See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/250796969

On Brualdi's Theorem

Article  in  Journal of Mathematical Sciences · June 2004

DOI: 10.1023/B:JOTH.0000026284.29231.b4

CITATIONS READS

3 79

1 author:

L. Yu. Kolotilina
Russian Academy of Sciences
111 PUBLICATIONS   1,522 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Nekrasov matrices and generalizations View project

All content following this page was uploaded by L. Yu. Kolotilina on 28 July 2014.

The user has requested enhancement of the downloaded file.

Journal of Mathematical Sciences, Vol. 121, No. 4, 2004

ON BRUALDI’s THEOREM
L. Yu. Kolotilina UDC 512.643

This paper studies irreducible matrices A = (aij ) ∈ C n×n , n ≥ 2, satisfying Brualdi’s conditions



|aii | ≥ Ri (A), γ ∈ C(A),
i∈γ i∈γ


or, shortly, Brualdi matrices. Here, Ri (A) = |aij |, i = 1, . . . , n; C(A) is the set of circuits of length k ≥ 2 in the
j=i
directed graph of A, and γ is the support of γ.
Among the results obtained are a characterization of Brualdi matrices, implying, in particular, that they are
generalized diagonally dominant; the necessary and sufficient conditions of singularity of Brualdi matrices; explicit
expressions for the absolute values of the components of right null-vectors of a singular Brualdi matrix, and conditions
necessary and sufficient for a boundary point of Brualdi’s inclusion region to be an eigenvalue of an irreducible
matrix. Bibliography: 8 titles.

1. Introduction
In [3], R. Brualdi established the following important result.
Theorem 1.1 [3]. Let A = (aij ) ∈ C n×n , n ≥ 2, be an irreducible matrix and assume that



|aii| ≥ Ri(A), γ ∈ C(A), (1.1)
i∈γ i∈γ

with strict inequality for at least one circuit γ. Then A is nonsingular.

Here and below, we use the following notation:


n
Ri(A) = |aij |, i = 1, . . . , n;
j=1
j=i

C(A) denotes the set of circuits of length k ≥ 2 in the directed graph D(A) of the matrix A. Recall that a circuit
in D(A) is an ordered sequence γ of vertices i1 , . . . , ik , ik+1 = i1 , k ≥ 1, where i1 , . . . , ik are all distinct, and
(ij , ij+1 ), j = 1, . . . , k, are arcs of D(A). If γ is a circuit of D(A), then γ stands for the support of γ, i.e., for
the set {i1 , . . . , ik }.
Brualdi’s theorem generalizes the following classical results in matrix theory.
Theorem 1.2 (Lévy-Desplanques theorem). If a matrix A = (aij ) ∈ C n×n , n ≥ 2, is strictly (row) diagonally
dominant, i.e.,
|aii| > Ri (A), i = 1, . . . , n,
then A is nonsingular.
As is well known, the alternative formulation of the above result is as follows.
Theorem 1.3 (Gerschgorin’s theorem). The eigenvalues of a matrix A = (aij ) ∈ C n×n , n ≥ 2, lie in the union

n
G= Gi of the Gerschgorin circles
i=1
 
Gi = z ∈ C : |z − aii| ≤ Ri(A) , i = 1, . . . , n.

Translated from Zapiski Nauchnykh Seminarov POMI, Vol. 284, 2001, pp. 48–63. Original article submitted November
16, 2001.

1072-3374/04/1214-2465 
c 2004 Plenum Publishing Corporation 2465
Theorem 1.4 [8]. Let A = (aij ) ∈ C n×n , n ≥ 2, be an irreducible matrix. If

|aii| ≥ Ri (A), i = 1, . . . , n,

and at least one of these inequalities is strict, then A is nonsingular.

