A SHORT NOTE ON DOUBLY SUBSTOCHASTIC ANALOG OF

arXiv:1801.01232v1 [math.CO] 4 Jan 2018

BIRKHOFF’S THEOREM

LEI CAO∗

Abstract.

Let B be an n × n doubly substochastic matrix. We show that B can be written as a convex

combination of no more than σ(B)+t subpermutation matrices, where σ(B) is the number of nonzero
elements in B and t is the number of fully indecomposable components of B comp , the minimal doubly
stochastic completion of B obtained by a specific way.

Key words. Doubly Substochastic Matrices; Birkhoff-von Neumann Theorem; Subpermutation

Matrices;

AMS subject classifications. 15B51, 15A99.

1. Introduction. Let Ωn be the set all of n × n doubly stochastic matrices.

It is well-known that Ωn is a polytope and whose dimension is (n − 1)2 and whose
extreme points are the n × n permutation matrices (see [3] II.1, [5]). Let A = [aij ]
be an n × n (0,1) matrix. Define the face of Ωn corresponding to A, denoted by
F (A), as the set of all n × n doubly stochastic matrices X = [xij ] such that xij ≤
aij for all i, j = 1, 2, . . . , n. In [4], Brualdi and Gibson show that the dimension of
F (A) is σ(A) − 2n + t where σ(A) is the number of 1’s in A and t is the number of
fully indecomposable components of A. Therefore by Carathéodory’s theorem, every
doubly stochastic matrix in F (A) can be written as a convex combination of no more
than σ(A) − 2n + t + 1 permutation matrices. Recently we note that Dufossé and
Uçar showed that determining the minimal number of permutation matrices needed
is strongly NP-complete ([2]).

Let ωn be the set of all n × n doubly substochastic matrices. It is known that ωn

is a convex polytope as well and its vertices are the n × n subpermutation matrices
([6], p. 165). However, there is no upper bound or the number of subpermutation
matrices needed for a given doubly substochastic matrix is known. In this paper, we
give an upper bound via a combination of the results from [4] and [1].

In [1], Cao, Koyuncu and Parmer defined the sub-defect of doubly substochastic
matrices and provided a specific way to obtain a minimal doubly stochastic comple-

∗ Department of Mathematics, Georgian Court University, Lakewood, New Jersey, 08701, USA

(lcao@georgian.edu).
1
2

tion of any doubly substochastic matrix. Such a minimal completion contains the
most zeros among all completions and hence minimizes the number of nonzero ele-
ments which is the most influential part of the dimension of the face containing its
completion.

2. Properties and Notation. Definition 2.1. Let B ∈ ωn . Then the sub-

defect of B, denoted by sd(B), is defined to be the smallest integer k such that there
exists A ∈ Ωn+k containing B as a submatrix.

Let M be an n × n matrix. Denote the sum of all elements of M by σ(M ), the

sum of the ith row of M by ri (M ) and the sum of the ith column of M by ci (M ).
Then sd(B) can be computed easily by taking the difference between the size of B
and σ(B) as the follows.

Theorem 2.2. ([1], Theorem 2.1 ) Let B ∈ ωn . Then

sd(B) = ⌈n − σ(B)⌉,
where ⌈·⌉ is the ceiling function.

Denote ωn,k the set of all n × n doubly stochastic matrices with sub-defect k.
Then ωn,k partition ωn , and ωn,0 = Ωn . Let B ∈ ωn,k , then denote B comp ∈ Ωn+k
the minimal doubly stochastic completion of B obtained by the method above.

