ANNOUNCER: The following content
is provided under a Creative Commons license. Your support will help
MIT Open Courseware continue to offer high quality
educational resources for free. To make a donation or to
view additional materials from hundreds of
MIT courses, visit MITOpenCourseware@OCW.MIT.edu. PROFESSOR: So this is a big
day mathematically speaking, because we come to
this key idea, which is a little bit like eigenvalues. Well, a lot like
eigenvalues, but different because the matrix A now is
more usually rectangular. So for a rectangular matrix,
the whole idea of eigenvalues is shot because if I
multiply A times a vector x in n dimensions, out will
come something in m dimensions and it's not going
to equal lambda x. So Ax equal lambda x is not even
possible if A is rectangular. And even if A is square, what
are the problems, just thinking for a minute about eigenvalues? The case I wrote up
here is the great case where I have a symmetric
matrix and then it's got a full set of
eigenvalues and eigenvectors and they're
orthogonal, all good. But for a general square
matrix, either the eigenvectors are complex-- eigenvalues are complex
or the eigenvectors are not orthogonal. So we can't stay with
eigenvalues forever. That's what I'm saying. And this is the
right thing to do. So what are these pieces? So these are the left and these
are the right singular vectors. So this is the new
word is singular. And in between go the-- not the eigenvalues,
but the singular values. So we've got the
whole point now. You've got to pick up on this. There are two sets of
singular vectors, not one. For eigenvectors, we just
had one set, the Q's. Now we have a
rectangular matrix, we've got one set of left
eigenvectors in m dimensions, and we've got another
set of right eigenvectors in n dimensions. And numbers in between
are not eigenvalues, but singular values. So these guys are-- let me write what
that looks like. This is, again, a diagonal
matrix sigma 2 to sigma r, let's say. So it's again, a diagonal
matrix in the middle. But the numbers on the
diagonal are all positive or 0. And they're called
singular values. So it's just a different world. So really, the first step by
have to do, the math step, is to show that
any matrix can be factored into u times
sigma times v transpose. So that's the parallel
to the spectral theorem that any symmetric matrix
could be factored that way. So you're good for that part. We just have to do it to see
what are u and sigma and v? What are those vectors
and those singular values? Let's go. So the key is that A
transpose A is a great matrix. So that's the key to the
math is A transpose A. So what are the properties
of A transpose A? A is rectangular again. So maybe m by n A transpose. So this was m by n. And this was n by m. So we get a result
that's n by n. And what else can you tell
me about A transpose A? It's a metric. That's a big deal. And it's square. And well yeah, you
can tell me more now, because we talked
about something, a topic that's a little more
than symmetric last time. The matrix A transpose A
will be positive, definite. It's eigenvalues are
greater or equal to 0. And that will mean that we
can take their square roots. And that's what we will do. So A transpose A we'll
have a factorization. It's symmetric. It'll have a like, a Q
lambda Q transpose, but I'm going to call it V lambda-- no, yeah, lambda-- I'll still
call it lambda V transpose. So these V's-- what do we know
about eigenvectors of these V's or eigenvectors of this guy? Square, symmetric,
positive, definite matrix. So we're in good shape. And what do we know about the
eigenvalues of A transpose A? They are all positive. So the eigenvalues are--
well, or equal to 0. And these guys are orthogonal. And these guys are
greater or equal to there. So that's good. That's one of our-- We'll depend a lot on that. But also, you've got
to recognize that A, A transpose is a different
guy, A, A transpose. So what's the shape
of A, A transpose? How big is that? Now I've got-- what do I have? M by n times n by m. So this will be what size? N by m. Different shape but with
the same eigenvalues-- the same eigenvalues. So it's going to have some other
eigenvectors, u-- of course, I'm going to call
them u, because I'm going to go in over there. They'll be the same. Well, they're saying
yeah, let me-- I shouldn't-- I have to