Theorem 1.5 [7, 2]. If A = (aij ) ∈ C n×n , n ≥ 2, and

|aii| |ajj | > Ri(A) Rj (A), i = j, i, j = 1, . . . , n,

then A is nonsingular.
The latter theorem can obviously be reformulated in terms of the ovals of Cassini
 
Cij (A) = z ∈ C : |z − aii | |z − ajj | ≤ Ri (A) Rj (A) , i = j, i, j = 1, . . . , n,

as follows.
Theorem 1.6. All the eigenvalues of a matrix A = (aij ) ∈ C n×n , n ≥ 2, are contained in the domain


n
C(A) = Cij (A).
i,j=1
i=j

Note that Taussky’s Theorem 1.4 refers to the sparsity pattern of A only by requiring that A be irreducible,
whereas Brualdi’s theorem explicitly takes into account the matrix sparsity pattern.
On the other hand, it should also be noted that Theorems 1.1, 1.2, 1.4, and 1.5 provide sufficient conditions
for a matrix to be nonsingular but say nothing about their necessity.
The problem of singularity/nonsingularity for irreducible matrices satisfying nonstrict inequalities of general-
ized diagonal dominance was completely solved in [5], where the following result was established.
Theorem 1.7 [5]. An irreducible matrix A = (aij ) ∈ C n×n , n ≥ 2, satisfying the nonstrict inequalities


n
|aij |vj ≤ |aii|vi, i = 1, . . . , n, (1.2)
j=1
j=i

where v = (vi ) ∈ Rn is a positive vector, is singular if and only if all these inequalities hold with equality, and
there exists a unitary diagonal matrix D such that

D∗ (I − DA−1 A)D = |I − DA−1 A|. (1.3)

Further, if A is singular, then the eigenvalue λ = 0 is of geometric multiplicity one, and the corresponding right
eigenspace is spanned by the vector Dv.
Here and below, DB means the diagonal part of the matrix B, and |B| denotes the matrix composed of the
absolute values of the corresponding entries of B.
The lemma below shows that, instead of establishing the existence of a unitary diagonal matrix D for which
(1.3) holds true, one can verify a finite set of conditions specified by the circuit set.
Lemma 1.1 [4]. Let a matrix B = (bij ) ∈ C n×n , n ≥ 2, be irreducible. Then the following assertions are
equivalent:
(i) B = D∗ |B|D, where D is a unitary diagonal matrix;
(ii) for every circuit γ = i1 , . . . , ik , ik+1 = i1 , k ≥ 1, in the directed graph of B, the equality
 k 


k


bij ij+1 =  bij ij+1 
j=1 j=1

is satisfied.
Using Lemma 1.1, one can reformulate Theorem 1.7 as follows.

2466
Theorem 1.8 [5]. An irreducible matrix A = (aij ) ∈ C n×n , n ≥ 2, that satisfies inequalities (1.2) is singular if
and only if all the relations in (1.2) are equalities and, for every circuit i1 , . . . , ik , ik+1 = i1 of length k ≥ 2 in
the digraph D(A), the equality
 k 


k

−1 
(−1) k
a−1  aij ij aij ij+1 
ij ij aij ij+1 =
j=1 j=1

is valid.
The rest of this paper is organized as follows. In Sec. 2, we bring Brualdi’s theorem to its complete form
(see Theorem 2.2), i.e., we indicate a finite set of conditions necessary and sufficient for an irreducible matrix
satisfying inequalities (1.1) to be singular. These conditions are derived using a characterization of matrices that
satisfy the assumptions of Brualdi’s theorem, which, in particular, implies that such matrices possess generalized
diagonal dominance. Also we study the right null-space of a singular Brualdi matrix.
Section 3 considers two applications of the results obtained in Sec. 2. First, we consider inclusion regions
for eigenvalues of an irreducible matrix and derive conditions necessary and sufficient for a boundary point of
their union to be an eigenvalue of the matrix. Second, we show that the generalized Ostrowski–Brauer theorem
[6] (see Theorem 3.3) can easily be proved by using the results of Sec. 2.

