video_id
stringclasses
7 values
text
stringlengths
2
29.3k
rLlZpnT02ZU
to be actually looking at.
rLlZpnT02ZU
Yeah.
rLlZpnT02ZU
This parenthesis should really stop here.
rLlZpnT02ZU
I really wanted to put quadratic in parenthesis.
rLlZpnT02ZU
So the risk of this guy is what?
rLlZpnT02ZU
Well, it's the expectation of x bar n minus theta squared.
rLlZpnT02ZU
And we know it's the square of the variance,
rLlZpnT02ZU
so it's the square of the bias, which
rLlZpnT02ZU
we know is 0, so it's 0 squared plus the variance, which
rLlZpnT02ZU
is theta, 1 plus theta--
rLlZpnT02ZU
1 minus theta divided by n.
rLlZpnT02ZU
So it's just theta, 1 minus theta divided by n.
rLlZpnT02ZU
So this is just summarizing the performance of an estimator,
rLlZpnT02ZU
which is the random variable.
rLlZpnT02ZU
I mean, it's complicated.
rLlZpnT02ZU
If I really wanted to describe it,
rLlZpnT02ZU
I would just tell you the entire distribution
rLlZpnT02ZU
of this random variable.
rLlZpnT02ZU
But now what I'm doing is I'm saying, well,
rLlZpnT02ZU
let's just take this random variable, remove theta from it,
rLlZpnT02ZU
and see how small the fluctuations around theta--
rLlZpnT02ZU
the squared fluctuations around theta are in expectation.
rLlZpnT02ZU
So that's what the quadratic risk is doing.
rLlZpnT02ZU
And in a way, this decomposition,
rLlZpnT02ZU
as the sum of the bias square and the variance,
rLlZpnT02ZU
is really telling you that--
rLlZpnT02ZU
it is really accounting for the bias, which is, well,
rLlZpnT02ZU
even if I had an infinite amount of observations,
rLlZpnT02ZU
is this thing doing the right thing?
rLlZpnT02ZU
And the other thing is actually the variance,
rLlZpnT02ZU
so for finite number of observations,
rLlZpnT02ZU
what are the fluctuations?
rLlZpnT02ZU
All right.
rLlZpnT02ZU
Then you can see that those things, bias and variance,
rLlZpnT02ZU
are actually very different.
rLlZpnT02ZU
So I don't have any colors here, so you're
rLlZpnT02ZU
going to have to really follow the speed--
rLlZpnT02ZU
the order in which I draw those curves.
rLlZpnT02ZU
All right.
rLlZpnT02ZU
So let's find--
rLlZpnT02ZU
I'm going to give you three candidate estimators, so--
rLlZpnT02ZU
estimators for theta.
rLlZpnT02ZU
So the first one is definitely Xn bar.
rLlZpnT02ZU
That will be a good candidate estimator.
rLlZpnT02ZU
The second one is going to be 0.5, because after all,
rLlZpnT02ZU
why should I bother if it's actually going to be--
rLlZpnT02ZU
right?
rLlZpnT02ZU
So for example, if I ask you to predict
rLlZpnT02ZU
the score of some candidate in some election,
rLlZpnT02ZU
then since you know it's going to be very close to 0.5,
rLlZpnT02ZU
you might as well just throw 0.5 and you're not going
rLlZpnT02ZU
to be very far from reality.
rLlZpnT02ZU
And it's actually going to cost you 0 time and $0
rLlZpnT02ZU
to come up with that.
rLlZpnT02ZU
So sometimes maybe just a good old guess
rLlZpnT02ZU
is actually doing the job for you.
rLlZpnT02ZU
Of course, for presidential elections
rLlZpnT02ZU
or something like this, it's not very helpful
rLlZpnT02ZU
if your prediction is telling you this.
rLlZpnT02ZU
But if it was something different,
rLlZpnT02ZU
that would be a good way to generate some close to 1/2.
rLlZpnT02ZU
For a coin, for example, if I give you a coin,
rLlZpnT02ZU
you never know.
rLlZpnT02ZU
Maybe it's slightly biased.
rLlZpnT02ZU
But the good guess, just looking at it, inspecting it,
rLlZpnT02ZU
maybe there's something crazy happening
rLlZpnT02ZU
with the structure of it, you're going
rLlZpnT02ZU
to guess that it's 0.5 without trying to collect information.
rLlZpnT02ZU
And let's find another one, which is, well, you know,
rLlZpnT02ZU
I have a lot of observations.
rLlZpnT02ZU
But I'm recording couples kissing, but I'm on a budget.
rLlZpnT02ZU
I don't have time to travel all around the world
rLlZpnT02ZU
and collect some people.
rLlZpnT02ZU
So really, I'm just going to look at the first couple
rLlZpnT02ZU
and go home.
rLlZpnT02ZU
So my other estimator is just going to be X1.
rLlZpnT02ZU
I just take the first observation, 0, 1,
rLlZpnT02ZU
and that's it.
rLlZpnT02ZU
So now I'm going--
rLlZpnT02ZU
I want to actually understand what the behavior of those guys
rLlZpnT02ZU
is.
rLlZpnT02ZU
All right.
rLlZpnT02ZU
So we know-- and so we know that for this guy, the bias is 0
rLlZpnT02ZU
and the variance is equal to theta,
rLlZpnT02ZU
1 minus theta divided by n.
rLlZpnT02ZU
What is the bias of this guy, 0.5?
rLlZpnT02ZU
AUDIENCE: 0.5.
rLlZpnT02ZU
AUDIENCE: 0.5 minus theta?
rLlZpnT02ZU
PHILIPPE RIGOLLET: 0.5 minus theta, right.
rLlZpnT02ZU
So the bias, 0.5 minus theta.
rLlZpnT02ZU
What is the variance of this guy?
rLlZpnT02ZU
What is the variance of 0.5?
rLlZpnT02ZU
AUDIENCE: It's 0.
rLlZpnT02ZU
PHILIPPE RIGOLLET: 0.
rLlZpnT02ZU
Right.
rLlZpnT02ZU
It's just a deterministic number,
rLlZpnT02ZU
so there's no fluctuations for this guy.
rLlZpnT02ZU
What is the bias?
rLlZpnT02ZU
Well, X1 is actually--
rLlZpnT02ZU
just for simplicity, I can think of it