say when I say the same, I can't quite literally
mean the very same, because this has got n
eigenvalues and this has m eigenvalues. But the missing guys, the
ones that are in one of them and not in the other, depending
on the sizes, are zeros. So really, the heart of the
thing, the non-zero eigenvalues are the same. Well actually, I've
pretty much revealed what the SVD is going to use. It's going to use the U's from
here and the V's from here. But that's the story. You've got to see that story. So fresh start on the
singular value decomposition. What are we looking for? Well, as a factorization-- so we're looking for-- we want A. We want vectors v,
so that when I multiply by v-- so if it was an eigenvector,
it would be Av equal lambda v. But now for A, it's rectangular. It hasn't got eigenvectors. So Av is sigma, that the
new singular value times u. That's the first guy and the
second guy and the rth guy. I'll stop at r, the rank. Oh, yeah. Is that what I want? A-- let me just see. Av is sigma u. Yeah, that's good. So this is what takes the
place of Ax equal lambda x. A times one set of
singular vectors gives me a number of times the
other set of singular vectors. And why did I stop
at r the rank? Because after that,
the sigmas are 0. So after that, I could
have some more guys, but they'll be in the null
space 0 on down to of Vn. So these are the important ones. So that's what I'm looking for. Let me say it now in words. I'm looking for a bunch
of orthogonal vectors v so that when I
multiply them by A I get a bunch of
orthogonal vectors u. That is not so clearly possible. But it is possible. It does happen. I'm looking for one set
of orthogonal vectors v in the input
space, you could say, so that the Av's in the output
space are also orthogonal. In our picture of
the fundamental-- the big picture
of linear algebra, we have v's in this space, and
then stuff in the null space. And we have u's over
here in the columns space and some stuff in the
null space over there. And the idea is that I
have orthogonal v's here. And when I multiply by A-- so multiply by A-- then I get orthogonal u's over
here, orthogonal to orthogonal. That's what makes the
V's and they u's special. Right? That's the property. And then when we
write down-- well, let me write down
what that would mean. So I've just drawn a
picture to go with this-- those equations. That picture just goes
with these equations. And let me just write
down what it means. It means in matrix--
so I've written it. Oh yeah, I've written it here
in vectors one at a time. But of course,
you, know I'm going to put those vectors into
the columns of a matrix. So A times v1 up to
let's say vr will equal-- oh yeah. It equals sigma as times u. So this is what I'm
after is u1 up to ur multiplied by sigma
1 along to sigma r. What I'm doing now
is just to say I'm converting these
individual singular vectors, each v going
into a u to putting them all together into a matrix. And of course, what I've written
here is Av equals u sigma, Av equals u sigma. That's what that amounts to. Well, then I'm going to put
a v transpose on this side. And I'm going to get to A
equals u sigma v transpose, multiplying both sides
there by v transpose. I'm kind of writing the same
thing in different forms, matrix form, vector
at a time form. And now we have to find them. Now I've used up boards
saying what we're after, but now we've got to get there. So what are the v's
and what are the u's? Well, the cool idea is to
think of A transpose A. So you're with me
what we're for. And now think about
A transpose A. So if this is what
I'm hoping for, what will A transpose
A turn out to be? So big moment that's going
to reveal what the v's are. So if I form A transpose A-- so A transpose-- so I got
to transpose this guy. So A transpose is V sigma
transpose U transpose, right? And then comes A, which is
this, U sigma V transpose. So why did I do that? Why is it that A transpose A
is the cool thing to look at to make the problem simpler? Well, what becomes simpler
in that line just written? U transpose U is the
identity, because I'm looking for orthogonal, in
fact orthonormal U's. So that's the identity. So this is V sigma
transpose sigma V transpose. And I'll put parentheses
around that because that's a diagonal matrix. What does that tell me? What does that tell all of us? A transpose A has this form. Now we've seen that form before. We know that this is a