2. Main results
The first theorem of this section provides a description of irreducible matrices satisfying Brualdi’s conditions



|aii| ≥ Ri(A), γ ∈ C(A). (2.1)
i∈γ i∈γ

Theorem 2.1. An irreducible matrix A = (aij ) ∈ C n×n , n ≥ 2, satisfies inequalities (2.1) if and only if there
exist vi > 0, i = 1, . . . , n, such that

Ri (A) ≤ |aii|vivj−1 for all i = j such that aij = 0. (2.2)

Further, all the relations in (2.1) hold with equality if and only if all the inequalities in (2.2) are equalities.
Proof. Sufficiency. Assume that inequalities (2.2) are satisfied and let γ = i1 , . . . , ik , ik+1 = i1 , k ≥ 2, be a
circuit in D(A). Then aij ij+1 = 0, j = 1, . . . , k, and from (2.2) it follows that

Rij (A) ≤ |aij ij |vij vi−1

j+1
, j = 1, . . . , k. (2.3)

By multiplying the inequalities in (2.3) for all j = 1, . . . , k, we immediately obtain (2.1).

Furthermore, if all the relations in (2.2) are equalities, then, obviously, all the inequalities in (2.1) also are
equalities. Vice versa, if at least one of the inequalities in (2.2) is strict, say,

Ri (A) < |aii|vivj−1 , i = j, aij = 0,

then, for any circuit γ ∈ C(A) that contains the arc (i, j), the corresponding inequality in (2.1) will be strict,
whereas the existence of such circuits follows from the almost trivial Lemma 2.1 below. Thus, if all the Brualdi
conditions hold with equality, then all relations (2.2) are equalities. This proves the assertion concerning the
equality case.
Necessity. Assume that Brualdi’s conditions (2.1) are fulfilled and define the positive numbers vi , i =
1, . . . , n, as follows:

v1 = 1;


s−1
 
vi = min aij ij  Rij (A), i = 2, . . . , n, (2.4)
1=i1 ,... ,is =i
j=1

where minimum is taken over all simple paths in D(A) going from 1 to i.

2467
 First we will show that (2.2) holds for i = 1. Indeed, if a1j = 0, then the arc (1, j) is a simple path in D(A),
whence, by (2.4),
vj ≤ |a11 |/R1(A) = |a11|v1 /R1 (A).
Now let i > 1, j = i, and assume that aij = 0. Assume that, in accordance with (2.4),


s−1
 
vi = ai i  Ri (A), (2.5)
l l l
l=1

where 1 = i1 , . . . , is = i is a simple path in D(A). If il = j, l = 2, . . . , s − 1, then 1 = i1 , . . . , is = i, is+1 = j is

a simple path from 1 to j, and, in view of (2.5) and (2.4), we conclude that


s
vi |aii|/Ri(A) = |ailil |/Ril (A) ≥ vj ,
l=1

which proves (2.2).

Finally, if j = ik for some k, 2 ≤ k ≤ s − 1, then 1 = i1 , . . . , ik = j is a simple path from 1 to j, whereas
j = ik , ik+1 , . . . , is = i, is+1 = j is a circuit of D(A). Therefore, by (2.4), we have


k−1
vj ≤ |ailil |/Ril (A), (2.6)
l=1

whereas from (2.1) it follows that

s
|ailil |/Ril (A) ≥ 1. (2.7)
l=k

Thus, taking into account (2.5), (2.6), and (2.7), we derive

s
vi |aii|/Ri(A) = |ailil |/Ril (A) ≥ vj ,
l=1

which amounts to (2.2). Theorem 2.1 is proved completely.

