Linear Algebra Lectures

Introduction to systems of linear equations
These slides are based on Section 1 in Linear Algebra and its Applications by David C. Lay.
Definition 1. A linear equation in the variables x1, ..., xn is an equation that can be
written as
a1x1 + a2x2 +
+ anxn = b.
Example 2. Which of the following equations are linear?
4x1 5x2 + 2 = x1
x2 = 2( 6 x1) + x3
linear: 3x1 5x2 = 2
linear: 2x1 + x2 x3 = 2 6
4x1 6x2 = x1x2
x2 = 2 x1 7
not linear: x1x2
not linear: x1
Definition 3.
A system of linear equations (or a linear system) is a collection of one or more
linear equations involving the same set of variables, say, x1, x2, ..., xn.
A solution of a linear system is a list (s1, s2, ..., sn) of numbers that makes each
equation in the system true when the values s1, s2, ..., sn are substituted for x1, x2,
..., xn, respectively.
Example 4. (Two equations in two variables)
In each case, sketch the set of all solutions.
x1 + x2 = 1
x1 2x2 = 3
x1 + x2 = 0
2x1 4x2 = 8
-3
-2
2x1 + x2 = 1
4x1 2x2 = 2
-1
3 -3
-2
-1
3 -3
-2
-1
-1
-1
-1
-2
-2
-2
-3
-3
-3
Theorem 5. A linear system has either

no solution, or
one unique solution, or
infinitely many solutions.

Definition 6. A system is consistent if a solution exists.
Armin Straub
[email protected]
How to solve systems of linear equations

Strategy: replace system with an equivalent system which is easier to solve
Definition 7. Linear systems are equivalent if they have the same set of solutions.
Example 8. To solve the first system from the previous example:
x1 + x2 = 1
x1 + x2 = 0
x1 +
R2R2+R1
>
x2 = 1
2x2 = 1
Once in this triangular form, we find the solutions by back-substitution:

x2 = 1/2,
x1 = 1/2
Example 9. The same approach works for more complicated systems.

x1 2x2 + x3 = 0
2x2 8x3 = 8
4x1 + 5x2 + 9x3 = 9
x1 2x2 +
x3 = 0
2x2 8x3 = 8
3x2 + 13x3 = 9
R3 R3 + 4R1
R3 R3 + 2 R2
x1 2x2 + x3 = 0
2x2 8x3 = 8
x3 = 3
By back-substitution:
x3 = 3,
x2 = 16,
x1 = 29.
It is always a good idea to check our answer. Let us check that (29, 16, 3) indeed solves
the original system:
x1 2x2 + x3 = 0
2x2 8x3 = 8
4x1 + 5x2 + 9x3 = 9

3 @

2 16 8 3 @
29 2 16 +
0
8

9
4 29 + 5 16 + 9 3 @
Matrix notation
x1 2x2 = 1
x1 + 3x2 = 3
1 2
1 3
(coefficient matrix)
1 2 1
3
1 3
(augmented matrix)
Armin Straub
[email protected]
Definition 10. An elementary row operation is one of the following:

(replacement) Add one row to a multiple of another row.
(interchange) Interchange two rows.
(scaling) Multiply all entries in a row by a nonzero constant.

Definition 11. Two matrices are row equivalent, if one matrix can be transformed
into the other matrix by a sequence of elementary row operations.
Theorem 12. If the augmented matrices of two linear systems are row equivalent, then
the two systems have the same solution set.
Example 13. Here is the previous example in matrix notation.
x1 2x2 + x3 = 0
1 2 1
0
2x2 8x3 = 8 0
2 8
8 ,
4x1 + 5x2 + 9x3 = 9
R3 R3 + 4R1
4
5 9 9
x1 2x2 +
x3 = 0
2x2 8x3 = 8
3x2 + 13x3 = 9
x1 2x2 + x3 = 0
2x2 8x3 = 8
x3 = 3
1 2 1
0
8
0
2 8
0 3 13 9
1 2 1
0
2 8
0
0 1
R3 R3 + 2 R2
0
8
3
Instead of back-substitution, we can continue with row operations.

After R2 R2 + 8R3, R1 R1 R3, we obtain:
x1 2x2
2x2
x3
= 3
= 32
= 3
1 2
0
2
0
0
0 3
0 32
1
3
1
0
0
0
0
1
Finally, R1 R1 + R2, R2 2 R2 results in:

x1
x2
x3
= 29
= 16
= 3
0
1
0
29
16
3
We again find the solution (x1, x2, x3) = (29, 16, 3).
Armin Straub
[email protected]
Row reduction and echelon forms

Definition 14. A matrix is in echelon form (or row echelon form) if:
(1) Each leading entry (i.e. leftmost nonzero entry) of a row is in a column to the right
of the leading entry of the row above it.
(2) All entries in a column below a leading entry are zero.
(3) All nonzero rows are above any rows of all zeros.
Example 15. Here is a representative
0
0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
matrix in echelon form.
0 0 0
0 0 0 0
0 0 0 0 0 0 0
( stands for any value, and for any nonzero value.)

Example

0
(a)
0
0
0

(b)
0
0

0
(c)
0
0

(d)

16. Are the following matrices in echelon form?
YES
0 0 0 0
0 0 0 0
NOPE (but it is after exchanging

0 0 0 0
0 0 0 0
YES
0
0 0
0 0
0
NO
0
0 0
the first two rows)
Related and extra material

In our textbook: parts of 1.1, 1.3, 2.2 (just pages 78 and 79)
However, I would suggest waiting a bit before reading through these parts (say, until we covered
things like matrix multiplication in class).
Suggested practice exercise: 1, 4, 5, 10, 11 from Section 1.3

Armin Straub
[email protected]
Pre-lecture trivia
Who are these four?
Artur Avila, Manjul Bhargava, Martin Hairer, Maryam Mirzakhani

Just won the Fields Medal!
analog to Nobel prize in mathematics
awarded every four years
winners have to be younger than 40
cash prize: 15,000 C$
Armin Straub
[email protected]
Review
Each linear system corresponds to an augmented matrix.
6
2 1
1 2 1 9
1 2 12
2x1 x2
= 6
x1 +2x2 x3 = 9
x2 +2x3 = 12
augmented matrix
To solve a system, we perform row reduction.
>
R2R2+ 2 R1
>
R3R3+ 3 R2
2 1 0 6
3
0 2 1 6
0 1 2 12
2 1 0 6
0 3 1 6
2
echelon form!
4
3
Echelon form in general:
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0

0
0
0
0

0
0
0
0
0
0
0
0
0

0
0

0
The leading terms in each row are the pivots.
Armin Straub
[email protected]
Row reduction and echelon forms, continued

Definition 1. A matrix is in reduced echelon form if, in addition to being in echelon
form, it also satisfies:
Each pivot is 1.
Each pivot is the only nonzero entry in its column.
Example 2. Our initial matrix in echelon form put into reduced echelon form:
0
0 0 0
0 0 0 0
0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
Note that, to be in reduced echelon form, the pivots also have to be scaled to 1.
Example 3. Are the following
0 1 0 0 0 0
0 0 0 1 0 0 0
(a)
0 0 0 0 1 0 0
0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 1
1 0 5 0 7
0 2 4 0 6
(b)
0 0 0 5 0
0 0 0 0 0
1 0 2 3 2 24
(c) 0 1 2 2 0 7
0 0 0 0 1
4
matrices in reduced echelon form?
YES

NO
NO
Theorem 4. (Uniqueness of the reduced echelon form) Each matrix is row equivalent to one and only one reduced echelon matrix.
Question. Is the same statement true for the echelon form?
Clearly not; for instance,
Armin Straub
[email protected]
1 2
0 1
1 0
0 1
are different row equivalent echelon forms.
Example 5. Row reduce to echelon form (often called Gaussian elimination) and then
to reduced echelon form (often called GaussJordan elimination):
0 3 6 6 4 5
3 7 8 5 8 9
3 9 12 9 6 15
Solution.
After R1 R3, we get:
(R1 R2 would b e another option; try it!)
3 9 12 9 6 15
3 7 8 5 8 9
0 3 6 6 4 5
Then, R2 R2 R1 yields:
3 9 12 9 6 15
0 2 4 4 2 6
0 3 6 6 4 5
3
Finally, R3 R3 2 R2 produces the echelon form:
3 9 12 9 6 15
0 2 4 4 2 6
0 0 0 0 1 4
To get the reduced echelon form, we first scale
1 3 4 3
0 1 2 2
0 0 0 0
all rows:
2 5
1 3
1 4
Then, R2 R2 R3 and R1 R1 2R3, gives:
1 3 4 3 0 3
0 1 2 2 0 7
0 0 0 0 1 4
Finally, R1 R1 + 3R2 produces the reduced echelon form:
1 0 2 3 0 24
0 1 2 2 0 7
0 0 0 0 1
4
Armin Straub
[email protected]
Solution of linear systems via row reduction

After row reduction to echelon form, we can easily solve a linear system.
(especially after reduction to reduced echelon form)
Example 6.
1 6 0 3 0 0
0 0 1 8 0 5
0 0 0 0 1 7
x1 +6x2
x3
+3x4
8x4
x5
= 0
= 5
= 7
The pivots are located in columns 1, 3, 5. The corresponding variables x1, x3, x5 are
called pivot variables (or basic variables).
The remaining variables x2, x4 are called free variables.
We can solve each equation for the pivot variables in terms of the free variables (if
any). Here, we get:
x1 = 6x2 3x4
x1 +6x2
+3x4
= 0
x2 free
x3 = 5 + 8x4
x3 8x4
= 5
x
free
x5 = 7
4
x5 = 7
This is the general solution of this system. The solution is in parametric form, with
parameters given by the free variables.
Just to make sure: Is the above system consistent? Does it have a unique solution?
Example 7. Find a parametric description of the solution set of:
3x1
3x1
3x2 6x3 +6x4 +4x5 = 5

7x2 +8x3 5x4 +8x5 = 9
9x2 +12x3 9x4 +6x5 = 15
Solution. The augmented matrix is
0 3 6 6 4 5
3 7 8 5 8
9 .
3 9 12 9 6 15
We determined earlier that its reduced echelon form is
1 0 2 3 0 24
0 1 2 2 0 7 .
0 0 0 0 1
4
Armin Straub
[email protected]
The pivot variables are x1, x2, x5.

The free variables are x3, x4.
Hence, we find the general solution as:
x1 = 24 + 2x3 3x4
x2 = 7 + 2x3 2x4
x3 free
x free
4
x5 = 4
Related and extra material

In our textbook: still, parts of 1.1, 1.3, 2.2 (just pages 78 and 79)
As before, I would suggest waiting a bit before reading through these parts (say, until we covered
things like matrix multiplication in class).
Suggested practice exercise:

Section 1.3: 13, 20; Section 2.2: 2 (only reduce A, B to echelon form)
Armin Straub
[email protected]
Review
We have a standardized recipe to find all solutions of systems such as:
3x1
3x1
3x2 6x3 +6x4 +4x5 = 5

7x2 +8x3 5x4 +8x5 = 9
9x2 +12x3 9x4 +6x5 = 15
The computational part is to start with the augmented matrix
0 3 6 6 4 5
3 7 8 5 8
9 ,
3 9 12 9 6 15
and to calculate its reduced echelon form (which is unique!). Here:
1 0 2 3 0 24
0 1 2 2 0 7 .
0 0 0 0 1
4
pivot variables (or basic variables): x1, x2, x5
free variables: x3, x4
solving each equation for the pivot variables in terms of the free variables:
x1
x2
2x3 +3x4
2x3 +2x4
x5
Armin Straub
[email protected]
= 24
= 7
= 4
x1 = 24 + 2x3 3x4
x2 = 7 + 2x3 2x4
x3 free
x free
4
x5 = 4
Questions of existence and uniqueness

The question whether a system has a solution and whether it is unique, is easier to
answer than to determine the solution set.
All we need is an echelon form of the augmented matrix.
Example 1. Is the following system consistent? If so, does it have a unique solution?
3x1
3x1
3x2 6x3 +6x4 +4x5 = 5

7x2 +8x3 5x4 +8x5 = 9
9x2 +12x3 9x4 +6x5 = 15
Solution. In the course of an earlier example, we obtained the echelon form:
3 9 12 9 6 15
0 2 4 4 2 6
0 0 0 0 1
4
Hence, it is consistent (imagine doing back-substitution to get a solution).

Theorem 2. (Existence and uniqueness theorem) A linear system is consistent if
and only if an echelon form of the augmented matrix has no row of the form
[ 0 ... 0 b ],
where b is nonzero.
If a linear system is consistent, then the solutions consist of either
a unique solution (when there are no free variables) or
infinitely many solutions (when there is at least one free variable).
Example 3. For what values of h will the following system be consistent?
3x1 9x2 = 4
2x1 +6x2 = h
Solution. We perform row reduction to find an echelon form:
"
#

2
3
9
4
R2R2+
R1
3 9 4
3
8
2 6 h
0 0 h+ 3
>
The system is consistent if and only if h = 3 .

Armin Straub
[email protected]
Brief summary of what we learned so far

Each linear system corresponds to an augmented matrix.
Using Gaussian elimination (i.e. row reduction to echelon form) on the augmented
matrix of a linear system, we can
read off, whether the system has no, one, or infinitely many solutions;
find all solutions by back-substitution.
We can continue row reduction to the reduced echelon form.
Solutions to the linear system can now be just read off.
This form is unique!
Note. Besides for solving linear systems, Gaussian elimination has other important uses,
such as computing determinants or inverses of matrices.
A recipe to solve linear systems
(GaussJordan elimination)
(1) Write the augmented matrix of the system.

(2) Row reduce to obtain an equivalent augmented matrix in echelon form.
Decide whether the system is consistent. If not, stop; otherwise go to the next step.
(3) Continue row reduction to obtain the reduced echelon form.

(4) Express this final matrix as a system of equations.
(5) Declare the free variables and state the solution in terms of these.
Questions to check our understanding

On an exam, you are asked to find all solutions to a system of linear equations. You
find exactly two solutions. Should you be worried?
Yes, because if there is more than one solution, there have to be infinitely many solutions. Can
you see how, given two solutions, one can construct infinitely many more?
True or false?
There is no more than one pivot in any row.
True, because a pivot is the first nonzero entry in a row.
There is no more than one pivot in any column.

True, because in echelon form (thats where pivots are defined) the entries below a pivot
have to zero.
There cannot be more free variables than pivot variables.

False, consider, for instance, the augmented matrix
Armin Straub
[email protected]
[ 1 7 5 3 ].
The geometry of linear equations

Adding and scaling vectors
Example 4. We have already encountered matrices such as
1 4 2 3
2 1 2 2 .
3 2 2 0
Each column is what we call a (column) vector.
In this example, each column vector has 3 entries and so lies in R 3.
Example 5. A fundamental property of vectors is that vectors of the same kind can be
added and scaled.

1
4
5
2 + 1 = 1 ,
3
2
5
x1
7x1
7 x2 = 7x2 .
x3
7x3
Example 6. (Geometric description of R2) A vector

(x1, x2) in the plane.
Given x =
1
3
and y =
2
1
x1
x2
represents the point
, graph x, y, x + y, 2y.
0
0
Armin Straub
[email protected]
Adding and scaling vectors, the most general thing we can do is:
Definition 7. Given vectors v1, v2,
, vm in Rn and scalars c1, c2,
, cm, the vector
c1v1 + c2v2 +
+ cmvm
is a linear combination of v1, v2,
, vm.
The scalars c1,
, cm are the coefficients or weights.
Example 8. Linear combinations of v1, v2, v3 include:

1
v ,
3 2
3v1 v2 + 7v3,
v2 + v3,
0.
Example 9. Express
1
5
as a linear combination of
2
1
and
1
1
Solution. We have to find c1 and c2 such that

c1
This is the same as:
2
1
+ c2
1
1

1
=
.
5
2c1 c2 = 1
c1 +c2 = 5
Solving, we find c1 = 2 and c2 = 3.
Indeed,

2
1
1
2
+3
=
.
1
5
1
Note that the augmented matrix of the linear system is

2 1 1
,
1 1 5
and that this example provides a new way of think about this system.
Armin Straub
[email protected]
The row and column picture

Example 10. We can think of the linear system
2x y = 1
x+ y = 5
in two different geometric ways.
Row picture.
Each equation defines a line in R2.
Which points lie on the intersection of these lines?
Column picture.

The system can be written as x
Which linear combinations of
2
1
2
1

+y
and
1
1
1
1

1
5
produce
1
5
This example has the unique solution x = 2, y = 3.

(2, 3) is the (only) intersection of the two lines 2x y = 1 and x + y = 5.

2
1
+3
1
1
is the (only) linear combination producing
Armin Straub
[email protected]
1
5
Pre-lecture:
the shocking state of our ignorance
Q: How fast can we solve N linear equations in N unknowns?

Estimated cost

0
of Gaussian elimination:
to create the zeros below the pivot:
on the order of N 2 operations

if there is N pivots:
on the order of N N 2 = N 3 ops

A more careful count places the cost at 3 N 3 ops.

For large N , it is only the N 3 that matters.
It says that if N 10N then we have to work 1000 times as hard.
Thats not optimal!
W e can do b etter than G aussian elim ination :
Strassen algorithm (1969): N lo g 27 = N 2 .8 0 7

CoppersmithWinograd algorithm (1990): N 2 .3 7 5
StothersWilliamsLe Gall (2014): N 2 .3 7 3
Is N 2 possible? We have no idea!

Good news for applications:
(b etter is im p ossib le; w hy?)
(w ill see an exam ple so o n)
Matrices typically have lots of structure and zeros

which makes solving so much faster.
Armin Straub
[email protected]
Organizational
Help sessions in 441 AH: MW 4-6pm, TR 5-7pm
Review
A system such as
2x y = 1
x+ y = 5
can be written in vector form as

2
1
1
x
+y
=
.
1
1
5
The left-hand side is a linear combination of the vectors
2
1
and
1
1
The row and column picture

Example 1. We can think of the linear system
2x y = 1
x+ y = 5
in two different geometric ways. Here, there is a unique solution: x = 2, y = 3.
Row picture.
Each equation defines a line in R2.
Which points lie on the intersection
of these lines?
(2, 3) is the (only) intersection of
the two lines 2x y = 1 and x +
y = 5.
5
4
3
2
1
Armin Straub
[email protected]
Column picture.
The system can be written as

2
1
1
x
+y
=
.
1
1
5
Which
linear
combinations
of

1
1
and 1 produce 5 ?
4

2
1
3
2
(2, 3) are the coefficients of the

(only) such linear combination.
-3
-2
-1
Example 2. Consider the vectors
1
a1 = 0 ,
3
4
a2 = 2 ,
14
3
a3 = 6 ,
10
1
b = 8 .
5
Determine if b is a linear combination of a1, a2, a3.

Solution. Vector b is a linear combination of a1, a2, a3 if we can find weights x1, x2,
x3 such that:

1
4
3
1
x1 0 + x2 2 + x3 6 = 8
3
14
10
5
This vector equation corresponds to the linear system:
+4x2 +3x3 = 1
+2x2 +6x3 = 8
+14x2 +10x3 = 5
x1
3x1
Corresponding augmented matrix:
1 4 3 1
0 2 6 8
3 14 10 5
Note that we are looking for a linear combination of the first three columns which
Armin Straub
[email protected]
produces the last column.

Such a combination exists
the system is consistent.
Row reduction to echelon form:
1 4 3 1
1 4 3 1
1 4 3 1
0 2 6 8 0 2 6 8 0 2 6
8
3 14 10 5
0 2 1 2
0 0 5 10
Since this system is consistent, b is a linear combination of a1, a2, a3.
[It is consistent, because there is no row of the form
[ 0 0 0 b ]
with b 0.]
Example 3. In the previous example, express b as a linear combination of a1, a2, a3.
Solution. The reduced echelon form is:
1 4 3 1
0 2 6
8
0 0 5 10
1 4 3 1
0 1 3 4
0 0 1 2
1 4 0 7
0 1 0 2
0 0 1 2
1 0 0 1
0 1 0 2
0 0 1 2
We read off the solution x1 = 1, x2 = 2, x3 = 2, which yields
1
4
3
1
0 2 2 + 2 6 = 8 .
3
14
10
5
Summary
A vector equation
x1a1 + x2a2 +
+ xmam = b
has the same solution set as the linear system with augmented matrix
| |
a1 a2
| |
| |
am b .
| |
In particular, b can be generated by a linear combination of a1, a2,

, am if and only
if this linear system is consistent.
Armin Straub
[email protected]
The span of a set of vectors

Definition 4. The span of vectors v1, v2,
, vm is the set of all their linear combinations. We denote it by span{v1, v2,
, vm }.
In other words, span{v1, v2,
, vm } is the set of all vectors of the form
c1v1 + c2v2 +
+ cmvm ,
where c1, c2,
, cm are scalars.
Example 5.
(a) Describe span
n
2
1
o
geometrically.
The span consists of all vectors of the form

As points in R2, this is a line.
n o
(b) Describe span 21 , 41
geometrically.
2
1
The span is all of R2, a plane.
2.0
Thats because any vector

in R2 can

be written as x1 21 + x2 41 .
1.5
1.0
0.5
-2
-1
-0.5
-1.0
Lets show this without relying on our geometric intuition: let

Hence,
b1
b2
"
n
2
1
and
o
4
geometrically.
2

4
2
= 2 1 . Hence, the span
2
2
1
b1
2 4
1
0 1 b2 2 b1
is a linear combination of
(c) Describe span

Note that
2 4 b1
1 1 b2
b1
b2
any vector.
is consistent

2 4 b1
1 2 b2
is in the span of
2
1
"
and
So the span consists of vectors
Armin Straub
[email protected]
b1
b2
4
1
is as in (a).
Again, we can also see this after row reduction: let

"
b1
2 4
1
0 0 b2 2 b1
4
2
b1
1
b
2 1
#
1
b1
b2
any vector.
is not consistent for all
b1
b2
only if b2 2 b1 = 0 (i.e. b2 = 2 b1).

#
"
= b1
1
1
2
A single (nonzero) vector always spans a line, two vectors v1, v2 usually span a plane
but it could also be just a line (if v2 = v1).
We will come back to this when we discuss dimension and linear independence.
4
2
1 , 2
1
1
Example 6. Is span
a line or a plane?
Solution. The span is a plane unless, for some ,
4
2
2 = 1 .
1
1
Looking at the first entry, = 2, but that does not work for the third entry. Hence,
there is no such . The span is a plane.
Example 7. Consider
1 2
A = 3 1 ,
0 5
8
b = 3 .
17
Is b in the plane spanned by the columns of A?

Solution. b in the plane spanned by the columns of A if and only if
1 2 8
3 1 3
0 5 17
is consistent.
To find out, we row reduce to an echelon form:
1 2 8
8
8
1 2
1 2
3 1 3 0 5 21 0 5 21
0 5 17
0 5 17
0 0 4
From the last row, we see that the system is inconsistent. Hence, b is not in the plane
spanned by the columns of A.
Armin Straub
[email protected]
Conclusion and summary

The span of vectors a1, a2,
, am is the set of all their linear combinations.
Some vector b is in span{a1, a2,

, am } if and only if there is a solution to the
linear system with augmented matrix
| |
| |
a1 a2 am b .
| |
| |
Each solution corresponds to the weights in a linear combination of the a1, a2,
,
am which gives b.
This gives a second geometric way to think of linear systems!
Armin Straub
[email protected]
Pre-lecture:
the goal for today
We wish to write linear systems simply as Ax = b.

For instance:
2x1 +3x2 = b1
3x1 +x2 = b2
2 3
3 1

x1
b1
=
x2
b2
Why?
Its concise.
The compactness also sparks associations and ideas!
For instance, can we solve by dividing by A? x = A1b?
If Ax = b and Ay = 0, then A(x + y) = b.
Leads to matrix calculus and deeper understanding.
multiplying, inverting, or factoring matrices
Matrix operations
Basic notation
We will use the following notations for an m n matrix A (m rows, n columns).
In terms of the columns of A:
A = [ a1 a2
In terms of the entries of A:
a1,1 a1,2
a2,2
a
A = 2,1

am,1 am,2
| |
an ] = a1 a2
| |
a1,n
a2,n
,
am,n
ai,j =
|
an
|
entry in
i-th row,
j-th column
Matrices, just like vectors, are added and scaled componentwise.

Example 1.

1 0
2 3
3 3
(a)
+
=
5 2
3 1
8 3
Armin Straub
[email protected]
(b) 7
2 3
3 1

14 21
=
21 7
Matrix times vector

Recall that (x1, x2,
, xn) solves the linear system with augmented matrix
| |
[ A b ] = a1 a2
| |
if and only if
| |
an b
| |
x1a1 + x2a2 +
+ xnan = b.
It is therefore natural to define the product of matrix times vector as
x1
Ax = x1a1 + x2a2 +
+ xnan ,
x = .
xn
The system of linear equations with augmented matrix [ A b ] can be written in
matrix form compactly as Ax = b.
The product of a matrix A with a vector x is a linear combination of the columns of
A with weights given by the entries of x.
Example 2.