Given B = (bij ) ∈ ωn,k to obtain a minimal doubly stochastic completion of B,

one can append rows and columns to B as the following:

1. Construct
 
b11 b12 . . . b1n x1
b b22 . . . b2n x2 
 21 
 . .. .. .. .. 
 ..
B̃ =  . . . . 

 
bn1 bn2 . . . bnn xn 
y1 y2 . . . yn z
where
x1 = 1 − r1 (B), y1 = 1 − c1 (B)
and
i−1
X i−1
X
xi = min{1 − ri (B), 1 − xj }, yi = min{1 − ci (B), 1 − yj }
j=1 j=1

for i = 2, 3, . . . , n and
n
X n
X
z = min{1 − xi , 1 − yi }.
i=1 i=1
 3

2. If B is doubly stochastic, then one doesn’t want to do the construction. If

B̃ is a doubly stochastic matrix, then it implies that sd(B) = 1 and B̃ is a
minimal completion. Otherwise, once show that B̃ ∈ ωn+1,k+2 and repeat
the previous step.

This algorithm gives a minimal completion because each row and column ap-
pended except the last row and the last column has sum 1 due to the way to choose
xi and yi for i = 1, 2, . . . , n in step 1 and this is the largest step we can have to
complete it.

Another important ingredient is a known formal for the dimension of a face of

Ωn . A matrix is called a (0, 1) matrix or a binary matrix if all its entries are either
0 or 1. An n × n binary matrix A = [aij ] has total support if ars = 1 implies there
exists an n × n permutation matrix P = [pij ] with prs = 1 and P ≤ A ([8]). Let
S = [sij ] ∈ ωn , then the binary matrix A corresponding to S, denoted by B(S), is
defined as aij = 1 if sij 6= 0. A fact is that for any S ∈ Ωn , the corresponding binary
matrix B(S) has total support due to Birkhoff’s theorem. A doubly stochastic matrix
A is fully indecomposable, provided there do not exist permutation matrices P and
Q such that P AQ = A1 ⊕ A2 ([7]).

Theorem 2.3. [9] A binary matrix A has total support if and only if there exist
permutation matrices P and Q such that P AQ is a direct sum of fully indecomposable
matrices.

Given an n × n binary matrix A = [aij ] which has total support, the face of Ωn
corresponding to A, denoted by F (A), is the set of all matrices S = [sij ] ∈ Ωn with
sij ≤ aij for i, j = 1, 2, . . . , n.

Theorem 2.4. (Corollary 2.6, [4]) Let A be an n × n binary matrix with total
support, let P and Q be permutation matrices such that P AQ = A1 ⊕ · · · ⊕ At , where
Ai is a fully indecomposable matrix for i = 1, . . . , t. Then

dim F (A) = σ(A) − 2n + t.

where dim F (A) is the dimension of F (A).

3. Main Results. Lemma 3.1. Let D ∈ ωn,k and write

 
D X
Dcomp =
Y Z
where X, Y and Z have dimension n × k, k × n and k × k respectively. Then

(i) k − 1 < σ(X) = σ(Y ) ≤ k;

(ii) each column of X contains at least one positive element and ci (X) = 1 for
i = 1, . . . , k − 1;
4

(iii) each row of Y contains at least one positive element and ri (Y ) = 1 for i =
1, . . . , k − 1.

Proof. (i) Since D ∈ ωn,k , n − k ≤ σ(D) < n − k + 1. Since Dcomp ∈ Ωn+k ,

σ(D) + σ(X) = n because D and X partition the first n rows of a doubly stochastic
matrix Dcomp . Therefore

k − 1 < σ(X) = n − σ(D) ≤ k.

Similarly, D and Y partition the first n columns of a doubly stochastic matrix Dcomp ,
so

k − 1 < σ(X) = σ(Y ) ≤ k.

(ii) Since D ∈ ωn,k , n − k ≤ σ(D) < n − k + 1; Since X is a submatrix of a

doubly stochastic matrix Dcomp , ci (X) ≤ 1; Since D and X compose the first n rows
of a doubly stochastic matrix, σ(D) + σ(X) = n and hence σ(X) > k − 1; Since
X has k columns, ci (X) > 0 for all i = 1, 2, . . . , k. Suppose for some 1 ≤ t < k,
ct (X) < 1. Then all rows of D have been completed, otherwise one should increase
the values of some element in this column to complete those rows. Hence ci (X) = 0
for i = t + 1, t + 2, . . . , k. Therefore

t
X
σ(X) = ci (X) > k − 1
i=1

which is impossible since there are only t < k non-zero columns and all of them have
sum at most 1.

