WEBVTT

00:00:00.500 --> 00:00:01.720
GILBERT STRANG: OK, thanks.

00:00:01.720 --> 00:00:08.290
Here's a second video
that involves the matrix

00:00:08.290 --> 00:00:10.040
exponential.

00:00:10.040 --> 00:00:13.650
But it has a new idea
in it, a basic new idea.

00:00:13.650 --> 00:00:19.860
And that idea is two matrices
being called "similar."

00:00:19.860 --> 00:00:23.520
So that word "similar"
has a specific meaning,

00:00:23.520 --> 00:00:27.820
that a matrix A, is
similar to another matrix

00:00:27.820 --> 00:00:33.420
B, if B comes from A this way.

00:00:33.420 --> 00:00:35.480
Notice this way.

00:00:35.480 --> 00:00:40.290
It means there's some matrix M--
could be any invertible matrix.

00:00:40.290 --> 00:00:43.240
So that I take A,
multiply on the right

00:00:43.240 --> 00:00:45.920
by M and on the
left by M inverse.

00:00:45.920 --> 00:00:48.130
That'd probably give
me a new matrix.

00:00:48.130 --> 00:00:53.320
Call it B. That matrix
is called "similar" to B.

00:00:53.320 --> 00:00:57.720
I'll show you examples of
matrices that are similar.

00:00:57.720 --> 00:01:02.570
But first is to get
this definition in mind.

00:01:02.570 --> 00:01:06.450
So in general, a
lot of matrices are

00:01:06.450 --> 00:01:13.810
similar to-- if I have a certain
matrix A, I can take any M,

00:01:13.810 --> 00:01:16.860
and I'll get a similar
matrix B. So there

00:01:16.860 --> 00:01:18.900
are lots of similar matrices.

00:01:18.900 --> 00:01:22.510
And the point is all
those similar matrices

00:01:22.510 --> 00:01:24.420
have the same eigenvalues.

00:01:24.420 --> 00:01:26.500
So there's a little
family of matrices

00:01:26.500 --> 00:01:29.920
there, all similar
to each other and all

00:01:29.920 --> 00:01:31.610
with the same eigenvalues.

00:01:31.610 --> 00:01:34.570
Why do they have the
same eigenvalues?

00:01:34.570 --> 00:01:37.430
I'll just show you, one line.

00:01:37.430 --> 00:01:42.360
Suppose B has an
eigenvalue of lambda.

00:01:42.360 --> 00:01:45.800
So B is M inverse AM.

00:01:45.800 --> 00:01:46.720
So I have this.

00:01:46.720 --> 00:01:49.560
M inverse AMx is lambda x.

00:01:49.560 --> 00:01:51.634
That's Bx.

00:01:51.634 --> 00:01:53.520
B has an eigenvalue of lambda.

00:01:53.520 --> 00:01:56.970
I want to show that A has
an eigenvalue of lambda.

00:01:56.970 --> 00:01:57.590
OK.

00:01:57.590 --> 00:01:59.560
So I look at this.

00:01:59.560 --> 00:02:05.600
I multiply both sides
by M. That cancels this.

00:02:05.600 --> 00:02:10.440
So when I multiply by M,
this is gone, and I have AMx.

00:02:10.440 --> 00:02:12.890
But the M shows up on
the right-hand side,

00:02:12.890 --> 00:02:14.120
I have lambda Mx.

00:02:17.100 --> 00:02:20.300
Now I would just look
at that, and I say, yes.

00:02:20.300 --> 00:02:28.370
A has an eigenvector-- Mx
with eigenvalue lambda.

00:02:28.370 --> 00:02:33.770
A times that vector is
lambda times that vector.

00:02:33.770 --> 00:02:39.440
So lambda is an eigenvalue of A.
It has a different eigenvector,

00:02:39.440 --> 00:02:40.620
of course.

