Des. Codes Cryptogr.

(2012) 65:223–232
DOI 10.1007/s10623-011-9548-3

The maximum order of adjacency matrices of graphs

with a given rank

W. H. Haemers · M. J. P. Peeters

Received: 28 October 2010 / Revised: 19 May 2011 / Accepted: 14 July 2011 /

Published online: 2 August 2011
© The Author(s) 2011. This article is published with open access at Springerlink.com

Abstract We look for the maximum order m(r ) of the adjacency matrix A of a graph G
with a fixed rank r , provided A has no repeated rows or all-zero row. Akbari, Cameron and
Khosrovshahi conjecture that m(r ) = 2(r +2)/2 − 2 if r is even, and m(r ) = 5 · 2(r −3)/2 − 2 if
r is odd. We prove the conjecture and characterize G in the case that G contains an induced
subgraph r2 K 2 or r −3
2 K2 + K3.

Keywords Graph · Adjacency matrix

Mathematics Subject Classification (2000) 05B20 · 05C50

1 Introduction

In this paper we discus the problem of determining the maximum number of vertices of a
graph in terms of the rank of its adjacency matrix. This problem is only properly defined if
the matrix has no repeated rows. Given a graph we can duplicate a vertex arbitrarily often
and add isolated vertices as many as we like without changing the rank. This motivates the
following definition [2].
Definition 1 A graph is reduced if it has no isolated vertices and no two vertices have the
same set of neighbors.
Although we find this definition more convenient, some authors [1,3] use ‘reduced’ purely
for the condition that no two vertices have the same set of neighbors and allow the graph to
have at most one isolated vertex. This explains why there is sometimes a difference of one
in their results and ours.

This is one of several papers published together in Designs, Codes and Cryptography on the special topic:
“Combinatorics – A Special Issue Dedicated to the 65th Birthday of Richard Wilson”.

W. H. Haemers (B) · M. J. P. Peeters

Department of Econometrics and Operations Research, Tilburg University, Tilburg, The Netherlands
e-mail: [email protected]

123
224 W. H. Haemers, M. J. P. Peeters

Let m(r ) be the maximum number of vertices in a reduced graph with rank r over R.
There are clearly no graphs of rank 1, and the only reduced graphs of rank 2 and 3 are K 2
and K 3 , respectively. The number m(r ) can be defined for any field and for all these cases
there is an upper bound of 2r − 1 (see Sect. 3). Godsil and Royle [2] prove that for the field
F2 , m(r ) is equal to this upper bound. In this paper we only consider m(r ) for the field of
real numbers.
Kotlov and Lovász [3] improve the upper bound on m(r ) to O(2r/2 ). They also give a
construction (Construction (a) in the next section) that transforms a reduced graph of rank r
on n vertices into a reduced graph of rank r + 2 on 2n + 2 vertices.
Starting with K 2 (with n = r = 2) or K 3 (with n = r = 3) and recursively applying the
construction leads to graphs of rank r with n(r ) vertices where
 (r +2)/2
2 −2 if r is even,
n(r ) = .
5 · 2(r −3)/2 − 2 if r is odd, r > 1.
So we have m(r ) ≥ n(r ). Akbari et al. [1] give a second construction (Construction (b) in
the next section) that transforms a reduced regular graph of rank r on n vertices with degree
2 into a reduced regular graph of rank r + 2 on 2n + 2 vertices with degree n + 1. For this
n

construction one can start with K 2 or the complement of cube (which has r = 5). In fact, at
any step in this construction process, one can change to Construction (a). This leads to  r2 
nonisomorphic graphs of rank r with n(r ) vertices. The authors of [1] conjecture that these
and only these graphs have the maximum number of vertices m(r ).
Conjecture If r ≥ 2, then m(r ) = n(r ), and the extremal graphs can be obtained from
K 2 , K 3 or the complement of the cube by Constructions (a) and (b).
The conjecture has been verified by computer up to r = 8. In this paper we shall prove
the conjecture for reduced graphs with rank r that contain an induced subgraph r2 K 2 or
r −3
2 K 2 + K 3 . We also present two new constructions. These constructions, however, do not
lead to a counter example for the conjecture.

