International Journal of Engineering and Technical Research (IJETR)

ISSN: 2321-0869, Volume-3, Issue-4, April 2015

An Easy New Approach for Matrix Decomposition

Mohammed Hassan Elzubair
A=WH
Abstract In the mathematical discipline of linear algebra Matrix multiplication can be implemented as computing the
there are many operations defined on matrices such as addition , columns vectors of A as linear combinations of the columns
multiplication and matrix inverse etc. Actually we note that in vector in W using coefficients supplied by the columns of H,
the real world it is not feasible for most of the matrix
so that each column of A can be computed as follows
computations to be calculated in an optimal explicit way, such as
matrix inversion . Matrix decomposition is a fundamental theme ai wi hi ,
in linear algebra and applied statistics which has both scientific Where a i is the ith column vector of the matrix A and hi is
and engineering significance. Now suppose we are given a
matrix A and we have to transform it into a product of two the ith column vector of the matrix H .
matrices ,say we want to write A=BC therefore in this case a To make it clear we will give a common example which is
numerically stable and fairly fast scheme is desired to compute facial image decomposition[2].
the two matrices A and B. In general we mean by the sentence Let A a matrix with a real-valued, nonnegative entries of
matrix decomposition or matrix factorization the factorization
of a matrix into a product of two or more matrices, so the size m n such that
process of converting or transforming a a given matrix in one A WH
side to product of two or more matrices in the other side is Where W is m r , H is r n , and both factors have none
called matrix decomposition or matrix factorization. Really negative entries . If the columns of A are pixels of a facial
matrix factorization or matrix decomposition has numerous image, the columns of W may be facial features such as eyes
algorithm like LU factorization ,QR decomposition ,rank or ears , and the coefficients in H represents the intensity of
factorization, singular value factorization etc . But most of these
these features . Non-negative matrix factorization is used in
algorithms are complicated ,so our goal here is to find a simple
algorithm with which we can factorize or decompose the abroad range of applications including graph clustering [3],
specified matrix and this was done by making use of the inverse Protein sequence motif discovery [4], and hyper spectra
of the matrix . Sometimes we need to factorize a matrix unmixing [5] .
according to different constraints like non- negative matrix For factor analyses which is a model for real- valued data we
factorization and this happened if we have a non-negative can construct an equation of four matrices in the form
matrix and we want to factorize it also into non- negative ones, A WH E where is E called a noisy matrix or the error
or any other constraints due to the nature of the problem which
matrix ,and we can use the least square method to find the
was represented by the matrices ,some methods of
decomposition factorizes approximately not exactly but here in
minimum error so as to get the approximate equation [7].
our easy new approach the factorization will be exact . A WH
In this paper we will introduce a very simple algorithm with
which we can factorize any matrix exactly to a product of any II. PRELIMINARIES
numbers of matrices by factorizing again each matrix of Sometimes matrix decomposition has a benefit that it makes
matrices which form the product .
to reduce dataset and summarization .For example if we have
Index Terms Invertible matrix, matrix decomposition, an 1000 50 matrix A this matrix has 50000 entities ,if we
noisy matrix, the entries of the matrix, singular matrix, success to factorize the matrix A into a product of two
constraints ,matrix multiplication ,associative law, the rank of matrices W and H such that W is an 1000 2 matrix and H is
the matrix. 2 50 matrix , we note that the sum of the entities in the two
matrices is 2100 ,
Theorem (1)
Any square n n
I. INTRODUCTION
matrix A which was written
This paper propose an easy new method for matrix
as A BC such that , B is an n k matrix and C is an
decomposition and it contains four sections, in section 2 there
k n matrix , k n then A is singular matrix
is a theorem with its proof using the notion of the determinant
and of the matrix and its properties, in section 3the invertible
matrix play an important role for matrix decomposition Proof
besides the associative law of matrix multiplication, it also b11 . . . b1k
contains some numerical example to make the idea very clear, Let , . . . . .

section 4 is the conclusion of the paper. B . . . .. .

We note matrix factorization refers to transformation of a . . . . .
given matrix into a given multiplication form [1]. bn1 . . . bnk
Now let matrix A be the product of the matrices W and H, c11 . . . c1n
. . . . .
Manuscript received April 20, 2015.

Mohammed Hassan Elzubair , Taif University - Raniah Branch ,
C . . . .. .

Department of Mathematics . . . . .
ck1 . . . ckn

294 www.erpublication.org
 An Easy New Approach for Matrix Decomposition

Since k n we can add n k zero columns to the matrix Every m n matrix A can be written as A=BC where B and
B to get a new square n n matrix W ,and we can also add m k and k n respectively
C are two matrices of sizes
n k zero rows or any n k rows to get another new square
n n matrix H where m k
Now we have Proof
We can form a none singular matrix D=WH where W is of
b11 . . . b1k 0...0 type m k and H of type k m
. . . . . Now we have an invertible square matrix D of type m m
,
W . . . .. . .. So we can write

. . . . . DA WHA (!)
bn1 . . . bnk 0 ...0 1
Multiplying both sides with D we get
D 1 DA D 1WHA
c11

... . c1n


A D 1W HA (2)
. . . .
So by taking B D W , C HA theorem (2) will be
1

H .c
.
k1 . . .. .c kn proved.
0 .0
Now in equation (2) we have

. . .

0 o
A D 1W HA

. . .
so substituting for the last A in by A D 1W HA ,
Now also we have A WH . then using associative law of matrix multiplication , we get
Also we have det W 0 , det H 0 .
A D 1W H D 1WH A
So we get det A det W det H 0 0 0 A is singular. Again we can substitute for the last A to get

III. METHODOLOGY

A D 1W H D 1W H D 1W HA , and so on
A D W H D W H ...D W HA EHEH ...EHA
1 1 1
3
where E D W and obviously D W H I ,
In this paper we want to factorize a matrix into two matrix or 1 1
more using the inverse matrix, so we make use the above m

theorem where I m is the identity matrix of rank m.