1 0
(a)
5 2

2 3
(b)
3 1

2 3
(c)
3 1

1
0
2
=2
+1
=
5
2
12

0
3
=
1
1

x1
2
3
2x1 + 3x2
= x1
+ x2
=
x2
3
1
3x1 + x2
2
1
This illustrates that linear systems can be simply expressed as Ax = b:

2x1 +3x2 = b1
2 3
x1
b1
=
3x1 +x2 = b2
3 1
x2
b2

2 3
5
1
(d) 3 1
= 4
1
1 1
0
Armin Straub
[email protected]
Example 3. Suppose A is m n and x is in R p. Under which condition does Ax make

sense?
We need n = p.
(G o th rough th e d efinition o f Ax to m ake su re you see w hy!)
Matrix times matrix

If B has just one column b, i.e. B = [ b ], then AB = [ Ab ].
In general, the product of matrix times matrix is given by
AB = [ Ab1 Ab2
Ab p ],
B = [ b1 b2
b p ].
Example 4.

1 0
2 3
2 3
(a)
=
5 2
1 2
12 11

1 0
2
1
0
2
=2
+1
=
because
5 2
1
5
2
12

1 0
3
1
0
3
and
.
= 3
+2
=
5 2
2
5
2
11

1 0
2 3 1
2 3 1
(b)
=
5 2
1 2 0
12 11 5
Each column of AB is a linear combination of the columns of A with weights given
by the corresponding column of B.
Remark 5. The definition of the matrix product is inevitable from the multiplication of
matrix times vector and the fact that we want AB to be defined such that (AB)x =
A(Bx).
A(Bx) = A(x1b1 + x2b2 + )
= x1Ab1 + x2Ab2 +
= (AB)x if the columns of AB are Ab1, Ab2,
Example 6. Suppose A is m n and B is p q.

(a) Under which condition does AB make sense?
We need n = p.
(G o throu gh the b oxed characterization o f AB to m ake su re you see w hy!)
(b) What are the dimensions of AB in that case?

AB is a m q matrix.
Armin Straub
[email protected]
Basic properties
Example 7.

2 3
(a)
3 1

1 0
(b)
0 1

2
=
3

2 3
2
=
3 1
3
1 0
0 1
3
1
3
1
This is the 2 2 identity matrix.

Theorem 8. Let A, B , C be matrices of appropriate size. Then:
A(BC) = (AB)C
associative
A(B + C) = AB + AC
left-distributive
(A + B)C = AC + BC
right-distributive
Example 9. However, matrix multiplication is not commutative!

2 3
1 1
2 5
=
(a)
3 1
0 1
3 4

1 1
2 3
5 4
(b)
=
0 1
3 1
3 1
Example 10. Also, a product can be zero even though none of the factors is:

2 0
0 0
0 0
=
3 0
2 1
0 0
Transpose of a matrix
Definition 11. The transpose AT of a matrix A is the matrix whose columns are
formed from the corresponding rows of A.
rows columns
Example
(a)
3
1
12.

0 T
2
3
1
1 =
0 1 4
4
x1
(b) [ x1 x2 x3 ]T = x2
x3

T
2 3
2 3
(c)
=
3 1
3 1
A matrix A is called symmetric if A = AT .
Armin Straub
[email protected]
Practice problems
True or false?
AB has as many columns as B.
AB has as many rows as B.
The following practice problem illustrates the rule (AB)T = B TAT .
Example 13. Consider the matrices
1 2
A = 0 1 ,
2 4
Compute:

1 2
1
2
(a) AB = 0 1
=
3 0
2 4

T
(b) (AB) =

1
0
2
(c) B TAT =
=
2 1 4

1 0 2
1 3
T T
(d) A B =
=
2 1 4
2 0
1 3
2 0
B=

1 2
.
3 0

Armin Straub
[email protected]
W hats that fishy sm ell?
Review: matrix multiplication

Ax is a linear combination of the columns of A with weights given by the entries
of x.

2 3
2
2
3
7
=2
+1
=
3 1
1
3
1
7
Ax = b is the matrix form of the linear system with augmented matrix [ A b ].

2 3
3 1

x1
b1
=
x2
b2
2x1 +3x2 = b1
3x1 +x2 = b2
Each column of AB is a linear combination of the columns of A with weights given

by the corresponding column of B.

2 3
3 1

2 1
7 2
=
1 0
7 3
Matrix multiplication is not commutative: usually, AB BA.
A comment on lecture notes

My personal suggestion:
before lecture: have a quick look (15min or so) at the pre-lecture notes to see where
things are going
during lecture: take a minimal amount of notes (everything on the screens will be
in the post-lecture notes) and focus on the ideas
after lecture: go through the pre-lecture notes again and fill in all the blanks by
yourself
then compare with the post-lecture notes
Since I am writing the pre-lecture notes a week ahead of time, there is usually some minor
differences to the post-lecture notes.
Armin Straub
[email protected]
Transpose of a matrix
Definition 1. The transpose AT of a matrix A is the matrix whose columns are formed
from the corresponding rows of A.
rows columns
Example
2
(a) 3
1
2.

0 T
2
3
1
1 =
0 1 4
4
x1
(b) [ x1 x2 x3 ]T = x2
x3

T
2 3
2 3
=
(c)
3 1
3 1
A matrix A is called symmetric if A = AT .
Theorem 3. Let A, B be matrices of appropriate size. Then:
(AT )T = A
(A + B)T = AT + B T
(AB)T = B TAT
(illustrated by last practice pro blem s)
Example 4. Deduce that (ABC)T = C TB TAT .

Solution. (ABC)T = ((AB) C)T = C T (AB)T = C TB TAT
Armin Straub
[email protected]
Back to matrix multiplication

Review. Each column of AB is a linear combination of the columns of A with weights
given by the corresponding column of B.
Two more ways to look at matrix multiplication
Example 5. What is the entry (AB)i,j at row i and column j?
The j-th column of AB is the vector A (col j of B).
Entry i of that is (row i of A) (col j of B). In other words:
(AB)i, j = (row i of A) (col j of B)

Use this row-column rule to compute:

2 3

2 3 6
16
3
0 1 =
1 0 1
0 3
2 0
2
0

2
2 3 6
16
1 0 1
0
3
1
0
3
3
[Can you see the rule (AB)T = B TAT from here?]

Observe the symmetry between rows and columns in this rule!
It follows that the interpretation
Each column of A B is a linear combination of the columns of A with
weights given by the corresponding column of B.
has the counterpart
Each row of AB is a linear combination of the rows of B with weights given
by the corresponding row of A.
Example 6.

1 2 3

1 0 0
1
2
3
(a)
4 5 6 =
0 0 1
7 8 9
7 8 9
Armin Straub
[email protected]
LU decomposition
Elementary matrices
Example 7.

1 0
(a)
0 1

0 1
(b)
1 0
1 0 0
(c) 0 2 0
0 0 1
1 0 0
(d) 0 1 0
3 0 1

a b
=
c d

a b
c d
=
c d
a b
a b c
a b c
d e f = 2d 2e 2f
g h i
g h i
a b c
a
b
c
d e f =
d
e
f
g h i
3a + g 3b + h 3c + i
a b
c d
Definition 8. An elementary matrix is one that is obtained by performing a single

elementary row operation on an identity matrix.
The result of an elementary row operation on A is EA
where E is an elementary matrix (namely, the one obtained by performing the same row operation
on the appropriate identity matrix).
Example
(a)
0
3
9.
0 0
1 0 0
1
1 0 0 1 0 =
1
3 0 1
0 1
1
1 0 0
1 0 0 1
We write 0 1 0 = 0 1 0 , but more on inverses soon.
3 0 1
3 0 1
Elementary matrices are invertible because elementary row operations are reversible.
1 0 0 1
1 0 0
(b) 2 1 0 = 2 1 0
0 0 1
0 0 1
1
1 0 0 1
1
(c) 0 2 0 =
2
0 0 1
1
1 0 0
1 0 0
(d) 0 0 1 = 0 0 1
0 1 0
0 1 0
Armin Straub
[email protected]
Practice problems
Example 10. Choose either column or row interpretation to see the result of the
following products.
1 2 0
1 0 0
(a) 2 1 2 0 1 0 =
0 2 1
0 1 1
1 2 0
1 0 0
(b) 0 1 0 2 1 2 =
0 2 1
0 1 1
Armin Straub
[email protected]
Review
Example 1. Elementary matrices in action:
0 0 1
a b c
g h i
(a) 0 1 0 d e f = d e f
1 0 0
g h i
a b c

a b c
1 0 0
a b c
(b) 0 1 0 d e f = d e f
7g 7h 7i
0 0 7
g h i
1 0 0
a b c
a
b
c
(c) 0 1 0 d e f =
d
e
f
3 0 1
g h i
3a + g 3b + h 3c + i
a b c
1 0 0
a + 3c b c
(d) d e f 0 1 0 = d + 3f e f
3 0 1
g h i
g + 3i h i
1 0 0
1 0 0
(e) 2 1 0 = 2 1 0
0 0 1
0 0 1
Armin Straub
[email protected]
LU decomposition, continued
Gaussian elimination revisited
Example 2. Keeping track of the elementary matrices during Gaussian elimination on A:

2 1
A=
4 6

R2 R2 2R1
1 0
2 1
2 1
EA =
=
2 1
4 6
0 8
Note that:
A=E
2 1
0 8
1 0
2 1

2 1
0 8
We factored A as the product of a lower and upper triangular matrix!

We say that A has triangular factorization.
A = LU is known as the LU decomposition of A.

L is lower triangular, U is upper triangular.
Definition 3.
lower triangular
0 0 0 0
0 0 0
0 0
Armin Straub
[email protected]
upper

triangular

m issing entries are 0
2 1 1
Example 4. Factor A = 4 6 0 as A = LU .
2 7 2
Solution. We begin with R2 R2 2R1 followed by R3 R3 + R1:
1
E1A = 2
0
E2(E1A) = 0
1
1
E3E2E1A = 0
0
0
1
0
0
1
0
0
1
1
2 1 1
0
0 4 6 0
1
2 7 2
0
2 1 1
0
0 8 2
1
2 7 2
0
2 1 1
0 0 8 2
1
0 8 3
2 1 1
= 0 8 2
0 0 1
=U
The factor L is given by:

L =E11E21E31
1
= 2 1
1
= 2 1
1
= 2 1
1 1 1
no te th at E3E2E1A = U
1
1
1
1
A=E
1
1
1
1 1
1
1 1 1
In conclusion, we found the following LU decomposition of A:
2 1 1
1
2 1 1
4 6 0 = 2 1
8 2
2 7 2
1 1 1
1
Note: The extra steps to compute L were unnecessary! The entries in L are precisely
the negatives of the ones in the elementary matrices during elimination. C an you see it?
Armin Straub
[email protected]
Once we have A = LU , it is simple to solve Ax = b.

Ax = b
L(Ux) = b
Lc = b and
Ux = c
Both of the final systems are triangular and hence easily solved:
Lc = b by forward substitution to find c, and then
Ux = c by backward substitution to find x.
Important practical point: can be quickly repeated for many different b.
2 1 1
4
Example 5. Solve 4 6 0 x = 10 .
2 7 2
3
Solution. We already found the LU decomposition A = LU :
2 1 1
1
2 1 1
4 6 0 = 2 1
8 2
2 7 2
1 1 1
1
Forward substitution to solve Lc = b for c:
1
4
2 1
c = 10
1 1 1
3
4
c = 2
3
Backward substitution to solve Ux = c for x:

1 1
4
8 2 x = 2
1
3
1
x = 1
3
Its always a good idea to do a quick check:
2 1 1
4
4 6 0 x = 10
2 7 2
3
Armin Straub
[email protected]
Triangular factors for any matrix

Can we factor any matrix A as A = LU ?
Yes, almost! Think about the process of Gaussian elimination.
In each step, we use a pivot to produce zeros below it.
The corresponding elementary matrices are lower diagonal!
The only other thing we might have to do, is a row exchange.

Namely, if we run into a zero in the position of the pivot.
All of these row exchanges can be done at the beginning!

Definition 6. A permutation matrix is one that is obtained by performing row
exchanges on an identity matrix.
1 0 0
Example 7. E = 0 0 1 is a permutation matrix.
0 1 0
EA is the matrix obtained from A by permuting the last two rows.
Theorem 8. For any matrix A there is a permutation matrix P such that PA = LU .
In other words, it might not be possible to write A as A = LU , but we only need to permute the
rows of A and the resulting matrix PA now has an LU decomposition: PA = LU .
Armin Straub
[email protected]
Practice problems
Is
2 0
0 1
upper triangular? Lower triangular?
Is
0 1
1 0
upper triangular? Lower triangular?
True or false?
A permutation matrix is one that is obtained by performing column exchanges
on an identity matrix.
Why do we care about LU decomposition if we already have Gaussian elimination?
Example 9.
5
2
1 1
Solve 4 6 0 x = 2
9
2 7 2
using the factorization we already have.
Example 10. The matrix
0 0 1
A = 1 1 0
2 1 0
cannot be written as A = LU (so it doesnt have a LU decomposition). But there is a
permutation matrix P such that PA has a LU decomposition.
0 1 0
1 1 0
Namely, let P = 0 0 1 . Then PA = 2 1 0 .
1 0 0
0 0 1
PA can now be factored as PA = LU . Do it!!
(By the way, P
0 0 1
= 0 1 0
1 0 0
Armin Straub
[email protected]
would work as well.)
Review
Elementary matrices performing row operations:
1 0 0
a b c
a
b
c
2 1 0 d e f = d 2a e 2b f 2c
0 0 1
g h i
g
h
i
Gaussian elimination on A gives an LU decomposition A = LU :
2 1 1
1
2 1 1
4 6 0 = 2 1
8 2
2 7 2
1 1 1
1
U is the echelon form, and L records the inverse row operations we did.
LU decomposition allows
1 0 0 1
1 0
= a 1
a 1 0
0 0 1
0 0
Already not so clear: a

0
us to solve Ax = b for
0
0
1
0 0 1
1 0
= a 1
1 0
b 1
ab b
many b.
0
0
1
Goal for today: invert these and any other matrices (if possible)
Armin Straub
[email protected]
The inverse of a matrix

1
Example 1. The inverse of a real number a is denoted as a1. For instance, 71 = 7 and
7 71 = 71 7 = 1.
In the context of n n matrix multiplication, the role of 1 is taken by the n n identity
matrix
In =
.
1
Definition 2. An n n matrix A is invertible if there is a matrix B such that
AB = BA = In.
In that case, B is the inverse of A and we write A1 = B.
Example 3. We already saw that elementary matrices are invertible.
1 0 0 1
1
2 1 0 = 2 1
0 0 1
1
Armin Straub
[email protected]
Note.
The inverse of a matrix is unique. Why?
S o A 1 is well-d efined.
Assume B and C are both inverses of A. Then:

C = CIn = CAB = InB = B
Do not write
A
.
B
Why?
Because it is unclear whether it should mean AB 1 or B 1A.
If AB = I, then BA = I (and so A1 = B).

Example 4. The matrix A =
0 1
0 0
N ot easy to show at this stage.
is not invertible. Why?
Solution.

Example 5. If A =
a b
c d
0 1
0 0

a b
c d

c d
1
=

0 0
1
, then

1
d
b
A1 =
ad bc c a
provided that ad bc 0.
Lets check that:

1
1
d b
a b
ad bc
0
=
= I2
c d
0
cb + ad
ad bc c a
ad bc
Note.
A 1 1 matrix [ a ] is invertible
a 0.

a b
A 2 2 matrix
is invertible
ad bc 0.
c d
We will encounter the quantities on the right again when we discuss determinants.
Armin Straub
[email protected]
Solving systems using matrix inverse

Theorem 6. Let A be invertible. Then the system Ax = b has the unique solution
x = A1b.
Proof. Multiply both sides of Ax = b with A1 (from the left!).
Example 7. Solve
7x1 +3x2 = 2
using matrix inversion.
5x1 2x2 = 1
Solution. In matrix form Ax = b, this system is

7 3
2
x=
.
5 2
1
Computing the inverse:

Recall that
a b
c d
1
7 3
5 2
1

1 2 3
2 3
=
=
5 7
1 5 7

1
d b
=
.
ad bc c a
Hence, the solution is:

x=A
Armin Straub
[email protected]
b=
2 3
5 7

2
1

7
=
17
Recipe for computing the inverse

To solve Ax = b, we do row reduction on [ A b ].
To solve AX = I, we do row reduction on [ A I ].
To compute A1:
GaussJordan method
Form the augmented matrix [ A I ].

Compute the reduced echelon form.
(i.e. G auss Jordan elim ination)
If A is invertible, the result is of the form
I A1

.
2 0 0
Example 8. Find the inverse of A = 3 0 1 , if it exists.
0 1 0
Solution. By row reduction:
[A I]
Hence, A
1
2

1
2
0 0
1 0 0
0 1 0 0 0 1
3
0 0 1 2 1 0
2 0 0 1 0 0
3 0 1 0 1 0
0 1 0 0 0 1
I A1
0 0
=
0 0 1 .
3
1 0
2
Example 9. Lets do the previous example step by step.

[A I]
I A1
2 0 0 1 0 0
3 0 1 0 1 0
0 1 0 0 0 1
>
R2R2+ 2 R1
>
R1 2 R1
R2R3
Armin Straub
[email protected]
2 0 0 1 0 0
3
0 0 1 2 1 0
0 1 0 0 0 1
1
2
0 0
1 0 0
0 1 0 0 0 1
3
0 0 1 2 1 0
Note. Here is another way to see why this algorithm works:

Each row reduction corresponds to multiplying with an elementary matrix E:
[A I]
[ E 1A E 1I ]
[ E 2E 1 A E 2 E 1 ]
So at each step:
[A I]
[ FA F ]
with F = Er E2E1
If we manage to reduce [ A I ] to [ I F ], this means

FA = I
and hence A1 = F .
Some properties of matrix inverses

Theorem 10. Suppose A and B are invertible. Then:
A1 is invertible and (A1)1 = A.
Why? Because AA1 = I
AT is invertible and (AT )1 = (A1)T .

Why? Because (A1)T AT = (AA1)T = I T = I
(R ecall th at (AB)T = B TAT .)
AB is invertible and (AB)1 = B 1A1.

Why? Because (B 1A 1)(AB) = B 1IB = B 1B = I
Armin Straub
[email protected]
Review
The inverse A1 of a matrix A is, if it exists, characterized by
AA1 = A1A = In.
a b
c d
1

1
d b
=
ad bc c a
If A is invertible, then the system Ax = b has the unique solution x = A1b.

GaussJordan method to compute A1:

bring to RREF [ A I ]
I A1
(A1)1 = A
(AT )1 = (A1)T
(AB)1 = B 1A1
Why? Because (B 1A 1)(AB) = B 1IB = B 1B = I
Further properties of matrix inverses

Theorem 1. Let A be an n n matrix. Then the following statements are equivalent:
(i.e., for a given A, th ey are eith er all true or all false)
(a) A is invertible.
(b) A is row equivalent to In.
(c) A has n pivots.
(E asy to check!)
(d) For every b, the system Ax = b has a unique solution.

Namely, x = A1b.
(e) There is a matrix B such that AB = In.
(A has a right inverse.)
(f) There is a matrix C such that CA = In.
(A has a left inverse.)
Note. Matrices that are not invertible are often called singular.
The book uses singular for n n matrices that do not have n pivots. As we just saw, it doesnt
make a difference.
Example 2. We now see at once that A =
0 1
0 0
is not invertible.
Why? Because it has only one pivot.

Armin Straub
[email protected]
Application: finite differences

Let us apply linear algebra to the boundary value problem (BVP)
d 2u
2 = f (x),
dx
0 6 x 6 1,
u(0) = u(1) = 0.
f (x) is given, and the goal is to find u(x).

Physical interpretation: models steady-state temperature distribution in a bar (u(x) is temperature
at point x) under influence of an external heat source f (x) and with ends fixed at 0 (ice cube at
the ends?).
Remark 3. Note that this simple BVP can be solved by integrating f (x) twice. We get
two constants of integration, and so we see that the boundary condition u(0) = u(1) = 0
makes the solution u(x) unique.
Of course, in the real applications the BVP would be harder. Also, f (x) might only be known at
some points, so we cannot use calculus to integrate it.
u(x)
x
We will approximate this problem as follows:
replace u(x) by its values at equally spaced points in [0, 1]
u0
0
approximate
u1
u2
2h
h
d 2u
dx2
)
3h
)
2h
)
(h
=
u(
u3
3h
u(
un
...
nh
h
(n
)
1
+
un
at these points (finite differences)
replace differential equation with linear equation at each point

solve linear problem using Gaussian elimination
Armin Straub
[email protected]
Finite differences
Finite differences for first derivative:
du u
u(x + h) u(x)
=
dx x
h
or u(x) u(x h)
h
or u(x + h) u(x h)
2h
@
@
symmetric and most accurate
Note. Recall that you can always use LHospitals rule to determine the limit of such
quantities (especially more complicated ones) as h 0.
Finite difference for second derivative:
u(x + h) 2u(x) + u(x h)
d2 u
h2
dx2
the only symmetric choice involving only u(x), u(x h)
d2 u
Question 4. Why does this approximate
as h 0?
dx2
Solution.
d2 u
dx2
du
du
(x + h) dx (x)
dx
h
u(x + h) u(x)
h
u(x) u(x h)
h
h
u(x + h) 2u(x) + u(x h)
h2
Armin Straub
[email protected]
Setting up the linear equations
u0
0
d2u
0
u1
)
(h
u2
0 6 x 6 1,
h
(2
u3
2h
Using dx2
d 2u
= f (x),
dx2
...
3h
u(2h) 2u(h) + u(0)
h2
2u1 u2 = h2 f (h)
h)
)
un
u(x + h) 2u(x) + u(x h)

,
h2
at x = h:
h
(3
u(0) = u(1) = 0.
n
u(
+
un
nh
we get:
= f (h)
(1)
u(3h) 2u(2h) + u(h)
= f (2h)
at x = 2h:
h2
2
u1 + 2u2 u3 = h f (2h)
(2)
at x = 3h:
u2 + 2u3 u4 = h2 f (3h)
(3)
u((n + 1)h) 2u(n h) + u((n 1)h)
at x = nh:
h2
un1 + 2un = h2 f (nh)
= f (nh)
(n)
1
Example 5. In the case of six divisions (n = 5, h = 6 ), we get:
2 1
1 2 1
1 2 1
1 2 1
1 2
A
Armin Straub
[email protected]
u1
u2
u3
u4
u5
x
h2 f (h)
h2 f (2h)
h2 f (3h)
h2 f (4h)
h2 f (5h)
Such a matrix is called a band matrix. As we will see next, such matrices always have
a particularly simple LU decomposition.
Gaussian elimination:
2 1
1 2 1
1 2 1
1 2 1
1 2
>
R2R2+ 2 R1
1
1
1
1
1
>
R3R3+ 3 R2
1
1
2
3
1
1
1
>
R4R4+ 4 R3
1
1
1
3
4
1
1
>
R5R5+ 5 R4
1
1
1
1
4
5
2 1
3
0 2 1
1 2 1
1 2 1
1 2
2 1
0 3 1
2
1
0
1 2 1
1 2
2 1
3
0 2 1
1
0
0
1
4
1 2
2 1
3
0 2 1
1
0
1
0
4
6
0
5
In conclusion, we have the LU decomposition:
2 1
1
2 1

1 2 1
2
=
3 1
1 2 1

3
1 2 1
4
1 2
1
4
5 1
2 1
3
2
1
4
3
4
6
5
Thats how the LU decomposition of band matrices always looks like.
Armin Straub
[email protected]
Review
Goal: solve for u(x) in the boundary value problem (BVP)
d 2u
2 = f (x),
dx
0 6 x 6 1,
u(0) = u(1) = 0.
replace u(x) by its values at equally spaced points in [0, 1]
u0
0
u1
0
d 2u
dx2
h
u(
)
u2
2
u(
h)
u3
2h
3
u(
3h
u(x + h) 2u(x) + u(x h)

h2
h)
h)
un
...
n
u(
un
nh
at these points (finite differences)
get a linear equation at each point x = h, 2h,

, nh; for n = 5, h = 6 :
1
2 1
1 2 1
1 2 1
1 2 1
1 2
u1
u2
u3
u4
u5
h f (h)
h2 f (2h)
h2 f (3h)
h2 f (4h)
h2 f (5h)
Compute the LU decomposition:
1
2 1
1
1

1 2 1
2

3 1
1 2 1
=
3
1 2 1
4
1 2
1
4
5 1
2 1
3
2
1
4
3
4
6
5
Thats how the LU decomposition of band matrices always looks like.
Armin Straub
[email protected]
LU decomposition vs matrix inverse

In many applications, we dont just solve Ax = b for a single b, but for many different
b (think millions).
Note, for instance, that in our example of steady-state temperature distribution in a bar the matrix
A is always the same (it only depends on the kind of problem), whereas the vector b models the
external heat (and thus changes for each specific instance).
Thats why the LU decomposition saves us from repeating lots of computation in

comparison with Gaussian elimination on [ A b ].
What about computing A1?
We are going to see that this is a bad idea. (It usually is.)
Example 1. When using LU decomposition to solve Ax = b, we employ forward and

backward substitution:
Ax = b
A=L U
Lc = b
and
Ux = c
Here, we have to solve, for each b,
1
1

2 1
3 1
3
4
1
4
5 1
3
2
c = b,
2 1
1
4
3
x = c
1
5
1
6
5
by forward and backward substitution.