2 Constructions

The lower bound n(r ) for m(r ) mentioned above, is based upon constructions that transform
a reduced graph on n vertices with rank r into a reduced graph on 2n + 2 vertices with rank
r + 2. Here we present four such constructions. The first construction is due to Kotlov and
Lovász [3] and the second one to Akbari, Cameron and Khosrovshahi [1]. The other two
seem to be new.
Construction (a) [1,3] Let G be a reduced graph with n vertices, adjacency matrix A and
rank r . Construct the graph G a with 2n + 2 vertices and adjacency matrix:
⎡ ⎤
A A 0 0
⎢ A A 1 0⎥
⎢ ⎥
⎣ 0 1 0 1⎦,
0 0 1 0
where 0 and 1 are the all-zero and the all-one row or column vectors, respectively. It is
straightforward to check that G a is reduced and has rank r + 2.
Construction (b) [1] Let G be a reduced regular graph with n vertices with degree n2 ,
adjacency matrix A and rank r . Construct the graph G b with 2n + 2 vertices and adjacency
matrix:

123
The maximum order of adjacency matrices of graphs 225

⎡ ⎤
A A 1 0
⎢A A 0 1⎥
⎢ ⎥.
⎣1 0 0 1⎦
0 1 1 0

It is straightforward to check that G b is a reduced regular graph of rank r + 2 with degree

n + 1.
Construction (c) Let G be a reduced graph with n vertices, adjacency matrix A and rank
r . Construct the graph G c with 2n + 2 vertices and adjacency matrix:
⎡ ⎤
A A 1 0
⎢ A A 0 1⎥
⎢ ⎥
⎣ 1 0 0 1⎦,
0 1 1 0

where A = J − A for the all-one matrix J . It is straightforward to check that G c is a reduced

regular graph of rank r + 2 with degree n + 1.
Construction (d) Let G be a reduced regular graph with n vertices with degree n2 , adja-
cency matrix A, and rank r . Construct the graph G d with 2n + 2 vertices and adjacency
matrix:
⎡ ⎤
A A 0 0
⎢ A A 1 0⎥
⎢ ⎥
⎣ 0 1 0 1⎦.
0 0 1 0

It is straightforward to check that G d is a reduced graph with rank r + 2.

Notice that Constructions (a) and (c) can be applied for any reduced graph, whereas Con-
structions (b) and (d) can only be applied to reduced regular graphs of degree n2 . On the other
hand Constructions (b) and (c) produce reduced regular graphs of degree half the number
of vertices, whilst Constructions (a) and (d) produce non-regular graphs. In fact, a stronger
property holds for the outcome of Construction (c).

Definition 2 We say that a graph G is 1-closed if it is reduced and for any column of its
adjacency matrix there is another column such that the two columns add up to the all-one
vector 1.

The following results follow straightforwardly from the definitions.

Proposition 1 Let G be a graph with n vertices.

(i) If G is reduced, then G c is 1-closed.

(ii) If G is 1-closed, then G is regular of degree n2 .
(iii) If G is 1-closed, then so is G b .
(iv) If G is 1-closed, then G b and G c are isomorphic.
(v) If G is 1-closed, then G a and G d are isomorphic.

Although the four constructions are different they often give the same outcome. For example
Constructions (b) and (c) produce isomorphic graphs if they are applied to a 1-closed graph,
and the same is true for Constructions (a) and (d). Another example is the following lemma.

123
226 W. H. Haemers, M. J. P. Peeters

Lemma 1 Let G be a reduced graph. Then (G a )c and (G c )b are isomorphic.

