebra, Addison-Wcsley, Reading, Mass., 1971.

76 MATHEMATICS MAGAZINE