symmetric matrix, symmetric and even positive definite. So what are the v's? The v's are the eigenvectors
of A transpose A. This is the Q lambda Q transpose
for that symmetric matrix. So we know the v's
are the eigenvectors, v is the eigenvectors
of A transpose A. I guess we're also going
to get the singular values. So the sigma transpose sigma,
which will be the sigma squared are the eigenvalues of
A transpose A. Good! Sort of by looking for the
correct thing, U sigma V transpose and then just
using the U transpose U equal identity, we got
it back to something we perfectly recognize. A transpose A has that form. So now we know what the V's are. And if I do it the other way,
which, what's the other way? Instead of A transpose
A, the other way is to look at A, A transpose. And if I write all
that down, that a is the U sigma V transpose, and
the A transpose is the V sigma transpose U transpose. And again, this stuff
goes away and leaves me with U sigma, sigma
transpose U transpose. So I know what the U's are too. They are eigenvectors
of A, A transpose. Isn't that a beautiful symmetry? You just-- A transpose
A and A, A transpose are two different guys now. So each has its own
eigenvectors and we use both. It's just right. And I just have to
take the final step, and we've established the SVD. So the final step is to remember
what I'm going for here. A times a v is supposed
to be a sigma times a u. See, what I have
to deal with now is I haven't quite finished. It's just perfect
as far as it goes, but it hasn't gone to
the end yet because we could have double eigenvalues
and triple eigenvalues, and all those horrible
possibilities. And if I have triple eigenvalues
or double eigenvalues, then what's the deal
with eigenvectors if I have double eigenvalues? Suppose a matrix has
a symmetric matrix, has a double eigenvalue. Let me just take an example. So symmetric matrix like
say, 1, 1, 5, make it. Why not? What's the deal
with eigenvectors for that matrix 1, 1, 5? So 5 has got an eigenvector. You can see what it is, 0, 0, 1. What about eigenvectors
that go with lambda equal 1 for that matrix? What's up? What would be eigenvectors
for a lambda equal 1? Unfortunately, there was
a whole plane of them. Any vector of the form x, y, 0. Any vector in the x, y
plane would produce x, y, 0. So I have a whole
plane of eigenvectors. And I've got to pick two that
are orthogonal, which I can do. And then they have to be-- in the SVD those
two orthogonal guys have to go to two
orthogonal guys. In other words, it's a
little bit of detail here, a little getting into
this exactly what is-- well, actually, let
me tell you the steps. So I use this to conclude that
the V's the singular vectors should be eigenvalues. I concluded those
guys from this step. Now I'm not going to
use this step so much. Of course, it's in the back of
my mind but I'm not using it. I'm going to get
the u's from here. So u1 is A v1 over sigma
1 ur is Avr over sigma r. You see what I'm doing here? I'm picking in a possible
plane of things the one I want, the u's I want. So I've chosen the v's. I've chosen the sigmas. They were fixed
for A transpose A. The eigenvectors are
v's, the things-- the eigenvalues
are sigma squared. And now then this
is the u I want. Are you with me? So I want to get
these u's correct. And if I have a whole
plane of possibilities, I got to pick the right one. And now finally, I have to
show that it's the right one. So what is left to show? I should show that these u's are
eigenvectors of A, A transpose. And I should show that
they're orthogonal. That's the key. I would like to show that
these are orthogonal. And that's what goes
in this picture. The v's-- I've got
orthogonal, guys, because they're the eigenvectors
of a symmetric matrix. Pick them orthogonal. But now I'm multiplying
by A, so I'm getting the u which is Av over
sigma for the basis vectors. And I have to show
they're orthogonal. So this is like
the final moment. Does everything
come together right? If I've picked the v's as the
eigenvectors of A transpose A, and then I take these for
the u, are they orthogonal? So I would like to think
that we can check that fact and that it will come out. Could you just help
me through this one? I'll never ask for anything