Lemma 2.1 [6]. Let A = (aij ) ∈ C n×n , n ≥ 2, be an irreducible matrix. If aij = 0 for some i = j, then there
is a circuit γ ∈ C(A), γ = i1 , . . . , ik , ik+1 = i1 , k ≥ 2, such that i1 = i and i2 = j.
Remark

n 2.1. If all the inequalities in (2.2) hold with equality, then, in the terminology of [3], the vector
v = vi i=1 is a D(A)-vector, i.e., for any i, 1 ≤ i ≤ n, if air = 0 and ais = 0, then vr = vs .
The following simple lemma shows that matrices satisfying conditions (2.2) are nonstrictly generalized
diagonally dominant.
Lemma 2.2. Let a matrix A = (aij ) ∈ C n×n , n ≥ 2, such that
Ri (A) > 0, i = 1, . . . , n,
satisfy conditions (2.2) for some vi > 0, i = 1, . . . , n. Then

|aij |vj ≤ |aii|vi, i = 1, . . . , n, (2.8)
j=i

and all the inequalities in (2.8) are equalities if and only if all the inequalities in (2.2) are equalities.
Proof. By multiplying both sides of the inequalities in (2.2) corresponding to aij = 0, j = i, j = 1, . . . , n, by
|aij | and summing the resulting inequalities over j, we obtain the relations
  
Ri(A) |aij |vj = Ri (A) |aij |vj ≤ |aii|vi |aij | = |aii |viRi(A), i = 1, . . . , n,
j=i: j=i j=i:
aij =0 aij =0

which imply (2.8) because Ri(A) > 0.

The assertion concerning the case of all equalities is trivial.

Lemma 2.2 and Theorem 2.1 imply the following result.

2468
Corollary 2.1. Let A = (aij ) ∈ C n×n , n ≥ 2, be an irreducible matrix that satisfies Brualdi’s conditions (2.1).
Then
(i) A is generalized diagonally dominant, i.e., it satisfies (2.8) for some vi > 0, i = 1, . . . , n;
(ii) the following assertions are equivalent:
(a) all the inequalities in (2.1) are equalities;
(b) all the inequalities in (2.2) are equalities;
(c) all the inequalities in (2.8) are equalities.
Remark 2.2. From Corollary 2.1 it follows that if (2.1) holds with at least one strict inequality, then (2.8) also
holds with at least one strict inequality. This fact was first mentioned in [1].
Remark 2.3. Corollary 2.1 means, in particular, that an irreducible matrix A satisfies the equalities



|aii| = Ri(A), γ ∈ C(A),
i∈γ i∈γ

n
if and only if its comparison matrix M (A) = mij i,j=1 , where

|aij |, i = j,
mij =
−|aij |, i =
 j,
is a singular M -matrix.
Using Theorem 1.7 and Remark 2.1, from Corollary 2.1 one can readily derive the following result.
Corollary 2.2. Under

nthe assumptions of Corollary 2.1, the matrix A is singular if and only if there exist a
positive vector v = vi i=1 such that

|aij |vj = |aii|vi, i = 1, . . . , n,
j=i

and a unitary diagonal matrix D such that

  
−1
D∗ I − DA A D = I − DA
−1 
A.
Further, if A is singular, then v is a D(A)-vector, satisfying the relations
Ri (A) = |aii|vivj−1 , aij = 0, i = j; (2.9)

the eigenvalue λ = 0 is of geometric multiplicity one, and the corresponding right eigenspace (the right null-space
of A) is spanned by the vector Dv.
Remark 2.4. Irreducible matrices satisfying Brualdi’s conditions (2.1) form a proper subclass of the class of
irreducible matrices with generalized diagonal dominance. Indeed, consider the irreducible matrix
 
4 −4 −4
A = −1 4 −2  .
0 −2 4
The equality    
2 0
A1 = 0
1 2
shows that A possesses the property of generalized diagonal dominance. However, for the circuit 1,2,1 we have
R1 (A)R2 (A) = 24 > 16 = |a11 | |a22|,
which means that Brualdi’s conditions (2.1) are not fulfilled.
Now we are ready to present the complete version of Brualdi’s theorem, providing conditions necessary and
sufficient for an irreducible matrix A satisfying (2.1) to be singular. Note that these conditions form a finite
set of equalities, corresponding to all the circuits from C(A) and involving the diagonal entries of A, the sums
Ri(A), and the complex counterparts of the signs of its nonzero off-diagonal entries.