00:02:40.620 --> 00:02:43.060
If matrices have
the same eigenvalues

00:02:43.060 --> 00:02:47.100
and the same eigenvectors,
that's the same matrix.

00:02:47.100 --> 00:02:51.920
But if I do this,
allow an M matrix

00:02:51.920 --> 00:02:56.310
to get in there, that
changes the eigenvectors.

00:02:56.310 --> 00:02:59.630
Here they were
originally x for B.

00:02:59.630 --> 00:03:02.350
And now for A,
they're M times x.

00:03:02.350 --> 00:03:06.930
It does not change the
eigenvalues because of this M

00:03:06.930 --> 00:03:10.320
on both sides allowed
me to bring M over

00:03:10.320 --> 00:03:13.175
to the right-hand side
and make that work.

00:03:13.175 --> 00:03:13.675
OK.

00:03:16.310 --> 00:03:18.980
Here are some similar matrices.

00:03:18.980 --> 00:03:20.570
Let me take some.

00:03:20.570 --> 00:03:23.335
So these will be all similar.

00:03:26.930 --> 00:03:31.200
Say 2, 3, 0, 4.

00:03:31.200 --> 00:03:31.870
OK?

00:03:31.870 --> 00:03:36.580
That's a matrix A. I can see
its eigenvalues are 2 and 4.

00:03:36.580 --> 00:03:43.170
Well, I know that it will be
similar to the diagonal matrix.

00:03:43.170 --> 00:03:47.800
So there is some matrix
M that connects this one

00:03:47.800 --> 00:03:52.050
with this one, connects
this A with that B. Well,

00:03:52.050 --> 00:03:56.660
that B is really capital lambda.

00:03:56.660 --> 00:03:59.440
And we know what matrix
connects the original A

00:03:59.440 --> 00:04:04.020
to its eigenvalue matrix.

00:04:04.020 --> 00:04:05.820
What is the M that does that?

00:04:05.820 --> 00:04:08.110
It's the eigenvector matrix.

00:04:08.110 --> 00:04:12.140
So to get this particular--
to get this guy,

00:04:12.140 --> 00:04:18.450
starting from here, I use
M is V for this example

00:04:18.450 --> 00:04:21.220
to produce that.

00:04:21.220 --> 00:04:23.540
Then B is lambda.

00:04:23.540 --> 00:04:25.820
But there are other
possibilities.

00:04:25.820 --> 00:04:27.520
So let me see.

00:04:27.520 --> 00:04:33.230
I think probably a
matrix is-- there

00:04:33.230 --> 00:04:34.645
is the matrix, A transpose.

00:04:37.190 --> 00:04:39.170
Is that similar to A?

00:04:39.170 --> 00:04:41.030
Is A transpose similar to A?

00:04:41.030 --> 00:04:43.610
Well, answer-- yes.

00:04:43.610 --> 00:04:48.720
The transpose matrix has those
same eigenvalues, 2 and 4,

00:04:48.720 --> 00:04:51.850
and different eigenvectors.

00:04:51.850 --> 00:04:58.600
And those eigenvectors
would connect the original A

00:04:58.600 --> 00:05:02.610
and this A or that A transpose.

00:05:02.610 --> 00:05:06.040
So the transpose of a matrix
is similar to the matrix.

00:05:06.040 --> 00:05:09.470
What about if I
change the order?

00:05:09.470 --> 00:05:11.660
4, 0, 0, 2.

00:05:14.240 --> 00:05:18.084
So I've just flipped
the 2 and the 4,

00:05:18.084 --> 00:05:20.083
but of course I haven't
changed the eigenvalues.

00:05:22.700 --> 00:05:25.620
You could find the
M that does that.

00:05:25.620 --> 00:05:29.180
You can find an M so that
if I multiply on the right

00:05:29.180 --> 00:05:32.900
by M and on the left by M
inverse, it flips those.

00:05:32.900 --> 00:05:35.060
So there's another
matrix similar.

