A SUMMARY OF MATRICES

Total horizontal velocity vectors from three GPS stations can establish how the
triangular area enclosed by the stations deforms or strains over time. Although you can
get a sense of strain with a graphical analysis, you can do a robust analysis using
matrices. Matrices and computers make the math feasible. This document summarizes
matrices; for more explanation, see “Two faces of vectors.”

DEFINITIONS
It will be useful to remember (or perhaps learn) a few things about matrices so we can
use them to solve problems. We will start by defining a matrix we will call a.

é a11 ù
ê ú
a =ê a21 ú
êa ú
ë 31 û
Matrix a is called a column matrix, because its three elements are arrayed in a vertical
column. Matrix a might also be called a 3x1 matrix, because its elements are arrayed in
three rows and 1 column. The subscripts associated with each of the elements in the
array indicate the position of the element in the array, so a 21 is the element in the 2nd row
and 1st column. The first subscript indicates what row the element is located in, and the
second subscript indicates the column. We often associate the subscript “1” with an x-
coordinate axis, “2” with a y-axis, and “3” with a z-axis.

A matrix with one row and three columns (a row matrix or a 1x3 matrix) would look like
this,

é m11 m12 m13 ù

ë û

and a matrix with three rows and two columns (a 3x2 matrix) would look like this

é m11 m12 ù
ê ú
ê m21 m22 ú
êm m32 ú
ë 31 û
Matrix b is a 3x3 matrix because it has three rows and three columns. It is also known
as a square matrix, because it has the same number of rows as columns.

é b11 b12 b13 ù

ê ú
b=ê b21 b22 b23 ú
êb b b33 ú
ë 31 32 û

A column matrix (e.g. 3×1) or a row matrix (e.g. 1×3) is also referred to as a vector.

Some simple matrix mathematics

We can add, subtract, multiply or divide each of the components of a matrix by a

scalar. For example, if s is a scalar and b is a 2x2 matrix,

é b11 b12 ù é s+ b11 s+ b12 ù

s+ b=s+ ê ú =ê ú
ê b b ú s+ b21 s+ b22
ë 21 22 û êë ú
û

We can also add, subtract, multiply or divide two matrices that have the same
dimensions, component-by-component. For example, if b and c are both 2 x 2 matrices,

é b11 b12 ù é c11 c12 ù é b11 / c11 b12 / c12 ù

b¸ c =ê ú¸ ê ú=ê ú
ê b b ú c c b /c b /c
ë 21 22 û êë 21 22 ú
û ê
ë 21 21 22 22 ú
û

Multiplication of two matrices b and c by components is only possible if the two

matrices have the same number of rows and columns (that is, they have the same
dimensions), and might be indicated by b*c

é b11 b12 ù é c11 c12 ù é b11 * c11 b12 * c12 ù

b* c =ê ú* ê ú=ê ú
ê b b ú c c b *c b *c
ë 21 22 û êë 21 22 ú
û ê
ë 21 21 22 22 ú
û

Probably the more commonly used mode of matrix multiplication, distinct from simple
multiplication by component as described above, involves the use of dot products. Two
matrices can be multiplied using dot products if the number of rows in one matrix equals
the number of columns in the other matrix. The matrices to be multiplied in this manner
do not need to have the same dimensions or number of components; however, one
matrix must have the same number of rows as the other matrix has columns. Let's
multiply matrices a and b together using dot products to yield a product: matrix d.

c =b×a

or, represented another way,

é d11 ù é b11 b12 b13 ùé a11 ù

ê ú ê úê ú
ê d21 ú=ê b21 b22 b23 úê a21 ú
êd ú êb b b33 úê a31 ú
ë 31 û ë 31 32 ûë û

We can think of matrix b as consisting of three 3-component vectors: {b11, b12, b13} in
the top row, {b21, b22, b23} in the middle row, and {b31, b32, b33} in the bottom row.
Each element of the resultant matrix d is the dot product of matrix a with one row-vector
of matrix b.