again, just get the SVD one. So I would like
to show that u1-- so let me put up what I'm doing. I'm trying to show that
u1 transpose u2 is 0. They're orthogonal. So u1 is A v1 over sigma 1. That's transpose. That's u1. And u2 is A v2 over sigma 2. And I want to get 0. The whole conversation
is ending right here. Why is that thing 0? The v's are orthogonal. We know the v's are orthogonal. They're orthogonal
eigenvectors of A transpose A. Let me repeat that. The v's are orthogonal
eigenvectors of A transpose A, which I know we can find them. Then I chose the u's to be this. And I want to get the answer 0. Are you ready to do it? We want to compute
that and get 0. So what do I get? We just have to do it. So I can see that the
denominator is that. So is it v1 transpose
A transpose times A v2. And I'm hoping to get 0. Do I get 0 here? You hope so. v1 is orthogonal v2. But I've got A transpose A
stuck in the middle there. So what happens here? How do I look at that? v2 is an eigenvector of
A transpose A. Terrific! So this is v1 transpose. And this is the matrix times v2. So that's sigma 2
transpose v2, isn't it? It's the eigenvector
with eigenvalue sigma 2 squared times v2. Yeah, divided by
sigma 1 sigma 2. So the A's are out of there now. So I've just got these
numbers, sigma 2 squared. So that would be
sigma 2 over sigma 1-- I've accounted for these numbers
here-- times v1 transpose v2. And now what's up? They're orthonormal. We got it. That's 0. That is 0 there, yeah. So not only are the v's
orthogonal to each other, but because they're eigenvectors
of A transpose A, when I do this, I discover
that the Av's are orthogonal to each other
over in the column space. So orthogonal v's in the
row space, orthogonal Av's over in column space. That was discovered late--
much long after eigenvectors. And it's a interesting history. And it just comes out right. And then it was discovered,
but not much used, for oh, 100 years probably. And then people saw that it
was exactly the right thing, and data matrices
became important, which are large rectangular matrices. And we have not-- oh, I better say a word, just
a word here about actually computing the v's and
sigmas and the u's So how would you
actually find them? What I most want to
say is you would not go this A transpose A route. Why is it like it? Is that a big mistake? If you have a matrix
A, say 5,000 by 10,000, why is it a mistake
to actually use A transpose A in
the computation? We used it heavily in the proof. And we could find another proof
that wouldn't use it so much. But why would I not
multiply these two together? It's very big, very expensive. It adds in a whole
lot of round off-- you have a matrix that's now-- its vulnerability to round
off errors is squared-- that's called its condition
number-- gets squared. And you just don't go there. So the actual computational
methods are quite different. And we'll talk about those. But the A transpose
A, because it's symmetric positive definite,
made the proof so nice. You've seen the nicest
proof, I'd say, of the-- Now I should think
about the geometry. So what does A
equal A for u sigma? Maybe I take another
board, but it will fill it. But it's a good U
sigma V transpose. So it's got three factors there. And I would like
then each factor is kind of a special matrix. U and V are orthogonal matrix. So I think of
those as rotations. Sigma is a diagonal matrix. I think of it as stretching. So now I'm just going
to draw the picture. So here's unit vectors. And the first thing--
so if I multiply by x, this is the first
thing that happens. So that rotates. So here's x's. Then V transpose x's. That's still a circle
length and change for those, when I multiply
by an orthogonal matrix. But the vectors turned. It's a rotation. Could be a reflection, but
let's keep it as a rotation. Now what does sigma do? So I have this unit circle. I'm in 2D. So I'm drawing a
picture of the vectors. These are the unit
vectors in 2D, x,y. They got turned by
the orthogonal matrix. What does sigma do
to that picture? It stretches, because
sigma multiplies by sigma 1 in the
first component, sigma 2 in the second. So it stretches these guys. And let's suppose this is
number 1 and this is number 2, this is number 1 and number 2. So sigma 1, our
convention is sigma 1-- we always take sigma 1
greater or equal to sigma 2, greater or equal whatever,