Proof Let A be the adjacency matrix of G then (G a )c has adjacency matrix

1 2 3 4 5 6 7 8 9 10
1 A A 0 0 A A 1 1 1 0
2 A A 1 0 A A 0 1 1 0
3 0 1 0 1 1 0 1 0 1 0
4 0 0 1 0 1 1 0 1 1 0
5 A A 1 1 A A 0 0 0 1 ,
6 A A 0 1 A A 1 0 0 1
7 1 0 1 0 0 1 0 1 0 1
8 1 1 0 1 0 0 1 0 0 1
9 1 1 1 1 0 0 0 0 0 1
10 0 0 0 0 1 1 1 1 1 0

If the rows and columns are re-ordered as indicated by the numbers we get the following
matrix:
2 5 9 4 1 6 8 10 3 7
2 A A 1 0 A A 1 0 1 0
5 A A 0 1 A A 0 1 1 0
9 1 0 0 1 1 0 0 1 1 0
4 0 1 1 0 0 1 1 0 1 0
1 A A 1 0 A A 1 0 0 1.
6 A A 0 1 A A 0 1 0 1
8 1 0 0 1 1 0 0 1 0 1
10 0 1 1 0 0 1 1 0 0 1
3 1 1 1 1 0 0 0 0 0 1
7 0 0 0 0 1 1 1 1 1 0

which is the adjacency matrix of (G c )b . 



By Proposition 1, G c is 1-closed and hence (G c )c = (G c )b . Therefore:

Corollary 1 If G is a reduced graph, then (G a )c and (G c )c are isomorphic.

For the even case, one can start with G = K 2 of rank 2 which is 1-closed. For each k ∈
{0, 1, 2, . . . , r −2
2 }, one can first apply k times Construction (c) (or (b), which gives the same
outcome), and then apply r2 − 1 − k times Construction (a) (or (d) once, followed by r2 − 2 − k
times Construction (a)). In this way one finds r2 non-isomorphic reduced graphs of rank r on
n(r ) = 2(r +2)/2 − 2 vertices if r is even.
For the odd case, we start with G = K 3 of rank 3. For each k ∈ {0, 1, 2, . . . , r −32 } one can
r −3
first apply k times Construction (c), and then apply 2 −k times Construction (a). In this way
one finds r −1 2 non-isomorphic reduced graphs of rank r on n(r ) = 5 · 2
(r −3)/2 − 2 vertices.
c
Note that if G is K 3 , then G is the complement of the cube. Therefore these graphs can also
be obtained by use of Constructions (a) and (b), if we start with K 3 or the complement of the
cube. So we can conclude that the new constructions give no new reduced graphs of rank r
with n(r ) vertices.

123
The maximum order of adjacency matrices of graphs 227

3 Linear extensions

Let G be a reduced graph with n vertices and adjacency matrix A of rank r . Then A contains
a principle r × r submatrix B of rank r which corresponds to an induced subgraph H of G.
We call G a linear extension of H . We have

B Y B  
A=   −1 =  B −1 B Y .
Y Y B Y Y

Observe that a zero column or two identical columns in [ B Y ] would imply that G is not
reduced. Therefore [ B Y ] has at most 2n − 1 columns, so m(r ) ≤ 2n − 1.
The above structure of A can be used to design a method to derive all linear extensions of
a given graph of full rank r . This method is also described in a slightly different form in [1]:
Step 1 Let H be a graph with r vertices and adjacency matrix B of full rank r . Check
for all (0, 1)-vectors v
= 0 whether v  B −1 v = 0 and keep the vector v if so. Let
(v 1 , v 2 , . . . , v k ) be the list of such vectors. Clearly this list contains all columns of
B. Assume without loss of generality that B = [ v 1 v 2 · · · vr ].
Step 2 Construct a graph G on the vertex set {v 1 , v 2 , . . . , v k } as follows. For each two
distinct indices i and j, put an edge from v i to v j if and only if v i B −1 v j ∈ {0, 1}.
Notice that any of the vertices v 1 , v 2 , . . . , vr is adjacent to all other vertices, because
B −1 v i = ei for i ∈ {1, . . . , r }. For a clique C that contains v 1 , v 2 , . . . , vr , con-
struct a matrix X that contains the vertices of C as columns. If we suppose that
X = [ B Y ], then

 −1 B −1
  B Y
A=X B X= B B Y =
Y Y  Y B −1 Y

is the adjacency matrix of a reduced graph of rank r containing H as a subgraph.