(iii) Take the transpose of Dcomp to convert to (ii).

Lemma 3.2. Let D ∈ ωn,k and write

 
D X
Dcomp =
Y Z

where X, Y and Z have dimension n × k, k × n and k × k respectively. Then

(i) σ(B(X)) ≤ n + k − 1;
(ii) σ(B(Y )) ≤ n + k − 1;
(iii) σ(B(Z)) ≤ 1.

and hence σ(B(Dcomp )) ≤ σ(B(D)) + 2(n + k) − 1 ≤ n2 + 2(n + k) − 1.

 5

Proof. (i) Denote

x11 x12 . . . x1k

 
 x21 x22 . . . x2k 
X= . ..  .
 
.. ..
 .. . . . 
xn1 xn2 . . . xnk

Due to the way that Dcomp , x11 = 1 − r1 (D) is constructed and x1j = 0 for j =
2, 3, . . . , k since Dcomp ∈ Ωn+k . So X must has the form

1 − r1 (D) 0 ... 0
 
 x21 x22 . . . x2k 
X = ..  .
 
.. .. ..
 . . . . 
xn1 xn2 . . . xnk

Unless the first row of D is zero for which the first column of X is zero except x11 = 1,
we will choose x21 = min{1 − r2 (D), 1 − x11 } ≥ 0. If x21 = 1 − r2 (D) < 1 − x11 , then
the second row is completed and x2j = 0 for j = 2, 3, . . . , k. If x21 = 1 − x11 <
1 − r2 (D), then the first column of X is completed, so xj1 = 0 for j = 3, 4, . . . , n.
If x21 = 1 − r2 (D) < 1 − x11 , then both second row and the first column of X are
completed. Namely, X must have the forms
 
1 − r1 (D) 0 ... 0
1 − r (D) 0 ... 0 
 2 
x x . . . x
 
X=  31 32 3k 
.. .. .. .. 
. . . . 
 

xn1 xn2 . . . xnk

for which x31 will be chosen next; or

 
1 − r1 (D) 0 ... 0
1 − (1 − r (D)) x22 . . . x2k 
 1 
0 x32 . . . x3k 
 
X= 
.. .. .. .. 
. . . . 
 

0 xn2 . . . xnk

for which x22 will be chosen next; or

 
1 − r1 (D) 0 ... 0
1 − r (D) = 1 − (1 − r (D)) 0 ... 0 
 2 1 
0 x32 . . . x3k 
 
X=  .. .. .. .. 
. . . . 
 

0 xn2 . . . xnk
6

for which x32 will be chosen next.

In general, once xst is determined, it may complete the row in which case xs+1,t
is determined next; or it may complete the column in which case xs,t+1 is determined
next; or it may complete both the row and column in which case xs+1,t+1 is determined
next.

Therefore X must look like

 
∗
∗ ∗ 
 

 ∗ 

∗
 
 
∗ ∗
 
 
 

 ∗ ∗ 


 ∗ 


 ∗ 

∗
 
 
∗
 
 
 
 ∗ ∗
∗

where ∗ are non-negative entries and other entries are zero. Each row contains at
most two positive entries and at most k − 1 row contains two positive entries which
implies

σ(B(X)) ≤ n + k − 1.

(ii) Take the transpose of Dcomp , then it is converted to (i).

(iii) Due to the way Dcomp is constructed, once xst is determined, one determine
Ps
xs,t+1 only when the column containing xst is completed. This means that i=1 xit =
1, so xjt = 0 for j = s + 1, . . . , n. Note that the first k − 1 columns in X must be
completed implying that all the first k − 1 columns in Z are zero. Similarly, the first
k − 1 rows are zero due to the structure of Y. Hence the only entry in Z which can
comp
be positive is Zkk = Dn+k,n+k . Namely σ(B(Z)) ≤ 1.