2469
Theorem 2.2. Let A = (aij ) ∈ C n×n , n ≥ 2, be an irreducible matrix and assume that



|aii| ≥ Ri(A), γ ∈ C(A).
i∈γ i∈γ

The matrix A is singular if and only if, for every circuit γ = i1 , . . . , ik , ik+1 = i1 , k ≥ 2, of D(A), the equality

k
aij ij+1

k
aij ij
(−1)k = (2.10)
|aij ij+1 | Rij (A)
j=1 j=1

is valid.
Proof. By Corollary 2.1, there exist vi > 0, i = 1, . . . , n, such that inequalities (2.8) are valid. Applying Theorem
1.8, we arrive at the conclusion that the matrix A is singular if and only if

k

k
 −1   
(−1)k a−1 a  aij ij+1 
ij ij aij ij+1 = ij ij (2.11)
j=1 j=1

for every circuit γ = i1 , . . . , ik , ik+1 = i1 , k ≥ 2, of D(A), and


|aij |vj = |aii|vi, i = 1, . . . , n.
j=i

In view of assertion (ii) of Corollary 2.1, the latter equalities are equivalent to the conditions



|aii| = Ri(A), γ ∈ C(A), (2.12)
i∈γ i∈γ

implying that equalities (2.11) can also be written as (2.10) if A is singular.

Thus, it only remains to ascertain that (2.10) implies both (2.11) and (2.12). Indeed, equalities (2.12) im-
mediately follow from (2.10) if one passes to absolute values, whereas (2.11) readily stems from (2.10) and (2.12).


The final theorem of this section yields explicit expressions for the absolute values of the components zi ,
i = 1, . . . , n, of a right eigenvector z associated with the zero eigenvalue of a singular irreducible matrix A
satisfying the conditions


|aii| = Ri(A), γ ∈ C(A). (2.13)
i∈γ i∈γ

Theorem 2.3. Let A = (aij ) ∈ C n×n, n ≥ 2, be a singular irreducible matrix that satisfies conditions (2.13).
If
n
Az = 0, z = zi i=1 , z = 0,
then the values vi = |zi |, i = 1, . . . , n, can be determined as follows:

v1 = 1, (2.14)


l(j)−1

vj = ai  Ri (A), j = 2, . . . , n, (2.15)
r ir r
r=0

where l(j) is the distance in D(A) from vertex 1 to vertex j, and 1 = i0 , i1 , . . . , il(j) = j is a shortest path in
D(A) from 1 to j.
Proof. By Corollary 2.2, we have

Ri(A) = |aii|vivj−1 , aij = 0, i = j. (2.16)

2470
Therefore, if a1j = 0, i.e., l(j) = 1, then necessarily

vj = |a11 |/R1(A),

and such components vj are uniquely determined.

For j such that a1j = 0, we must show that the above definition is correct, i.e., that if k = l(j) > 1, and

1 = i0 , i1 , . . . , ik = j and 1 = j0 , j1 , . . . , jk = j

are two distinct shortest paths in D(A) from 1 to j, then

k−1

k−1
|air ir |/Rir (A) = |ajr jr |/Rjr (A). (2.17)
r=0 r=0

Using induction, we assume that all the components vi for which l(i) < k are uniquely determined by (2.15).
Then, from (2.16) it follows that

|aik−1 ik−1 | |ajk−1 jk−1 |

vj = vi = vjk−1 , (2.18)
Rik−1 (A) k−1 Rjk−1 (A)

and equality (2.17) stems from (2.18) and equalities (2.15) for vik−1 and vjk−1 , where l(ik−1 ) = l(jk−1 ) = k − 1 <
k = l(j).

Remark 2.5. Obviously, the components of the vector v could be determined starting from an arbitrary vi = 1,
1 ≤ i ≤ n, rather than from v1 = 1. In addition, if a certain row of A contains r nonzero off-diagonal entries,
then, by Theorem 2.3, at least r components of the vector v must coincide. In particular, if the matrix A has a
row in which all the off-diagonal entries are nonzero, then the vector v has no more than two distinct components.