By determining all cliques of G , we determine all linear extensions of H .
We will use this method in the next section for the special cases H = r
2 K2, and H =
K 3 + r −3
2 K2.

4 Main result

Let m be even. Consider the following m × m matrix pattern.

⎡ ⎤
0 ∗ ∗ ∗ ··· ··· ∗ ∗ ∗ 1
⎢0 0 ∗ ∗ ··· ··· ∗ ∗ 1 0⎥
⎢ .. ⎥
⎢ . · · · · · · ..
. ⎥
⎢0 0 0 1 0 0⎥
⎢ ..
⎥
⎢ . ∗ ∗ ..
. ⎥
⎢0 0 0 0 0 0⎥
⎢. . . . ⎥
⎢. .. . .. .. ⎥
∗ ⎢. 0 1 .. ⎥
Mm = ⎢ . . . . ⎥
⎢. . . . . ⎥
⎢. . .. 1 0 . . ⎥
⎢. ⎥
⎢. .
.. . . . . . . . . . .. .
..
. ⎥
.. ⎥
⎢. .
⎢ ⎥
⎢ . . .. . ⎥
. ⎥
⎢0 0 1 0 . . .
⎣0 1 0 0 ··· ··· ··· ··· 0 0
⎦
1 0 0 0 ··· ··· ··· ··· 0 0

123
228 W. H. Haemers, M. J. P. Peeters

Every column of this matrix stands for a type of (0, 1)-vector where two (0, 1)-vectors
have the same type if they only differ in the ∗-positions. So a column with k ∗-positions
m
defines 2k (0, 1)-vectors. So in total we get 2(2 2 − 1) (0, 1)-vectors. We define Mr to be the
r × (2(r +2)/2 − 2) matrix whose columns are the vectors defined by the pattern of Mr∗ . Let
R be the reverse identity matrix of order r , that is the adjacency matrix of r2 K 2 , and define
Ar = Mr R Mr . It is straightforward to check that Ar is a symmetric (0, 1)-matrix with rank
r
r and zeros on the diagonal. So Ar defines a reduced graph of rank r on n(r ) = 2(2 2 − 1)
vertices. We call this graph G r .

Theorem 1 Let r be even and let G be a linear extension of r2 K 2 . Then G has at most n(r )
vertices. In case of equality G is isomorphic to G r .

Proof Let A be the adjacency matrix of G. Then

R Y R  
A= = R −1 R Y = Z  R Z ,
Y  Y  R −1 Y Y
 
where Z = R Y . Now every column z of Z is a (0, 1)-vector of length r satisfying
z Rz = 0. Let V be the set of all (0, 1)-vectors z
= 0 of length r satisfying z Rz = 0.
Then the columns of Z are elements of V . Define a graph G with vertex set V where two
vertices x and y are adjacent if x  Ry ∈ {0, 1}. Clearly the columns of Z form a clique in
G . We now prove that G can be properly colored with n(r ) colors proving that G has at most
n(r ) vertices.
Let z = [z 1 · · · zr ] be a column of Z . We know that z Rz = 0. So if z i = 1, then
zr +1−i = 0, which means that z i + zr +1−i ≤ 1. Hence z has at most r2 coefficients that are
equal to 1. We call the set

Iz = {i ∈ {1, 2, . . . , r2 } : z i + zr +1−i = 1}

the footprint of z. Now if I ⊆ {1, 2, . . . , r2 } and I has cardinality k > 0, then there are 2k
r 
different vectors z ∈ V with footprint Iz = I . Furthermore, there are k2 different subsets
of {1, 2, . . . , r2 } of cardinality k. As a consequence,

2 r 
r
 r
|V | = 2 2k = 3 2 − 1.
k
k=1

Out of the 2k different vectors z ∈ V with footprint Iz = I , there are 2k−1 vectors for which
|{i ∈ {1, 2, . . . , r2 } : z i = 1}| is even and 2k−1 vectors for which |{i ∈ {1, 2, . . . , r2 } : z i = 1}|
is odd. Define
V (I ) = {z ∈ V : Iz = I };
V (I )0 = {z ∈ V (I ) : |{i ∈ {1, 2, . . . , r2 } : z i = 1}| is even};
V (I )1 = {z ∈ V (I ) : |{i ∈ {1, 2, . . . , r2 } : z i = 1}| is odd}.
We claim that the sets V (I )0 and V (I )1 are cocliques in G . For any x, y ∈ V (I ), one can
easily check that

x  Ry = |{i ∈ I : x i
= y i }| =

|{i ∈ I : x i = 1}| + |{i ∈ I : y i = 1}| − 2|{i ∈ I : x i = y i = 1}|.

