Scribd
Upload
129 views

Lesson03 PDF

The document discusses LU decomposition, which factors a matrix A into lower triangular matrix L and upper triangular matrix U such that A = LU. It describes how elementary row operations can transform any square matrix into upper triangular form. For nonsingular matrices, the LU decomposition and additional row operations can yield the identity matrix, demonstrating the matrix is invertible. Permutations may be required if initial row exchanges are needed before factoring.

Uploaded by

sal27adam
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
129 views

Lesson03 PDF

The document discusses LU decomposition, which factors a matrix A into lower triangular matrix L and upper triangular matrix U such that A = LU. It describes how elementary row operations can transform any square matrix into upper triangular form. For nonsingular matrices, the LU decomposition and additional row operations can yield the identity matrix, demonstrating the matrix is invertible. Permutations may be required if initial row exchanges are needed before factoring.

Uploaded by

sal27adam
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

MA 511, Session 3

LU Decomposition

Remark: It is easy to “undo” what an elementary


matrix does.

Pij Pij =I
Ei (c)Ei ( 1c ) =I
Eij (c)Eij (−c) = I

Go back to the system




 a11 x1 + a12 x2 + · · · + a1n xn = b1


 a21 x1 + a22 x2 + · · · + a2n xn = b2

 ............ ...



am1 x1 + am2 x2 + · · · + amn xn = bm

It can be written using matrices as

Ax = b
where
 
a11 a12 ... a1n
 a21 a22 ... a2n 
A=
 ...
,
... ... ... 
am1 am2 ... amn

x  b 
1 1
 x2   b2 
   
 ·   · 
x= 
 · , b= 
 · .
   
 ·   · 
xn bm

Consider now Gaussian elimination applied to a


(square) n × n matrix A.
By elementary row operations, A can be brought
to upper triangular form
 
a11 a12 ... a1n
 0 ā22 ... ā2n 
 
... ... ... ... 
0 0 ... ānn
Example:
   
1 2 3 1 2 3
 4 5 6  −→  0 −3 −6 
7 8 9 0 0 0

Using elementary matrices:


 
1 2 3
E32 (−2)E31 (−7)E21 (−4)  4 5 6 
7 8 9
 
1 2 3
=  0 −3 −6 
0 0 0

Thus,
 
1 2 3
4 5 6
7 8 9
 
1 2 3
= E21 (4)E31 (7)E32 (2)  0 −3 −6 
0 0 0
Remark: In our algorithm, we began with the top
row and worked our way down. This implies that
the elementary matrices E32 (−2), E31 (−7), E21 (−4)
(and hence also E21 (4), E31 (7), E32 (2)) are obtained
from I introducing no changes above the diagonal.
Therefore, their product, just like themselves, is
lower triangular:

L = E21 (4)E31 (7)E32 (2)

   
1 0 0 1 0 0 1 0 0
= 4 1 00 1 00 1 0
0 0 1 7 0 1 0 2 1

 
1 0 0
= 4 1 0
7 2 1

Note that the entries l21 , l31 , l32 of L below the di-
agonal are exactly the multipliers 4, 7, and 2 in the
elementary matrices E21 (4), E31 (7), E32 (2) above.
Therefore, if no row exchanges are needed, we may
write
A = LU
where L is lower triangular with 1’s on the diagonal
and U is upper triangular. This is the LU decompo-
sition of A.

From here we can derive a more symmetric fac-


torization of A. Suppose none of the coefficients on
the diagonal of U are zero. Then,
  
1 0 ... 0 d1 u12 . . . u1n
 l21 1 ... 0   d2 . . . u2n 
A=  0 
... ... ... ...   ... ... ... ... 
ln1 ln2 ... 1 0 0 ... dn
 
1 0 ... 0
 l21 1 ... 0 
= ... ... ... ... 

ln1 ln2 . . . 1
  
d1 0 ... 0 1 u12 /d1 . . . u1n /d1
 0 d2 ... 0   . . . u2n /d2 
  0 1 
... ... ... ...   ... ... ... ... 
0 0 . . . dn 0 0 ... 1

This is the LDU decomposition of A.


Suppose now that we do need row exchanges.
These are accomplished multiplying by Pij matrices.
A product P of such elementary matrices is called a
permutation matrix.

Example:
     
1 1 1 1 1 1 1 1 1
A = 1 1 3 → 0 0 2 → 0 3 6
2 5 8 0 3 6 0 0 2

We can do the necessary row exchanges first,


then the LU factorization:
   
1 1 1 1 1 1
P23 A =  2 5 8  →  0 3 6
1 1 3 0 0 2
Showing the elimination steps:
E31 (−1)E21 (−2)P23 A = U
Now P A = LU , where P = P23 , L = E21 (2)E31 (1).
This can always be done since, for i, k, l > j,
Pkl Eij (c) equals either Elj (c)Pkl , when k = i, or
Ekj (c)Pkl , when l = i, or else Eij (c)Pkl .
Definition: We say A is nonsingular if there is a
sequence of row operations on A that yields U with
nonzero coefficients on the diagonal. Otherwise A is
singular.

Suppose A is nonsingular. Then, there is a se-


quence of row operations taking A to U . Since the
coefficients on the diagonal of U are nonzero, we can
use them as pivots to make every coefficient above
the diagonal of U become zero using more elemen-
tary row operations (multiplication from the left by
upper triangular elementary matrices). Finally, di-
viding each row by the coefficient on the diagonal
(pivot), we can make all the coefficients on the diag-
onal equal to 1.

In summary: If A is nonsingular, there is a se-


quence of elementary row operations that take A to
the identity matrix.

You might also like