é d11 ù é topRow×a ù
ê ú ê ú
ê d21 ú=ê middleRow×a ú
êd ú ê bottomRow×a ú
ë 31 û ë û, or

é d11 ù é ( b11a11 ) + ( b12 a21 ) + ( b13a31 ) ù

ê ú ê ú
ê d21 ú=ê ( b21a11 ) + ( b22 a21 ) + ( b23a31 ) ú
êd ú ê ú
ë 31 û ê ( b a ) + ( b32 a21 ) + ( b33a31 )
ë 31 11
ú
û

For example, the top element in matrix d is found by solving the following equation

d11 = (b11 x a11) + (b12 x a21) + (b13 x a31).

And so, the product (matrix d) of multiplying a 3x1 vector (matrix a) by a 3x3 matrix
(matrix b) is a matrix with three elements: another vector. What if we take vectors a, b
and e and multiply them together to yield a matrix f, where a and b are the same as
before and matrix e is
 é e11 e12 e13 ù
ê ú
e=ê e21 e22 e23 ú
êe e e33 ú
ë 31 32 û?

It turns out that the equation

f =e×b×a

is functionally the same as;

f =e×d

where

d =b×a

as we defined matrix d above, so

é f11 ù é eTopRow×d ù é ( e11d11 ) + ( e12 d21 ) + ( e13d31 ) ù

ê ú ê ê ú
ú ê ú
ê f21 ú=ê eMiddleRow×d ú= ( e21d11 ) + ( e22 d21 ) + ( e23d31 )
ê f ú ê eBottomRow×d ú ê ú
ë 31 û ë û ê ( e31d11 ) + ( e32 d21 ) + ( e33d31 ) ú
ë û

In these examples, we start the process of multiplying from the right side of the
equation, where we will find a 3x1 matrix representing a vector. Multiplying a 3x1 matrix
by the next matrix to the left (a 3x3 matrix) yields another 3-component matrix
representing a vector. If there is another 3x3 matrix to the left, repeat the process, and
keep repeating the process until you reach the = sign.
Example. Find the result of the following matrix multiplication:

é 3 4 6 ùé 0.2 0.6 0.5 ùé 11 ù

ê úê úê ú
ê 1 2 8 úê 0.8 0.9 0.7 úê 15 ú
ê
ë9 7 5 úûê
ë 0.4 0.1 0.3 ú
ûê
ë 18 ú
û

Solution,

step 1. Start by multiplying the 3x1 vector matrix on the right by the 3x3 matrix next to it,
in the middle of the sequence.

é ù
é 20.2 ù ê ( 0.2 ´ 11) + ( 0.6 ´ 15 ) + ( 0.5 ´ 18 ) ú é 0.2 0.6 0.5 ùé 11 ù
ê ú =ê 0.8 ´ 11) + ( 0.9 ´ 15 ) + ( 0.7 ´ 18 ) ú=ê 0.8 0.9 0.7 úê 15 ú
ê 34.9 ú ê ( ú ê 0.4 0.1 0.3 úê 18 ú
ê
ë 11.3 ú
û ê ( 0.4 ´ 11) + ( 0.1 ´ 15 ) + ( 0.3 ´ 18 ) ú ê
ë ú
ûê
ë ú
û
ë û

Step 2. Use the results of step 1 as the 3x1 vector matrix on the right.

é ù
é 268 ù ê ( 3 ´ 20.2 ) + ( 4 ´ 34.9 ) + ( 6 ´ 11.3) ú é 3 4 6 ùé 20.2 ù
ê ú=ê 1´ 20.2 ) + ( 2 ´ 34.9 ) + ( 8 ´ 11.3) ú=ê 1 2 8 úê 34.9 ú
ê 180.4 ú ê ( ú ê 9 7 5 úê 11.3 ú
ê
ë 482.6 ú
û ê ( 9 ´ 20.2 ) + ( 7 ´ 34.9 ) + ( 5 ´ 11.3) ú ê
ë ú
ûê
ë ú
û
ë û

The product of the three matrices is the following 3-component vector:

{268, 180.4, 482.6}.

 Recognizing different parts of a matrix and different types of matrices

In this square matrix,

é A 0 0 ù
ê ú
ê 0 B 0 ú
ë 0 0 C ú
ê û

the part of the matrix that has all of the capital letters (A, B, C) is called the diagonal or
axis of the matrix. Values that are in the positions occupied by the 0s are said to be off-
axis terms.