123
The maximum order of adjacency matrices of graphs 229

So x  Ry = 0 if and only if x = y. Moreover, if either x, y ∈ V (I )0 or x, y ∈ V (I )1 , then

x  Ry is even. Hence, if x
= y, then x  Ry
∈ {0, 1}. This proves the claim. Now, the vertex
set of G can be partitioned in the right number of cocliques:

2 r 
r

2 2 = n(r ).
k
k=1

So the chromatic number of G is at most n(r ), as desired.

Finally we have to prove that all cliques of maximum size in G correspond with a graph
that is isomorphic to G r . For this, we make use of the automorphisms of r2 K 2 . Note that one
can permute the r2 edges arbitrarily and that every edge can be flipped. These automorphisms
imply transformations on elements of V that do not change their mutual inner products:
φ(x) Rφ(y) = x  Ry. We show that each maximum clique in V can be transformed to the
maximum clique defined by the pattern of Mr∗ .
A clique of maximum size should have one vertex in every coclique of the described
r
coloring. V contains 2 2 vectors of weight r2 , two of which are in the clique. Since all these
vectors correspond to all possible ways to assign 0’s and 1’s to the vertices of r2 K 2 such that
every edge has a 1 and a 0, these are all isomorphic. So without loss of generality the two
vectors of weight r2 in the maximum clique are the characteristic vectors of {1, 2, . . . , r2 } and
{1, 2, . . . , r2 − 1, r2 + 1}, since their inner product should be 0 or 1. Now consider all vectors
with footprint {1, 2, . . . , r2 − 1}. Two of these are in the clique and since the inner product
with the previous two vectors should be 0 or 1, without loss of generality these two vectors
are the characteristic vectors of {1, 2, . . . , r2 − 1} and {1, 2, . . . , r2 − 2, r2 + 2}. Now consider
all vectors in V with footprint {1, 2, . . . , r2 − 2}, etcetera.
So without loss of generality and by using the automorphisms of r2 K 2 , the columns of
∗
Mr , where every ∗ is replaced by a 1 are part of the maximum clique. All other vertices of
the maximum clique are fixed, since there is only one candidate left in each coclique that can
be added to this set. Let I ⊆ {1, 2, . . . , r2 } and let k be the largest element of I . Consider the
vectors with footprint I . Then the only two vectors in V with footprint I that can be in the
clique are the characteristic vectors of I and I ∪ {r − k + 1}\{k}. The resulting maximum
clique is the one defined by the pattern of Mr∗ . 


Theorem 2 Let r ≥ 5 be odd and let G be a linear extension of K 3 + r −3 2 K 2 . Then G has

at most n(r ) vertices and there is a unique graph for which equality holds.

Proof Let A be the adjacency matrix of G, let R be the reverse identity matrix of order r − 3,
and let  be the adjacency matrix of the triangle K 3 . Then

R O
N=
O 

is the adjacency matrix of K 3 + r −3

2 K 2 and A = Z N
 −1 Z for some r ×n (0, 1)-matrix Z . Let

V be the set of all (0, 1)-vectors z

= 0 of length r satisfying z N −1 z = 0, and define G to be
the graph with vertex set V where two vertices x and y are adjacent if x  N −1 y ∈ {0, 1}. Then
clearly the columns of Z form a clique in G . We shall prove that G can be properly colored
with n(r ) colors proving that G has at most n(r ) vertices. Let z = [xy] ∈ V with y ∈ {0, 1}3
and x ∈ {0, 1}r −3 , then z is a (0, 1)-vector satisfying z N −1 z = x  Rx + 21 y  (− I )y = 0,
which can only happen if x  Rx = 0 and y  ( − I )y = 0. So