Combine all of (i), (ii) and (iii), we have σ(B(Dcomp )) = σ(B(D)) + σ(B(X)) +
σ(B(Y )) + σ(B(Z)) ≤ σ(B(D)) + 2(n + k) − 1 ≤ n2 + 2(n + k) − 1.

The following example illustrates the procedure of obtaining Dcomp for a given
 7

doubly substochastic matrix D.

0.1 0 0.2 0.1
 
 0 0.2 0.1 0
Example 3.3. Let D =  0.2 0
 . Since σ(D) = 1.8, D ∈ ω4,3 and
0 0.1
0.1 0.2 0.3 0.2
hence

 
0.1 0 0.2 0.1 x11 x12 x13
 0 0.2 0.1 0 x21 x22 x23 
 
0.3 0 0 0.1 x31 x32 x33 
 
Dcomp = 0.1 0.2 0.3 0.2 x41 x42 x43 
 
 
y11 y12 y13 y14 z11 z12 z13 
 
y21 y22 y23 y24 z21 z22 z23 
y31 y32 y33 y34 z31 z32 z33

We start by letting x11 = 1 − r1 (D) = 0.6. Since x11 completes the first row, all
elements on the right to x11 must be 0.

 
0.1 0 0.2 0.1 0.6 0 0
 0 0.2 0.1 0 x21 x22 x23 
 
0.2 0 0 0.1 x31 x32 x33 
 
Dcomp = 0.1 0.2 0.3 0.2 x41 x42 x43 
 
 
y11 y12 y13 y14 z11 z12 z13 
 
y21 y22 y23 y24 z21 z22 z23 
y31 y32 y33 y34 z31 z32 z33

Next, let x21 = min{1 − r2 D, 1 − x11 } = min{0.7, 0.4} = 0.4 which completes the
fourth column of Dcomp , so all elements below x21 must be 0.

 
0.1 0 0.2 0.1 0.6 0 0
 0 0.2 0.1 0 0.4 x22 x23 
 
0.2 0 0 0.1 0 x32 x33 
 
Dcomp = 0.1 0.2 0.3 0.2 0 x42 x43 
 
0
 
y11 y12 y13 y14 z12 z13 
 
y21 y22 y23 y24 0 z22 z23 
y31 y32 y33 y34 0 z32 z33

Next, let x22 = 1 − r2 (D) − x21 = 0.3 which completes the second row and hence
x23 = 0.
8

 
0.1 0 0.2 0.1 0.6 0 0
 0 0.2 0.1 0 0.4 0.3 0 
 
0.2 0 0 0.1 0 x32 x33 
 
Dcomp = 0.1 0.2 0.3 0.2 0 x42 x43 
 
0
 
y11 y12 y13 y14 z12 z13 
 
y21 y22 y23 y24 0 z22 z23 
y31 y32 y33 y34 0 z32 z33
Next, let x32 = min{1 − r3 (D) − x31 , 1 − x12 − x22 } = 0.7 which completes both the
third row and the fifth column, so all elements below and on the right to x32 must be
0.

 
0.1 0 0.2 0.1 0.6 0 0
 0 0.2 0.1 0 0.4 0.3 0 
 
 0.2 0 0 0.1 0 0.7 0 
 
Dcomp =  0.1 0.2 0.3 0.2 0 0 x43 
 
0 0
 
y11 y12 y13 y14 z13 
 
y21 y22 y23 y24 0 0 z23 
y31 y32 y33 y34 0 0 z33
Next, let x43 = 1 − r4 (D) = 0.2 to complete the fourth row.

 
0.1 0 0.2 0.1 0.6 0 0
 0 0.2 0.1 0 0.4 0.3 0 
 
0.2 0 0 0.1 0 0.7 0 
 
Dcomp = 0.1 0.2 0.3 0.2 0 0 0.2
 
0 0 z13 
 
y11 y12 y13 y14
 
y21 y22 y23 y24 0 0 z23 
y31 y32 y33 y34 0 0 z33

Next, take the transpose of Dcomp to obtain Y and Z by the same procedure and
then take the transpose again to obtain Dcomp completely.
 