greater equal, sigma rank. And they're all positive. And the rest are 0. So sigma 1 will be
bigger than sigma 2. So I'm expecting a
circle goes to an ellipse when you stretch-- I didn't get it quite
perfect, but not bad. So this would be sigma
2 v2, sigma 1 v1, and this would be sigma 2 v2. And we now have an ellipse. So we started with
x is in a circle. We rotated. We stretched. And now the final step
is take these guys and multiply them by u. So this was the
sigma V transpose x. And now I'm ready for the
u part which comes last because it's at the left. And what happens? What's the picture now? What does u do to the ellipse? It rotates it. It's another orthogonal matrix. It rotates it
somewhere, maybe there. And now we see the
u's, u2 and u1. Well, let me think about that. Basically, that's
not that's right. So this SVD is telling us
something quite remarkable that every linear
transformation, every matrix
multiplication factors into a rotation times a stretch
times a different rotation, but possibly different. Actually, when would the
u be the same as a v? Here's a good question. When is u the same as v
when are the two singular vectors just the same? AUDIENCE: A square. PROFESSOR: Because A
would have to be square. And we want this to be
the same as Q lambda Q transpose if they're the same. So the U's would be
the same as the V's when the matrix is symmetric. And actually we need it
to be positive definite. Why is that? Because our convention is these
guys are greater or equal to 0. It's going to be
the same, then-- so far a positive
definite symmetric matrix, the S that we started
with is the same as the A on the next line. Yeah, the Q is the U, the Q
transpose is the V transpose, the lambda is the sigma. So those are the good matrices. And they're the ones that
you can't improve basically. They're so good you can't make
a positive definite symmetric matrix better than it is. Well, maybe diagonalize
it or something, but OK. Now I think of like,
one question here that helps me anyway to
keep this figure straight, how I want to count
the parameters in this factorization. So I am 2 by 2. I'm 2 by 2. So A has four
numbers, a, b, c, d. Then I guess I feel
that four numbers should appear on the right hand side. Somehow the U and the
sigma and the V transpose should use up a total
of four numbers. So we have a counting
match between the left side that's got four numbers a,
b, c, d, and the right side that's got four numbers
buried in there somewhere. So how can we dig them out? How many numbers in sigma? That's pretty clear. Two, sigma 1 and sigma 2. The two eigenvalues. How many numbers
in this rotation? So if I had a
different color chalk, I would put 2 for the number of
things I counted for by sigma. How many parameters does a
two by two rotation require? One. And what's a good
word for that one? Is that one parameter? It's like I have our
cos theta, sine theta, minus sine theta, cos theta. There's a number theta. It's the angle it rotates. So that's one guy to tell
the rotation angle, two guys to tell the stretchings, and
one more to tell the rotation from you, adding up to four. So those count--
that was a match up with the four numbers, a,
b, c, d that we start with. Of course, it's a complicated
relation between those four numbers and rotations
and stretches, but it's four
equals four anyway. And I guess if you
did three by threes-- oh, three by threes. What would happen then? So let me take three. Do you want to care
for three by threes? Just, it's sort of satisfying
to get four equal four. But now what do we
get three by three? We got how many numbers here? Nine. So where are those nine numbers? How many here? That's usually the easy-- three. So what's your guess for
the how many in a rotation? And a 3D rotation, you take
a sphere and you rotate it. How many how many numbers
to tell you what's what-- to tell you what you did? Three. We hope three. Yeah, it's going to be three,
three, and three for the three dimensional world
that we live in. So people who do rotations for a
living understand that rotation in 3D, but how do you see this? AUDIENCE: Roll, pitch, and yaw. PROFESSOR: Sorry? AUDIENCE: Roll, pitch and yaw. PROFESSOR: Roll, pitch, and yaw. That sounds good. I mean, it's three words
and we've got it, right? OK, yeah. Roll, pitch and yaw. Yeah, I guess a pilot hopefully,