How many operations
(additions and m ultiplications)
are needed in the n n case?
2(n 1) for Lc = b, and 1 + 2(n 1) for Ux = c.

So, roughly, a total of 4n operations.
On the other hand,
1
A1 =
6
5
4
3
2
1
4
8
6
4
2
3
6
9
6
3
2
4
6
8
4
1
2
3
4
5
How many operations are needed to compute A1b?

This time, we need roughly 2n2 additions and multiplications.
Armin Straub
[email protected]
Conclusions
Large matrices met in applications usually are not random but have some structure
(such as band matrices).
When solving linear equations, we do not (try to) compute A1.
It destroys structure in practical problems.
As a result, it can be orders of magnitude slower,
and require orders of magnitude more memory.
It is also numerically unstable.
LU decomposition can be adjusted to not have these drawbacks.
A practice problem
Example 2. Above we computed the LU decomposition for n = 5. For comparison,
here are the details for computing the inverse when n = 3.
Do it for n = 5, and appreciate just how much computation has to be done.
Invert
2 1
A = 1 2 1 .
1 2
Solution.
>
>
1
2 1 0 1 0 0
2 1
R2R2+
R1
2
1 2 1 0 1 0
3
0 2
0 1 2 0 0 1
0 1
2 1
2
R3R3+ 3 R2
0 3
2
0 0
1
1 2 0
1
R1 2 R1
2
0 1 3
2
R2 3 R2
0 0
1
3
R3 4 R3
1 0 0
2
R2R2+ 3 R3
0 1 0
1
R1R1+ 2 R2
0 0 1
>
>
1
2 1
Hence, 1 2 1 =
1 2
Armin Straub
[email protected]
3
4
1
2
1
4
1
2
1
1
2
1
4
1
2
3
4
0 1 0 0
1
1 2 1 0
2 0 0 1
0 1 0 0
1
1 2 1 0
4
1 2
1
3
3 3
1
0
0
1 2
0
3 3
1
4
3
4
1
2
1
4
1
2
1
2
1
1
2
3
4
1
4
1
2
3
4
Vector spaces and subspaces

We have already encountered vectors in Rn. Now, we discuss the general concept of
vectors.
In place of the space Rn, we think of general vector spaces.
Definition 3. A vector space is a nonempty set V of elements, called vectors, which
may be added and scaled (multiplied with real numbers).
The two operations of addition and scalar multiplication must satisfy the following
axioms for all u, v, w in V , and all scalars c, d.
(a) u + v is in V
(b) u + v = v + u
(c) (u + v) + w = u + (v + w)
(d) there is a vector (called the zero vector) 0 in V such that u + 0 = u for all u in V
(e) there is a vector u such that u + (u) = 0
(f) cu is in V
(g) c(u + v) = cu + cv
(h) (c + d)u = cu + du
(i) (cd)u = c(du)
(j) 1u = u
tl;dr
A vector space is a collection of vectors which can be added and scaled
(without leaving the space!); subject to the usual rules you would hope for.
namely: associativity, commutativity, distributivity
Example 4. Convince yourself that M22 =
n
a b
c d
Solution. In this context, the zero vector is 0 =
: a, b, c, d in R is a vector space.
0 0
0 0
Addition is componentwise:

a b
c d

e f
a+e b+ f
+
=
g h
c+ g d+h
Scaling is componentwise:

a b
ra rb
r
=
c d
rc rd
Armin Straub
[email protected]
Addition and scaling satisfy the axioms of a vector space because they are defined
component-wise and because ordinary addition and multiplication are associative, commutative, distributive and what not.
Important note: we do not use matrix multiplication here!
Note: as a vector space, M22 behaves precisely like R4; we could translate between
the two via

a

a b
b .
c
c d
d
A fancy person would say that these two vector spaces are isomorphic.
Example 5. Let Pn be the set of all polynomials of degree at most n > 0. Is Pn a
vector space?
Solution. Members of Pn are of the form
p(t) = a0 + a1t +
+ antn ,
where a0, a1,
, an are in R and t is a variable.
Pn is a vector space.
Adding two polynomials:
[a0 + a1t +
+ antn] + [b0 + b1t +
+ bntn]
= [(a0 + b0) + (a1 + b1)t +
+ (an + bn)tn]
So addition works component-wise again.
Scaling a polynomial:
r[a0 + a1t +
+ antn]
= [(ra0) + (ra1)t +
+ (ran)tn]
Scaling works component-wise as well.
Again: the vector space axioms are satisfied because addition and scaling are defined
component-wise.
As in the previous example, we see that Pn is isomorphic to Rn+1:
a0
a1
a0 + a1t +
+ antn

an
Armin Straub
[email protected]
Example 6. Let V be the set of all polynomials of degree exactly 3. Is V a vector space?
Solution. No, because V does not contain the zero polynomial p(t) = 0.
Every vector space has to have a zero vector; this is an easy necessary (but not sufficient) criterion
when thinking about whether a set is a vector space.
More generally, the sum of elements in V might not be in V :

[1 + 4t2 + t3] + [2 t + t2 t3] = [3 t + 5t2]
Armin Straub
[email protected]
Review
A vector space is a set of vectors which can be added and scaled (without leaving
the space!); subject to the usual rules.
The set of all polynomials of degree up to 2 is a vector space.
[a0 + a1t + a2t2] + [b0 + b1t + b2t2] = [(a0 + b0) + (a1 + b1)t + (a2 + b2)t2]
r[a0 + a1t + a2t2] = [(ra0) + (ra1)t + (r a2)t2]
Note how it works just like R3.

The set of all polynomials of degree exactly 2 is not a vector space.
[1 + 4t + t2] + [3 t t2] = [4 + 3t]
degree 2
degree 2
NOT degree 2
An easy test that often works is to check whether the set contains the zero vector.
(Works in the previous case.)
Example 1. Let V be the set of all functions f : R R. Is V a vector space?
Solution. Yes!
Addition of functions f and g:
(f + g)(x) = f (x) + g(x)
Note that, once more, this definition is component-wise.
Likewise for scalar multiplication.
Armin Straub
[email protected]
Subspaces
Definition 2. A subset W of a vector space V is a subspace if W is itself a vector space.
Since the rules like associativity, commutativity and distributivity still hold, we only need
to check the following:
W V is a subspace of V if
W contains the zero vector 0,
W is closed under addition,
(i.e. if u, v W then u + v W )
W is closed under scaling.
(i.e. if u W an d c R then cu W )
Note that 0 in W (first condition) follows from W closed under scaling (third condition). But it
is crucial and easy to check, so deserves its own bullet point.
Example 3. Is W = span
Solution. Yes!
W contains
a
a

a
a
+

b
b

ca
ca
0
0

=0
a+b
a+b
n
1
1
1
1
o
is in W .
is in W .
(
a
0 : a, b
b
Example 4. Is W =
a subspace of R2?
in R
a subspace of R3?
Solution. Yes!
0
W contains 0 .
0
a1 + a2
a2
a1
0 + 0 =
0
b1 + b2
b2
b1
ca
a
c 0 = 0 is in W .
cb
b
is in W .
The subspace W is isomorphic to R2 (translation:

same!
Example 5. Is W =
a
1 : a, b
b
in R

a
0 a )
b
b
but they are not the
a subspace of R3?
Solution. No! Missing 0.

Armin Straub
[email protected]
Note: W
(
0
a
= 1 + 0 : a, b
b
0
in R
is close to a vector space.
Geometrically, it is a plane, but it does not contain the origin.
Example 6. Is W =
Solution. Yes!
W contains
0
0

0
0
+

0
0

0
0
0
0

n
o
0
0
a subspace of R2?
0
0
is in W .
is in W .
o
n

x
Example 7. Is W = x + 1 : x in R a subspace of R2?
Solution. No! W does not contain
0
0
[If 0 is missing, some other things always go wrong as well.

For instance, 2
1
2
Example 8. Is W =
2
4
n
0
0
or
o
1
2
n
2
3
x
x+1
3
5
: x in R
are not in W .]
o
a subspace of R2?
[In other words, W is the set from the previous example plus the zero vector.]

Solution. No! 2
1
2
2
4
not in W .
Spans of vectors are subspaces

Review. The span of vectors v1, v2,
, vm is the set of all their linear combinations.
We denote it by span{v1, v2,
, vm }.
In other words, span{v1, v2,
, vm } is the set of all vectors of the form
c1v1 + c2v2 +
+ cmvm ,
where c1, c2,
, cm are scalars.
Theorem 9. If v1,
, vm are in a vector space V , then span{v1,
, vm } is a subspace
of V .
Why?
0 is in span{v1,
, vm }
[c1v1 +
+ cmvm] + [d1v1 +
+ dmvm]
= [(c1 + d1)v1 +
+ (cm + dm)vm]
r[c1v1 +
+ cmvm] = [(rc1)v1 +
+ (r cm)vm]
Armin Straub
[email protected]
Example 10. Is W =
n
a + 3b
2a b
o
: a, b in R a subspace of R2?
Solution. Write vectors in W in the form

a + 3b
2a b
a
2a

3b
1
3
+
=a
+b
b
2
1
to see that

1
3
W = span
,
.
2
1
By the theorem, W is a vector space. Actually, W = R2.
Example 11. Is W =
matrices?
n
a 2b
a + b 3a
o
: a, b in R a subspace of M22, the space of 2 2

a 2b
a + b 3a

1 0
0 2
+b
= a
1 3
1 0
to see that

1 0
0 2
W = span
,
.
1 3
1 0
By the theorem, W is a vector space.
Armin Straub
[email protected]
Practice problems
Example 12. Are the following sets vector spaces?
o
n

a b
: a + 3b = 0, 2a c = 1
(a) W1 =
c d
No, W1 does not contain 0.
(b) W2 =
n
a + c 2b
b + 3c
c
Yes, W2 = span
n
: a, b, c in R
1 0
0 0

0 2
1 0

1 0
3 1
o
Hence, W2 is a subspace of the vector space Mat22 of all 2 2 matrices.
(c) W3 =
n
a
b
: ab > 0
No. For instance,
3
1
o

2
4
1
3
is not in W3.
(d) W4 is the set of all polynomials p(t) such that p (2) = 1.

No. W4 does not contain the zero polynomial.
(e) W5 is the set of all polynomials p(t) such that p (2) = 0.

Yes. If p (2) = 0 and q (2) = 0, then (p + q) (2) = 0. Likewise for scaling.
Hence, W5 is a subspace of the vector space of all polynomials.
Armin Straub
[email protected]
Midterm!
Midterm 1: Thursday, 78:15pm
in 23 Psych if your last name starts with A or B
in Foellinger Auditorium if your last name starts with C, D, ..., Z
bring a picture ID and show it when turning in the exam
Review
A vector space is a set V of vectors which can be added and scaled (without leaving
the space!); subject to the usual rules.
W V is a subspace of V if it is a vector space itself; that is,
W contains the zero vector 0,
W is closed under addition,
(i.e. if u, v W th en u + v W )
W is closed under scaling.
(i.e. if u W and c R th en cu W )
span{v1,
, vm } is always a subspace of V .
n
Example 1. Is W =
matrices?
2a b 0
b
3
: a, b in R
(v1,
, vm are vectors in V )
a subspace of M22, the space of 2 2
Solution. No, W does not contain the zero vector.

Example 2. Is W =
matrices?
n
2a b 0
b
3a
: a, b in R
a subspace of M22, the space of 2 2

2a b 0
2 0
1 0
= a
+b
b
3a
0 3
1 0
to see that

2 0
1 0
W = span
,
.
0 3
1 0
Like any span, W is a vector space.
Armin Straub
[email protected]
Example 3. Are the following sets vector spaces?

o
n

a b
: a + 3b = 0, 2a c = 1
(a) W1 =
c d
No, W1 does not contain 0.
(b) W2 =
n
a + c 2b
b + 3c
c
Yes, W2 = span
n
: a, b, c in R

1 0
0 0
0 2
1 0

1 0
3 1
o
Hence, W2 is a subspace of the vector space Mat22 of all 2 2 matrices.
(c) W3 =
n
a + c 2b
b + 3c c + 7
We still have W3 =
: a, b, c in R
0 0
0 7
+ span
n
1 0
0 0

Hence, W3 is a subspace if and only if
(more complicated)

0 0
0 7
0 2
1 0

1 0
3 1
o
is in the span.
Equivalently (why?!), we have to check whether
(W e can an swer such q uestions!)

a + c 2b
0 0
=
has solutions a, b, c.
b + 3c c + 7
0 0
There is no solution (2b = 0 implies b = 0, then b + 3c = 0 implies c = 0; this contradicts

c + 7 = 0).
(d) W4 =
n
a
b
: ab > 0
No. For instance,
3
1
o

2
4
1
3
is not in W4.
(e) W5 is the set of all polynomials p(t) such that p (2) = 1.

No. W5 does not contain the zero polynomial.
(f) W6 is the set of all polynomials p(t) such that p (2) = 0.

Yes. If p (2) = 0 and q (2) = 0, then (p + q) (2) = p (2) + q (2) = 0. Likewise for scaling.
Hence, W6 is a subspace of the vector space of all polynomials.
Armin Straub
[email protected]
What we learned before vector spaces

Linear systems
Systems of equations can be written as Ax = b.
x1 2x2 = 1
x1 + 3x2 = 3

1 2
1
x=
1 3
3
Sometimes, we represent the system by its augmented matrix.

1 2 1
1 3 3
A linear system has either

no solution (such a system is called inconsistent),

echelon form contains row [ 0 ... 0 b ] with b 0
one unique solution,

system is consistent and has no free variables
infinitely many solutions.

system is consistent and has at least one free variable
We know different techniques for solving systems Ax = b.

Gaussian elimination on [ A b ]
LU decomposition A = LU
using matrix inverse, x = A1b
Armin Straub
[email protected]
Matrices and vectors

A linear combination of v1, v2,
, vm is of the form
c1v1 + c2v2 +
+ cmvm.
span{v1, v2,
, vm } is the set of all such linear combinations.
Spans are always vector spaces.
For instance, a span in R3 can be {0}, a line, a plane, or R3.
The transpose AT of a matrix A has rows and columns flipped.

2 0 T
3 1 = 2 3 1
0 1 4
1 4
(A + B)T = AT + B T
(AB)T = B TAT
An m n matrix A has m rows and n columns.
The product Ax of matrix times vector is
| |
a1 a2
| |
|
x1
an = x1a1 + x2a2 +
+ xnan.
|
xn
Different interpretations of the product of matrix times matrix:

column interpretation
a b c
1 0 0
a + 3c b c
d e f 0 1 0 = d + 3f e f
3 0 1
g h i
g + 3i h i
row interpretation

1 0 0
a b c
a
b
c
0 1 0 d e f =
d
e
f
3 0 1
g h i
3a + g 3b + h 3c + i
row-column rule
(AB)i,j = (row i of A) (col j of B)
Armin Straub
[email protected]
The inverse A1 of A is characterized by A1A = I (or AA1 = I).

1
a b
d b
=
c d
ad bc c a
Can compute A1 using GaussJordan method.
>
RREF
[A I]
(AT )1 = (A1)T
I A1
(AB)1 = B 1A1

An n n matrix A is invertible
A has n pivots
Ax = b has a unique solution
(if true for one b, then true for all b)
Gaussian elimination
Gaussian elimination can bring any matrix into an echelon form.
0
0
0
0
0
0

0
0
0
0
0
0
0
0
0
0

0
0
0
0

0
0
0
0
0
0
0
0
0

0
0

0
It proceeds by elementary row operations:

(scaling) Multiply all entries in a row by a nonzero constant.
Each elementary row operation can be encoded as multiplication with an elementary matrix.
a b c d
1 0 0
a
b
c
d
1 1 0 e f g h = e a f b g c h d
0 0 1
i j k l
i
j
k
l
We can continue row reduction to obtain the (unique) RREF.
Armin Straub
[email protected]
Using Gaussian elimination

Gaussian elimination and row reductions allow us:
solve systems of linear systems
0
3 6 4 5
3 7
9
8 8
3 9 12 6 15
1 0 2 0 24
0 1 2 0 7
4
0 0
0 1
compute the LU decomposition A = LU
x = 24 + 2x3
1
x2 = 7 + 2x3
x
free
3
x4 = 4
1
2
1 1
2 1
1
4 6 0 = 2 1
8 2
1 1 1
2 7 2
1
compute the inverse of a matrix
2 0 0
to find 3 0 1
0 1 0
0 0
=
0 0 1 , we use GaussJordan:
3
1 0
2
1
1 0 0 2 0 0
2 0 0 1 0 0
3 0 1 0 1 0 R R E F 0 1 0 0 0 1
3
0 1 0 0 0 1
0 0 1 2 1 0
1
2
>
determine whether a vector is a linear combination of other vectors
1
2
3
is a linear combination of
the system corresponding to

(Each solution
x1
x2
Armin Straub
[email protected]
1
1
1
and
1 1 1
1 2 2
1 0 3
1
2
0
if and only if
is consistent.
gives a linear combination
1
1
1
2 = x1 1 + x2 2 .)
3
1
0
Organizational
Interested in joining class committee?
meet 3 times to discuss ideas you may have for improving class
Next: bases, dimension and such
http://abstrusegoose.com/235
Armin Straub
[email protected]
Solving Ax = 0 and Ax = b
Column spaces
Definition 1. The column space Col(A) of a matrix A is the span of the columns of A.
If A = [ a1
an ], then Col(A) = span{a1,

, an }.
In other words, b is in Col(A) if and only if Ax = b has a solution.

Why? Because Ax= x1a1 +
+ xnan is the linear combination of columns of A with coefficients
given by x.
If A is m n, then Col(A) is a subspace of Rm.

Why? Because any span is a space.
Example 2. Find a matrix A such that W = Col(A) where
2x y
W=
: x, y in R .
3y
7x + y
Solution. Note that
2x y
2
1
3y = x 0 + y 3 .
7x + y
7
1
Hence,
1
2 1
2
W = span 0 , 3 = Col 0 3 .
7
1
7 1
Armin Straub
[email protected]
Null spaces
Definition 3. The null space of a matrix A is
Nul(A) = {x : Ax = 0}.
In other words, if A is m n, then its null space consists of those vectors x R n which solve the
homogeneous equation Ax = 0.
Theorem 4. If A is m n, then Nul(A) is a subspace of Rn.

Proof. We check that Nul(A) satisfies the conditions of a subspace:
Nul(A) contains 0 because A0 = 0.
If Ax = 0 and Ay = 0, then A(x + y) = Ax + Ay = 0.
Hence, Nul(A) is closed under addition.
If Ax = 0, then A(cx) = cAx = 0.

Hence, Nul(A) is closed under scalar multiplication.
Armin Straub
[email protected]
Solving Ax = 0 yields an explicit description of Nul(A).

By that we mean a description as the span of some vectors.
Example 5. Find an explicit description of Nul(A) where

A=

3 6 6 3 9
.
6 12 13 0 3
Solution.

3 6 6 3 9
6 12 13 0 3
>
R2R22R1
>
R1R12R2
>
1
R1 3 R1
3
0

1
0

1
0
6
0
2
0
2
0
6
1
2
1
0
1

3
9
6 15

1
3
6 15

13 33
6 15
From the RREF we read off a parametric description of the solutions x to Ax = 0. Note
that x2, x4, x5 are free.
x =
x1
x2
x3
x4
x5
2x2 13x4 33x5
x2
=
6x4 + 15x5
x4
x
2
13
33
1
0
0
= x2 0 + x4 6 + x5
15
0
1
0
0
0
1
In other words,
Nul(A) = span
2
1
0
0
0
13
0
6
1
0
33
0
15
0
1
Note. The number of vectors in the spanning set for Nul(A) as derived above (which
is as small as possible) equals the number of free variables in Ax = 0.
Armin Straub
[email protected]
Another look at solutions to Ax = b

Theorem 6. Let x p be a solution of the equation Ax = b.
Then every solution to Ax = b is of the form x = x p + xn, where xn is a solution to
the homogeneous equation Ax = 0.
In other words, {x : Ax = b} = x p + Nul(A).
We often call x p a particular solution.
The theorem then says that every solution to Ax = b is the sum of a fixed chosen particular
solution and some solution to Ax = 0.
Proof. Let x be another solution to Ax = b.

We need to show that xn = x x p is in Nul(A).
A(x x p) = Ax Ax p = b b = 0

1 3 3 2
1
Example 7. Let A = 2 6 9 7 and b = 5 .

1 3 3 4
5
Using the RREF, find a
1 3 3
2 6 9
1 3 3
parametric description of
R2R22R1
1
2 1
R3R3+R1
0
7 5
0
4 5
R3R32R2
1
0
0
1
R2 3 R2
1
0
0
R1R13R2
1
0
0
>
>
>
>
the solutions to Ax = b:
3 3 2 1
0 3 3 3
0 6 6 6
3 3 2 1
0 3 3 3
0 0 0 0
3 3 2 1
0 1 1 1
0 0 0 0
3 0 1 2
0 1 1 1
0 0 0 0
Every solution to Ax = b is therefore of the form:

x1
2 3x2 + x4
x2
x2

x =
x 3 =
1 x4
x4
x4
Armin Straub
[email protected]
2
3
0
1
=
1 + x2 0
0
0
+ x4 0
1
elements of Nul(A)
xp
We can see nicely how every solution is the sum of a particular solution x p and solutions
to Ax = 0.
Note. A convenient way to just find a particular solution is to set all free variables to
zero (here, x2 = 0 and x4 = 0).
Of course, any other choice for the free variables will
in a particular solution.
result
For instance, x2 = 1 and x4 = 1 we would get
xp =
4
1
.
0
1
Practice problems
True or false?
The solutions to the equation Ax = b form a vector space.
No, with the only exception of b = 0.
The solutions to the equation Ax = 0 form a vector space.

Yes. This is the null space Nul(A).
Example 8. Is the given set W a vector space?

If possible, express W as the column or null space of some matrix A.
(a) W =
(b) W =
(c) W =
x
y
z
x
y
z
: 5x = y + 2z
: 5x 1 = y + 2z
x
y
x+ y
: x, y in R
Example 9. Find an explicit description of Nul(A) where

A=
Armin Straub
[email protected]

1 3 5 0
.
0 1 4 2
Review
Every solution to Ax = b is the sum of a fixed chosen particular solution and some
solution to Ax = 0.
1
3 3 2
1
For instance, let A = 2
6 9 7 and b = 5 .
1 3 3 4
5
Every solution to Ax = b is of the form:

2 3x2 + x4
x1
x2
x2

x =
x3 =
1 x4
x4
x4
Is span
1
1
1
1 , 2 , 1
1
3
3
2
0
1 + x2
0

xp
1
+ x4
0
0
1
0
1
1
elem ents of N u l(A)
equal to R3?
Linear independence
Review.
span{v1, v2,
, vm } is the set of all linear combinations
c1v1 + c2v2 +
+ cmvm.
span{v1, v2,
, vm } is a vector space.
)
1
1
1
1 , 2 , 1
3
3
1
Example 1. Is span
equal to R3?
Solution. Recall that the span is equal to
1 1 1
1 2 1 x : x in R3 .
1 3 3
Hence, the span is equal to R3 if and only if the system with augmented matrix
is consistent for all b1, b2, b3.

Armin Straub
[email protected]
1 1 1 b1
1 2 1 b2
1 3 3 b3
Gaussian elimination:
1 1 1 b1
1 2 1 b2
1 3 3 b3
1
0
0
1
0
0
b1
1 1
1 2 b2 b1
2 4 b3 b1
1 1
b1
1 2
b2 b1
0 0 b3 2b2 + b1
The system is only consistent if b3 2b2 + b1 = 0.

Hence, the span does not equal all of R3.
)
1
1
1
1 , 2 , 1
3
3
1
What went wrong?
span
Well, the three vectors in the span satisfy

1
1
1
1 = 3 1 + 2 2 .
3
1
3
( )
)
1
1
1
1
1
1 , 2 , 1 = span 1 , 2 .
3
1
3
3
1
Hence, span
We are going to say that the three vectors are linearly dependent because they
satisfy

1
1
1
3 1 + 2 2 1 = 0.
3
3
1
Definition 2. Vectors v1,
, v p are said to be linearly independent if the equation
x1v1 + x2v2 +
+ x pv p = 0
has only the trivial solution (namely, x1 = x2 =
= x p = 0).
Likewise, v1,
, vp are said to be linearly dependent if there exist coefficients x1,
, x p, not all zero,
such that
x1v1 + x2v2 +
+ x pvp = 0.
Armin Straub
[email protected]
Example 3.
Are the
1
1
1
vectors 1 , 2 , 1
1
3
3
independent?
If possible, find a linear dependence relation among them.

Solution. We need to check whether

1
x1 1 + x2
1
the equation

1
1
0
= 0
2 + x3 1
3
3
0
has more than the trivial solution.