00:05:35.060 --> 00:05:37.270
Oh, there could be plenty more.

00:05:37.270 --> 00:05:41.580
All I want to do is have
the eigenvalues be 4 and 2.

00:05:41.580 --> 00:05:43.700
Shall I just create some more?

00:05:43.700 --> 00:05:46.930
Here is a 0, 6.

00:05:46.930 --> 00:05:49.280
I wanted to get the trace right.

00:05:49.280 --> 00:05:52.210
4 plus 2 matches 0 plus 6.

00:05:52.210 --> 00:05:54.000
Now I have to get the
determinant right.

00:05:54.000 --> 00:05:56.700
That has a determinant of 8.

00:05:56.700 --> 00:06:01.440
What about a 2 and
a minus 4 there?

00:06:01.440 --> 00:06:04.430
I think I've got the
trace correct-- 6.

00:06:04.430 --> 00:06:07.260
And I've got the
determinant correct-- 8.

00:06:07.260 --> 00:06:08.890
And there the determinant is 8.

00:06:08.890 --> 00:06:10.620
So that would be
a similar matrix.

00:06:10.620 --> 00:06:12.180
All similar matrices.

00:06:12.180 --> 00:06:17.610
A family of similar matrices
with the eigenvalues 4 and 2.

00:06:17.610 --> 00:06:25.200
So I want to do another
example of similar matrices.

00:06:25.200 --> 00:06:27.130
What will be different
in this example

00:06:27.130 --> 00:06:29.900
is there'll be
missing eigenvectors.

00:06:29.900 --> 00:06:36.060
So let me say, 2, 2, 0, 1.

00:06:36.060 --> 00:06:44.770
So that has eigenvalues 2 and
2 but only one eigenvector.

00:06:44.770 --> 00:06:48.940
Here is another
matrix like that.

00:06:48.940 --> 00:06:52.420
Say, so the trace should be 4.

00:06:52.420 --> 00:06:54.660
The determinant should be 4.

00:06:54.660 --> 00:06:58.360
So maybe I put a 2
and a minus 2 there.

00:06:58.360 --> 00:07:01.510
I think that has the
correct trace, 4,

00:07:01.510 --> 00:07:04.540
and the great
determinant, also 4.

00:07:04.540 --> 00:07:11.260
So that will have eigenvalues 2
and 2 and only one eigenvector,

00:07:11.260 --> 00:07:12.560
so it's similar to this.

00:07:12.560 --> 00:07:14.250
Now here's the point.

00:07:14.250 --> 00:07:21.580
You might say, what
about 2, 2, 0, 0.

00:07:21.580 --> 00:07:24.490
That has the
correct eigenvalues,

00:07:24.490 --> 00:07:26.050
but it's not similar.

00:07:26.050 --> 00:07:31.040
There's no matrix M that
connects that diagonal matrix

00:07:31.040 --> 00:07:34.250
with these other matrices.

00:07:34.250 --> 00:07:37.610
That matrix has no
missing eigenvectors.

00:07:37.610 --> 00:07:40.770
These matrices have one
missing eigenvector.

00:07:40.770 --> 00:07:43.380
What's called the Jordan form.

00:07:43.380 --> 00:07:44.860
The Jordan form.

00:07:48.110 --> 00:07:51.130
So that didn't belong.

00:07:51.130 --> 00:07:52.840
That's not in that family.

00:07:52.840 --> 00:07:58.620
The Jordan form is--
you could say-- well,

00:07:58.620 --> 00:08:01.070
that'll be the Jordan form.

00:08:01.070 --> 00:08:06.330
The most beautiful member of
the family is the Jordan form.

00:08:06.330 --> 00:08:09.210
So I have a whole lot of
matrices that are similar.

00:08:09.210 --> 00:08:12.960
That is the most beautiful,
but it's not in the family.

00:08:12.960 --> 00:08:19.480
It's related but
not in the family.