In a symmetric matrix, like the one below, the values above the diagonal are equal to
the values below and directly across the diagonal.

é A d e ù
ê ú
ê d B f ú
ê e f C ú
ë û

In an antisymmetric matrix, the values across the diagonal from each other have the
same magnitude but different sign.

é A d e ù
ê ú
ê -d B f ú
ê -e - f C ú
ë û

An asymmetric matrix, like

éA d e ù
ê ú
ê g B h ú
ê n - f C ú
ë û

lacks at least some of the symmetries we have just examined.

If we define a matrix M as follows

 éA d e ù
ê ú
M =ê g B h ú
ê n - f C ú
ë û,

The transpose of matrix M is represented by MT and is

éA g n ù
ê ú
M T =ê d B - f ú
ê e h C ú
ë û.

The values along the diagonal of the transposed matrix are unchanged from the original
matrix, but the values across the diagonal from each other are swapped.

The inverse of a matrix M is symbolized by M-1. If a matrix is multiplied by its inverse,

the result is the identity matrix whose diagonal terms are all 1s and whose off-axis
terms are all 0s.

é1 0 0 ù
M ×M =ê 0 1 0 ú
-1
ê ú
ê
ë 0 0 1 ú
û.

If the transpose of a matrix is the same as the inverse of the matrix, that matrix is called
an orthogonal matrix.
 Matrix Norm

Given a square complex or real matrix  , a matrix norm   is a nonnegative number

associated with   having the properties

1.   when   and   iff  ,

2.   for any scalar  ,

3.  ,

4.  .

Let  , ...,   be the eigenvalues of  , then

(1)

The matrix  -norm is defined for a real number   and a matrix   by

(2)

where   is a vector norm. The task of computing a matrix  -norm is difficult for   
since it is a nonlinear optimization problem with constraints.

Matrix norms are implemented as Norm[m, p], where   may be 1, 2, Infinity,

or "Frobenius".

The maximum absolute column sum norm   is defined as

(3)
The spectral norm  , which is the square root of the maximum eigenvalue of   
(where   is the conjugate transpose),

(4)

is often referred to as "the" matrix norm.

The maximum absolute row sum norm is defined by

(5)

,  , and   satisfy the inequality

Block Multiplication

Example:
Suppose that matrices A, B, P, X and Y are defined as

 1 2
 1 0 0 0 0 
 3 4 
0 1 0 0 0   I O23    X 
A     2 and B   2 0   
 0 0 2 0 3  O22 P    Y 
   0 3
 0 0 0 3 1
 1 2 

then

 I O23   X    I X  O23Y    X 
AB   2      2    
 O22 P  Y    O22 X  PY    PY 
 2 0
2 0 3   7 6 
PY    0 3  1 11 
 0 3 1     
 1 2 

 1 2
 3 4 
 X  
 AB     
 PY  7 6
 
 1 11 

This is somewhat faster than the direct evaluation of

 1 2
 1 0 0 0 0 1 2
 3 4 
0 1 0 0 0   3 4
AB    2 0    
 0 0 2 0 3  7 6
  0 3  
 0 0 0 3 1  1 11 
 1 2 

The partitioning of the matrices in a matrix multiplication must be such that all matrix
products are defined.
Example :

0 1
P   
Find the complete set of (22) matrices that commute with  0 0 .

a b
A   
Let the general (22) matrix be  c d  . Then

a b 0 1 0 a
AP      0 c
c d 0 0  

and

0 1a b  c d 
PA        
0 0c d  0 0

AP = PA if and only if c = 0 and d = a.

0 1
P   
Therefore, the complete set of (22) matrices that commute with  0 0  is

 a b  
A   
  0 a   , where a and b are any real numbers.
Example:

 I X 
A   
For the matrix  O  I  , where X, I and O are all square matrices of the same

size (kk), find an expression for any natural number power of A, An.

 I X   Ik X   I Okk 
A2   k   k  I 2k 
 Okk  I k   Okk 
 Ik   Okk I k 

 A 3  A 2 A  IA  A , A4  A2A2  I , A 5  A 4 A  A , etc.

Therefore

 I  n even 
An  
 A  n odd 