123
230 W. H. Haemers, M. J. P. Peeters
⎧ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤⎫
⎨ 0 0 1 1 ⎬
y ∈ y0 = ⎣ 0 ⎦ , y1 = ⎣ 1 ⎦ , y2 = ⎣ 0 ⎦ , y3 = ⎣ 1 ⎦
⎩ ⎭
0 1 1 0
r−3
and hence V has 4 · 3 2 − 1 elements. Let W be the set of all (0, 1)-vectors x
= 0 of
length r − 3 satisfying x  Rx = 0. For a vector x = [x1 · · · xr −3 ] ∈ W , we define its
footprint Ix as
Ix = {i ∈ {1, 2, . . . , r −3
2 } : x i + xr −2−i = 1}.

For I ⊆ {1, 2, . . . , r −3
2 }, I
= ∅ we define the following sets:

W (I ) = {x ∈ W : Ix = I };
W (I )0 = {x ∈ W (I ) : |{i ∈ {1, 2, . . . , r −3
2 } : x i = 1}| is even };
W (I )1 = {x ∈ W (I ) : |{i ∈ {1, 2, . . . , r −3
2 } : x i = 1}| is odd }.
Now the following partition of V defines a proper coloring of G .
First the three sets

 
 

0 0 0
, , .
y1 y2 y3

Next, for every I ⊆ {1, 2, . . . , r −3

2 }, I
= ∅ we take the following five sets:

 

x x
: x ∈ W (I )0 , : x ∈ W (I )1 ,
y0 y0

 

x x
: x ∈ W (I )0 ∪ : x ∈ W (I )1 ,
y1 y2

 

x x
: x ∈ W (I )0 ∪ : x ∈ W (I )1 ,
y2 y3

 

x x
: x ∈ W (I )0 ∪ : x ∈ W (I )1 ,
y3 y1
r−3
Similarly as in the previous proof it can be checked that the 5 · (2 2 − 1) + 3 = n(r ) sets
are cocliques in G . Thus we found a proper coloring with n(r ) colors, as promised.
To show that all cocliques in G of size n(r ) lead to isomorphic graphs, we will also use
the following proper coloring of G .
First the three sets

 
 

0 0 0
, , .
y1 y2 y3

Next, for every I ⊆ {1, 2, . . . , r −3

2 }, I
= ∅ we take the following five sets:

 

x x
: x ∈ W (I )0 , : x ∈ W (I )1 ,
y0 y0

 

x x
: x ∈ W (I )0 ∪ : x ∈ W (I )1 ,
y1 y3

 

x x
: x ∈ W (I )0 ∪ : x ∈ W (I )1 ,
y3 y2

 

x x
: x ∈ W (I )0 ∪ : x ∈ W (I )1 .
y2 y1

123
The maximum order of adjacency matrices of graphs 231

Now suppose we have a clique in G of size n(r ). Then this clique contains one vertex of each
of the cocliques of the two proper colorings. So it contains the three vectors

0 0 0
, and .
y1 y2 y3
r−3
The clique also contains 2 · (2 2 − 1) vertices of the form


x
: x ∈ W (I ) .
y0
Like in the proof of Theorem 1, it follows that without loss of generality we can take
the vertices for which x is a vector that fits the pattern of Mr∗−3 . Next for every I ⊆
{1, 2, . . . , r −3
2 }, I
= ∅ we have three vectors

x1 x2 x3
, and
y1 y2 y3
in the clique. Because the clique contains a vertex in each color class of both given color-
ings, it follows that either x 1 , x 2 , x 3 ∈ W (I )0 or x 1 , x 2 , x 3 ∈ W (I )1 . It can be verified
that without loss of generality x 1 , x 2 and x 3 are all three equal to the characteristic vector
of I . 


5 Concluding remarks

In the previous section we considered maximum linear extensions of r2 K 2 and r −3 2 K2 + K3.