00:08:19.480 --> 00:08:21.470
It's not similar to those.

00:08:21.470 --> 00:08:23.210
And the best one
would be this one.

00:08:23.210 --> 00:08:28.850
So the Jordan form would be
that one with the eigenvalues

00:08:28.850 --> 00:08:30.920
on the diagonal.

00:08:30.920 --> 00:08:33.789
But because there's a
missing eigenvector,

00:08:33.789 --> 00:08:36.429
there has to be a
reason for that.

00:08:36.429 --> 00:08:40.820
And it's in the 1 there,
and I can't have a 0 there.

00:08:40.820 --> 00:08:41.700
OK.

00:08:41.700 --> 00:08:44.860
So that's the idea
of similar matrices.

00:08:44.860 --> 00:08:51.120
And now I do have one more
important note, a caution

00:08:51.120 --> 00:08:54.180
about matrix exponentials.

00:08:54.180 --> 00:08:58.065
Can I just tell you this
caution, this caution?

00:09:03.240 --> 00:09:08.340
If I look at e to the
A times e to the B.

00:09:08.340 --> 00:09:11.920
The exponential of A times
the exponential of B.

00:09:11.920 --> 00:09:22.130
My caution is that usually that
is not e to the B, e to the A.

00:09:22.130 --> 00:09:24.890
If I put B and A in
the opposite order,

00:09:24.890 --> 00:09:26.120
I get something different.

00:09:26.120 --> 00:09:32.780
And it's also not
e to the A plus B.

00:09:32.780 --> 00:09:35.900
Those are all different.

00:09:35.900 --> 00:09:40.550
Which, if I had 1 by 1, just
numbers here, of course,

00:09:40.550 --> 00:09:43.310
that's the great rule
for exponentials.

00:09:43.310 --> 00:09:49.580
But for matrix exponentials,
that rule doesn't work.

00:09:49.580 --> 00:09:53.480
That is not the same
as e to the A plus B.

00:09:53.480 --> 00:09:55.380
And I can show you why.

00:09:55.380 --> 00:10:03.760
e to the A is I plus A plus
1/2 A squared and so on.

00:10:03.760 --> 00:10:10.790
e to the B is I plus B plus
1/2 B squared and so on.

00:10:10.790 --> 00:10:13.130
And I do that multiplication.

00:10:13.130 --> 00:10:22.100
And I get I. And I get an
A. And I get a B times I.

00:10:22.100 --> 00:10:30.200
And now I get 1/2 B squared
and an AB and 1/2 A squared.

00:10:30.200 --> 00:10:31.690
Can I put those down?

00:10:31.690 --> 00:10:36.680
1/2 A squared, and
there's an A times a B.

00:10:36.680 --> 00:10:40.420
And there's a 1/2 B squared.

00:10:40.420 --> 00:10:42.210
OK.

00:10:42.210 --> 00:10:43.780
This makes the point.

00:10:43.780 --> 00:10:47.370
If I multiply the
exponentials in this order,

00:10:47.370 --> 00:10:52.100
I get A times B. What
if I multiply them

00:10:52.100 --> 00:10:54.960
in the other order,
in that order?

00:10:54.960 --> 00:10:57.200
If I multiply e to the
B times e to the A,

00:10:57.200 --> 00:11:01.500
then the B's will be
out in front of the A's.

00:11:01.500 --> 00:11:08.190
And this would become a
BA, which can be different.

00:11:08.190 --> 00:11:12.880
So already I see that
the two are different.

00:11:12.880 --> 00:11:15.545
Here is e to the
A, e to the B. It

00:11:15.545 --> 00:11:19.450
has A before B. If
I do it this way,

00:11:19.450 --> 00:11:22.755
it'll have B before A.
If I do it this way,

00:11:22.755 --> 00:11:23.630
it'll have a mixture.