knows about those three. Yeah, yeah, yeah. Which is roll? When you are like
forward and back? Does anybody, anybody? Roll, pitch, and yaw? AUDIENCE: Pitch is
the up and down one. PROFESSOR: Pitch is
the up and down one. OK. AUDIENCE: Roll is like,
think of a barrel roll. And yaw is your
side-to-side motion. PROFESSOR: Oh, yaw, you
stay in a plane and you-- OK, beautiful. Right, right. And that leads us to our
four-- four dimensions. What's your guess on 4D? Well, we could do
the count again. If it was 4 by 4, we would
have 16 numbers there. And in the middle, we always
have an easy time with that. That would be 4. So we've got 12
left to share out. So six somehow-- six-- six angles in four dimensions. Well, we'll leave it there. Yeah, yeah, yeah. OK. So there is the SVD
but without an example. Examples, you know, I would
have to compute A transpose A and find it. So the text will do that-- does it for a particular matrix. Oh! Yeah, the text does it
for a matrix 3, 4, 0, 5 that came out pretty well. A few facts we
could learn though. So if I multiply all
the eigenvalues together for a matrix A, what do I get? I get the determinant. What if I multiply the
singular values together? Well again, I get
the determinant. You can see it right away
from the big formula. Take determinant--
take determinant. Well, assuming the
matrix A is square. So it's got a determinant. Then I take determinant
of this product. I can take the
separate determinants. That has determinant
equal to one. An orthogonal matrix,
the determinant is one. And similarly, here. So the product of the sigmas
is also the determinant. Yeah. Yeah, so the product of the
sigmas is also the determinant. The product of the
sigmas here will be 15. But you'll find that sigma
one is smaller than lambda 1. So here are the
eigenvalues, lambda 1 less or equal to lambda 2, say. But the singular values
are outside them. Yeah. But they still multiply. Sigma 1 times sigma
2 will still be 15. And that's the same as
lambda 1 times lambda 2. Yeah. But overall, computing
the examples of the SVD take more time because-- well, yeah, you just compute
A transpose A and you've got the v's. And you're on your way. And you have to take the
square root of the eigenvalues. So that's the SVD as
a piece of pure math. But of course, what we'll do
next time starting right away is use SVD. And let me tell you
even today, the most-- yeah, yeah most important
pieces of the SVD. So what do I mean by
pieces of the SVD? I've got one more blackboard
still to write on. So here we go. So let me write out A is
the u's times the sigmas-- sigmas 1 to r times the v's-- v transpose v1 transpose
down to vr transpose. So those are across. Yeah. Actually what I've
written here-- so you could say there
is a big economies. There is a smaller size SVD
that has the real stuff that really counts. And then there's a larger SVD
that has a whole lot of zeros. So this it would be the
smaller one, m by r. This would be r by r. And these would all be positive. And this would be r by n. So that's only using
the r non-zeros. All these guys are
greater than zero. Then the other one
we could fill out to get a square
orthogonal matrix, the sigmas and square v's v1
transpose to vn transpose. So what are the shapes now? This shape is m by m. It's a proper orthogonal matrix. This one also n by n. So this guy has to be--
this is the sigma now. So it has to be what size? m by m. That's the remaining space. So it starts with the sigmas,
and then it's all zeros, accounting for null space stuff. Yeah. So you should really see
that these two are possible. That all these zeros
when you multiply out, just give nothing, so
that really the only thing that non-zero is in these bits. But there is a complete one. So what are these extra u's
that are in the null space of A, A transpose or A transpose A? Yeah, so two sizes, the large
size and the small size. But then the things that
count are all in there. OK. So I was going to
do one more thing. Let me see what it was. So this is section
1.8 of the notes. And you'll see examples there. And you'll see a second
approach to the finding the u's and v's and sigmas. I can tell you what that is. But maybe with just do
something nice at the end, let me tell you about another
factorization of A that's famous in engineering, and
it's famous in geometry. So this is NEA is a