If the conjecture of Akbari, Cameron and Khosrovshahi is correct, then a linear extension of
any other graph of full rank r has at most n(r ) vertices. For a linear extension of a graph H
with adjacency matrix B to be large, we need many vectors v for which v  B −1 v = 0. For
most choices of H the number of such vectors is rather small. Especially if B −1 contains few
zeros, one doesn’t expect many such vectors. If H is the complete graph K r , for example,
the only such vectors are the columns of B. Intuitively one expects that the two choices of
H considered in this paper have the maximum number of vertices in their maximum linear
extensions. Proving this intuitive result, would prove the conjecture. Therefore we strongly
believe that the conjecture is true.
The following results are straightforward consequences of the constructions of Sect. 2.
Proposition 2 If G is a linear extension of H , then G a is a linear extension of H + K 2 .
As a consequence, it follows that the graphs G r with n(r ) vertices, mentioned in Theorem 1
and 2, are isomorphic to the ones obtained by repeatedly applying Construction (a) starting
from K 2 if r is even, or K 3 if r is odd. The other constructions are linear extensions of graphs
H different from r2 K 2 and r −3 2 K2 + K3.

Proposition 3 Let G be a linear extension of H = (V, E), and let H  = (V, E  ) with
V  = V ∪ {v1 , v2 } and E  = E ∪ {{v, v1 } : v ∈ V } ∪ {v1 , v2 }. Then G a , G b , G c and G d (if
defined) are linear extensions of H  .
If G is a reduced graph of rank r , then any induced subgraph of full rank r can be linearly
extended to G. In particular, for the unique graph G r of Theorem 1, every induced subgraph
of full rank r has a linear extension with at least n(r ) vertices. This leads to the following
result.

123
232 W. H. Haemers, M. J. P. Peeters

Theorem 3 Let r be even and let H be a graph on r vertices having an adjacency matrix B
of the following form.
⎡ ⎤
0 1
⎢ 0 ⎥
⎢ C 1 ⎥
⎢ .. . . ⎥
⎢ . . ⎥
⎢ ⎥
⎢  0 1 ⎥
B=⎢
⎢
C
1 0
O ⎥.
⎥
⎢ ⎥
⎢ . . . .. ⎥
⎢ . ⎥
⎢ ⎥
⎣ 1 O 0 ⎦
1 0
Then H has rank r and G r is a linear extension of H .

Proof Clearly rank B = r . Let Ar be the adjacency matrix of G r , and recall that R is the
adjacency matrix of r2 K 2 . Define
⎡ ⎤
0 1
⎢ 0 ⎥
⎢ C 1 ⎥
⎢ .. ..
. ⎥
⎢ . ⎥
⎢ ⎥
⎢ 0 1 ⎥
N =⎢ ⎢
O
1 0
O ⎥.
⎥
⎢ ⎥
⎢ .. .. ⎥
⎢ . . ⎥
⎢ ⎥
⎣ 1 O 0 ⎦
1 0

Then B = N R N . Since each column of N is also a column of Mr , we get that H is a

subgraph of G r . 


Thus, because of Proposition 3 and Theorem 3, we have many graphs of full rank r whose
maximum linear extension has at least n(r ) vertices. To show that these linear extensions
have at most n(r ) vertices could be an important step towards a proof of the conjecture.

Acknowledgment We thank the referees for many important remarks, which lead to a considerable improve-
ment of the manuscript.

Open Access This article is distributed under the terms of the Creative Commons Attribution Noncommer-
cial License which permits any noncommercial use, distribution, and reproduction in any medium, provided
the original author(s) and source are credited.

References

1. Akbari S., Cameron P.J., Khosrovshahi G.B.: Ranks and signatures of adjacency matrices (preprint 2004).
http://www.maths.qmw.ac.uk/~pjc/preprints/ranksign.pdf.
2. Godsil C.D., Royle G.F.: Chromatic number and the 2-rank of a graph. J. Combin. Theory Ser. B 81,
142–149 (2001).
3. Kotlov A., Lovász L.: The rank and size of graphs. J. Graph Theory 23, 185–189 (1996).

123