00:11:29.210 --> 00:11:34.900
So e to the A plus B
will have a I and an A

00:11:34.900 --> 00:11:40.050
and a B and a 1/2
A plus B squared.

00:11:40.050 --> 00:11:47.120
So that'll be 1/2 of A squared
plus AB plus BA plus B squared.

00:11:50.260 --> 00:11:51.680
Different again.

00:11:51.680 --> 00:11:57.860
Now I have a sort of
symmetric mixture of A and B.

00:11:57.860 --> 00:12:01.420
In this case, I had A
before B. In this case,

00:12:01.420 --> 00:12:04.700
I had B on the left side of A.

00:12:04.700 --> 00:12:06.490
So all three of
those are different,

00:12:06.490 --> 00:12:11.140
even in this term
of the series that

00:12:11.140 --> 00:12:13.600
defines those exponentials.

00:12:13.600 --> 00:12:20.050
And that means that
systems of equations,

00:12:20.050 --> 00:12:28.140
if the coefficients change over
time, are definitely harder.

00:12:28.140 --> 00:12:33.520
We were able to solve
dy dt equals, say,

00:12:33.520 --> 00:12:38.520
cosine of t times y.

00:12:38.520 --> 00:12:44.020
Do you remember how-- that
this was solvable for a 1 by 1.

00:12:44.020 --> 00:12:49.450
We put the exponent--
the solution was y

00:12:49.450 --> 00:12:52.900
is e to the-- we
integrated cosine t

00:12:52.900 --> 00:12:58.690
and got sine t times y of 0.

00:13:01.280 --> 00:13:04.340
e to the sine t--

00:13:04.340 --> 00:13:06.590
Can I just think of putting
that into the differential

00:13:06.590 --> 00:13:08.290
equation-- its derivative.

00:13:08.290 --> 00:13:13.720
The derivative of e to the
sine t will be e to the sine t.

00:13:13.720 --> 00:13:15.590
I'm using the chain rule.

00:13:15.590 --> 00:13:19.430
The derivative of e to the
sine t will be e to the sine t

00:13:19.430 --> 00:13:25.040
again, times the derivative
of sine t, which is cos t,

00:13:25.040 --> 00:13:26.590
so it works.

00:13:26.590 --> 00:13:29.500
That's fine as a solution.

00:13:29.500 --> 00:13:34.000
But if I have matrices
here-- if I have matrices,

00:13:34.000 --> 00:13:37.980
then the whole thing goes wrong.

00:13:37.980 --> 00:13:40.290
You could say that the
chain rule goes wrong.

00:13:43.490 --> 00:13:47.590
You can't put the
integral up there

00:13:47.590 --> 00:13:52.530
and then take the derivative
and expect it to come back down.

00:13:52.530 --> 00:13:56.670
The chain rule will not work
for matrix exponentials,

00:13:56.670 --> 00:13:58.180
the simple chain rule.

00:13:58.180 --> 00:14:01.380
And the fact is
that we don't have

00:14:01.380 --> 00:14:06.640
nice formulas for the
solutions to linear systems

00:14:06.640 --> 00:14:10.610
with time-varying coefficients.

00:14:10.610 --> 00:14:15.600
That has become a harder problem
when we went from one equation

00:14:15.600 --> 00:14:18.010
to a system of an equation.

00:14:18.010 --> 00:14:25.770
So this is the caution slide
about matrix exponentials.

00:14:25.770 --> 00:14:27.530
They're beautiful.

00:14:27.530 --> 00:14:31.120
They work perfectly if you
just have one matrix A.

00:14:31.120 --> 00:14:33.670
But if somehow two
matrices are in there

00:14:33.670 --> 00:14:36.250
or a bunch of
different matrices,

00:14:36.250 --> 00:14:43.880
then you lose the good rules,
and you lose the solution.

00:14:43.880 --> 00:14:44.380
OK.

00:14:44.380 --> 00:14:46.190
Thank you.