In other words, the three vectors are independent if and only if the system
1 1 1
1 2 1 x = 0
1 3 3
has no free variables.
To find out, we reduce the matrix to echelon form:
1 1 1
1 1 1
1 1 1
1 2 1 0 1 2 0 1 2
1 3 3
0 2 4
0 0 0
Since there is a column without pivot, we do have a free variable.
Hence, the three vectors are not linearly independent.
To find a linear dependence relation, we solve this system.
Initial steps of Gaussian elimination are as before:
1 0 3 0
1 1 1 0
1 1 1 0
1 2 1 0
0 1 2 0 0 1 2 0
1 3 3 0
0 0 0 0
0 0 0 0
x3 is free. x2 = 2x3, and x1 = 3x3. Hence, for any x3,

1
1
1
0
3x3 1 2x3 2 + x3 1 = 0 .
3
3
1
0
Since we are only interested in one linear combination, we can set, say, x3 = 1:

1
1
1
0
3 1 2 2 + 1
= 0
1
3
3
0
Armin Straub
[email protected]
Linear independence of matrix columns

Note that a linear dependence relation, such as

1
1
1
3 1 2 2 + 1 = 0,
1
3
3
can be written in matrix form as
1 1 1
3
1 2 1 2 = 0.
1 3 3
1
Hence, each linear dependence relation among the columns of a matrix A corresponds to a nontrivial solution to Ax = 0.
Theorem 4. Let A be an m n matrix.
The columns of A are linearly independent.
Ax = 0 has only the solution x = 0.
Nul(A) = {0}
A has n pivots.

Example 5. Are the
1
1
1
vectors 1 , 2 , 2
1
3
3
Solution. Put the vectors in a matrix, and

1 1 1
1 1
1 2 2 0 1
1 3 3
0 2
(one in each colum n)
independent?
produce an echelon form:

1
1 1 1
3 0 1 3
4
0 0 2
Since each column contains a pivot, the three vectors are independent.
Example 6. (once again, short version)
Are the
1
1
1
vectors 1 , 2 , 1
1
3
3
independent?
Solution. Put the vectors in a matrix, and

1 1 1
1 1
1 2 1 0 1
1 3 3
0 2
produce an echelon form:

1
1 1 1
2 0 1 2
4
0 0 0
Since the last column does not contain a pivot, the three vectors are linearly dependent.
Armin Straub
[email protected]
Special cases
A set of a single nonzero vector {v1} is always linearly independent.
Why? Because x1v1 = 0 only for x1 = 0.
A set of two vectors {v1, v2} is linearly independent if and only if neither of the
vectors is a multiple of the other.
Why? Because if x1v1 + x2v2 = 0 with, say, x2 0, then v2 = x1 v1.
x
A set of vectors {v1,

, v p } containing the zero vector is linearly dependent.
Why? Because if, say, v1 = 0, then v1 + 0v2 +
+ 0vp = 0.
If a set contains more vectors than there are entries in each vector, then the set is
linearly dependent. In other words:
Any set {v1,
, v p } of vectors in Rn is linearly dependent if p > n.
Why?
Let A be the matrix with columns v1,

, vp. This is a n p matrix.
The columns are linearly independent if and only if each column contains a pivot.
If p > n, then the matrix can have at most n pivots.
Thus not all p columns can contain a pivot.
In other words, the columns have to be linearly dependent.
Example 7. With the least amount of work possible, decide which of the following sets
of vectors are linearly independent.
( )
(a)
3
9
2 , 6
1
4
Linearly independent, because the two vectors are not multiples of each other.
(b)
)
3
2
1
Linearly independent, because it is a single nonzero vector.
(c) columns
1 2 3 4
of 5 6 7 8
9 8 7 6
Linearly dependent, because these are more than 3 (namely, 4) vectors in R 3.
(d)
0
9
3
2 , 6 , 0
0
4
1
Linearly dependent, because the set includes the zero vector.
Armin Straub
[email protected]
Review
Vectors v1,
, v p are linearly dependent if
x1v1 + x2v2 +
+ x pv p = 0,
and not all the coefficients are zero.
The columns of A are linearly independent

each column of A contains a pivot.
Are the
1
1
1
vectors 1 , 2 , 1
1
3
3
independent?
1 1 1
1 2 1
1 3 3
1 1 1
0 1 2
0 2 4
1 1 1
0 1 2
0 0 0
So: no, they are dependent! (Coeffs x3 = 1, x2 = 2, x1 = 3)

Any set of 11 vectors in R10 is linearly dependent.
A basis of a vector space

Definition 1. A set of vectors {v1,
, v p } in V is a basis of V if
V = span{v1,
, v p }, and
the vectors v1,

, v p are linearly independent.
In other words, {v1,
, vp } in V is a basis of V if and only if every vector w in V can be uniquely
expressed as w = c1v1 +
+ c pvp.
Example 2. Let
1
e1 = 0 ,
0
0
e2 = 1 ,
0
0
e3 = 0 .
1
Show that {e1, e2, e3} is a basis of R3.
It is called the standard basis.
Solution.
Clearly, span{e1, e2, e3} = R3.
{e1, e2, e3} are independent, because
has a pivot in each column.
1 0 0
0 1 0
0 0 1
Definition 3. V is said to have dimension p if it has a basis consisting of p vectors.

Armin Straub
[email protected]
This definition makes sense because if V has a basis of p vectors, then every basis of V has p vectors.
Why? (Think of V = R 3.)
A basis of R 3 cannot have more than 3 vectors, because any set of 4 or more vectors in R 3 is linearly
dependent.
A basis of R 3 cannot have less than 3 vectors, because 2 vectors span at most a plane (challenge:
can you think of an argument that is more rigorous?).
Example 4. R3 has dimension 3.

Indeed, the standard
1
0
0
basis 0 , 1 , 0
0
0
1
has three elements.
Likewise, Rn has dimension n.
Example 5. Not all vector spaces have a finite basis. For instance, the vector space of
all polynomials has infinite dimension.
Its standard basis is 1, t, t2, t3,
This is indeed a basis, because any polynomial can be written as a unique linear combination:
p(t) = a0 + a1t +
+ antn for some n.
Recall that vectors in V form a basis of V if they span V and if they are linearly
independent. If we know the dimension of V , we only need to check one of these two
conditions:
Theorem 6. Suppose that V has dimension d.
A set of d vectors in V are a basis if they span V .
A set of d vectors in V are a basis if they are linearly independent.
Why?
If the d vectors were not independent, then d 1 of them would still span V . In the end, we
would find a basis of less than d vectors.
If the d vectors would not span V , then we could add another vector to the set and have d + 1
independent ones.
Example 7. Are the following sets a basis for R3?

( )
(a)
0
1
2 , 1
1
0
No, the set has less than 3 elements.
(b)
1
1
0
1
2 , 1 , 0 , 2
0
3
1
0
No, the set has more than 3 elements.

Armin Straub
[email protected]
(c)
1
0
1
2 , 1 , 0
3
1
0
The set has 3 elements.
1 0
2 1
0 1
Hence, it is

1
1
0 0
0
3
a basis if and only if the vectors are independent.

1 0 1
0 1
1 2 0 1 2
0 0 5
1 3
Since each column contains a pivot, the three vectors are independent.
Hence, this is a basis of R 3.
Example 8. Let P2 be the space of polynomials of degree at most 2.

What is the dimension of P2?
Is {t, 1 t, 1 + t t2} a basis of P2?
Solution.
The standard basis for P2 is {1, t, t2}.
This is indeed a basis because every polynomial
a0 + a1t + a2t 2
can clearly be written as a linear combination of 1, t, t2 in a unique way.
Hence, P2 has dimension 3.
The set {t, 1 t, 1 + t t2} has 3 elements. Hence, it is a basis if and only if the
three polynomials are linearly independent.
We need to check whether
x1t + x2(1 t) + x3(1 + t t2) = 0
(x2 +x3)+(x1 x2 +x3)tx3t2
has only the trivial solution x1 = x2 = x3 = 0.

We get the equations
x2 + x3 = 0
x1 x2 + x3 = 0
x3 = 0
which clearly only have the trivial solution. (If you dont see it, solve the system!)
Hence, {t, 1 t, 1 + t t2} is a basis of P2.
Armin Straub
[email protected]
Shrinking and expanding sets of vectors

We can find a basis for V = span{v1,
, v p } by discarding, if necessary, some of the
vectors in the spanning set.
Example 9. Produce a basis of R2 from the vectors

1
2
1
v1 =
, v2 =
, v3 =
.
2
4
1
Solution. Three vectors in R2 have to be linearly dependent.
Here, we notice that v2 = 2v1.
The remaining vectors {v1, v3} are a basis of R2, because the two vectors are clearly
independent.
Checking our understanding

Example 10. Subspaces of R3 can have dimension 0, 1, 2, 3.
The only 0-dimensional subspace is {0}.
A 1-dimensional subspace is of the form span{v } where v 0.
These subspaces are lines through the origin.
A 2-dimensional subspace is of the form span{v, w} where v and w are not

multiples of each other.
These subspaces are planes through the origin.
The only 3-dimensional subspace is R3 itself.

True or false?
Suppose that V has dimension n. Then any set in V containing more than n vectors
must be linearly dependent.
Thats correct.
The space Pn of polynomials of degree at most n has dimension n + 1.

True, as well. A basis is {1, t, t2,
, tn }.
The vector space of functions f : R R is infinite-dimensional.

Yes. A still-infinite-dimensional subspace are the polynomials.
Consider V = span{v1,
, v p }. If one of the vectors, say vk, in the spanning set is
a linear combination of the remaining ones, then the remaining vectors still span V .
True, vk is not adding anything new.
Armin Straub
[email protected]
Review
{v1,
, v p } is a basis of V if the vectors
span V , and
are independent.
The dimension of V is the number of elements in a basis.
The columns of A are linearly independent
each column of A contains a pivot.
Warmup
Example 1. Find a basis and the dimension of
a
+
b
+
2c
2a
+
2b
+
4c
+
d
W=
: a, b, c, d real.
b+c+d
3a + 3c + d
Solution.
First, note that

1
1

2 2
W = span
,
0 1
3
0
2
4
,
1
3
0
1
,
1
1
Is dim W = 4? No, because the third vector is the sum of the

Suppose we did not notice
1 1 2 0
1 1 2 0
2 2 4 1 0 0 0 1

A =
0 1 1 1 0 1 1 1
3 0 3 1
0 3 3 1
1 1 2 0
0 1 1 1
0 3 3 1
0 0 0 1

1 1 2 0
1 1
0 1 1 1 0 1

0 0 0 4 0 0
0 0 0 1
0 0
first two.
2
1
0
0
0
1
4
0
Not a pivot in every column, hence the 4 vectors are dependent.
Armin Straub
[email protected]
[Not necessary here, but:

To get a relation, solve Ax = 0. Set free variable x3 = 1.
Then x4 = 0, x2 = x3 = 1 and x1 = x2 2x3 = 1. The relation is

1
1
2
2
0 1
3
0

2
0
4 1
+ + 0
1 1
3
1
= 0.
Precisely, what we noticed to begin with.]
Hence, a basis for W
1
1
2 2
is 0 , 1
0
3
0
1
,
1
1
and dim W = 3.
It follows from the echelon form that these vectors are independent.
Every set of linearly independent vectors can be extended to a basis.

In other words, let {v1,
, vp } be linearly independent vectors in V . If V has dimension d, then we
can find vectors vp+1,
, vd such that {v1,
, vd } is a basis of V .
Example 2. Consider

1
1
H = span
0 , 1 .
0
1
Give a basis for H. What is the dimension of H?

Extend the basis of H to a basis of R3.
Solution.
The vectors are independent. By definition, they span H.
( )
Therefore,
1
1
0 , 1
1
0
is a basis for H.
In particular, dim H = 2.
( )
1
1
0 , 1
1
0
is not a basis for R3. Why?
to contain 3 vectors.
Because a basis for R3 needs
Or, because, for instance,
0
0
1
is not in H.
So: just add this (or any other) missing vector!

( )
By construction,
0
1
1
0 , 1 , 0
1
1
0
is independent.
Hence, this automatically is a basis of R3.
Armin Straub
[email protected]
Bases for column and null spaces

Bases for null spaces
To find a basis for Nul(A):
find the parametric form of the solutions to Ax = 0,
express solutions x as a linear combination of vectors with the free variables as
coefficients;
these vectors form a basis of Nul(A).
Example 3. Find a basis for Nul(A) with

3 6 6 3 9
A=
.
6 12 15 0 3
Solution.

3 6 6 3 9
6 12 15 0 3
The solutions toAx = 0 are:
x =
2x2 5x4 13x5

x2
2x4 + 5x5
x4
x5
Hence, Nul(A) = span
2
1
0
0
0
5
0
2
1
0
3
0

1
0

1
0
= x2
13
0
5
0
1
These vectors are clearly independent.
If you dont see it, do compute an echelon form!

6 3
9
3 6 15

2 1 3
1 2 5

0 5 13
1 2 5
6
0
2
0
2
0
2
1
0
0
0
+ x4
5
0
2
1
0
+ x5
13
0
5
0
1
(p erm ute first and th ird row to the b ottom )
Better yet: note that the first vector corresponds to the solution with x2 = 1 and the other free
variables x4 = 0, x5 = 0. The second vector corresponds to the solution with x4 = 1 and the other
free variables x2 = 0, x5 = 0. The third vector
Hence,
2
1
0
0
0
5
0
2
1
0
Armin Straub
[email protected]
13
0
5
0
1
is a basis for Nul(A).
Bases for column spaces

Recall that the columns of A are independent

Ax = 0 has only the trivial solution (namely, x = 0),
A has no free variables.
A basis for Col(A) is given by the pivot columns of A.
Example 4. Find a basis for Col(A) with
1
2
A =
3
4
Solution.
1
2
3
4
2 0 4
4 1 3
6 2 22
8 0 16
2 0 4
4 1 3
6 2 22
8 0 16
1
0
0
0
1
0
0
0
2 0 4
0 1 5
0 2 10
0 0 0
2 0 4
0 1 5
0 0 0
0 0 0
The pivot columns are the first and third.

0
1
.
Hence, a basis for Col(A) is 23 , 1
Warning: For the basis of Col(A), you have to take the columns of A, not the columns
of an echelon form.
Row operations do not preserve the column space.
[For instance,
1
0
>
R1R2
Armin Straub
[email protected]
0
1
have different column spaces (of the same dimension).]
Review
{v1,
, v p } is a basis of V if the vectors span V and are independent.
To obtain a basis for Nul(A), solve Ax = 0:
>
1 2 0 5
0 0 1 2
2x2 5x4
x2
= x2
x =
2x4
x4
1
+ x4
0
0
5
0
2
1
3 6 6 3
6 12 15 0
RREF
Hence,
5
2
1 0
,
0 2
1
0
form a basis for Nul(A).
To obtain a basis for Col(A), take the pivot columns of A.
2 0 4
4 1 3
6 2 22
8 0 16
1
2
3
4
Hence,
0
1
2 1
3 , 2
0
4
>
1
0
0
0
2 0
4
0 1 5
0 0
0
0 0
0
form a basis for Col(A).
Row operations do not preserve the column space.

For instance,
1
0
>
R1R2
0
1
On the other hand: row operations do preserve the null space.

Why? Recall why/that we can operate on rows to solve systems like Ax = 0!
Dimension of Col(A) and Nul(A)

Definition 1. The rank of a matrix A is the number of its pivots.
Theorem 2. Let A be an m n matrix of rank r. Then:
dim Col(A) = r
Why? A basis for Col(A) is given by the pivot columns of A.
dim Nul(A) = n r is the number of free variables of A

Why? In our recipe for a basis for Nul(A), each free variable corresponds to an element in the
basis.
dim Col(A) + dim Nul(A) = n

Why? Each of the n columns either contains a pivot or corresponds to a free variable.
Armin Straub
[email protected]
The four fundamental subspaces

Row space and left null space
Definition 3.
The row space of A is the column space of AT .
Col(AT ) is spanned by the columns of AT and these are the rows of A.
The left null space of A is the null space of AT .
Why left? A vector x is in Nul(AT ) if and only if ATx = 0.

Note that ATx = 0
(ATx)T = xTA = 0T .
Hence, x is in Nul(AT ) if and only if xTA = 0.
Example 4. Find a basis for Col(A) and Col(AT ) where
1
2
A =
3
4
2 0 4
4 1 3
6 2 22
8 0 16
Solution. We know what to do for Col(A) from an echelon form of A, and we could
likewise handle Col(AT ) from an echelon form of AT .
But wait!
Instead of doing twice the work, we only need an echelon form of A:
1
2
3
4
Armin Straub
[email protected]
2 0 4
4 1 3
6 2 22
8 0 16
1
0
0
0
2 0 4
0 1 5
0 2 10
0 0 0
1
0
0
0
2 0 4
0 1 5
= B
0 0 0
0 0 0
Hence, the rank of A is 2.

A basis for Col(A)
0
1
2 1
is 3 , 2
0
4
Recall that Col(A) Col(B). Thats because we performed row operations.

However, the row spaces are the same! Col(AT ) = Col(B T )
The row space is preserved by elementary row operations.
In particular: a basis for Col(AT ) is given
Theorem 5.
0
1
2 0
by 0 , 1
5
4
(Fundamental Theorem of Linear Algebra, Part I)
Let A be an m n matrix of rank r.

dim Col(A) = r
(subspace of R m)
dim Col(AT ) = r
(subspace of R n)
dim Nul(A) = n r
(subspace of R n)
dim Nul(AT ) = m r
(# of free variables of A)
(subspace of R m)
In particular:
The column and row space always have the same dimension!
In other words, A and AT have the same rank.
[i.e. sam e num b er of pivots]
Easy to see for a matrix in echelon form
but not obvious for a random matrix.
Armin Straub
[email protected]
2 1 3 0
0 0 1 2 ,
0 0 0 7
Linear transformations
Throughout, V and W are vector spaces.
Definition 6. A map T : V W is a linear transformation if
T (cx + dy) = cT (x) + dT (y)
for all x, y in V and all c, d in R.
In other words, a linear transformation respects addition and scaling:

T (x + y) = T (x) + T (y)
T (cx) = cT (x)
It also sends the zero vector in V to the zero vector in W :
T (0) = 0
[because T (0) = T (0 0) = 0 T (0) = 0]
Example 7. Let A be an m n matrix.

Then the map T (x) = Ax is a linear transformation T : Rn Rm.
Why?
Because matrix multiplication is linear:
A(cx + dy) = cAx + dAy
The LHS is T (cx + dy) and the RHS is cT (x) + dT (y).
Example 8. Let Pn be the vector space of all polynomials of degree at most n. Consider
the map T : Pn Pn1 given by
T (p(t)) =
d
p(t).
dt
This map is linear! Why?

Because differentiation is linear:
d
d
d
[ap(t) + bq(t)] = a dt p(t) + b dt q(t)
dt
The LHS is T (ap(t) + bq(t)) and the RHS is aT (p(t)) + bT (q(t)).
Representing linear maps by matrices

Let x1,
, xn be a basis for V .
A linear map T : V W is determined by the values T (x1),

, T (xn).
Armin Straub
[email protected]
Why?
Take any v in V .
It can be written as v = c1x1 +

+ cnxn because {x1,
, xn } is a basis and hence spans V .
Hence, by the linearity of T ,
T (v) = T (c1x1 +
+ cnx) = c1T (x1) +
+ cnT (xn).
Definition 9. (From linear maps to matrices)

Let x1,
, xn be a basis for V , and y1,
, ym a basis for W .
The matrix representing T with respect to these bases
has n columns (one for each of the x j ),
the j-th column has m entries a1,j ,
, am, j determined by
T (x j ) = a1, jy1 +
+ am,jym.
Example 10. Let V = R2 and W = R3. Let T be the linear map such that
T

1
0

1
= 2 ,
3

0
1

4
= 0 .
7
What is the matrix A representing T with respect to the standard bases?

Solution. The standard bases are

1
0
x1

0
1
for R2,
x2
1
0
0
0 , 1 , 0
0
0
1
y1
y2
for R3.
y3

0
0
1
1
T (x1) = 2 = 1 0 + 2 1 + 3 0
1
0
0
3
= 1y1 + 2y
3
2 + 3y
1
A = 2
3

4
T (x2) = 0 = 4y1 + 0y2 + 7y3
7
1 4
A = 2 0
3 7

Armin Straub
[email protected]
(We did not have time yet to discuss the next example in class, but it will be helpful if
your discussion section already meets Tuesdays.)
Example 11. As in the previous example, let V = R2 and W = R3. Let T be the (same)
linear map such that
T

1
0

1
= 2 ,
3

0
1
4
= 0 .
7
What is the matrix B representing T with respect to the following bases?

1
1
x1

1
2
x2
for R2,
1
0
0
1 , 1 , 0
1
0
1
y1
y2
for R3.
y3
Solution. This time:

T (x1)
T (x2)

1
1
0
=T
=T
+T
1
0
1

1
4
5
= 2 + 0 = 2
3
10
7

0
0
1
= 5 1 3 1 + 5 0
1
0
1
5
= 3
5

1
1
0
=T
= T
+ 2T
2
0
1

7
4
1
= 2 + 2 0 = 2
11
7
3

0
0
1
= 7 1 9 1 + 4 0
1
0
1
5
7
= 3 9
5
4
can you see it?

otherw ise: d o it!
Tedious, even in this simple example! (But we can certainly do it.)
A matrix representing T encodes in column j the coefficients of T (x j ) expressed as

a linear combination of y1,
, ym.
Armin Straub
[email protected]
Practice problems
Example 12. Suppose A =
1 2 3 4 1
2 4 7 8 1
. Find the dimensions and a basis for all four
fundamental subspaces of A.
Example 13. Suppose A is a 5 5 matrix, and that v is a vector in R5 which is not

a linear combination of the columns of A.
What can you say about the number of solutions to Ax = 0?
Solution. Stop reading, unless you have thought about the problem!
Existence of such a v means that the 5 columns of A do not span R 5.
Hence, the columns are not independent.
In other words, A has at most 4 pivots.
So, at least one free variable.
Which means that Ax = 0 has infinitely many solutions.
Armin Straub
[email protected]
A map T : V W between vector spaces is linear if
T (x + y) = T (x) + T (y)
T (cx) = cT (x)
Let A be an m n matrix.
T : Rn Rm defined by T (x) = Ax is linear.
T : Pn Pn1 defined by T (p(t)) = p (t) is linear.
The only linear maps T : R R are T (x) = x.
Recall that T (0) = 0 for linear maps.

x
Linear maps T : R2 R are of the form T
= x + y.
y

2x
For instance, T (x, y) = xy is not linear: T
2y
2T (x, y)
Example 1. Let V = R2 and W = R3. Let T be the linear map such that
T
What is T

0
4
=2

0
4

1
1
0
4

1
1

1
= 0 ,
4

1
1

1
= 2 .
0

+2

=T 2

1
1

2
4
2
1
1
1
1
+2 1
= 2T 1
+ 2T
= 0 + 4 = 4
1
1
8
0
8
Let x1,
, xn be a basis for V .
A linear map T : V W is determined by the values T (x1),

, T (xn).
Why?
Take any v in V .
Write v = c1x1 +
+ cnxn.
(Possible, because {x1,

, xn } spans V .)
By linearity of T ,
T (v) = T (c1x1 +
+ cnx) = c1T (x1) +
+ cnT (xn).
Armin Straub
[email protected]
Important geometric examples

We consider some linear maps R2 R2, which are defined by matrix multiplication,
that is, by x Ax.
In fact: all linear maps R n R m are given by x
Ax, for some matrix A.
Example 2.
The matrix A =
c 0
0 c
gives the map x
cx, i.e.
stretches every vector in R2 by the same factor c.
Example 3.
The matrix A =
gives the map
0 1
1 0

x
y
y
x
, i.e.
reflects every vector in R2 through the line y = x.
Example 4.
The matrix A =
1 0
0 0
gives the map xy x0 , i.e.
projects every vector in R2 through onto the x-axis.

Armin Straub
[email protected]
Example 5.
The matrix A =
gives the map
0 1
1 0

x
y

y
x
, i.e.
rotates every vector in R2 counter-clockwise by

90.
Representing linear maps by matrices

Definition 6. (From linear maps to matrices)
Let x1,
The matrix representing T with respect to these bases
has n columns (one for each of the x j ),
the j-th column has m entries a1,j ,

, am, j determined by
T (x j ) = a1, jy1 +
+ am,jym.
Example 7.
Recall the map T given by
x
y
y
x
(reflects every vector in R 2 through the line y = x)
Which matrix A represents T with respect to the

standard bases?
Which matrix
B represents T with respect to the

1
1
basis 1 , 1 ?
Solution.

T 10 =

T 01 =
0
1
0
1
1
0
0 1
1 0
. Hence, A =
. Hence, A =
If a linear map T : Rn Rm is represented by the matrix A with respect to the

standard bases, then T (x) = Ax.
Armin Straub
[email protected]
Matrix multiplication corresponds to function composition!

That is, if T1, T2 are represented by A1, A2, then T1(T2(x)) = (A1A2)x.