0.1 0 0.2 0.1 0.6 0 0
 0 0.2 0.1 0 0.4 0.3 0
 
0.2 0 0 0.1 0 0.7 0
 
Dcomp = 0.1 0.2 0.3 0.2 0 0 0.2
 
0.6 0.4 0 0 0 0 0
 
 
 0 0.2 0.4 0.4 0 0 0
0 0 0 0.2 0 0 0.8
 9

Theorem 3.4. Let A ∈ ωn,k , then A can be written as a convex combination of

no more than

σ(B(A)) + t

subpermutation matrices, where t is the number of fully indecomposable components

of Acomp .

Proof. Since Acomp ∈ Ωn+k ,

dim F (B(Acomp )) = σ(B(Acomp )) − 2(n + k) + t

due to Theorem 2.4. Apply Lemma 3.2, we have

dim F (B(Acomp )) = σ(B(Acomp )) − 2(n + k) + t ≤ σ(B(A)) + t − 1,

meaning that Acomp can be written as a convex combination of no more than

σ(B(A)) + t

permutation matrices. Therefore A can be written as a convex combination of no

more than σ(B(A)) + t subpermutation matrices.

The bound given by Theorem 3.4 is actually tight. Here is an example in which
the equality holds.

Example 3.5.
7 
12 0
Let A = 1 1 ∈ ω2,1 and its minimal completion is
6 2

7 5 
12 0 12
Acomp = 1
6
1
2
1 
3
∈ Ω3
1 1 1
4 2 4

and
       
0 0 1 1 0 0 0 0 1 1 0 0
1  1 1 1
Acomp = · 1 0 0  + · 0 1 0 + · 0 1 0 + · 0 0 1
6 4 4 3
0 1 0 0 0 1 1 0 0 0 1 0

and hence a convex combination of subpermutation matrices of A is

       
1 0 0 1 1 0 1 0 0 1 1 0
A= · + · + · + · . (3.1)
6 1 0 4 0 1 4 0 1 3 0 0
10

In addition, suppose that

       
1 0 0 1 0 0 0 0
A = x1 + x2 · + x3 + x4 ·
0 0 0 0 1 0 0 1
     
1 0 0 1 0 0
+ x5 + x6 · + x7
0 1 1 0 0 0
 
x1 + x5 x2 + x6
= .
x3 + x6 x4 + x5
P7
With i=1 xi = 1, we have a system corresponding to the augmented matrix
 7

1 0 0 0 1 0 0 12
0 1 0 0 0 1 0 0 
 
0 0 1 0 0 1 0 16 
 
0 0 0 1 1 0 0 12 
 
1 1 1 1 1 1 1 1
which is equivalent to
 5

1 0 0 0 0 −1 1 6
0 1 0 0 0 1 0 0
 
1
0 0 1 0 0 1 0 .

6
0 3
0 0 1 0 −1 1

4

1
0 0 0 0 1 1 −1 4

Each row has a pivot position, so x6 and x7 are free variables. Since xi ≥ 0 for
i = 1, . . . , 7, the second row implies that x2 = x6 = 0. The solution to this system is
that

 5
 x1 = − x7
6




= 0



 x2

1




 x3 =


 6
3
x4 = − x7


 4
1




 x5 = + x7
4




x = 0



 6


x7 = x7
It is clear that x3 6= 0 and one can choose x7 to make one of x1 , x4 , x5 and x7 to be
zero.Therefore at least four of these seven coefficients are non-zero. If one let x7 = 0,
then this specific solution gives (3.1).
 11

According to Theorem 3.4, A contains 3 non-zero elements and Acomp is not

decomposable meaning that Acomp is 1 fully indecomposable component, so A can be
written as a convex combination of no more than 4 subpermutation matrices while at
least 4 subpermutation matrices are needed as we showed above, so the equality holds.

