NOTES
ON LINEAR ALGEBRA
CONTENTS:
[12]
EIGENVALUES / EIGENVECTORS
[12] EIGENVALUES / EIGENVECTORS
In
the last section, we examined when two vectors are a basis for the plane. Let’s
recall what this means. The plane is 2-dimensional. So, we expect that we
should be able to specify any vector with two pieces of information, say East
component and North component. This corresponds to the standard basis (x-axis,
y-axis).
In
[11] we saw that, as long as W1 and W2 are not in the same direction, they are
a basis for the plane. This means we can write any vector V as V = a W1 + b W2,
where a and b are numbers that can be determined, and depend on V.
However,
nowhere in [11] did we discuss why we would want to use a basis other than the
standard x-axis, y-axis.
The
reason is geometry. Often we’ll be studying certain matrices that model
physical systems. Those physical systems may have certain axes of symmetry,
which will often manifest itself in the matrix. And what we will find is that
the matrix looks more ‘natural’, more ‘symmetric’, if we change basis.
Let
A be a matrix acting on a vector v. What can we say about the vector Av?
In
general, not much, unless I give you information about A and v. A vector
encodes two pieces of information: magnitude, and direction. When we apply a
matrix to a vector, we get a new vector. Usually, that vector will have a
different magnitude and a different direction.
In
terms of computations, this is often unfortunate. We may be interested in some
iterative system, where we might have A100 v or A6022045
v. If Av is in a different direction than v, we have no quick and easy way to
calculate A2 v. Why? We know the magnitude and direction of Av. But
we know nothing about A(Av).
If
Av is in the same direction as v, however, it’s a different ballgame. Let’s say
Av = 3v. Then we can calculate Aany
power v easily.
For
example:
A2
v = A
(Av)
= A
(3 v)
= 3
(A v)
= 3
(3 v)
= 32
v
A3
v = A
(A2 v)
= A
(32 v) by the
previous calculation
= 32
(A v)
= 32
(3 v)
= 33
v
A4
v = A
(A3 v)
= A
(33 v) by the
previous calculation
= 33
(A v)
= 33
(3 v)
= 34
v
A5
v = A
(A4 v)
= A
(34 v) by the
previous calculation
= 34
(A v)
= 34
(3 v)
= 35
v
And
a similar argument shows A6 = 36 v, ..., An v
= 3n v.
DEFINITION OF
EIGENVALUE / EIGENVECTOR:
We say v is an
eigenvector and l is its eigenvalue if
(1)
A v = l v
(2)
v is not the zero vector
Note
that a vector is an eigenvector relative to a given matrix. For example,
(2 1) (1) (3) (1)
(1 2) (1) = (3) = 3 (1)
(2 1) (1) (1) (1)
(1 2) (-1) = (-1) = 1 (-1)
So
(1,1) is an eigenvector with eigenvalue 3, while
(1,-1) is an eigenvector with eigenvalue 1. But if we
consider a different matrix
(1 2) (1) (2)
(3 4) (1) = (7)
we
see (1,1) is not an eigenvector.
Why
do we exclude the zero vector? The reason is the zero vector would be an
eigenvector for ANY matrix, and ANY number would be its eigenvalue.
Let Z be the zero vector. Then
A Z = Z = 2 Z = 3 Z = -2342.324 Z
and
Z would not have a unique eigenvalue. Again, this is
a matter of notation / convenience. We will see later its just easier if all eigenvalues are non-zero, for we’ll prove certain nice
facts about them. For example, a beautiful theorem of Linear Algebra (The
Principal Axis Theorem) states that if you have a symmetric real matrix, you
can find a basis of mutually perpendicular vectors Vi such that each Vi is an
eigenvector of the matrix A! Wow! This means we can compute the action of large
powers of symmetric matrices with very little work.
Let’s
see now how eigenvectors can make life liveable.
We’ll work with the matrix
A = (2 1)
(1 2)
We
saw above that if V = (1,1) and W = (1,-1) then
A V = 3 V
A W = 1 W
V
and W are not in the same direction, so they’re a basis for the plane. So, if
you give me any vector (x,y), I can find a and b
(depending on x and y) such that
(x,y) = a
V + b W.
Why
is this helpful?
Let’s
build up in stages.
A(V + W) = AV + AW
= 3V + 1W = 3V + W
A(2V + 11 W) = A(2V) + A(11 W)
= 2 (A V) + 11 (A W)
= 2 (3 V) + 11 W
=
6 V + 11 W
A4 (2V + 11 W) = B
(2 V + 11 W), where B = A4
= B (2V) + B (11 W)
= 2 (B V) + 11 (B W)
= 2 (A4 V) + 11 (A4
W)
= 2 (34 V) + 11 W
NOTE:
we have the rule
A (X + Y) = A
X + AY
This
is still true for powers of A, as A4, A5, etc. are also
matrices:
A4 (X + Y) = A4 X + A4 Y
So,
in full generality:
Eq 12.1 A n (aV + bW) = An
(aV) + An (bW)
= a (An V) + b (An
W)
= a (3n V) + b W
So
we see how easy it is to calculate. Let’s take n = 2,000,000, and consider A U,
where U = (2,3). Now, A U is not in the same direction as U, nor is A2
U in the same direction as U or A U, etc., so we have 2,000,000 matrix
multiplications to do! That’s a lot.
The
other way is, FIRST, we express U in terms of our nice basis V, W. We do this
by Gaussian Elimination:
U = a V + b W
(2) (1) (1)
(3) = a (1) + b(-1)
(2) (1a) (1b)
(3) = (1a) + (-1b)
(2) (1a
+ 1b)
(3) = (1a + -1b)
(2) (1 1) (a)
(3) = (1 -1) (b)
And
then we do Gaussian Elimination. And now we are done! By Eq 12.1
we are done – all we have to do is put in the values of a and b, and that n =
2,000,000.
Doing
the calculation this way is about two steps; the other way is 2,000,000. This
is a phenomenal savings. The long way is beyond the strength of the computers
-–the numbers are just too large.
Why
are we able to have such a savings? This is not a trivial point – it’s one of
the strengths of Linear Algebra. Linear Algebra is an efficient way of
arranging computations. The long way involves lots of hidden cancellation,
cancellation that never survives to the end when we group everything.
An
example might help. Consider doing the following addition:
1
+5 -
5
+6 -6
+7 -7
+8 -8
+9 -9
If
we add horizontally, each row is zero but the first. It doesn’t matter how many
rows I’ve got, the final answer is just going to be one. There’s only one
computation to do.
What
if we add vertically?
The
we get 6 + 1 +1 + 1 + 1 - 9 = 1. We get the same answer, but it’s many more
steps. The reason is we add 5 then subtract 5. We add 6 then subtract 6. We add
7 then subtract 7. We add 8 then subtract 8. We add 9 then subtract 9. We keep
doing things that cancel, but we don’t realize it, and hence have to go thru
all the steps.