1
1

1
1
= 11 = 1 11 + 0 1
. Hence, B = 10
1

1
= 0 11 1
= 1
.
Hence,
B
=
1
1 0
0 1
Example 8. Let T : R2 R3 be the linear map such that

T

1
0

1
= 2 ,
3

0
1
4
= 0 .
7
What is the matrix B representing T with respect to the following bases?

1
1
x1

1
2
x2
for R ,

0
0
1
1 , 1 , 0
0
1
1
y1
y2
for R3.
y3
Solution. This time:

1
1
0
=T
+T
1
0
1

1
4
5
= 2 + 0 = 2
3
7
10

1
0
0
= 5 1 3 1 + 5 0
1
0
1
5
B = 3
5

1
1
0
T (x2) = T
= T
+ 2T
2
0
1

1
4
7
= 2 + 2 0 = 2
3
7
11

1
0
0
= 7 1 9 1 + 4 0
1
0
1
5
7
B = 3 9
5
4
T (x1) = T
can you see it?

otherw ise: d o it!
Tedious, even in this simple example! (But we can certainly do it.)
A matrix representing T encodes in column j the coefficients of T (x j ) expressed as

a linear combination of y1,
, ym.
Armin Straub
[email protected]
Practice problems
Example 9. Let T : R2 R2 be the map which rotates a vector counter-clockwise by
angle .
Which matrix A represents T with respect to the standard bases?
Verify that T (x) = Ax.
Solution. Only keep reading if you need a hint!
The first basis vector
1
0
gets send to
Hence, the first column of A is
Armin Straub
[email protected]
cos
sin
Review
A linear map T : V W satisfies T (cx + dy) = cT (x) + dT (y).
T : Rn Rm defined by T (x) = Ax is linear.
(A an m n m atrix)
A is the matrix representing T w.r.t. the standard bases

For instance: T (e1) = Ae1
= 1st
0
e1 =
column of A
Let x1,
The matrix representing T w.r.t. these bases encodes in column j the coefficients of T (xj )
expressed as a linear combination of y1,
, ym.
For instance: let T : R 3 R 3 be reflection through the x-y-plane, that is, (x, y, z)
z).
The matrix representing T w.r.t. the basis
T
!
1
1 =
1
!
0
1 =
0

0
0
1
1
= 1 +0 1 2 0
1
1
0
1
1
0
0
0
1 , T 0 = 0
0
1
1
0
0
1
1 , 1 , 0
1
0
1
is
(x, y,
1 0 0
0 1 0 .
2 0 1
Example 1. Let T : P3 P2 be the linear map given by

T (p(t)) =
d
p(t).
dt
What is the matrix A representing T with respect to the standard bases?

Solution. The bases are
1, t, t2, t3 for P3,
1, t, t2 for P2.
The matrix A has 4 columns and 3 rows.
Armin Straub
[email protected]

0
The first column encodes T (1) = 0 and hence is 0 .
0
1
For the second column, T (t) = 1 and hence it is 0 .
0
0
2
For the third column, T (t ) = 2t and hence it is 2 .
0
0
For the last column, T (t3) = 3t2 and hence it is 0 .
3
In conclusion, the matrix representing T is
0 1 0 0
A = 0 0 2 0 .
0 0 0 3
Note: By the way, what is the null space of A?
The null space has basis
1
0
0 .
0
The corresponding polynomial is p(t) = 1.
No surprise here: differentation kills precisely the constant polynomials.
Note: Let us differentiate 7t3 t + 3 using the matrix A.

First: 7t3 t + 3 w.r.t. standard basis:
3
0 1 0 0
1
0 0 2 0 1 = 0
0
21
0 0 0 3
7
1
0 in the standard basis
21
Armin Straub
[email protected]
3
1
0 .
7
is 1 + 21t2.
Orthogonality
The inner product and distances
Definition 2. The inner product (or dot product) of v, w in Rn:
v w = v Tw = v1w1 +
+ vnwn.
Because we can think of this as a special case of the matrix product, it satisfies the basic rules like
associativity and distributivity.
In addition: v w = w v.
Example 3. For instance,
1
1
1
2 1 = [ 1 2 3 ] 1 = 1 2 6 = 7.
2
2
3
Definition 4.
The norm (or length) of a vector v in Rn is
p
kv k =
v v = v12 +
+ vn2 .
This is the distance to the origin.
vw
The distance between points v and w in R is
dist(v, w) = kv wk.
w
Example 5. For instance, in R2,

x1 x2
x2
x1
=
,
dist
y1 y2
y2
y1
Armin Straub
[email protected]

p
= (x1 x2)2 + (y1 y2)2 .

Orthogonal vectors
Definition 6. v and w in Rn are orthogonal if
v w = 0.
How is this related to our understanding of right angles?
Pythagoras:

v and w are orthogonal

kv k2 + kwk2 = kv wk2
vw
v v + w w = (v w) (v w)
vv 2vw+ww
vw=0
Example 7. Are the following vectors orthogonal?

(a)
1
2

2
1
1
2

2
1
= 1 (2) + 2 1 = 0. So, yes, they are orthogonal.
2
1
(b) 2 , 1
1
1

1
2
2 1 = 1 (2) + 2 1 + 1 1 = 1.
1
1
Armin Straub
[email protected]
So not orthogonal.
Review
v w = v Tw = v1w1 +
+ vnwn, the inner product of v, w in Rn
p
Length of v: kv k = v v = v12 +
+ vn2
Distance between points v and w: kv wk
v and w in Rn are orthogonal if v w = 0.
This simple criterion is equivalent to Pythagoras theorem.
Example 1. The
0
0
1
vectors 0 , 1 , 0
1
0
0
are orthogonal to each other, and

have length 1.
We are going to call such a basis orthonormal soon.
Theorem 2. Suppose that v1,

, vn are nonzero and (pairwise) orthogonal. Then v1,
,
vn are independent.
Proof. Suppose that
c1v1 +
+ cnvn = 0.
Take the dot product of v1 with both sides:
0 = v1 (c1v1 +
+ cnvn)
= c1v1 v1 + c2v1 v2 +
+ cnv1 vn
= c1v1 v1 = c1kv1k2
But kv1k 0 and hence c1 = 0.
Likewise, we find c2 = 0,
, cn = 0.
Example 3. Let us consider
Hence, the vectors are independent.
1 2
A = 2 4 .
3 6
Find Nul(A) and Col(AT ). Observe!

Solution.
o
2
1
n o
Col(AT ) = span 12
Nul(A) = span
n
The two basis vectors are orthogonal!

Armin Straub
[email protected]
can you see it?

if n ot, do it!
2
1

1
2
=0
Example 4. Repeat for
1 2 1
A = 2 4 0 .
3 6 0
Solution.
1 2 1
2 4 0
3 6 0
)
Nul(A) = span
1 2 1
0 0 2
0 0 3
1 2 0
RREF
0 0 1
0 0 0
>
2
1
0
Col(AT ) = span
)
1
0
2 , 0
0
1
The 2 vectors form a basis.
Again, the vectors are orthogonal!

0
2
1 0 = 0.
1
0

1
2
1 2 = 0,
0
0
Note:
2
Because 1 is
0
orthogonal to both basis vectors, it is orthogonal to every vector
in the row space.
Vectors in Nul(A) are orthogonal to vectors in Col(AT ).
The fundamental theorem, second act

Definition 5. Let W be a subspace of Rn, and v in Rn.
v is orthogonal to W , if v w = 0 for all w in W .

(
v is orthogonal to each vector in a basis of W )
Another subspace V is orthogonal to W , if every vector in V is orthogonal to W .

The orthogonal complement of W is the space W of all vectors that are orthogonal to W .
Exercise: show that the orthogonal complement is indeed a vector space.
Example 6. In the previous example,

We found that
1 2 1
A = 2 4 0 .
3 6 0
2
Nul(A) = span 1 ,
0
Armin Straub
[email protected]

0
1
Col(AT ) = span 2 , 0
1
0
are orthogonal subspaces.

Indeed, Nul(A) and Col(AT ) are orthogonal complements.
Why?
Because
0
1
2
, 2 , 0 are
1
1
0
0
orthogonal, hence independent, and hence a basis of all of R 3.
Remark 7. Recall that, for an m n matrix A, Nul(A) lives in Rn and Col(A) lives
in Rm. Hence, they cannot be related in a similar way.
In the previous example, they happen to be both subspaces of R3:
2
Nul(A) = span 1 ,
0
But these spaces are not
Theorem 8.
orthogonal:

1
1
Col(A) = span 2 , 0
3
0

1
2
0
1
0
0
0
(Fundamental Theorem of Linear Algebra, Part I)
Let A be an m n matrix of rank r.
dim Col(A) = r (subspace of Rm)
dim Col(AT ) = r (subspace of Rn)
dim Nul(A) = n r (subspace of Rn)
dim Nul(AT ) = m r (subspace of Rm)

Theorem 9.
(Fundamental Theorem of Linear Algebra, Part II)
Nul(A) is orthogonal to Col(AT ).
(both subspaces of R n)
Note that dim Nul(A) + dim Col(AT ) = n.
Hence, the two spaces are orthogonal complements in R n.
Nul(AT ) is orthogonal to Col(A).
Again, the two spaces are orthogonal complements.
Armin Straub
[email protected]
Review
v and w in Rn are orthogonal if v w = v1w1 +
+ vnwn = 0.
This simple criterion is equivalent to Pythagoras theorem.
Nonzero orthogonal vectors are independent.

1 2
1 2 T
2
1
Nul 2 4 = span
, Col 2 4 = span
1
2
3 6
3 6
Fundamental Theorem of Linear Algebra:

(A an m n matrix)
(both subspaces of R n)
Moreover, dim Col(AT ) + dim Nul(A) = n

= r (ran k of A)
= nr
Hence, they are orthogonal complements in R n.
Nul(AT ) and Col(A) are orthogonal complements.
(in R m)

Why? Suppose that x is in Nul(A). That is, Ax = 0.
But think about what Ax = 0 means (row-column rule).
It means that the inner product of every row with x is zero.
But that implies that x is orthogonal to the row space.
Example 1. Find all vectors orthogonal
1
to 1
1
0
and 1 .
1
Solution. (FTLA, no thinking) In other words:

!
find the orthogonal complement of Col

!

1 0 T
FTLA: this is Nul 1 1

= Nul 10
1 1
which has
(
span
0
basis: 1 .
1
)
0
1
1
1 0
1 1
1 1
1 1
1 1

are the vectors orthogonal
Armin Straub
[email protected]
1
to 1
1
0
and 1 .
1
Solution. (a little thinking) The FTLA is not magic!

0
1
1 x = 0 and 1 x = 0
1
1

1 1 1
0
x=
0 1 1
0

1 1 1
x in Nul
0 1 1
This is the same null space we obtained from the FTLA.
Example 2. Let V =
a
b
c
: a + b = 2c .
Find a basis for the orthogonal complement of V .

Solution. (FTLA, no thinking) We note that V = Nul([ 1 1 2 ]).
FTLA: the orthogonal complement is Col([ 1 1 2 ]T ).
Basis for the orthogonal
1
complement: 1
2
Solution. (a little thinking) a + b = 2c

1
a
b 1 = 0.
2
c
)
1
1 .
2
So: V is actually defined as the orthogonal complement of span
A new perspective on Ax = b

Ax = b is solvable
b is in Col(A)
b is orthogonal to Nul(AT )
(direct approach)
(indirect approach)
The indirect approach means: if yTA = 0 then yTb = 0.
Armin Straub
[email protected]
Example 3. Let
1 2
A = 3 1 .
0 5
For which b does Ax = b have a solution?
Solution. (old)
1 2 b1
3 1 b2
0 5 b3
So, Ax = b is consistent
1 2
b1
0 5 3b1 + b2
b3
0 5
1 2
b1
0 5
3b1 + b2
0 0 3b1 + b2 + b3
3b1 + b2 + b3 = 0.
Solution. (new) Ax = b solvable

to find Nul(AT ):
b orthogonal to Nul(AT )
1 3 0
2 1 5
>
RREF
1 0 3
0 1 1
3
We conclude that Nul(A ) has basis 1 .
1
3
Ax = b is solvable
b 1 = 0. As above!
1
Motivation
Example 4. Not all linear systems have solutions.
In fact, for many applications, data needs to be fitted and there is
no hope for a perfect match.
For instance, Ax = b with

1 2
1
x=
2 4
2
has no solution:
n o

1
is not in Col(A) = span 12

2
Ax
b
Instead of giving up, we want the x which makes Ax and b as

close as possible.
Such x is characterized by Ax being orthogonal to the error b Ax (see picture!)
Armin Straub
[email protected]
Application: directed graphs

1

2
1
Graphs appear in network analysis

(e.g. internet) or circuit analysis.
arrow indicates direction of flow
no edges from a node to itself
at most one edge between nodes
Definition 5. Let G be a graph with m edges and n nodes.

The edge-node incidence matrix of
1,
Ai,j = +1,
0,
G is the m n matrix A with

if edge i leaves node j ,
if edge i enters node j ,
otherwise.
Example 6. Give the edge-node incidence matrix of our graph.

Solution.

2
1
A =
1 1
0
1 0
1
0 1 1
0 1 0
0
0 1
0
0
0
1
1
each column represents a node
Armin Straub
[email protected]
each row represents an edge
Review
A graph G can be encoded by the edge-node incidence matrix:
1 1
0
0

1
2
1 0
1
0
A = 0 1 1
0
0 1 0
1
0
0 1 1
2
3
4
each column represents a node
each row represents an edge
If G has m edges and n nodes, then A is the m n matrix with
1, if edge i leaves node j ,

Ai,j = +1, if edge i enters node j ,
0, otherwise.
Meaning of the null space
The x in Ax is assigning values to each node.
You may think of assigning potentials to each node.
1 1
0
1 0
1
0 1 1
0 1 0
0
0 1
0
0
0
1
1
x1 + x2
x1
x1 + x3
x2

x3 = x2 + x3
x2 + x4
x4
x3 + x4

1
1
So: Ax = 0
nodes connected by an edge are assigned the same value
Armin Straub
[email protected]
For our graph: Nul(A) has
basis 11 .
1
This always happens as long as the graph is connected.

Example 1. Give a basis for Nul(A) for the following graph.
Solution. If Ax = 0 then x1 = x3 (connected by edge) and x2 = x4

(connected by edge).
Nul(A) has the

1
0
0 1
basis: 1 , 0
0
1
Just to make sure: the edge-node incidence matrix is:

A=
1 0 1 0
0 1 0 1
In general:
dim Nul(A) is the number of connected subgraphs.
For large graphs, disconnection may not be apparent visually.
But we can always find out by computing dim Nul(A) using Gaussian elimination!
Meaning of the left null space

The y in y TA is assigning values to each edge.
You may think of assigning currents to
1 1
0
1 0
1
A =
0
1
1
0 1 0
0
0 1
1 1 0
0
0
1
0 1 1 0
0
1
1
0 1
0
0
0
1
1
y1
y2
y3
y4
y5
each edge.
T =
,
A
0
1
1
1 1 0
0
0
1
0 1 1 0
0
1
1
0 1
0
0
0
1
1
y1 y2
y1 y3 y4
=
y2 + y3 y5
y4 + y5
Armin Straub
[email protected]
So: ATy = 0
at each node, (directed) values assigned to edges add to zero
When thinking of currents, this is Kirchhoffs first law.
(at each node, incoming and outgoing currents balance)
What is the simplest way to balance current?

Assign the current in a loop!
Here, we have two loops: edge 1, edge 3, edge 2 and edge3, edge 5, edge 4.
Correspondingly,
1
1
1
0
0
and
0
0
1
1
1
are in Nul(AT ). Check!
Example 2. Suppose we did not see this.

Let us solve ATy = 0 for our graph:
>
1 1 0 0 0
1 0 1 1 0
0 1 1 0 1
0 0 0 1 1
The parametric solution is
So, a basis for Nul(AT ) is
y3 y5
y3 + y5
y3
y5
y5

1
1
1
1

1
, 0
1
0
0
1
0 1 0 1
1 1 0 1
0 0 1 1
0 0 0 0
RREF 0
0
0

2
1
Observe that these two basis vectors correspond to loops.
Note that we get the simpler loop
0
0
1
1
1
as
1
1
1
0
0
1
1
0
1
1
In general:
dim Nul(AT ) is the number of (independent) loops.
For large graphs, we now have a nice way to computationally find all loops.
Armin Straub
[email protected]
Practice problems
Example 3. Give a basis for Nul(AT ) for the following graph.
Example 4. Consider the graph with edge-node incidence matrix
1 1
1 0
A = 0 1
0
1
0
1
1
0
0
0
.
0
1
(a) Draw the corresponding directed graph with numbered edges and nodes.
(b) Give a basis for Nul(A) and Nul(AT ) using properties of the graph.
Armin Straub
[email protected]
Solutions to practice problems

Example 5. Give a basis for Nul(AT ) for the following graph.
Solution. This graph contains no loops,

so Nul(AT ) = {0}.
Nul(AT ) has the empty set as basis
(no basis vectors needed).
For comparison: the edge-node incidence matrix

1 0 1 0
A=
0 1 0 1
indeed has Nul(AT ) = {0}.
Example 6. Consider the graph with edge-node incidence

matrix
1 1
1 0
A = 0 1
0
1

2
1
0
1
1
0
0
0
.
0
1
Give a basis for Nul(A) and Nul(AT ).
Solution.
If Ax = 0, then x1 = x2 = x3 = x4 (all connected by edges).
Nul(A) has the
basis: 11 .
1
The graph is connected, so only 1 connected subgraph and dim Nul(A) = 1.
The graph has one loop: edge1, edge2, edge3

Assign values y1 = 1, y2 = 1, y3 = 1 along the edges of that loop.
Nul(AT ) has the
1
1
.
basis: 1
The graph has 1 loop, so dim Nul(AT ) = 1.
Armin Straub
[email protected]
Review for Midterm 2

As of yet unconfirmed:
final exam on Friday, December 12, 710pm
conflict exam on Monday, December 15, 710pm
Directed graphs
Go from directed graph to edge-node incidence matrix A and vice versa.
Basis for Nul(A) from connected subgraphs.
For each connected subgraph, get a basis vector x that assigns 1 to all nodes in that subgraph,
and 0 to all other nodes.
Basis for Nul(AT ) from (independent) loops.

For each (independent) loop, get a basis vector y that assigns 1 and 1 (depending on direction)
to the edges in that loop, and 0 to all other edges.
Example 1.
Basis for
Basis for
Nul(A): 11
1
1
1
Nul(AT ): 1
0
0

2
1
0
0
1
1
1
Fundamental notions
Vectors v1,
, vn are independent if the only linear relation
c1v1 +
+ cnvn = 0
is the one with c1 = c2 =
= cn = 0.
How to check for independence?
The columns of a matrix A are independent
Nul(A) = {0}.
Vectors v1,
, vn in V are a basis for V if
they span V , that is V = span{v1,

, vn }, and
they are independent.
In that case, V has dimension n.
Vectors v, w in Rm are orthogonal if v w = v1w1 +

+ vmwm = 0.
Armin Straub
[email protected]
Subspaces
From an echelon form of A, we get bases for:
Nul(A) by solving Ax = 0
Col(A) by taking the pivot columns of A
Col(AT ) by taking the nonzero rows of the echelon form
Example 2.
1
2
A =
3
4
Basis for
Basis for
Basis for
0
1
Col(A): 23 , 1
2
0
4
0
1
Col(AT ): 20 , 01
5
4
4
2
1 0
Nul(A): 0 , 5
1
0
2 0 4
4 1 3
RREF
6 2 22
8 0 16
>
1
0
0
0
2
0
0
0
0
1
0
0
4
5
0
0
Dimension of Nul(AT ): 2
The solutions to Ax = b are given by x p + Nul(A).

The fundamental theorem states that
Nul(A) and Col(AT ) are orthogonal complements
So: dim Nul(A) + dim Col(AT ) = n (number of columns of A)
Nul(AT ) and Col(A) are orthogonal complements

So: dim Nul(AT ) + dim Col(A) = m (number of rows of A)
In particular, if r = rank(A) (nr of pivots):

dim Col(A) = r
dim Col(AT ) = r
dim Nul(A) = n r
dim Nul(AT ) = m r
Example 3. Consider the following subspaces of R4:
(a) V = bc : a + 2b = 0, a + b + d = 0
Armin Straub
[email protected]
(b) V
: a, b, c in R
a+bc
= 2a +
3c
In each case, give a basis for V and its orthogonal complement.

Try to immediately get an idea what the dimensions are going to be!
Solution.
First step: express these subspaces as one of the four subspaces of a matrix.

1 2 0 0
(a) V = Nul 1 1 0 1
(b) V = Col
1
0
2
0
1 1
1 0
0 3
0 1
Give a basis for each.

(a) row reductions: 11
basis for V : 01 ,
0
1
0
row reductions: 2
0
(b)
1 1
1 0
0 3
0 1
2 0 0
1 0 1
2
1
0
1
1 2 0 0
0 1 0 1
1 0 0 2
0 1 0 1
1 1 1
0 1
0
0 2 5
0 0
1
(no need to continue; we already see that the columns are independent)
basis for V
1
1
0 1
: 2 , 0
0
0
1
0
,
3
1
Use the fundamental theorem to find bases for the orthogonal complements.

1 2 0 0 T
(a) V = Col 1 1 0 1
note the two rows are clearly independent.
1
1
basis for V : 20 , 10
1
0

1 1 1 T
V = Nul 02 10 03 = Nul
0 0 1
1 0 2 0
row reductions: 1 1 0 0
1 0 3 1
2/5
basis for V : 2/5

1/5
1
(b)
Armin Straub
[email protected]
!
1 0 2 0
1 1 0 0
1 0 3 1
1 0 2 0
0 1 2 0
0 0 5 1
1 0 0 2/5
0 1 0 2/5
0 0 1 1/5
Example 4. What does it mean for Ax = b if Nul(A) = {0}?

Solution. It means that if there is a solution, then it is unique.
Thats because all solutions to Ax = b are given by xp + Nul(A).
Example 5. Let T : R2 R3 be the linear map represented by the matrix
1 0
2 1
3 0
with respect to the bases

(a) What is T 11 ?
0
1

1
1
of
1
1
0
R2 and 1 , 0 , 1
0
1
1
of R3.
(b) Which matrix represents T with respect to the standard bases?

Solution.
The matrix tells us that:
0
1

1
1

T

= 1
= 0

1
1
1 + 2 0 + 3
1
0

1
1
1 + 1 0 + 0
0
1
1
(a) Note that 11 = 2 01 + 1
.

1
= 2T 01 + T
Hence, T 1

1
(b) Note that 10 = 01 + 1
.

0
1
Hence, T 0 = T 1 + T
We already know that T

So, T is represented
Armin Straub
[email protected]

1
1
0
1

4 3
by 4 4
6 5
1
1

3
= 4 .
5

0
1 =
1

0
1 =
1
3
4
5
1
0
1

7
1
3
= 2 4 + 0 = 8 .
11
1
5

4
1
3
= 4 + 0 = 4 .
6
1
5
with respect to the standard bases.
Check your understanding

Think about why each of these statements is true!
Ax = b has a solution x if and only if b is in Col(A).
Thats because Ax are linear combinations of the columns of A.
A and AT have the same rank.

Recall that the rank of A (number of pivots of A) equals dim Col(A).
So this is another way of saying that dim Col(A) = dim Col(AT ).
The columns of an n n matrix are independent if and only if the rows are.

Let r be the rank of A, and let A be m n for now.

The columns are independent
But also: the rows are independent
r = n (so that dim Nul(A) = 0).

r = m.
In the case m = n, these two conditions are equivalent.
Ax = b has a solution x if and only if b is orthogonal to Nul(AT ).

This follows from Ax = b has a solution x if and only if b is in Col(A)
together with the fundamental theorem, which says that Col(A) is the orthogonal complement
of Nul(AT ).
The rows of A are independent if and only if Nul(AT ) = {0}.

Recall that elements of Nul(A) correspond to linear relations between the columns of A.
Likewise, elements of Nul(AT ) correspond to linear relations between the rows of A.
Armin Straub
[email protected]
We can deal with complicated linear systems, but what to do if there is no solutions
and we want a best approximate solution?
This is important for many applications, including fitting data.
Suppose Ax = b has no solution. This means b is not in Col(A).

Idea: find best approximate solution by replacing b with its projection onto Col(A).
Recall: if v1,
, vn are (pairwise) orthogonal:
v1 (c1v1 +
+ cnvn) = c1v1 v1
Implies: the v1,
, vn are independent (unless one is the zero vector)
Orthogonal bases
Definition 1. A basis v1,
, vn of a vector space V is an orthogonal basis if the
vectors are (pairwise) orthogonal.
Example 2. The standard
Example 3. Are the
0
0
1
basis 0 , 1 , 0
1
0
0
0
1
1
vectors 1 , 1 , 0
1
0
0
is an orthogonal basis for R3.
an orthogonal basis for R3?
Solution.
Armin Straub
[email protected]
So this is an orthogonal basis.
1
1
0
1
1
0
1
1
0
1
1
0
0
0
1
0
0
1
=0
=0
=0
Note that we do not need to check that the three vectors are independent. That follows
from their orthogonality.
Example 4. Suppose v1,
, vn is an orthogonal basis of V , and that w is in V . Find
c1,
, cn such that
w = c1v1 +
+ cnvn.
Solution. Take the dot product of v1 with both sides:
v1 w = v1 (c1v1 +
+ cnvn)
= c1v1 v1 + c2v1 v2 +
+ cnv1 vn
= c1v1 v1
Hence, c1 =
v1 w
v w
. In general, c j = j
.
v1 v1
vj vj
If v1,
, vn is an orthogonal basis of V , and w is in V , then
w = c1v1 +
+ cnvn
Example 5.
3
Express 7
4
in terms of the
with
cj =
w vj
.
vj vj
0
1
1
basis 1 , 1 , 0 .
1
0
0
Solution.