Although Theorem 3.4 is proven via minimal doubly stochastic completion of

doubly substochastic matrices, the upper bound of the number of subpermutation
matrices needed does not depend on the sub-defect at all. Let B = [bij ] ∈ ωn
and b = maxi {ri (B)}. Note that 1b B ∈ ωn . If b < 1, then the sub-defect of 1b B
is possibly much less than the sub-defect of B. But 1b B needs the same number of
subpermutation matrices or at most one more, the zero matrix, as B. If there exists
a bij = 1, then one can permute B as a direct sum of an identity matrix and a
smaller doubly substochastic matrix for which one can play the trick to the doubly
substochastic matrix. So sub-defect does not affect the number of subpermutation
matrices needed. Here are two examples.
7 
0
Example 3.6. Let A = 12 1 1 be the same matrix in Example 3.5 and let
6 2
7 7
   
1 1 12 0 36 0
B= A= · 1 1 = 1 1 .
3 3 6 2 18 6

Although sd(A) = 1 6= 2 = sd(B), since

       
1 0 0 1 1 0 1 0 0 1 1 0
A= · + · + · + · ,
6 1 0 4 0 1 4 0 1 3 0 0
         
1 1 0 0 1 1 0 1 0 0 1 1 0 1 0 0
B= A= · + · + · + · + · .
3 18 1 0 12 0 1 12 0 1 9 0 0 3 0 0
Therefore, if two doubly substochastic matrices are scalar multiple of each other, the
difference between the numbers of subpermutation matrices needed for their convex
expansion is at most 1.

Let B = [bij ] ∈ ωn . If max{bij }ni,j=1 = 1, then there exist permutation matrices

P and Q such that P BQ = Ik ⊕ B̃ where Ik is a k × k identity for some k ≤ n.
Hence B̃ and B need the same number of subpermutation matrices for their convex
expansions.
7 
0
Example 3.7. Let Let A = 12 1 1 be the same matrix in Example 3.5 and let
6 2

1 0
 
0 1 
B = I2 ⊕ A =  7
.

12 0
1 1
6 2
12

Since
       
1 0 0 1 1 0 1 0 0 1 1 0
A= · + · + · + · ,
6 1 0 4 0 1 4 0 1 3 0 0

1 0 1 0 1 0 1 0
       
1  0 1  1 0 1  1 0 1  1 0 1 
B = · + · + · + · .
6  0 0 4  1 0 4  0 0 3  1 0
1 0 0 1 0 1 0 0

4. Acknowledgement. The author would thank the anonymous reviewers for

critically reading the manuscript and suggesting substantial improvements.

REFERENCES

[1] Cao, Lei and Koyuncu, Selcuk and Parmer, Timmothy. A minimal completion of doubly
substochastic matrix. Linear Multilinear Algebra, 64(11): 2313–2334, 2016.
[2] Fanny Dufossé and Bora Uçar. Notes on Birkhoff-von Neumann decomposition of doubly
stochastic matrices. Linear Algebra and its Applications, 497: 108 – 115, 2016.
[3] M. Marcus and H. Minc. A Survey of Matric Theory and Matrix Inequalities. Dover Publica-
tions, 1992.
[4] Richard A Brualdi and Peter M Gibson. Convex polyhedra of doubly stochastic matrices.
I. Applications of the permanent function. Journal of Combinatorial Theory, Series A,
22(2): 194 – 230, 1977.
[5] Birkhoff, Garrett. Three observations on linear algebra. Univ. Nac. Tucumán. Revista A., 5:
147–151, 1946.
[6] Horn, Roger A. and Johnson, Charles R. Topics in Matrix Analysis. Cambridge University
Press, New York, 1994.
[7] Trevor I. Fenner and Georghios Loizou. On fully indecomposable matrices. Journal of Computer
and System Sciences, 5(6): 607 – 622, 1971.
[8] Sinkhorn, Richard and Knopp, Paul. Concerning nonnegative matrices and doubly stochastic
matrices Pacific J. Math., 21(2):343–348, 1967.
[9] H. Perfect and L. Mirsky. The Distribution of Positive Elements in Doubly-Stochastic Matrices.
J. London Math. Soc., 40: 689–698, 1965.