Binary Factorizations of the Matrix of All Ones Maguy Trefois Paul - - PowerPoint PPT Presentation

binary factorizations of the matrix of all ones
SMART_READER_LITE
LIVE PREVIEW

Binary Factorizations of the Matrix of All Ones Maguy Trefois Paul - - PowerPoint PPT Presentation

Binary Factorizations of the Matrix of All Ones Maguy Trefois Paul Van Dooren Jean-Charles Delvenne Universit catholique de Louvain, Belgium ILAS, June 2013 1 Motivation : the finite-time average consensus problem We have: n


slide-1
SLIDE 1

1

Binary Factorizations of the Matrix of All Ones

Maguy Trefois Paul Van Dooren Jean-Charles Delvenne

Université catholique de Louvain, Belgium

ILAS, June 2013

slide-2
SLIDE 2

2

Motivation: the finite-time average consensus problem We have:

  • n communicating agents with an initial position
  • a communication topology

At each time step:

  • each agent sends its current position to some other agents

according to the communication pattern

  • with the received information, each agent changes its position

The goal: after a finite time, all the agents meet at the average of their initial positions

slide-3
SLIDE 3

3

Vector of initial positions:

  • x(0)

Dynamics:

  • x(t + 1) = At+1.

x(0) The matrix A respects the communication topology: A =     ? ? ? ? ?    

slide-4
SLIDE 4

4

  • x(t + 1) = At+1.

x(0) The matrix A is a solution to the consensus if:

  • A is of the form

    ? ? ? ? ?    

  • After a finite time m, Am = 1

4.

   1 . . . 1 . . . . . . . . . 1 . . . 1   

slide-5
SLIDE 5

4

  • x(t + 1) = At+1.

x(0) The matrix A is a solution to the consensus if:

  • A is of the form

    ? ? ? ? ?    

  • After a finite time m, Am = 1

4.

   1 . . . 1 . . . . . . . . . 1 . . . 1    Question: for which communication patterns is it possible to reach the consensus ?

slide-6
SLIDE 6

5

We should study the solutions to the equation: Am = 1 n.    1 . . . 1 . . . . . . . . . 1 . . . 1    , where A ∈ Rn×n.

slide-7
SLIDE 7

5

We should study the solutions to the equation: Am = 1 n.    1 . . . 1 . . . . . . . . . 1 . . . 1    , where A ∈ Rn×n. ⇒ difficult to tackle directly

slide-8
SLIDE 8

5

We should study the solutions to the equation: Am = 1 n.    1 . . . 1 . . . . . . . . . 1 . . . 1    , where A ∈ Rn×n. ⇒ difficult to tackle directly Simpler problem: study the solutions to: Am =    1 . . . 1 . . . . . . . . . 1 . . . 1    , where A ∈ {0, 1}n×n.

slide-9
SLIDE 9

6

Outline

Factorization problem The De Bruijn matrices Factorizations into commuting factors General form of a root of In with minimum rank A root class of In Conclusion

slide-10
SLIDE 10

7

Factorization problem The De Bruijn matrices Factorizations into commuting factors General form of a root of In with minimum rank A root class of In Conclusion

slide-11
SLIDE 11

8

We are looking for the solutions to

m

  • i=1

Ai = A1A2...Am = In, where

  • In is the n × n matrix with all ones
  • each factor Ai is an n × n binary matrix.
slide-12
SLIDE 12

8

We are looking for the solutions to

m

  • i=1

Ai = A1A2...Am = In, where

  • In is the n × n matrix with all ones
  • each factor Ai is an n × n binary matrix.