0
1
1
3
7 = c1 1 + c2 1 + c3 0
1
0
4
0

3
1
7 1
0
4

=
1
1
1 1
0
0
1
1
0

3
7
+ 4
1
1
0
1
1
0
4
10
4
=
1 + 1 + 0
2
2
1
0
0
1
Armin Straub
[email protected]
1
1
0
1
1
0
1
1
0

3
7
+ 4
0
0
1
0
0
1
0
0
1
0
0
1
Definition 6. A basis v1,

, vn of a vector space V is an orthonormal basis if the
vectors are orthogonal and have length 1.
Example 7. The standard
1
0
0
basis 0 , 1 , 0
0
0
1
is an orthonormal basis for R3.
If v1,
, vn is an orthonormal basis of V , and w is in V , then
w = c1v1 +
+ cnvn
Example 8.
3
Express 7
4
in terms of the
with
c j = v j w.
0
0
1
basis 0 , 1 , 0 .
1
0
0
Solution. Thats trivial, of course:

0
0
1
3
7 = 3 0 + 7 1 + 4 0
1
0
0
4
But note that the coefficients are

3
1
7 0 = 3,
4
0
Example 9. Is the basis

3
0
7 1 = 7,
4
0
0
1
1
1 , 1 , 0
1
0
0

3
0
7 0 = 4.
4
1
orthonormal? If not, normalize the vectors
to produce an orthonormal basis.

Solution.
1
1
0
1
1
0
0
0
1
has length
1
1
1 1 = 2
0
0
has length
has length

1
1
1 1 = 2
0
0

0
0
0 0 = 1
1
1
The corresponding orthonormal basis
Armin Straub
[email protected]
normalized:
normalized:
1
1
1
2
0
1
1
1
2 0
0
is already normalized: 0
1
0
1
1
1
1
is 1 , 1 , 0 .
2 0
2
1
0
Example 10.
3
Express 7
4
in terms of the basis
1 1 1 1 0
1 ,
1 , 0 .
2 0
2
1
0
Solution.
3
1
4
7 1 1 =
,
2
2
4
0
Hence, just as in Example 5:

3
1
10
7 1 1 =
,
2 0
2
4

0
3
7 0 = 4.
1
4
0
3
1
1
4 1
10 1
7 =
1 + 1 + 4 0
2 2
2 2 0
1
4
0
Orthogonal projections
Definition 11. The orthogonal projection of vector x onto

vector y is
x y
x =
y.
yy
The vector x is the closest vector to x, which is in
span{y }.
Characterized by: the error x = x x is orthogonal to

span{y }.
To find the formula for x, start with x = cy.

(x x) y = (x cy) y = x y cy y
xy
It follows that c = y y .
wanted
x is also called the component of x orthogonal to y.
Example 12. What is the orthogonal projection of x =
8
4
onto y =
3
1
Solution.

x y
8 3 + 4 1 3
3
6
x =
y=
= 2
=
1
1
2
yy
32 + 12
The component of x orthogonal to y is

8
6
2
x x =
=
.
4
2
6
(Note that, indeed
2
6
Armin Straub
[email protected]
and
3
1
are orthogonal.)
Example
13. What
are the orthogonal projections of
0
1
1
1 , 1 , 0 ?
1
0
0
2
1
1
onto each of the vectors
Solution.
2
1
1
2
1
1
2
1
1
1
1
1
2 1 + 1 (1) + 1 0
1
on 1 : 12 + (1)2 + 02 1 = 2 1
0
0
0
1
3 1
21+11+10 1
on 1 : 12 + 12 + 02 1 = 2 1
0
0
0
0
0
20+10+11 0
on 0 : 02 + 02 + 12 0 = 0
1
1
1

1
3 1
1
Note that these sum up to 2 1 + 2 1 +
0
0

2
0
0 = 1 !
1
1
Thats because the three vectors are an orthogonal basis for R 3.
Recall: If v1,
, vn is an orthogonal basis of V , and w is in V , then
w = c1v1 +
+ cnvn
with
cj =
w vj
.
vj vj
w decomposes as the sum of its projections onto each basis vector
Armin Straub
[email protected]
Comments on midterm
Suppose V is a vector space, and you are asked to give a basis.
CORRECT: V has
1
1
basis 1 , 0
1
0
CORRECT: V has basis
1
1
1 , 0
1
0
OK: V = span
1
1
1 , 0
1
0
(but you really should point out that the two vectors are independent)
( )
1
1 ,
0
1 1
basis = 1 0
0 1
INCORRECT: V =
INCORRECT:
1
0
1
Review
Orthogonal projection of x onto y:
x =
Example 1.
x y
y.
yy
Error x = x x is orthogonal to y.
If y1,
, yn is an orthogonal basis of V , and x is in V ,
then
x yj
x = c1 y1 +
+ cn yn
with c j =
.
yj yj
x decomposes as the sum of its projections onto each vector
in the orthogonal basis.
2
Express 1
1
x
in terms of the
1
0
1
basis 1 , 1 , 0 .
0
0
1
y1
y2
y3
Solution. Note that y1, y2, y3 is an orthogonal basis of R3.
Armin Straub
[email protected]

2
1
1
0
1 = c1 1 + c2 1 + c3 0
1
0
0
1

1
2
1 1
0
1

=
1
1
1 1
0
0
1
1 +
0
projection of x onto y1

1
2
1 1
0
1

1
1
1 1
0
0
projection
1
1
0
3
1
= 1 + 1 + 0
2
2
0
0
1

0
2
1 0
1
1

0
0
0 0
1
1
1
1 +
0
of x onto y2
0
0
1
projection of x onto y3
Orthogonal projection on subspaces

Theorem 2. Let W be a subspace of Rn. Then, each x in Rn can be uniquely written as
x = x + x .
in W
in W
x
x
v1
x
v2
is the orthogonal projection of x onto W .
x
is the point in W closest to x. For any other y in W , dist(x, x
) < dist(x, y).
x
If v1,
, vm is an orthogonal basis of W , then
=
x

x v1
x vm
v1 +
+
vm.
v1 v1
vm vm
is determined, x = x x
.
Once x
(This is also the orthogonal projection of x onto W .)
)
0
3
0 , 1 ,
0
1
Example 3. Let W = span
and
0
x = 3 .
10
Find the orthogonal projection of x onto W .
(or: find the vector in W which is closest to x)

Write x as a vector in W plus a vector orthogonal to W .
Armin Straub
[email protected]
Solution.
Note that
3
w1 = 0
1
and
0
w2 = 1
0
are an orthogonal basis for W .
[We will soon learn how to construct orthogonal bases ourselves.]
Hence, the orthogonal projection of x onto W is:

=
x
x w1
w1 +
w1 w1
3
0
3 0
10
1
x w2
w2 =
3
3
w2 w2
0 0
1
1
3
0
1
0
0
3 1
+ 1 0 0
0
0
1 1
0
0
0
1
0

3
0
3
10
=
0 +3 1 = 3
10
1
0
1
x is the vector in W which best approximates x.
Orthogonal projection of x onto the orthogonal complement of W :

3
3
0
x = 3 3 = 0 .
9
1
10
Note: Indeed,
3
0
9
Hence,
3
0
3
x = 3 = 3 + 0 .
1
10
9
in W
is orthogonal to w1 =
3
0
1
and
in W
0
w2 = 1 .
0
Definition 4. Let v1,

, vm be an orthogonal basis of W , a subspace of Rn. Note that
the projection map W : Rn Rn, given by
x
x =

x v1
x vm
v1 +
+
vm
v1 v1
vm vm
is linear. The matrix P representing W with respect to the standard basis is the
corresponding projection matrix.
Example 5. Find
matrix P which corresponds to orthogonal projection
( the projection
)
0
3
0 , 1
0
1
onto W = span
Solution. Standard basis
in R3.
0
0
1
: 0 , 1 , 0 .
1
0
0
The first column of P encodes the projection

3
1
0 0
1
0
3
3
0 0
1
1
3
0 +
1

0
1
0 1
0
0
0
0
1 1
0
0
Armin Straub
[email protected]
0
3
3
1 = 0 .
10
0
1
1
of 0 :
0
Hence P =
9
10
3
10

0
The second column of P encodes the projection of 1 :
0
0
0
3
0
1 1

9
1 0
0
0
3
0
10
0
0
1
0
1 = 1 . Hence P = 0

0 +

0
0
3
3
3
1 1 0
0 0 1
0
0
10
0
0
1
1
0
The third column of P encodes the projection of 0 :
1
0
0
3
0
9
0 1
0 0
0
3
0
3
10
0
1
1
1
1
0 . Hence P = 0 1
1 =

0 +

0
0
3
3
10
3
1 1 0
0 0 1
1
0
10
3
10
0
.
1
10
Example 6. (again)
Find the orthogonal projection of
Solution. x = Px =
9
10
0
3
10
0
1
0
0
3
0 , 1 .
0
1
0
x = 3
10
onto W = span
3
0
= 3 ,
0
3
1
1
10
3
10
as in the previous example.
10
Example 7. Compute P 2 for the projection matrix we just found. Explain!

Solution.
9
10
3
10
9
10
3
10
9
10
0
0

0 1 0 0 1 0 = 0

1
1
3
3
3
0 10
0 10
10
10
10
0
1
0
3
10
1
10
Projecting a second time does not change anything anymore.
Practice problems
Example 8. Find the closest point to x in span{v1, v2}, where
2
4
x =
0 ,
2
1
1
v1 =
0 ,
0
1
0
v2 =
1 .
1
Solution. This is the orthogonal projection of x onto span{v1, v2}.
Armin Straub
[email protected]
Review
Let W be a subspace of Rn, and x in Rn
(but maybe not in W ).
x
x
v1
x
v2
Let x be the orthogonal projection of x onto W .

(vector in W as close as possible to x)
If v1,
, vm is an orthogonal basis of W , then
=
x

x v1
x vm
v1 +
+
vm.
v1 v1
vm vm
proj. o f x on to v1
pro j. of x on to vm
The decomposition x = x + x is unique.

in W
in W
Least squares
Definition 1. x is a least squares solution of the system Ax = b if x is such that
Ax b is as small as possible.
If Ax = b is consistent, then a least squares solution x is just
an ordinary solution.
b = 0)
(in that case, Ax
Interesting case: Ax = b is inconsistent.

(in other words: the system is overdetermined)
Idea. Ax = b is consistent
b is in Col(A)
Ax
So, if Ax = b is inconsistent, we
replace b with its projection b onto Col(A),
and solve Ax = b.
(consistent
Armin Straub
[email protected]
b
by construction!)
Example 2. Find the least squares solution to Ax = b, where
1 1
A = 1 1 ,
0 0
2
b = 1 .
1
Solution. Note that the columns of A are orthogonal.

[Otherwise, we could not proceed in the same way.]
Hence, the projection b of b onto Col(A) is

1
2
1 1
0
1

b =
1
1
1 1
0
0
1
1
0

1
2
1 1
+ 1 0
1
1
1 1
0
0
1
1
1
2
1
3
1 = 1 + 1 = 1 .
2
2
0
0
0
0
We have already solved Ax = b in the process: x =
1/2
3/2
The normal equations

The following result provides a straightforward recipe (thanks to the FTLA) to find least
squares solutions for any matrix.
[The previous example was only simple because the columns of A were orthogonal.]
Theorem 3. x is a least squares solution of Ax = b

ATAx = ATb
(the normal equations)
Proof.

G

x is a least squares solution of Ax = b
FTLA
Ax b is as small as possible
Ax b is orthogonal to Col(A)
Ax b is in Nul(AT )
AT (Ax b) = 0
ATAx = ATb
Armin Straub
[email protected]
Example 4. (again) Find the least squares solution to Ax = b, where
1 1
A = 1 1 ,
0 0
2
b = 1 .
1
Solution.
AT A =
1 1
1 1
ATb =
1
1
1
0
1
0
0

1 0
1 0
1
1
0
2
1
1
2 0
0 2
1
3
The normal equations ATAx = ATb are

Solving, we find (again) x =
1/2
3/2

2 0
1
=
x
.
0 2
3
Example 5. Find the least squares solution to Ax = b, where
2
b = 0 .
11
4 0
A = 0 2 ,
1 1
What is the projection of b onto Col(A)?

Solution.
4 0
4 0 1
ATA =
0 2
0 2 1
1 1

2
4 0 1
ATb =
0
0 2 1
11

17 1
1 5
19
11
The normal equations ATAx = ATb are

Solving, we find x =
1
2

17 1
19
=
x
.
1 5
11
The projection of b onto Col(A) is

4
4 0
1
Ax = 0 2 2 = 4 .
3
1 1
the projection of b onto Col(A)?

Just to make sure: why is Ax
, Ax
b is as small as possible.
Because, for a least squares solution x
Armin Straub
[email protected]
The projection b of b onto Col(A) is

b = Ax,
with x such that ATAx = ATb.
If A has full column rank, this is
(columns of A independent)
b = A(ATA)1ATb.
Hence, the projection matrix for projecting onto Col(A) is
P = A(ATA)1AT .
Application: least squares lines

Experimental data: (xi , yi)
Wanted: parameters 1, 2 such that yi 1 + 2xi for all i
4
2
0
0
This approximation should be so that

SSres =
[yi (1 + 2xi)]2 is as small as possible.
residual sum of squares
Example 6. Find 1, 2 such that the line y = 1 + 2x best fits the data points (2, 1),
(5, 2), (7, 3), (8, 3).
Solution. The equations yi = 1 + 2xi in matrix form:
1
1
1
1
x1
y1

x2
y2
x3
2
y3
x4
y4
d esign m atrix X
Armin Straub
[email protected]
ob servatio n
vector y
Here, we need to find a least-squares solution to
1
1
1
1
X TX
1 1
2 5
X Ty =
Solving
4 22
22 142
9
57
1
2
2
5
7
8
1

2
1

2 = 3 .
3
1 2
1 1
1 5 =
7 8
1 7
1 8

1
1 1 1
2 =
5 7 8
3
3
, we find

2
1
2
2/7
5/14
4 22
22 142
9
57
Hence, the least squares line is y = 7 + 14 x.
4
2
0
0
Armin Straub
[email protected]
perimeter = 4
perimeter = 4
perimeter =
perimeter =
perimeter = 4
perimeter = 4
perimeter = 4
perimeter =
perimeter =
perimeter =
=4
Happy Halloween!
Armin Straub
[email protected]
Review

G
x is a least squares solution of the system Ax = b

FTLA
x is such that Ax b is as small as possible

ATAx = ATb
Application: least squares lines

Example 1. Find 1, 2 such that the line y = 1 + 2x best fits the data points (2, 1),
(5, 2), (7, 3), (8, 3).
4
2
0
0
Comment . As usual in practice, we are minimizing the (sum of squares of the) vertical offsets:
http://mathworld.wolfram.com/LeastSquaresFitting.html
Solution. The equations yi = 1 + 2xi in matrix form:
Armin Straub
[email protected]
1
1
1
1
x1
y1

y2
x2
=
y3
x3
2
x4
y4
d esign m atrix X
ob servatio n
vector y
Here, we need to find a least squares solution to
1
1
1
1
X TX =
1 1
2 5
X Ty =
Solving
4 22
22 142
9
57
1
2
2
5
7
8

1

1
2

2 = 3 .
3
1
2

1 1
1 5 =
7 8
1 7
1 8

1
1 1 1
2 =
5 7 8
3
3
, we find

2
1
2
2/7
5/14
4 22
22 142
9
57
Hence, the least squares line is y = 7 + 14 x.
4
2
0
0
How well does the line fit the data (2, 1), (5, 2), (7, 3), (8, 3)?
How small is the sum of squares of the vertical offsets?
Armin Straub
[email protected]
residual sum of squares: SSres =
(yi (1 + 2xi))2
error at (xi ,yi)
The choice of 1, 2 from least squares, makes SS res as small as possible.
total sum of squares: SStot =

1P
where y = n
(yi y)2,
yi is the mean of the observed data

SS
coefficient of determination: R2 = 1 SS r e s
tot
General rule: the closer
R2
is to 1, the better the regression line fits the data.
Here, y = 9/4:
R2 =

1
=1
2
7
(2, 1), (5, 2), (7, 3), (8, 3)
+ 14 2
2
0.075
= 0.974
2.75
very close to 1
2
2
2

5
5
5
2
2
2
+ 2 7 + 14 5
+ 3 7 + 14 7
+ 3 7 + 14 8

9 2
9 2
9 2
9 2
1 4 + 2 4 + 3 4 + 3 4
good fit
Other curves
We can also fit the experimental data (xi , yi) using other curves.
Example 2. yi 1 + 2xi + 3x2i with parameters 1, 2, 3.
The equations yi = 1 + 2xi + 3x2i in matrix form:
1 x1 x21
y1
1
y2
1 x2 x22
2 =
y3
1 x3 x23
3
desig n m atrix X
observation
vector y
Given data (xi , yi), we then find the least squares solution to X = y.
Multiple linear regression
In statistics, linear regression is an approach for modeling the relationship between a scalar dependent variable and one or more explanatory
variables.
The case of one explanatory variable is called simple linear regression.
For more than one explanatory variable, the process is called multiple
linear regression.
http://en.wikipedia.org/wiki/Linear_regression
The experimental data might be of the form (vi , wi , yi), where now the dependent
variable yi depends on two explanatory variables vi , wi (instead of just one xi).
Armin Straub
[email protected]
Example 3. Fitting a linear relationship yi 1 + 2vi + 3wi, we get:
y1
1 v 1 w1
y2
1 v 2 w2 1

1 v 3 w3 2 = y 3
3
design m atrix
o bservation
vector
And we again proceed by finding a least squares solution.
Review
x
x
v1
x
v2
Suppose v1,
, vm is an orthonormal basis of W .
The orthogonal projection of x onto W is:
=
x
hx, v1iv1
+
+
proj. o f x on to v1
hx, vm ivm
pro j. of x on to vm
(To stay agile, we are writing hx, v1i = x v1 for the inner product.)
GramSchmidt
Example 4. Find an orthonormal basis for V = span
2
1
0
, 1
0 0
0
0
Recipe. (GramSchmidt orthonormalization)
1
1
,
1 .
Given a basis a1,

, an, produce an orthonormal basis q1,
, qn.
b1 = a1,
b2 = a2 ha2, q1iq1,
b3 = a3 ha3, q1iq1 ha3, q2iq2,

Armin Straub
[email protected]
b1
kb1k
b
q2 = 2
kb2k
b
q3 = 3
kb3k
q1 =
Example 5. Find an orthonormal basis for V = span
2
1
0
, 1
0 0
0
0
Solution.

2
2
1 1

b2 =
0 h 0
0
0
1
1
1
1
, q iq h
1 1 1 1
1
1

1
1

b3 =
1 h
1
1
0
b1 =
0
0
0
1
, q1iq1 =
0
0
0
, q2iq2 =
1
1
1
,
1 .
q1 =
q2 =
1
0
0
0
0
1
0
0
0
1 0
q3 =
2 1
1
We have obtained an orthonormal basis for V :
1
0
0 1
,
0 0
0
0
Armin Straub
[email protected]
0
1 0
, .
2 1
1
Review
Vectors q1,
, qn are orthonormal if
qiTqj =
0, if i j ,
1, if i = j.
GramSchmidt orthonormalization:
b2
Input: basis a1,

, an for V .
Output: orthonormal basis q1,

, qn for V .
b1
kb1k
b
q2 = 2
kb2k
b
q3 = 3
kb3k
b1 = a1,
q1 =
b2 = a2 ha2, q1iq1,
Example 1. Apply GramSchmidt to the
a1
a2
1
1
1
vectors 2 , 1 , 1 .
1
0
2
Solution.
1
b1 = 2 ,
2

1
1
2
1
1
1
1
1
b2 = 1 h 1 , 2 i 2 = 1 ,
3
3
3
2
2
2
0
0
1
1
1
2
1
b3 = 1 h 1 , q1iq1 h 1 , q2iq2 =
= 2 ,
9
1
1
1
1
We obtained the orthonormal vectors
q
1
=
1
1
2
3
2
q
=
2
2
1
1
3
2
=
q
3
2
1
2
3
1
2
2
1 1
1
1
2 , 1 , 2 .
3
3
3
2
2
1
Theorem 2. The columns of an m n matrix Q are orthonormal

QTQ = I (the n n id entity)
Proof. Let q1,
, qn be the columns of Q.
They are orthonormal if and only if qiTq j =
Armin Straub
[email protected]
0, if i j ,
1, if i = j.
All these inner products are packaged in QTQ = I:
T
q1
1 0 0
| |
q1 q2 = 0 1 0
q2T
| |
0 0

Definition 3. An orthogonal matrix is a square matrix Q with orthonormal columns.

It is historical convention to restrict to square matrices, and to say orthogonal matrix even
though orthonormal matrix might be better.
An n n matrix Q is orthogonal
In other words, Q1 = QT .
Example 4. P
0 0 1
= 1 0 0
0 1 0
QTQ = I
is orthogonal.
In general, all permutation matrices P are orthogonal.

Why? Because their columns are a permutation of the standard basis.
And so we always have P TP = I.
Example 5. Q =
cos sin
sin cos
Q is orthogonal because:
cos
sin

sin
cos
is an orthonormal basis of R2

Just to make sure: why length 1? Because
Alternatively: QTQ =
Example 6. Is H =
1 1
1 1
cos sin
sin cos

cos
sin
p 2
= cos + sin 2 = 1.
cos sin
sin cos
1 0
0 1
orthogonal?
No, the columns are orthogonal but not normalized.

But
1 1
1 1
is an orthogonal matrix.
Just for fun: a n n matrix with entries 1 whose columns are orthogonal is called a Hadamard
matrix of size n.
A size 4 example:
H H
H H
1 1
1
1
1 1 1 1
= 1 1 1 1
1 1 1 1
Continuing this construction, we get examples of size 8, 16, 32,
It is believed that Hadamard matrices exist for all sizes 4n.

But no example of size 668 is known yet.
Armin Straub
[email protected]
The QR decomposition (flashed at you)

Gaussian elimination in terms of matrices: A = LU
GramSchmidt in terms of matrices: A = QR
Let A be an m n matrix of rank n.
(columns independent)
Then we have the QR decomposition A = QR,

where Q is m n and has orthonormal columns, and
R is upper triangular, n n and invertible.
Idea: GramSchmidt on the columns of A, to get the columns of Q.
Example 7. Find the QR decomposition of
1 2 4
A = 0 0 5 .
0 3 6
Solution.
We apply GramSchmidt to the columns of A:
1
0
0
2
0
3
4
5
6
q1
2
h 0 ,
3
4
h 5 ,
6
0
q1iq1 = 0 ,
3
4
q1iq1 h 5 ,
6
0
0 = q2
1
0
q2iq2 = 5 ,
0
0
1 =
0
q3
1 0 0
Hence: Q = [ q1 q2 q3 ] = 0 0 1
0 1 0
To find R in A = QR,
note that QTA = QTQR = R.
1 0 0
1 2 4
1 2 4
R = 0 0 1 0 0 5 = 0 3 6
0 1 0
0 3 6
0 0 5
Summarizing, we have
1 2 4
1 0 0
1 2 4
0 0 5 = 0 0 1 0 3 6 .
0 3 6
0 1 0
0 0 5
Recipe. In general, to obtain A = QR:
GramSchmidt on (columns of) A, to get (columns of) Q.
Then, R = QTA.
Armin Straub
[email protected]
The resulting R is indeed upper triangular, and we get:
|
|
a1 a2
|
|
| |
= q1 q2
| |
q1Ta1 q1Ta2 q1Ta3
q2Ta2 q2Ta3
q3Ta3
It should be noted that, actually, no extra work is needed for computing R:

all the inner products in R have been computed during GramSchmidt.
(Just like the LU decomposition encodes the steps of Gaussian elimination,
the QR decomposition encodes the steps of GramSchmidt.)
Practice problems
Example 8. Complete

1 1
1 2
2 , 1
3
3
2
2
to an orthonormal basis of R3.
(a) by using the FTLA to determine the orthogonal complement of the span you already
have
(b) by using GramSchmidt after throwing in an independent vector such
Armin Straub
[email protected]
1
as 0
0
1 1 2
A = 0 0 1 .
1 0 0
Review
Let A be an m n matrix of rank n.
(columns independent)
Then we have the QR decomposition A = QR,

where Q is m n with orthonormal columns, and
R is upper triangular and invertible.
To obtain
| |
a1 a2
| |
| |
= q1 q2
| |
q1Ta1 q1Ta2 q1Ta3
q2Ta2 q2Ta3
q3Ta3
GramSchmidt on (columns of) A, to get (columns of) Q.