3. Applications
This section considers two applications of the results presented in Sec. 2.
First we consider the location of the eigenvalues of an irreducible matrix.
In [3], the following results were established.

Theorem 3.1.∗ Let A = (aij ) ∈ C n×n , n ≥ 2, be an irreducible matrix.

(i) Any eigenvalue of A belongs to the domain

B(A) = Bγ (A),
γ∈C(A)

where  



Bγ (A) = z ∈ C : |z − aii| ≤ Ri(A) , γ ∈ C(A).
i∈γ i∈γ

(ii) A boundary point of the domain B(A) can be an eigenvalue of the matrix A only if it is a boundary point
of every Bγ (A), γ ∈ C(A).

Assertion (ii) of Brualdi’s theorem is just a necessary condition for a boundary point of the domain B(A) to
be an eigenvalue of the matrix A. By applying Theorem 2.2 and Corollary 2.2 to the shifted matrix A − λI, one
immediately obtains the corresponding necessary and sufficient conditions, which are provided by the following
theorem.
∗ Note that assertion (i) was proved in [3] for the class of weakly irreducible matrices, which strictly contains the class of
irreducible matrices.

2471
Theorem 3.2. Under the assumptions of Theorem 3.1, a boundary point of the domain B(A) is an eigenvalue
of A if and only if, for every circuit γ = i1 , i2 , . . . , ik , ik+1 = i1 , k ≥ 2, of D(A), the equality

k
aij ij+1

k
aij ij − λ
(−1)k =
|aij ij+1 | Rij (A)
j=1 j=1

is valid.
Further, any eigenvalue λ of the matrix A that is a boundary point of the domain B(A) is of geometric
multiplicity one, and the corresponding eigenspace is spanned by a vector z = (zi ) ∈ C n such that

|aii − λ|
|zj | = |zi| , aij = 0, i = j.
Ri (A)

The second application of the results established in Sec. 2 concerns the following generalization of the
Ostrowski–Brauer Theorem 1.5.
Theorem 3.3 [6]. Let A = (aij ) ∈ C n×n , n ≥ 2, be an irreducible matrix and assume that

|aii| |ajj | ≥ Ri (A) Rj (A) if aij = 0, i = j. (3.1)

Then A is singular if and only if there exists a unitary diagonal matrix D ∈ C n×n such that
−1 −1
D∗ (I − DA A)D = |I − DA A|, (3.2)

and either
|aii| = Ri (A), i = 1, . . . , n, (3.3)
or the set  
S = i, 1 ≤ i ≤ n : |aii| < Ri(A)
is nonempty and the following conditions are fulfilled:
(i) both principal submatrices A[S] and A[ S ] of A, where S = {1, 2, . . . , n} \ S, are diagonal;
(ii) there is a constant α > 1 such that

α, i ∈ S,
Ri (A)/|aii| = (3.4)
α−1 , i ∈ S.

Further, if A is singular, then the geometric multiplicity

n of its zero eigenvalue is one, and the corresponding
eigenspace is spanned by the vector Dv, where v = vi i=1 and

α if Ri(A) > |aii|,
vi = (3.5)
1 if Ri(A) ≤ |aii|,

whereas D and α are the same as in (3.2) and (3.4), respectively.

Below, we present a new shorter proof of Theorem 3.3 that is based on the results of the previous section.
Proof. If k ≥ 2 and γ = i1 , i2 , . . . , ik , ik+1 = i1 ∈ C(A), then, by (3.1), the following inequalities are valid:
 
ai1i1 ai2 i2  ≥ Ri1 (A) Ri2 (A),
 
ai i ai i  ≥ Ri (A) Ri (A),
2 2 3 3 2 3

....................... .....
 
aikik ai1 i1  ≥ Rik (A) Ri1 (A).