U sigma V transpose. We've got that. Now the other one
that I'm thinking of, I'll tell you its name. It's called the polar
decomposition of a matrix. And all I want you to see
is that it's virtually here. So a polar means-- what's polar in--
for a complex number, what's the polar form
of a complex number? AUDIENCE: e to the i theta. PROFESSOR: Yeah, it's e
to the i theta times r. Yeah. A real guy-- so
the real guy r will translate into a symmetric guy. And the e to the i theta
will translate into-- what kind of a matrix reminds
you of e to the i theta? AUDIENCE: Orthogonal. PROFESSOR: Orthogonal, size 1. So orthogonal. So that's a very,
very kind of nice. Every matrix factors
into a symmetric matrix times an orthogonal matrix. And I of course, describe these
as the most important classes of matrices. And here, we're saying every
matrix is a S times a Q. And I'm also saying that
I can get that quickly out of the SVD. So I'm just want to do it. So I want to find an S
and find a Q out of this. So to get an S-- So let me just start it. U sigma-- but now
I'm looking for an S. So what shall I put in now? I better put in-- if I've got to U
sigma something, and I want it to
be a symmetric, I should put in U
transpose would do it. But then if I put
it in U transpose, I've got to put it in U.
So now I've got U sigma. U transpose U is the identity. Then I've got to
get V transpose. And have I got what
the polar decomposition is asking for in this line? So, yeah. What have I got here? Where's the where's
the S in this? So you see, I took the SVD and I
just put the identity in there, just shifted things a little. And now where's the S
that I can read off? For three, that's an S.
That's a symmetric matrix. And where's the Q? Well, I guess we can see
where the Q has to be. It's here, yeah. Yeah, so just by
sticking U transpose U and putting the
parentheses right, I recover that decomposition
of a matrix, which in mechanical engineering
language, is language tells me that any
strain can be-- which is like stretching
of elastic thing, has a symmetric kind of a
stretch and a internal twist. Yeah. So that's good. Well, this was a 3, 6, 9
boards filled with matrices. Well, it is 18 0, 6, 5. So maybe that's all right. But the idea is to use
them on a matrix of data. And I'll just tell
you the key fact. The key fact-- if I have
a big matrix of data, A, and if I want to pull
out of that matrix the important part,
so that's what data science has to be doing. Out of a big matrix, some part
of it is noise, some part of it is signal. I'm looking for the most
important part of the signal here. So I'm looking for the most
important part of the matrix. In a way, the biggest
numbers, but of course, I don't look at
individual numbers. So what's the biggest
part of the matrix? What are the
principal components? Now we're really getting in-- it could be data. And we want to do
statistics, or we want to see what has
high variance, what has low variance, we'll do these
connections with statistics. But what's the important
part of the matrix? Well, let me look at
U sigma V transpose. Here, yeah, let me look at it. So what's the one most
important part of that matrix? The right one? It's a rank one piece. So when I say a part, of course
it's going to be a matrix part. So the simple matrix
building block is like a rank one matrix, a
something, something transpose. And what should I
pull out of that as being the most
important rank one matrix that's in that product? So I'll erase the
1.8 while you think what do I do to pick out the big
deal, the thing that the data is telling me first. Well, these are orthonormal. No one is bigger
than another one. These are orthonormal, no one
is bigger than another one. But here, I look here, which
is the most important number? Sigma 1. Sigma 1. So the part I pick out is
this biggest number times it's row times it's column. So it's u 1 sigma 1 v1 transpose
is the top principal part of the matrix A.
It's the leading part of the matrix A.
It's the biggest rank one part of the matrix is there. So computing those three
guys is the first step to understanding the data. Yeah. So that's what's
coming next is-- and I guess tomorrow,
since they moved-- MIT declared Tuesday
to be Monday. They didn't change Wednesday. So I'll see you tomorrow for
the principal components. Good.