Then, R = QTA.
(actually unnecessary!)
Example 1. The QR decomposition is also used to solve systems of linear equations.

(we assum e A is n n, and A
exists)
Ax = b

QRx = b
Rx = QTb
The last system is triangular and is solved by back substitution.

QR is a little slower than LU but makes up in numerical stability.
If A is not n n and invertible, then Rx = QTb gives the least squares solutions!
Example 2. The QR decomposition is very useful for solving least squares problems:
ATAx = ATb

(QR)TQRx = (QR)Tb
=RTQTQR
T
T
R Rx = R QTb
Rx = QTb
Again, the last system is triangular and is solved by back substitution.
x is a least squares solution of Ax = b

Rx = QTb (where A = QR)
Armin Straub
[email protected]
Application: Fourier series

Review. Given an orthogonal basis v1, v2,
, we express a vector x as:
x = c 1v 1 + c 2v 2 +
,
civi =
hx, vi i
vi
hvi , vi i
projection
of x onto vi
A Fourier series of a function f (x) is an infinite expansion:

f (x) = a0 + a1cos(x) + b1sin(x) + a2cos(2x) + b2sin(2x) +
Example 3.
sin (x)
sin (2x)
sin (3x)
Armin Straub
[email protected]
Example 4. (just a preview)

1
1
4
1
blue
=
sin (x) + sin(3x) + sin(5x) + sin(7x) +
function
5
7
We are working in the vector space of functions R R.

More precisely, nice (say, piecewise continuous) functions that have period 2.
These are infinite dimensional vector spaces.
The functions
1, cos (x), sin (x), cos (2x), sin (2x),
are a basis of this space. In fact, an orthogonal basis!

Thats the reason for the success of Fourier series.
But what is the inner product on the space of functions?

Vectors in Rn: hv, wi = v1w1 +
+ vnwn
R 2
Functions: hf , g i = 0 f (x)g(x)dx
Why these limits? Because our functions have period 2.
Example 5. Show that cos (x) and sin (x) are orthogonal.
Solution.
hcos (x), sin (x)i =
cos (x)sin(x)dx =
0
1
(sin (x))2
2
2
=0
More generally, 1, cos (x), sin (x), cos (2x), sin (2x),
are all orthogonal to each other.
Armin Straub
[email protected]
Example 6. What is the norm of cos (x)?

Solution.
hcos (x), cos (x)i =
cos (x)cos(x)dx =
0
Why? Theres many ways to evaluate this integral. For instance:

you could use integration by parts,
you could use a trig identity,
heres a simple way:
R 2
R 2
cos2 (x)dx = 0 sin 2 (x)dx
(cos and sin are just a shift apart)
cos2 (x) + sin 2 (x) = 1

R 2
1 R 2
So: 0 cos 2 (x)dx = 2 0 1dx =
Hence, cos (x) is not normalized. It has norm kcos (x)k =
Example 7. The same calculation shows that cos (kx) and sin (kx) have norm
well.
as
Fourier series of f (x):

Example 8. How do we find a1?
Or: how much cosine is in a function f (x)?
Solution.
hf (x), cos (x)i
1
a1 =
=
hcos (x), cos (x)i
f (x)cos(x)dx
0
f (x) has the Fourier series

where
hf (x), cos (kx)i
ak =
=
hcos (kx), cos (kx)i
hf (x), sin (kx)i
bk =
=
hsin (kx), sin (kx)i
hf (x), 1i
=
a0 =
h1, 1i
Armin Straub
[email protected]
Z
1 2
f (x)cos(kx)dx,
0
Z 2
1
f (x)sin(kx)dx,
0
Z 2
1
f (x)dx.
2 0
Example 9. Find the Fourier series of the 2-periodic function f (x) defined by
f (x) =
Solution. Note that

a0 =
an =
=
bn =
=
=
=
=
=
1, for x (, 0),
+1, for x (0, ).
2
0
and
are the same here.
(why?!)
1
f (x)dx = 0
2
Z
1
f (x)cos(nx)dx

Z 0

Z
1
cos (nx)dx +
cos (nx)dx = 0
0
Z
1
f (x)sin(nx)dx

Z 0

Z
1
sin (nx)dx +
sin (nx)dx
0

Z
2
sin (nx)dx
0

2
1
cos(nx)
n
0
2
[1 cos (n)]
n
(
Z
2
[1 (1)n] =
n
4
n
if n is odd
if n is even
In conclusion,

4
1
1
1
f (x) =
.
5
7
3
Armin Straub
[email protected]
Review
An inner product for 2-periodic functions:
R 2
hf , gi = 0 f (x)g(x)dx
(in R n: hv, wi = v1w1 +

+ vnwn)

are orthogonal
An expansion in that basis is a Fourier series:

where
hf (x), cos (kx)i
=
hcos (kx), cos (kx)i
hf (x), sin (kx)i
=
bk =
hsin (kx), sin (kx)i
hf (x), 1i
a0 =
=
h1, 1i
ak =
Z
1 2
f (x)cos(kx)dx,
0
Z
1 2
f (x)sin(kx)dx,
0
Z
1 2
f (x)dx.
2 0
Example 1.

1
1
1
4
blue
=
function
3
5
7
Armin Straub
[email protected]
Example 2. Find the Fourier series of the 2-periodic function f (x) defined by
f (x) =
Solution. Note that

a0 =
an =
=
bn =
=
=
=
=
=
1, for x (, 0),
+1, for x (0, ).
2
0
and
are the same here.
(why?!)
Z
1
f (x)dx = 0
2
Z
1
f (x)cos(nx)dx

Z 0
Z
1
cos (nx)dx = 0
cos (nx)dx +
0
Z
1
f (x)sin(nx)dx

Z 0

Z
1
sin (nx)dx +
sin (nx)dx
0

Z
2
sin (nx)dx
0

1
2
cos(nx)
n
0
2
[1 cos (n)]
n
(
2
[1 (1)n] =
n
4
n
if n is odd
if n is even
In conclusion,

1
4
1
1
.
f (x) =
3
5
7
Armin Straub
[email protected]
Determinants
For the next few lectures, all matrices are square!
Recall that

1
a b 1
d b
=
.
c d
ad bc c a
The determinant of
a 2 2 matrix is det

a b
c d

= ad bc,
a 1 1 matrix is det ([ a ]) = a.
Goal: A is invertible
We will write both det

a
c
det (A) 0

a
b

and
c
d

b
d
for the determinant.
Definition 3. The determinant is characterized by:

the normalization det I = 1,
and how it is affected by elementary row operations:
Does not change the determinant.
Reverses the sign of the determinant.
(scaling) Multiply all entries in a row by s.
Multiplies the determinant by s.
Example 4. Compute
Solution.

1 0 0

0 2 0

0 0 7

R2 1 R2 1 0 0

2

2 0 1 0

0 0 7

Example 5. Compute
Solution.

1 2 3

0 2 4

0 0 7

1 0 0

0 2 0

0 0 7

1 2 0

14 0 1 0
0 0 1
Armin Straub
[email protected]

R3 1 R3
1 0 0

7

0 1 0
14

0 0 1

1 2 3

0 2 4

0 0 7

R2 1 R2 1 2 3

2

2 0 1 2

0 0 7
R1R13R3
R2R22R3

= 14

R3 1 R3
1 2 3

7

0 1 2
14

0 0 1

R1R12R2
1 0 0

0 1 0
14

0 0 1

= 14

The determinant of a triangular matrix is

the product of the diagonal entries.
Example 6. Compute
Solution.

1 2 0

3 1 2

2 0 1

1 2 0

3 1 2

2 0 1
@
@
R2R23R1
R3R32R1
R3R3 7 R2
=
Example 7. Discover the formula for
Solution.

a b

c d
1
0
0
0
2
2
0
0

R4R4 3 R3

2

1
0
0
0
2
2
0
0
Example 8. Compute
Solution.

1
0
0
0
2
2
0
0
3
1
2
3
4
5
1
5
3
1
2
3
4
5
1
5
1 2 0
0 7 2
0 4 1
1 2
0
0 7 2
1
0 0 7

1
1 (7) 7 = 1

a b

c d
a
b

c
0 d ab
R2R2 a R1

c

= a d b = ad bc

a

3 4
1 5
7
2 1 = 1 2 2 2
7
0 2
= 14
The following important properties follow from the behaviour under row operations.
det (A) = 0
A is not invertible
Why? Because det (A) = 0 if only if, in an echelon form, a diagonal entry is zero (that is, a pivot
is missing).
det (AB) = det (A)det(B)

1
det (A1) = det (A)

det (AT ) = det (A)
Armin Straub
[email protected]
Example 9. Recall that AB = 0, then it does not follow that A = 0 or B = 0. However,

show that det (A) = 0 or det (B) = 0.
Solution. Follows from det (AB) = det (0) = 0,
and det (AB) = det (A)det(B).
A bad way to compute determinants

Example 10. Compute

1 2 0

3 1 2

2 0 1
by cofactor expansion.
Solution. We expand by the first row:

+

1 2 0

2 3
3 1 2 = 1
1
2

2
2 0 1
0 1

i.e.

3 2

2
+ 0 3 1

1 1

2 0

2 1
0 1

+

2 + 0 3 1

2 0

1

= 1 (1) 2 (1) + 0 = 1

Each term in the cofactor expansion is 1 times an entry times a smaller determinant
(row and column of entry deleted).
+ +
+
.
The 1 is assigned to each entry according to
+ +

Solution. We expand by the second column:

1 2 0

3 1 2

2 0 1
0

= 2 3
+ (1)
2
+

2
2
1
1
= 2 (1) + (1) 1 0 = 1

1
0

0 3
2

Solution. We expand by the third column:

1 2 0

3 1 2

2 0 1

1 2
1 2
+

= 0 3 1
2
+ 1 3 1

2 0

2 0

+
= 0 2 (4) + 1 (7) = 1
Practice problems
Problem 1. Compute
1
0
2
2
2
5
7
9

3 4
0 0
.
6 10

7 11
Solution. The final answer should be 10.

Armin Straub
[email protected]
Review
The determinant is characterized by det I = 1 and the effect of row ops:
replacement: does not change the determinant
interchange: reverses the sign of the determinant
scaling row by s: multiplies the determinant by s
1
0
0
0
2
2
0
0
3
1
2
0
4
5
1
3

= 1 2 2 3 = 12

det (A) = 0
A is not invertible

1
det (A1) = det (A)

det (AT ) = det (A)
Whats wrong?!
det (A
) = det
1
ad b c
d b
c a
= a d bc (da (b)(c)) = 1
The corrected calculation is:

1
det a d bc
d b
c a
= (ad b c)2 (da (b)(c)) = ad bc

1
This is compatible with det (A1) = det (A) .

Example 1. Suppose A is a 3 3 matrix with det (A) = 5. What is det (2A)?
Solution. A has three rows.
Multiplying all 3 of them by 2 produces 2A.
Hence, det (2A) = 23 det (A) = 40.
Armin Straub
[email protected]
A bad way to compute determinants

Example 2. Compute

1 2 0

3 1 2

2 0 1
by cofactor expansion.
Solution. We expand by the second column:

1 2 0

3 1 2

2 0 1
0

= 2 3

2 + (1)
+

2
2
1
1
= 2 (1) + (1) 1 0 = 1

1
0

0 3
2

Solution. We expand by the third column:

1 2 0

3 1 2

2 0 1

1 2

1 2
+

= 0 3 1
2
+ 1 3 1

2 0

2 0

+
= 0 2 (4) + 1 (7) = 1
Why is the method of cofactor expansion not practical?

Because to compute a large n n determinant,
one reduces to n determinants of size (n 1) (n 1),
then n (n 1) determinants of size (n 2) (n 2),
and so on.
In the end, we have n! = n(n 1) 3 2 1 many numbers to add.

WAY TOO MUCH WORK! Already 25! = 15511210043330985984000000 1.55 10 2 5 .
Context: todays fastest computer, Tianhe-2, runs at 34 pflops (3.4 10 1 6 ops per second).
By the way: fastest is measured by computed LU decompositions!
Example 3.
First off, say hello to a new friend: i, the imaginary unit
It is infamous for i2 = 1 .

1

i

|1|

1 i

i 1

1 i

i 1 i
i 1

i

1 i

i 1 i
i 1
Armin Straub
[email protected]
=1
= 1 i2 = 2

1 i i 0
i
= 2 i2 = 3
= 1
i 1 i 1

1 i

= 1 i 1 i

i 1

i 0

i i 1 i

i 1

2
=3i 1 i

i 1

=5

1 i

i 1 i

i 1 i

i 1 i

i 1

1 i

= 1 i 1 i

i 1 i

i 1

1 i

i2 i 1 i

i 1

=5+3=8

The Fibonacci numbers!
Do you know about the connection of Fibonacci numbers and rabbits?
Eigenvectors and eigenvalues

Throughout, A will be an n n matrix.
Definition 4. An eigenvector of A is a nonzero x such that
Ax = x
for some scalar .
The scalar is the corresponding eigenvalue.

In words: eigenvectors are those x, for which Ax is parallel to x.
Example 5. Verify that
1
2
is an eigenvector of A =
0 2
4 2
Solution.
Ax =
0 2
4 2

1
2
4
8
= 4x
Hence, x is an eigenvector of A with eigenvalue 4.
Armin Straub
[email protected]
Example 6. Use your geometric understanding

to find

the eigenvectors and eigenvalues of A = 01 10 .

Solution. A
x
y
y
x
i.e. multiplication with A is reflection through the line y = x.
1
1
=1
1
1
1
1
is an eigenvector with eigenvalue = 1.
1
1
So: x =

1
1
So: x =
1
1
= 1
1
1
Practice problems
Problem 1. Let A be an n n matrix.
Express the following in terms of det (A):
det (A2) =
det (2A) =
Hint: (unless n = 1) this is not just 2 det (A)
Armin Straub
[email protected]
Review
If Ax = x, then x is an eigenvector of A with eigenvalue .
EG: x =
1
2
because Ax =
is an eigenvector of A =
0 2
4 2

1
2
Multiplication with A =

1
1
So: x =

=1

1
1
1
1
1
1
So: x =
0 1
1 0
4
8
0 2
4 2
with eigenvalue 4
= 4x
is reflection through the line y = x.
1
1
=

1
1
= 1
1
1
Example 1. Use your geometric understanding

to find

1 0
the eigenvectors and eigenvalues of A = 0 0 .

x
y
Solution. A
x
0
i.e. multiplication with A is projection onto the x-axis.
1
0
1
0
So: x =

0
1
=1
So: x =
1
0
0
0
0
1

=0
0
1
Armin Straub
[email protected]
Example 2. Let P be the projection matrix corresponding to orthogonal projection

onto the subspace V . What are the eigenvalues and eigenvectors of P ?
Solution.
For every vector x in V , Px = x.
These are the eigenvectors with eigenvalue 1.
For every vector x orthogonal to V , Px = 0.

These are the eigenvectors with eigenvalue 0.
Definition 3. Given , the set of all eigenvectors with eigenvalue is called the
eigenspace of A corresponding to .
Example 4. (continued) We saw that the projection matrix P has the two eigenvalues
= 0, 1.
The eigenspace of = 1 is V .
The eigenspace of = 0 is V .
How to solve Ax = x
Key observation:

Ax = x
Ax x = 0
(A I)x = 0
This has a nonzero solution
det (A I) = 0
Recipe. To find eigenvectors and eigenvalues of A.
First, find the eigenvalues using:

is an eigenvalue of A
det (A I) = 0
Then, for each eigenvalue , find corresponding eigenvectors by

solving (A I)x = 0.
Example 5. Find the eigenvectors and eigenvalues of
A=
Armin Straub
[email protected]

3 1
.
1 3
Solution.
A I =
3 1
1 3
det (A I) = 3 1
= 2 6 + 8 = 0
1 0
0 1
3
1
1
3

1
= (3 )2 1
3
1 = 2, 2 = 4
This is the characteristic polynomial of A. Its roots are the eigenvalues of A.

Find eigenvectors with eigenvalue 1 = 2:

1 1
A 1I = 1 1

Solutions to 11 11 x = 0 have basis 1
.
1

So: x1 = 1
is an eigenvector with eigenvalue
1

(A =
3 1
1 3
(A =
3 1
1 3
1 = 2.
All other eigenvectors with = 2 are multiples of x1.

n
o
span 1
is the eigenspace for eigenvalue = 2.
1
Find eigenvectors with eigenvalue 2 = 4:

1 1
1 1

1
1
Solutions to 1
have
basis
.
x
=
0
1 1
1

So: x2 = 11 is an eigenvector with eigenvalue
A 2I =
The eigenspace for eigenvalue = 4 is span
n
Example 6. Find the eigenvectors and the
A= 0
0
1
1
o
2 = 4.
eigenvalues of
2 3
6 10 .
0 2
Solution.
The characteristic polynomial is:

3
2
3

det (A I) = 0
6 10
0
0
2
A has eigenvalues 2, 3, 6.

= (3 )(6 )(2 )

3 2 3
A = 0 6 10
0 0 2
The eigenvalues of a triangular matrix are its diagonal entries.

1 = 2:
1 2 3
(A 1I)x = 0 4 10 x = 0
0 0 0
Armin Straub
[email protected]
2
x1 = 5/2
1
2 = 3:
0 2 3
(A 2I)x = 0 3 10 x = 0
0 0 1
3 = 6:
3 2 3
(A 3I)x = 0 0 10 x = 0
0 0 4

1
x2 = 0
0
2
x3 = 3
0
In summary,
A has eigenvalues 2, 3, 6 with corresponding
2/3
1 .
0
2
1
eigenvectors 5/2 , 0 ,
0
1
These three vectors are independent. By the next result, this is always so.
Theorem 7. If x1,
, xm are eigenvectors of A corresponding to different eigenvalues,
then they are independent.
Why?
Suppose, for contradiction, that x1,

, xm are dependent.
By kicking out some of the vectors, we may assume that there is (up to multiples) only one linear
relation: c1x1 +
+ cmxm = 0.
Multiply this relation with A:
A(c1x1 +
+ cmxm) = c11x1 +
+ cmmxm = 0
This is a second independent relation! Contradiction.
Practice problems
Example 8. Find the eigenvectors and eigenvalues of A =
Example 9. What are the eigenvalues of
2
1
A = 1
0
0
3
1
1
0
0
3
2
0 2
4 2
0
0
?
0
4
No calculations!
Armin Straub
[email protected]
Review

All eigenvectors (plus 0) with eigenvalue form the eigenspace of .
Why? Because Ax = x
det (A I)
= 0.
characteristic polynomial
(A I)x = 0.
By the way: this means that the eigenspace of is just Nul(A I).
E.g., if
3 2 3
A = 0 6 10
0 0 2
then det (A I) = (3 )(6 )(2 ).
Eigenvectors x1,
, xm of A corresponding to different eigenvalues are independent.
By the way:
product of eigenvalues = determinant
sum of eigenvalues = trace (sum of diagonal entries)
Example 1. Find the eigenvalues of A as well
spaces, where
2 0
A = 1 3
1 1
as a basis for the corresponding eigen

0
1 .
3
Solution.
The characteristic polynomial is:

2
0
0

det (A I) = 1 3
1
1
1
3

3
1

= (2 )
1
3
= (2 )[(3 )2 1]
=(2 )( 2)( 4)
A has eigenvalues 2, 2, 4.
2 0 0
A = 1 3 1
1 1 3
Since = 2 is a double root, it has (algebraic) multiplicity 2.
1 = 2:
0 0 0
(A 1I)x = 1 1 1 x = 0
1 1 1
Armin Straub
[email protected]
Two independent solutions:
1
x1 = 1 ,
0
0
x2 = 1
1
In other words: the eigenspace for = 2 is span

2 = 4:
2 0
0
(A 2I)x = 1 1 1 x = 0
1 1 1
In summary, A has eigenvalues 2 and 4:

eigenspace for = 2 has
eigenspace for = 4 has
0
1
1 , 1 .
1
0
0
x3 = 1
1
0
1
basis 1 , 1 ,
1
0
0
basis 1 .
1
An n n matrix A has up to n different eigenvalues.

Namely, the roots of the degree n characteristic polynomial det (A I).
For each eigenvalue , A has at least one eigenvector.

Thats because Nul(A I) has dimension at least 1.
If has multiplicity m, then A has up to m (independent) eigenvectors for .

Ideally, we would like to find a total of n (independent) eigenvectors of A.
Why can there be no more than n eigenvectors?!
Two sources of trouble: eigenvalues can be

complex numbers (that is, not enough real roots), or
repeated roots of the characteristic polynomial.
Example
2. Find the eigenvectors and eigenvalues of

A = 10 1
. Geometrically, what is the trouble?
0

Solution. A
x
y
y
x
i.e. multiplication with A is rotation by 90 (counter-clockwise).
Which vector is parallel after rotation by 90? Trouble.
Armin Straub
[email protected]
Fix: work with complex numbers!

= 2 + 1

1
det (A I) =
1

So, the eigenvalues are 1 = i and 2 = i.

1 = i:
Let us check:
2 = i:
x=
i 1
1 i
0 1
1 0

i 1
1 i
x=
0
0
i
1
0
0
=

1
i
x1 =

=i
i
1
x2 =
i
1
i
1
Example 3. Find the eigenvectors and eigenvalues of A =
1 1
0 1
. What is the trouble?
Solution.

det (A I) = 1 0

1
= (1 )2
1
So: = 1 is the only eigenvalue (it has multiplicity 2).
(A I)x =
0 1
0 0
x=0
So: the eigenspace is span
n
1
0
x1 = 10
o
. Only dimension 1!
Trouble: only 1 independent eigenvector for a 2 2 matrix

This kind of trouble cannot really be fixed.
We have to lower our expectations and look for generalized eigenvectors .
These are solutions to (A I)2x = 0, (A I)3x = 0,
Practice problems
Example 4. Find the eigenvectors and eigenvalues of
Armin Straub
[email protected]
1 2 1
A = 0 5 0 .
1 8 1
Review
Eigenvector equation: Ax = x
(A I)x = 0
det (A I)
= 0.
An n n matrix A has up to n different eigenvalues .

The eigenspace of is Nul(A I).
That is, all eigenvectors of A with eigenvalue .
If has multiplicity m, then A has up to m eigenvectors for .

At least one eigenvector is guaranteed (because det (A I) = 0).
Test yourself! What are the eigenvalues and eigenvectors?
1 0
0 1
= 1, 1
0 0
0 0
= 0, 0, eigenspace is R2
2 1
0 2
= 2, 2, eigenspace is span
eigenspace is R2
(ie. multiplicity 2),
n
1
0
o
Diagonalization
Diagonal matrices are very easy to work with.
Example 1. For instance, it is easy to compute their powers.
If
2 0 0
A = 0 3 0 ,
0 0 4
then A2 =
Example 2. If A =
6 1
2 3
22
32
4
and A100 =
21 0 0
31 0 0
4
100
, then A100 = ?
Solution.
Characteristic polynomial:
1 = 4:
2 1
2 1
2 = 5:
1 1
2 2
v=0

6 1

2
3
= ( 4)( 5)
eigenvector v1 =
eigenvector v2 =
1
1
1
2
100
Key observation: A100 v1 = 100
v2 = 100
1 v1 and A
2 v2
For A1 0 0 , we need A1 0 0
Armin Straub
[email protected]
1
0
and A1 0 0
0
1
1
0

= c1v1 + c2v2 = 12 + 2 11

100
100 1
=A
12 + 2
A
0
A
100
"
2 5 1 0 0 41 0 0
2 5 1 0 0 2 41 0 0

1
1

= 4100
1
2
+ 2 5100
1
1
We find the second column of A100 likewise. Left as exercise!

The key idea of the previous example was to work with respect to a basis given by the
eigenvectors.
Put the eigenvectors x1,
, xn as columns into a matrix P .
Axi = ixi
In summary: AP = PD
|
A x1
|
|
|
|
xn = 1x1 nxn
|
|
|
|
|
1
= x1 xn

n
|
|
Suppose that A is n n and has independent eigenvectors v1,

, vn.
Then A can be diagonalized as A = PDP 1.
the columns of P are the eigenvectors
the diagonal matrix D has the eigenvalues on the diagonal
Such a diagonalization is possible if and only if A has enough eigenvectors.
Armin Straub
[email protected]
Example 3.
Fibonacci numbers: 0, 1, 1, 2, 3, 5, 8, 13, 21, 34,
B y the way: not a u niversal law but only a fascinatingly prevalent tendency C oxeter
Did you notice:
13
8
= 1.625,
21
13
= 1.615,
34
21
= 1.619,
The golden ratio = 1.618

Wheres that coming from?
By the way, this is the most irrational number (in a precise sense).
Hence:
Fn+1
Fn
Fn+1
Fn
n

F1
= 11 10
F0

Fn+1 = Fn + Fn1
But we know how to compute

1 1
1 0

1 1 n
1 0
Solution. (Exercise to fill in all details!)