Multiplying these relations, we arrive at the inequality


2 
2
|aii| ≥ Ri(A) ,
i∈γ i∈γ

2472
 showing that the assumptions of Corollaries 2.1 and 2.2 are satisfied. Therefore, the matrix A is singular if and
only if there exists a unitary diagonal matrix D for which (3.2) holds true and



|aii| = Ri(A), γ ∈ C(A). (3.6)
i∈γ i∈γ
Further, the zero eigenvalue of A is of geometric multiplicity one, and the corresponding eigenvector Dv satis-
fies (2.9).
In view of Lemma 2.1, from (3.6) it follows that all the relations in (3.1) hold with equality, i.e.,
|aii| |ajj | = Ri(A)Rj (A), aij = 0, i = j. (3.7)
Denote    
S2 = i, 1 ≤ i ≤ n : |aii| > Ri(A) , S3 = i, 1 ≤ i ≤ n : |aii| = Ri(A) .
Conditions (3.7) readily imply that both A[S] and A[S2 ] are diagonal matrices, and
A[S ∪ S2 , S3 ] = 0.
Since, by assumption, the matrix A is irreducible, we conclude that either (3.3) holds true or S3 is empty,
whence S2 = S.
In the latter case, by Theorem 2.3 and (3.7), we have
v1 = 1,

 1, l(j) ≡ 0 (mod 2)
|ai0 i0 | |ai1i1 | . . . |ail(j)−1 il(j)−1 |  (3.8)
vj = =
 |a11 | , j = 2, . . . , n,
Ri0 (A)Ri1 (A) . . . Ril(j)−1 (A)  , l(j) ≡ 1 (mod 2)
R1 (A)
where i0 = 1, il(j) = j, and i0 , i1 , . . . , il(j) is a shortest path in D(A) from 1 to j. On the other hand, since both
A[S] and A[ S ] are diagonal, we obviously have
l(j) ≡ 0 (mod 2) ⇐⇒ either 1, j ∈ S or 1, j ∈ S. (3.9)
Now equalities (3.5) (which trivially hold if (3.3) is valid) readily follow from (3.8) and (3.9).
Thus, it only remains to prove (3.4). Indeed, Corollary 2.2 claims that

|aii |vi = |aij |vj , i = 1, . . . , n,
j=i

where (since both A[S] and A[ S ] are diagonal)

(3.5)  (3.5) 
|aii|α = |aii|vi = |aij |vj = |aij | = Ri(A), i ∈ S,
j∈S j∈S
and  
(3.5) (3.5)
|aii| = |aii|vi = |aij |vj = α |aij | = αRi(A), i ∈ S.
j∈S j∈S
This completes the proof of Theorem 3.3.

Translated by L. Yu. Kolotilina.

REFERENCES
1. Yu. A. Alpin, “Criteria for nonsingularity of matrices,” in: Algebra and Analysis [in Russian], Proceedings
of the Conference Dedicated to the Centenary of B. M. Gagaev, Kazan (1997), pp. 16–17.
2. A. Brauer, “Limits for the characteristic roots of a matrix. II,” Duke Math. J., 14, 21–26 (1947).
3. R. A. Brualdi, “Matrices, eigenvalues, and directed graphs,” Linear Multilinear Algebra, 11, 143–165 (1982).
4. M. Fiedler and V. Pták, “Cyclic products and an inequality for determinants,” Czechoslovak Math. J., 19,
428–450 (1969).
5. L. Yu. Kolotilina, “Nonsingularity/singularity criteria for nonstrictly block diagonally dominant matrices,”
Linear Algebra Appl. (to appear).
6. L. Yu. Kolotilina, “Generalizations of the Ostrowski–Brauer theorem,” Linear Algebra Appl. (to appear).
7. A. Ostrowski, “Über die Determinanten mit überwiegender Hauptdiagonale,” Comment. Math. Helv, 10,
69–96 (1937).
8. O. Taussky, “A recurring theorem on determinants,” Amer. Math. Monthly, 56, 672–676 (1949).

2473

View publication stats