video_id
stringclasses
7 values
text
stringlengths
2
29.3k
rLlZpnT02ZU
So in this chapter, we're going to talk about maximum
rLlZpnT02ZU
likelihood estimation.
rLlZpnT02ZU
Who has already seen maximum likelihood estimation?
rLlZpnT02ZU
OK.
rLlZpnT02ZU
And who knows what a convex function is?
rLlZpnT02ZU
OK.
rLlZpnT02ZU
So we'll do a little bit of reminders on those things.
rLlZpnT02ZU
So those things are when we do maximum likelihood estimation,
rLlZpnT02ZU
likelihood is the function, so we need to maximize a function.
rLlZpnT02ZU
That's basically what we need to do.
rLlZpnT02ZU
And if I give you a function, you
rLlZpnT02ZU
need to know how to maximize this function.
rLlZpnT02ZU
Sometimes, you have closed-form solutions.
rLlZpnT02ZU
You can take the derivative and set it equal to 0 and solve it.
rLlZpnT02ZU
But sometimes, you actually need to resort to algorithms
rLlZpnT02ZU
to do that.
rLlZpnT02ZU
And there's an entire industry doing that.
rLlZpnT02ZU
And we'll briefly touch upon it, but this is definitely
rLlZpnT02ZU
not the focus of this class.
rLlZpnT02ZU
OK.
rLlZpnT02ZU
So before diving directly into the definition
rLlZpnT02ZU
of the likelihood and what is the definition
rLlZpnT02ZU
of the maximum likelihood estimator, what
rLlZpnT02ZU
I'm going to try to do is to give you
rLlZpnT02ZU
an insight for what we're actually doing when we do
rLlZpnT02ZU
maximum likelihood estimation.
rLlZpnT02ZU
So remember, we have a model on a sample space E
rLlZpnT02ZU
and some candidate distributions P theta.
rLlZpnT02ZU
And really, your goal is to estimate a true theta
rLlZpnT02ZU
star, the one that generated some data, X1 to Xn,
rLlZpnT02ZU
in an iid fashion.
rLlZpnT02ZU
But this theta star is really a proxy for us
rLlZpnT02ZU
to know that we actually understand
rLlZpnT02ZU
the distribution itself.
rLlZpnT02ZU
The goal of knowing theta star is so that you can actually
rLlZpnT02ZU
know what P theta star.
rLlZpnT02ZU
Otherwise, it has-- well, sometimes we
rLlZpnT02ZU
said it has some meaning itself, but really you
rLlZpnT02ZU
want to know what the distribution is.
rLlZpnT02ZU
And so your goal is to actually come up with the distribution--
rLlZpnT02ZU
hopefully that comes from the family P theta--
rLlZpnT02ZU
that's close to P theta star.
rLlZpnT02ZU
So in a way, what does it mean to have two distributions that
rLlZpnT02ZU
are close?
rLlZpnT02ZU
It means that when you compute probabilities
rLlZpnT02ZU
on one distribution, you should have
rLlZpnT02ZU
the same probability on the other distribution pretty much.
rLlZpnT02ZU
So what we can do is say, well, now I
rLlZpnT02ZU
have two candidate distributions.
rLlZpnT02ZU
So if theta hat leads to a candidate distribution P theta
rLlZpnT02ZU
hat, and this is the true theta star,
rLlZpnT02ZU
it leads to the true distribution P theta star
rLlZpnT02ZU
according to which my data was drawn.
rLlZpnT02ZU
That's my candidate.
rLlZpnT02ZU
As a statistician, I'm supposed to come up
rLlZpnT02ZU
with a good candidate, and this is the truth.
rLlZpnT02ZU
And what I want is that if you actually give me
rLlZpnT02ZU
the distribution, then I want when
rLlZpnT02ZU
I'm computing probabilities for this guy,
rLlZpnT02ZU
I know what the probabilities for the other guys are.
rLlZpnT02ZU
And so really what I want is that if I compute a probability
rLlZpnT02ZU
under theta hat of some interval a, b,
rLlZpnT02ZU
it should be pretty close to the probability
rLlZpnT02ZU
under theta star of a, b.
rLlZpnT02ZU
And more generally, if I want to take
rLlZpnT02ZU
the union of two intervals, I want this to be true.
rLlZpnT02ZU
If I take just 1/2 lines, I want this to be true from 0
rLlZpnT02ZU
to infinity, for example, things like this.
rLlZpnT02ZU
I want this to be true for all of them at once.
rLlZpnT02ZU
And so what I do is that I write A for a probability event.
rLlZpnT02ZU
And I want that P hat of A is close to P star of A
rLlZpnT02ZU
for any event A in the sample space.
rLlZpnT02ZU
Does that sound like a reasonable goal
rLlZpnT02ZU
for a statistician?
rLlZpnT02ZU
So in particular, if I want those to be close,
rLlZpnT02ZU
I want the absolute value of their difference
rLlZpnT02ZU
to be close to 0.
rLlZpnT02ZU
And this turns out to be--
rLlZpnT02ZU
if I want this to hold for all possible A's, I
rLlZpnT02ZU
have all possible events, so I'm going to actually maximize over
rLlZpnT02ZU
these events.
rLlZpnT02ZU
And I'm going to look at the worst
rLlZpnT02ZU
possible event on which theta hat can depart from theta star.
rLlZpnT02ZU
And so rather than defining it specifically
rLlZpnT02ZU
for theta hat and theta star, I'm
rLlZpnT02ZU
just going to say, well, if you give me two probability
rLlZpnT02ZU
measures, P theta and P theta prime,
rLlZpnT02ZU
I want to know how close they are.
rLlZpnT02ZU
Well, if I want to measure how close they
rLlZpnT02ZU
are by how they can differ when I measure
rLlZpnT02ZU
the probability of some event, I'm
rLlZpnT02ZU
just looking at the absolute value of the difference
rLlZpnT02ZU
of the probabilities and I'm just
rLlZpnT02ZU
maximizing over the worst possible event that might
rLlZpnT02ZU
actually make them differ.
rLlZpnT02ZU
Agreed?
rLlZpnT02ZU
That's a pretty strong notion.
rLlZpnT02ZU
So if the total variation between theta and theta prime
rLlZpnT02ZU
is small, it means that for all possible A's that you give me,
rLlZpnT02ZU
then P theta of A is going to be close to P