The characteristic polynomial of A =
or
1 1
1 0
1+ 5
2
Fn
Fn 1

1 1 n 1
!
0
1 0

Fn+1
Fn
Hence,
1
0
1
0
= c1 v1 + c1v2.
1 5
2
(c1 = , c2 = )
5

= An 10 = n1 c1v1 + n2 c2v2
h
1
1+ 5
Fn = n1 c1 + n2 c2 =
2
5
Thats Binets formula.
But |2| < 1, and so Fn n1 c1 =
n
1+ 5
2
1 5
2
n i
.
n
.

1+ 5 n
1
In fact, Fn = round
. Dont you feel powerful!?
2
5
Practice problems
Problem 1. Find, if possible, the diagonalization of A =
Armin Straub
[email protected]
is 2 1.
The eigenvalues are 1 =

1.618 (the golden mean!) and 2 =
0.618.

Corresponding eigenvectors: v1 = 11 , v2 = 12
Write
F1
F0
0 2
4 2
Review for Midterm 3

Bring a number 2 pencil to the exam!
Extra help session: today and tomorrow, 47pm, in AH 441
Room assignments for Thursday, Nov 20, 7-8:15pm:
if your last name starts with A-E: 213 Greg Hall
if your last name starts with F-L: 100 Greg Hall
if your last name starts with M-Sh: 66 Library
if your last name starts with Si-Z: 103 Mumford Hall
Big topics:
Least squares
GramSchmidt
Determinants
Eigenvalues and eigenvectors
If v1,
, vn is an orthogonal basis of V , and x is in V , then
x = c1v1 +
+ cnvn
with
cj =
hx, v j i
.
hv j , v j i
Suppose that V is a subspace of W , and x is in W , then the orthogonal projection

of x onto V is given by
x = c1v1 +
+ cnvn
with
cj =
hx, v j i
.
hv j , v j i
The basis v1,

, vn has to be orthogonal for this formula!!
This decomposes x = x + x , where the error x is orthogonal to V .
in V
in V
(this
d ecom p osition is uniqu e)
x
x
v1
x
v2
Armin Straub
[email protected]
x with respect to the
The corresponding projection matrix represents x

standard basis.
Example 1.
(a) What is the orthogonal projection
Solution: The projection is
1
1 .
0
(b) What is the orthogonal projection

Solution: The projection is

1
1
h 1 , 1 i
0
0

1
1
h 1 , 1 i
0
0
Wrong approach!!
This is wrong because
(c)
0
0 .
0
1
1
1 , 1
1
0
1
of 1
0
1
of 1
0
1
1 +
0
0
1
0 , 1 ?
0
0
onto span
1
1
1 , 1 ?
1
0
onto span

1
1
h 1 , 1 i
1
0

1
1
h 1 , 1 i
1
1

0
1
1 = 0
0
1
are not orthogonal. (See next example!)
1
What is the orthogonal projection of 1
0
1
Solution: The projection is 1 .
1
1
1 , 1 ?
1
0
onto span
Wrong!!
1
1
h 1 , 1 i
0
0

1
1
h 1 , 1 i
0
0
1
1 +
0
1
1
Corrected : 1 , 1
1
0

1
1
h 1 , 1
0
0

1
1
h 1 , 1
0
0
1
1 +
i
0

1
1
h 1 , 1
1
0

1
1
h 1 , 1
1
1
0
1
1 , 0
1
0

0
1
h 1 , 0 i
1
0

0
1
h 1 , 0 i
1
1

1
1
1 = 1 +
i
0
1
1
2
1
3
1
(for instan ce, u sing G ram Schm idt)

0
1
0
0 = 1 + 0 0
1
0
1
(d) What( is the projection

matrix corresponding to orthogonal projection onto
)
span
1
0
1 , 1
0
0
1 0 0
Solution: The projection matrix is 0 1 0 .
0 0 0
1
1
0
0
1 , 0
What would GramSchmidt do? 1 , 1
0
0
0
0
(
1
What is the orthogonal projection of 1 onto span
1
(e)
Armin Straub
[email protected]
1
0
1 , 1 ?
0
0
Solution: The projection
1
1
1 0 0
is 0 1 0 1 = 1 .
0
1
0 0 0
The space of all nice functions with period 2 has the natural inner product hf ,
R 2
[in R n: hx, yi = x1 y1 +
+ xnyn]
gi = 0 f (x)g(x)dx.
The functions
are an orthogonal basis for this space.

Expanding a function f (x) in this basis produces its Fourier series
Example 2. How can we compute b2?
Solution.
b2sin(2x) is the orthogonal projection of f onto the span of sin (2x).
Hence:
b2
hf (x), sin (2x)i

=
=
hsin (2x), sin (2x)i
Least squares
2
f (x)sin(2x)dx
0
R 2
sin2 (2x)dx
0
x is a least squares solution of the system Ax = b.

x is such that Ax b is as small as possible.

ATAx = ATb
Example 3. Find the least squares line for the data points (2, 1), (5, 2), (7, 3), (8, 3).
Solution.
Looking for 1, 2 such that the line y = 1 + 2x best fits the data.
The equations yi = 1 + 2xi in matrix form:
1
1
1
1
y1
x1

y2
x2
=
y3
x3
2
y4
x4
d esign m atrix X
Armin Straub
[email protected]
ob servatio n
vector y
Here, we need to find a least squares solution to
1
1
1
1
X TX
1 1
2 5
X Ty =
Solving
4 22
22 142
9
57
1
2
2
5
7
8
1

1
2
2 = 3
3
1 2
1 1
1 5 =
7 8 1 7
1 8

1
1 1 1
2 =
5 7 8 3
3
, we find
1
2
2/7
5/14
4 22
22 142
9
57
GramSchmidt
Recipe. (GramSchmidt orthonormalization)
Given a basis a1,

, an, produce an orthonormal basis q1,
, qn.
b1 = a1,
b2 = a2 ha2, q1iq1,
b1
kb1k
b
q2 = 2
kb2k
b
q3 = 3
kb3k
q1 =
An orthogonal matrix is a square matrix Q with orthonormal columns.

Equivalently, QTQ = I (also true for non-square matrices).
Apply GramSchmidt to the (independent) columns of A to obtain the QR decomposition A = QR.
Q has orthonormal columns (the output vectors of GramSchmidt)
R = QTA is upper triangular
Armin Straub
[email protected]
1 1
A = 2 0 .
0 1
Solution. We apply GramSchmidt to the columns of A:
1 1
2 = q1
5 0
1
1
0 h 0 ,
1
1
1
4/5
1
1
1
q1iq1 = 0 2 = 2/5 ,
5 5 0
1
1
Hence: Q = [ q1
q2 ] =
5
4
45
And: R = QTA =
45
2
45
5
45
5
2
5
2
45
0
5
45
4/5
2/5 =
p
9/5
1

5
1 1
5
2 0 =
0
0 1
5
9
45
q2
Determinants
A is invertible
det (A) 0

1
det (A1) = det (A)

det (AT ) = det (A)
The determinant is characterized by:
the normalization det I = 1,
and how it is affected by elementary row operations:
Does not change the determinant.
Reverses the sign of the determinant.
(scaling) Multiply all entries in a row by s.
Multiplies the determinant by s.

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

R4R4 3 R3

2

Armin Straub
[email protected]
1
0
0
0
2
2
0
0

3 4
1 5
7
2 1 = 1 2 2 2
7
0 2
= 14
Cofactor expansion is another way to compute determinants.

1 2 0

3 1 2

2 0 1
0

= 2 3
+ (1)
2
+

2
2
1
1
= 2 (1) + (1) 1 0 = 1
Example 5. What is
1
1
0
2
1
2
3
0
1
2
3
0
4
5
1
5

1
0

0 3
2

Solution. The determinant is 0 because the matrix is not invertible (second and third
column are the same).
Eigenvalues and eigenvectors

Why? Because Ax = x
det (A I)
= 0.
(A I)x = 0.
The eigenspace of is Nul(A I).

It consists of all eigenvectors (plus 0) with eigenvalue .
Eigenvectors x1,
, xm of A corresponding to different eigenvalues are independent.
Useful for checking: sum of eigenvalues = sum of diagonal entries
Armin Straub
[email protected]
Welcome back!
The final exam is on Friday, December 12, 7-10pm
If you have a conflict (overlapping exam, or more than 2 exams within 24h), please email
Mahmood until Sunday to sign-up for the Monday conflict.

1 2
What are the eigenspaces of 0 3 ?
n o

= 1 has eigenspace Nul 00 22 = span 10
n o

2 2
= span 11
= 3 has eigenspace Nul
0 0
n o
INCORRECT: eigenspace span 10 , 11

Transition matrices
Powers of matrices can describe transition of a system.
Example 1. (review)
Fibonacci numbers Fn: 0, 1, 1, 2, 3, 5, 8, 13, 21, 34,
Hence:
Fn+1
Fn
Fn+1
Fn
n

F1
1 1
= 1 0
F0
Fn+1 = Fn + Fn1
1 1
1 0

Fn
Fn 1
Example 2. Consider a fixed population of people with or without a job. Suppose that,
each year, 50% of those unemployed find a job while 10% of those employed loose their
job.
What is the unemployment rate in the long term equilibrium?
Solution.
0.5
employed
0.9
0.1
no job
0.5
xt: proportion of population employed at time t (in years)

yt: proportion of population unemployed at time t

xt+1
0.9xt + 0.5yt
0.9 0.5
xt
=
=
yt+1
0.1xt + 0.5yt
0.1 0.5
yt
The matrix
0 .9 0 .5
0 .1 0 .5
is a Markov matrix. Its columns add to 1 and it has no negative entries.
Armin Straub
[email protected]
x
y
is an equilibrium if

x
y
0.9 0.5
0.1 0.5

x
y
x
y
is an eigenvector with eigenvalue 1.

n o

0.1 0.5
= span 51
Eigenspace of = 1: Nul
0.1 0.5
In other words,
Since x + y = 1, we conclude that
x
y
5/6
1/6
Hence, the unemployment rate in the long term equilibrium is 1/6.
Page rank
Googles success is based on an algorithm to rank webpages, the Page rank, named
after Google founder Larry Page.
The basic idea is to determine how likely it is that a web user randomly gets to a given
webpage. The webpages are then ranked by these probabilities.
Example 3. Suppose the internet consisted of only the four webpages A, B , C , D linked
as in the following graph:
Imagine a surfer following these links at random.
For the probability PR n(A) that she is at A (after n steps), we add:

the probability that she was at B (at exactly one time step before),
1
and left for A,
(thats PR n1(B) 2 )
the probability that she was at C, and left for A,
the probability that she was at D, and left for A.
1
Hence: PRn(A) = PRn1(B) 2 + PRn1(C) 1 + PRn1(D) 1
P R n(A)
P R n(B)
=
PR n(C)
P R n(D)
0
1
3
1
3
1
3
The PageRank
PR n 1(A)
0 0 0
PR n 1(B)
PR n 1(C)
0 0 1
PR n 1(D)
1
0 0
2

P R (A)
PR(A)
PR(B) P R (B)
vector PR(C) = P R (C)
P R (D)
PR(D)
1
2
1 0
is the long-term equilibrium.
It is an eigenvector of the Markov matrix with eigenvalue 1.
1
1
3
1
3
1
3
1
2
>
1 0 0 2
2
0 1 0 3
1 0
0
RREF
0 1 1
0 0 1 3
1
0 0 0 0
0 1
2
1
eigenspace of = 1 spanned by
Armin Straub
[email protected]
2
3
5
3
2
PR(A)
2
PR(B)
3
PR(C) = 16 5
3
PR(D)
1
0.375
0.125
=
0.313
0.188
This the PageRank vector.
The corresponding ranking of the webpages is A, C , D, B.
Practice problems
Problem 1. Can you see why 1 is an eigenvalue for every Markov matrix?
Problem 2. (just for fun) The real web contains pages which have no outgoing links.
In that case, our random surfer would get stuck (the transition matrix is not a Markov
matrix). Do you have an idea how to deal with this issue?
Armin Straub
[email protected]
Review
We model a surfer randomly clicking webpages.
Let PRn(A) be the probability that he is at A (after n steps).
1
1
PRn(A) = PRn1(B) 2 + PRn1(C) 1
0 2 1 0
PR n(A)
1
P R n 1(A)
PR n(B)
0 0 0
P R n 1(B)
3
=
PR n(C) 1
PR n 1(C)
3 0 0 1
PR n(D)
P R n 1(D)
1 1
0 0
3 2
+ PRn1(D)
0
1
=T
The transition matrix T is a Markov matrix.

Its columns add to 1 and it has no negative entries.
The Page rank of page A is PR(A) = PR(A).
(assum in g the lim it exists)
It is the probability that the surfer is at page A after n steps (with n ).
The PageRank
PR(A)
vector PR(B)
PR(C)
PR(D)
PR(A)
= T
satisfies PR(B)
PR(C)
PR(D)
PR(A)
PR(B)
.
PR(C)
PR(D)
It is an eigenvector of the transition matrix T with eigenvalue 1.
1
1
3
1
3
1
3

1
2
>
1 0 0 2
2
0 1 0 3
1 0
0
RREF
5
0 1 1
0
0
1
3
1
0
0
0
0
0 1
2
1
eigenspace of = 1 spanned by 35
3
1
2
0.375
PR(A)
2

PR(B)
3 0.125
PR(C) = 16 5 = 0.313 This
3
PR(D)
0.188
1
the PageRank vector.
The corresponding ranking of the webpages is A, C , D, B.

Remark 1. In practical situations, the system might be too large for finding the eigenvector by elimination.
Google reports having met about 60 trillion webpages
Googles search index is over 100,000,000 gigabytes
Number of Googles servers secret; about 2,500,000
More than 1,000,000,000 websites (i.e. hostnames; about 75% not active)
Armin Straub
[email protected]
Thus we have a gigantic but very sparse matrix.

An alternative to elimination is the power method:
If T is an (acyclic and irreducible) Markov matrix, then for any v0 the vectors T nv0
converge to an eigenvector with eigenvalue 1.
Here: T

1/4

=
T 1/4
1/4
1/4
1
2
1
3
1
3
1
3
0 0 0
0 0 1
1
0
0
2
1 0

3/8
1/12
=
1/3
5/24

P R(A)
P R(B)

PR(C) =
P R(D)
0.375
0.125
0.313
0.188
0.375
0.083
0.333
0.208
Note that the ranking of the webpages is already A, C , D, B if we stop here.

1/4

=
T 2 1/4
1/4
1/4

1/4

=
T 3 1/4
1/4
1/4
0.375
0.125
0.333
0.167
0.396
0.125
0.292
0.188
Remark 2.
If all entries of T are positive, then the power method is guaranteed to work.
In the context of PageRank, we can make sure that this is the case, by replacing T
with
0 2 1
1
0 0
3
(1 p)
1
3 0 0
1 1
0
3 2
0
+ p
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
1
4
Just to make sure: still a Markov matrix, now with positive entries
Google used to use p = 0.15.
Why does T nv0 converge to an eigenvector with eigenvalue 1?

Under the assumptions on T , its other eigenvalues satisfy || < 1.
Now, think in terms of a basis x1,
, xn of eigenvectors:
m
T m(c1x1 +
+ cnxn) = c1m
1 x1 +
+ cnn xn
As m increases, the terms with m

i for i 1 go to zero, and what is left over is an eigenvector
with eigenvalue 1.
Armin Straub
[email protected]
Linear differential equations

Example 3. Which functions y(t) satisfy the differential equation y = y?
Solution: y(t) = et and, more generally, y(t) = Cet.
(And nothing else.)
Recall from Calculus the Taylor series et = 1 + t + 2! + 3! +
t2
t3
Example 4. The differential equation y = ay with initial condition y(0) = C is solved

by y(t) = Ceat.
(T h is solution is u nique.)
Why? Because y (t) = aC ea t = ay(t) and y(0) = C.
Example 5. Our goal is to solve (systems of) differential equations like:

y1 = 2y1
y2 = y1 +3y2 +y3
y3 = y1 +y2 +3y3
y1(0) = 1
y2(0) = 0
y3(0) = 2
In matrix form:
2 0 0
y = 1 3 1 y,
1 1 3
1
y(0) = 0
2
Key idea: to solve y = Ay, introduce eA t
Armin Straub
[email protected]
Review of diagonalization
Put the eigenvectors x1,
, xn as columns into a matrix P .
Axi = ixi
|
A x1
|
In summary: AP = PD
|
|
|
xn = 1x1 nxn
|
|
|
1
|
|
= x1 xn

|
|
n
Let A be n n with independent eigenvectors x1,

, xn.

Example 6. Diagonalize the following matrix, if possible.
2 0 0
A = 1 3 1
1 1 3
Solution.
A has eigenvalues 2 and 4.
= 2:
0 0 0
1 1 1
1 1 1

eigenspace span
2 0
0
= 4: 1 1 1
eigenspace
1 1 1
1 1 0
2
= 1 0 1 and D = 2
4
0 1 1
(W e d id that in an earlier class!)
1
1
1 , 0
1
0
)
0
1
1
span
A = PDP 1
For many applications, it is not needed to compute P 1 explicitly.
We can check this by verifying AP = PD:
2 0 0
1 1 0
1 1 0
2
1 3 1 1 0 1 = 1 0 1
2
1 1 3
0 1 1
0 1 1
4
Armin Straub
[email protected]
Review
Let A be n n with independent eigenvectors x1,
, xn.
Why? We need to see that AP = PD:
Axi = ixi
A x1
|
|
|
|
xn = 1x1 nxn
|
|
|
1
|
|
= x1 xn

n
|
|
The differential equation y = ay with initial condition y(0) = C

is solved by y(t) = Ceat.
Recall from Calculus the Taylor series et = 1 + t + 2! + 3! +
t2
Goal: similar treatment
2

y = 1
1
t3
of systems like:

0 0
1
y(0) = 2
3 1 y,
1 3
1
Definition 1. Let A be n n. The matrix exponential is

eA = I + A +
Then:
1 2 1 3
A + A +
2!
3!
d At
e = AeA t
dt

1 22 1 33
d At
d
I + At + A t + A t +
e
=
2!
dt
dt
3!
1 2
1 32
= A + A t + A t + = AeA t
1!
2!
Why?
The solution to y = Ay, y(0) = y0 is y(t) = eAty0.

Why? Because y (t) = AeAty0 = Ay(t) and y(0) = e0Ay0 = y0.
Example 2. If A =
2 0
0 5
, then:
#
"
#
"

2 0
2 0
1
e
2
0
2
eA =
+
+
+ =
0 5
2! 0 52
0 e5
#
"
#
"

1 (2t)2
0
1 0
2t 0
e2t 0
At
+ =
e =
+
+
0 1
0 5t
2!
0
(5t)2
0 e5t

1 0
0 1
Clearly, this works to obtain eD for any diagonal matrix D.

Armin Straub
[email protected]
Example 3. Suppose A = PDP 1. Then, what is An?

Solution.
First, note that A2 = (PDP 1)(PDP 1) = PD 2P 1.
Likewise, An = PDnP 1.
(The point being that D n is trivial to compute because D is diagonal.)
Theorem 4. Suppose A = PDP 1. Then, eA = PeDP 1.

Why? Recall that An = PD nP 1.
1 2 1 3
A + A +
2!
3!
1
1
1
= I + PDP
+ PD 2P 1 + PD 3P 1 +
2!
3!

1 2 1 3
= P I + D + D + D + P 1 = P eDP 1
2!
3!
eA = I + A +
Example 5. Solve the differential equation

0 1
1
y =
y,
y(0) =
.
1 0
0
Solution. The solution to y = Ay, y(0) = y0 is y(t) = eA ty0.
Diagonalize A =

1

1
0 1
1 0

= 2 1,

so the eigenvalues are 1

n o

1 1
= span 11
= 1 has eigenspace Nul
1 1
n

o

= 1 has eigenspace Nul 11 11 = span 1
1
Hence, A = PDP 1 with P =

Compute the solution y = eAty0:
1 1
1 1
y = PeDtP 1"y0
#

t
1 1
e 0 1
=
1 1
0 et 2
#

" t
1 1 1
e 0
=
2 1 1
0 et
"
#

t
1 1 1
e
=
2 1 1
et
"
#
t
t
1 e +e
=
2 et et
Armin Straub
[email protected]
and D =
1 1
1 1

1
1

1 0
0 1
1
0
Example 6. Solve the differential equation
2 0 0
y = 1 3 1 y,
1 1 3
1
y(0) = 2 .
1
Solution.
Recall that the solution to y = Ay, y(0) = y0 is y = eA ty0.
A has eigenvalues 2 and 4.
= 2:
0 0 0
1 1 1
1 1 1
eigenspace span
2 0
0
= 4: 1 1 1
1 1 1
A = PDP
(W e d id that in an earlier class!)
with P
eigenspace span
1 1 0
= 1 0 1 ,
0 1 1
y = eAty0 = PeD tP 1 y0
2t
e
1 1 0
= 1 0 1
0 1 1
2t
e
1 1 0
= 1 0 1
0 1 1
1 1 0
e2t
= 1 0 1 0
0 1 1
e4t
e2t
y = e2t + e4t
e4t
1
1
1 , 0
1
0
Compute the solution y = eAty0:
Check (optional) that
D =
)
0
1
1
2
2
4

1 1 0 1 1
1 0 1 2
e2t
4t
0 1 1
1
e

1

0
e2t
1
e4t

e2t
2t
= e + e4t
e4t
indeed solves the original problem:
2e2t
e2t
2
0
0
y = 2e2t + 4e4t 1 3 1 e2t + e4t

1 1 3
4e4t
e4t
Remark 7. The matrix exponential shares many other properties of the usual exponential:
eA is invertible and (eA)1 = eA
eAeB = eA+B = eBeA if AB = BA
Armin Straub
[email protected]
Ending the Halloween torture
perimeter = 4
perimeter = 4
perimeter =
perimeter =
perimeter = 4
perimeter = 4
perimeter = 4
perimeter =
perimeter =
perimeter =
Length of the graph of y(x) on [a, b] is
b
a
1 + y (x)2 dx.
While the blue curve does converge to the circle,

its derivative does not converge!
In the language of functional analysis:
The linear map D: y y is not continuous!
(That is, two functions can be close without their derivatives being close.)
Even more extreme examples are provided by fractals. The Koch snowflake:
Armin Straub
[email protected]
Its perimeter is infinite!

Why? At each iteration, the perimeter gets multiplied by 4/3.
Its boundary has dimension log3 (4) 1.262!!

the effect of zooming in by a factor of 3
3
d = 1 = log3 (3)
d = 2 = log3 (3)
d = log3 (4)
Such fractal behaviour is also observed when attempting to measure the length of
a coastline: the measured length increases by a factor when using a smaller scale.
See: http://en.wikipedia.org/wiki/Coastline_paradox
Armin Straub
[email protected]

Linear Algebra Lectures

Uploaded by

Copyright:

Available Formats

Linear Algebra Lectures

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Linear Algebra Lectures

Uploaded by

Copyright:

Available Formats

Introduction to systems of linear equations

linear: 3x1 5x2 = 2

4x1 6x2 = x1x2

not linear: x1x2

Theorem 5. A linear system has either

one unique solution, or

infinitely many solutions.

How to solve systems of linear equations

Once in this triangular form, we find the solutions by back-substitution:

Example 9. The same approach works for more complicated systems.

Definition 10. An elementary row operation is one of the following:

(scaling) Multiply all entries in a row by a nonzero constant.

Instead of back-substitution, we can continue with row operations.

Finally, R1 R1 + R2, R2 2 R2 results in:

Row reduction and echelon forms

matrix in echelon form.

( stands for any value, and for any nonzero value.)

16. Are the following matrices in echelon form?

NOPE (but it is after exchanging

the first two rows)

Related and extra material

Suggested practice exercise: 1, 4, 5, 10, 11 from Section 1.3

Artur Avila, Manjul Bhargava, Martin Hairer, Maryam Mirzakhani

To solve a system, we perform row reduction.

Echelon form in general:

The leading terms in each row are the pivots.

Row reduction and echelon forms, continued

Example 3. Are the following

matrices in reduced echelon form?

are different row equivalent echelon forms.

(R1 R2 would b e another option; try it!)

Finally, R3 R3 2 R2 produces the echelon form:

Then, R2 R2 R3 and R1 R1 2R3, gives:

Solution of linear systems via row reduction

3x2 6x3 +6x4 +4x5 = 5

Solution. The augmented matrix is

The pivot variables are x1, x2, x5.

Related and extra material

Suggested practice exercise:

3x2 6x3 +6x4 +4x5 = 5

The computational part is to start with the augmented matrix

Questions of existence and uniqueness

3x2 6x3 +6x4 +4x5 = 5

Solution. In the course of an earlier example, we obtained the echelon form:

Hence, it is consistent (imagine doing back-substitution to get a solution).

The system is consistent if and only if h = 3 .

Brief summary of what we learned so far

(1) Write the augmented matrix of the system.

(3) Continue row reduction to obtain the reduced echelon form.

Questions to check our understanding

There is no more than one pivot in any column.

There cannot be more free variables than pivot variables.

The geometry of linear equations

Example 6. (Geometric description of R2) A vector

represents the point

Example 8. Linear combinations of v1, v2, v3 include:

Solution. We have to find c1 and c2 such that

The row and column picture

This example has the unique solution x = 2, y = 3.

Thats because any vector