In particular, we are investigating the solutions to: Am = In, where A is a binary matrix.

slide-13
SLIDE 13

9

Factorization problem The De Bruijn matrices Factorizations into commuting factors General form of a root of In with minimum rank A root class of In Conclusion

slide-14
SLIDE 14

10

Lemma

If A ∈ {0, 1}n×n is such that Am = In, then

  • A is p-regular, i.e A.1 = p.1 and 1T.A = p.1T
  • n = pm
slide-15
SLIDE 15

10

Lemma

If A ∈ {0, 1}n×n is such that Am = In, then

  • A is p-regular, i.e A.1 = p.1 and 1T.A = p.1T
  • n = pm

Definition

The De Bruijn matrix of order p and dimension n is a matrix of the form: D(p, n) := 1p ⊗ In/p ⊗ 1T

p ,

where

  • In/p is the identity matrix of dimension n/p
  • 1p is the p × 1 vector with all ones
  • ⊗ denotes the Kronecker product

Moreover, it is imposed that n = pm, for some integer m.

slide-16
SLIDE 16

11

D(2, 8) =             1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1            

slide-17
SLIDE 17

11

D(2, 8) =             1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1            

Proposition

The De Bruijn matrix D(p, n) with n = pm is such that D(p, n)m = In.

slide-18
SLIDE 18

11

D(2, 8) =             1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1            

Proposition

The De Bruijn matrix D(p, n) with n = pm is such that D(p, n)m = In. Question: Can we characterize all the roots from the De Bruijn matrices ?

slide-19
SLIDE 19

12

Factorization problem The De Bruijn matrices Factorizations into commuting factors General form of a root of In with minimum rank A root class of In Conclusion

slide-20
SLIDE 20

13

Factorization into commuting factors: Looking for the solutions to: AB = BA = In, where

  • A and B are binary matrices
  • A is p-regular
  • B is l-regular
slide-21
SLIDE 21

14

Factorization problem: AB = BA = In, where A is p-regular and B is l-regular.

Theorem

If A and B are commuting factors, then

  • p.l = n
  • rank(A) ≥ n/p and rank(B) ≥ n/l
  • if rank(A) = n/p (resp. rank(B) = n/l), then there exist

permutation matrices P1, P2 such that P1APT

2 = D(p, n)

(resp. P2BPT

1 = D(l, n)).

slide-22
SLIDE 22

15

Question: Is it possible that rank(A) > n/p ? A =     1 1 1 1 1 1 1 1     B =     1 1 1 1 1 1 1 1    

  • A and B are 2-regular
  • AB = BA = I4
  • BUT, rank(A) = 3 > 4/2
slide-23
SLIDE 23

16

Question: Can we choose P1 = P2 ? A =         1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1         B =         1 1 1 1 1 1 1 1 1 1 1 1        

  • A is 3-regular, B is 2-regular and AB = BA = I6
  • rank(A) = 6/3, rank(B) = 6/2
  • BUT, A is not isomorphic to D(3, 6) since

A2 =         3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3         , D(3, 6)2 =         2 2 2 1 1 1 1 1 1 2 2 2 2 2 2 1 1 1 1 1 1 2 2 2 2 2 2 1 1 1 1 1 1 2 2 2        

slide-24
SLIDE 24

17

Corollary

Let A be a binary matrix satisfying Am = In. Then,

  • A is p-regular
  • if rank(A) = n/p, then there are permutation matrices P1, P2

such that P1APT

2 = D(p, n).

As previously,

  • A may have a rank greater than n/p
  • A may not be isomorphic to D(p, n) even though rank(A) = n/p
slide-25
SLIDE 25

18

Factorization problem The De Bruijn matrices Factorizations into commuting factors General form of a root of In with minimum rank A root class of In Conclusion

slide-26
SLIDE 26

19

Theorem

Let A ∈ {0, 1}n×n such that Am = In, A is p-regular and pm = n. If rank(A) = n/p, then A is isomorphic to a matrix P1D(p, n), where P1 = diag(Q1, ..., Qp) with each Qi ∈ {0, 1}n/p×n/p is a permutation matrix.

slide-27
SLIDE 27

19

Theorem

Let A ∈ {0, 1}n×n such that Am = In, A is p-regular and pm = n. If rank(A) = n/p, then A is isomorphic to a matrix P1D(p, n), where P1 = diag(Q1, ..., Qp) with each Qi ∈ {0, 1}n/p×n/p is a permutation matrix. Not all the matrices of that form are solutions. Indeed, consider A =             1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1            

slide-28
SLIDE 28

20

Factorization problem The De Bruijn matrices Factorizations into commuting factors General form of a root of In with minimum rank A root class of In Conclusion

slide-29
SLIDE 29

21

A 2-circulant matrix:       1 1 1 1 1 1 1 1 1 1      

Theorem (Wu, 2002)