Now suppose we have an m n matrix A and we want to
Since we have I m I m I m ... I m we can deduce that
decompose it into a product of a certain m m invertible
matrix D,then we can write
I m D 1W H D 1W H ... D 1W H EHEH ...EH 4
A DD 1 A A DD 1C DC ,where D 1C C
So we have a matrix A equal to a certain square matrix D So we can also analyze the identity matrix of any rank using
multiplied by other matrix C , or we can write this method.

A AD 1 D ED where in this case D is an n n matrix,
The benefit of the easy approach is that ,if a matrix can be
since multiplication of matrices is associative.
decompose to certain matrices
Now suppose we have an m n matrix A which is not square by using this approach we can get the other factors ,to show
and we we want to decompose this matrix into a product of a this suppose we have a matrix
certain m k matrix w and other matrix , in this case we can
form an invertible m m matrix D by multiplying w by any 1
T
3 1 and we are given two matrices .
k m matrix h( i.e D=wh) 4 2 1

By this method we can factorize any matrix to a product of
1 0 2 , 1 0
A B 2
,
5
any matrices we want and this can be done by two ways either
a) we decompose the main matrix into two matrices and then 4 2 4

1 2

we decompose each one of these two matrices using this easy
method to get four matrices, then again we can decompose We want to find a matrix C such that T ABC .
each one of these four matrices by the same way, and so on We can do this as follows:
then by using the association law of matrix multiplication we
get the result (at every step of decomposition we use different 1 0
1 2 3 4
2
0
invertible matrices) , or AB 17 D
4 2 5 4
1 14


2
1 1
b) by writing A AD1 D1 ...Dn Dn where Di is invertible 7 2
n n matrix and using association law to get 1
13 13
3
D
A AD1 D11 D2 ...Dn1 Dn1 Dn
17
or any other combination Theorem
26 26
(2)
Next we have

295 www.erpublication.org
 International Journal of Engineering and Technical Research (IJETR)
ISSN: 2321-0869, Volume-3, Issue-4, April 2015
T DD 1T ABD 1T 1 2
1 1
1 0 7 2 1 1 1 0 1 1
1 2
13 1 1 W ,H , D WH ,
2
0 3
13 2 1 0 1 1 0 1 2
4 5 4
1
17 3 2 1
2
2 4
26 26

0 5
1 0 1 17 5 2 1
13 , D 1
1
T
0 2
2 13 13
4 5 7 1 1
4 2 5
45


1 2
26 call these four matrices standard matrices , or we can use any
26 13
1 17 5 other four matrices satisfying the conditions, in general for

C 13 13 13 every matrix A of type m n we can write A=BC where B
7
is m k matrix and C is k n one such that k m by taking
5 45

26 26 13
1
four standard matrices W,H,D, D where W is an
Example .1 m k matrix and H is k n matrix.
Given A 2 3 4 1 1 , write A= BC where B is 2 4
5 0 1 2 1
IV. CONCLUSION .
matrix and C is 4 5 one. This paper propose very effective and simple method for
Solution By taking matrix decomposition by using invertible matrices ,so by easy
method we can factorize any matrix to certain matrices we
1 2 want and other matrices ,although there are many methods for
1 1
1 1 0 1 matrix factorization ,this easy method will be the simplest one
W ,H
2 1 1 0 1 0 of them. Unlike the other approach we can decompose any

0 5 matrix to any number of factors (matrices). Easy approach
1 2 has the benefit that for every m n matrix A we can take
1 1 1 0 1 1
1 1 four standard matrices W,H,D, D where
1
W is an
D 1
2
2 1 0 1

1 0

m k matrix, H is k n to factorize A.
0 5
2 1 REFERENCES
D 1
1 1
[1] E.W Wesstein. Matrix decomposition. Math world-AWolfram Web
Now we have Resource.
[2] D.Gullamet and J.Vitria .Non-negative matrix factorization for face
1 2
region . In topics in an artificial intelligence ,Springer ,2002
1 1 1 0
1 1
2 3 4 1 1 [3] D.Kuang ,H.Park and C.H.Dig.Symmetric non-negative matrix
DA
1
1 0 1 factorization for graph clustering .In SDM,Volume12 ,2012
5
2 1 0 0 1 2
[4] W.Kim,B.Chen,JKim,Y.Pan,andH.Park. Sparse none negative
0 5 factorization for protein sequence motif discovery . Expert system with
D 1 we get
applications , 2011
Multiplying both sides with [5] S.Jia and Y Qian .Constrained non-negative matrix factorization for
hyperspectral unmixing .2009
1 2 [6] Abdelaziz hamad and Bahrom Sanugi (2011) .Neural Network and

1 2 3 4 1 1
Scheduling Germany : lap Lambert Academic publishing
1 2 1 1 1 1 0 1 [7] Mohammed Hassan Elzubair, Abdelaziz Hamad Elawad(2014),
D DA
1 1 2 1 0 1 1 0 5 0 1 2 1 International Journal of Engineering and Technical Research (IJETR)

0 5
1 2

2 1 1 1 1 0 1 1 2 3 4 1 1
A
0 5 0 1 2 1
1 1 2 1 0 1 1

0
5
12 3 6 5 3
2
4 3 2 1 7 3 5 3
A
3 2 1 1 2 3 4 1 1

25 0 5 10 5

Note
Since matrix factorization is not unique ,we can use more
simple matrices W , H , D , D 1
Also for writing any 2 5 matrix A, as A=BC where B
is 2 4 matrix and C is 4 5 one we can use these four
matrices

296 www.erpublication.org