Let A ∈ {0, 1}n×n be g-circulant and such that Am = In. If

  • gm ≡ 0 mod n
  • A is p-regular,

then A is isomorphic to D(p, n).

slide-30
SLIDE 30

22

Definition

A nice permutation matrix is built as follows: start with a p × p permutation matrix. Then, replace all the zeros by a zero p × p matrix and each one by a p × p permutation matrix. Repeat this m

  • times. You obtain a permutation matrix of dimension pm.
slide-31
SLIDE 31

22

Definition

A nice permutation matrix is built as follows: start with a p × p permutation matrix. Then, replace all the zeros by a zero p × p matrix and each one by a p × p permutation matrix. Repeat this m

  • times. You obtain a permutation matrix of dimension pm.

Theorem

Any matrix of the form P1D(p, n) (n = pm) with P1 a nice permutation matrix is a m-th root of In.

slide-32
SLIDE 32

22

Definition

A nice permutation matrix is built as follows: start with a p × p permutation matrix. Then, replace all the zeros by a zero p × p matrix and each one by a p × p permutation matrix. Repeat this m

  • times. You obtain a permutation matrix of dimension pm.

Theorem

Any matrix of the form P1D(p, n) (n = pm) with P1 a nice permutation matrix is a m-th root of In.

Theorem

Any nice permutation of the De Bruijn matrix D(p, n) is isomorphic to D(p, n).

slide-33
SLIDE 33

23

Factorization problem The De Bruijn matrices Factorizations into commuting factors General form of a root of In with minimum rank A root class of In Conclusion

slide-34
SLIDE 34

24

We have investigated the solutions to Am = In,

  • ver the n × n binary matrices.
slide-35
SLIDE 35

24

We have investigated the solutions to Am = In,

  • ver the n × n binary matrices.
  • If A is a solution, then
  • A is p-regular and n = pm
  • rank(A) ≥ n/p
  • rank(A) = n/p implies that A is essentially a De Bruijn matrix
slide-36
SLIDE 36

24

We have investigated the solutions to Am = In,

  • ver the n × n binary matrices.
  • If A is a solution, then
  • A is p-regular and n = pm
  • rank(A) ≥ n/p
  • rank(A) = n/p implies that A is essentially a De Bruijn matrix
  • Any root of In with minimum rank is isomorphic to a matrix of

the form diag(Q1, ..., Qp).D(p, n), where any Qi is a permutation matrix.

slide-37
SLIDE 37

24

We have investigated the solutions to Am = In,

  • ver the n × n binary matrices.
  • If A is a solution, then
  • A is p-regular and n = pm
  • rank(A) ≥ n/p
  • rank(A) = n/p implies that A is essentially a De Bruijn matrix
  • Any root of In with minimum rank is isomorphic to a matrix of

the form diag(Q1, ..., Qp).D(p, n), where any Qi is a permutation matrix.

  • Any nice permutation of the De Bruijn matrix is a root of In

isomorphic to the De Bruijn matrix

slide-38
SLIDE 38

24

We have investigated the solutions to Am = In,

  • ver the n × n binary matrices.
  • If A is a solution, then
  • A is p-regular and n = pm
  • rank(A) ≥ n/p
  • rank(A) = n/p implies that A is essentially a De Bruijn matrix
  • Any root of In with minimum rank is isomorphic to a matrix of

the form diag(Q1, ..., Qp).D(p, n), where any Qi is a permutation matrix.

  • Any nice permutation of the De Bruijn matrix is a root of In

isomorphic to the De Bruijn matrix

  • Future work: characterize all the roots of In with minimum

rank.

slide-39
SLIDE 39

25

Thank you for your attention !