video_id
stringclasses
7 values
text
stringlengths
2
29.3k
rLlZpnT02ZU
theta prime of A, because if--
rLlZpnT02ZU
let's say I just found the bound on the total variation
rLlZpnT02ZU
distance, which is 0.01.
rLlZpnT02ZU
All right.
rLlZpnT02ZU
So that means that this is going to be larger
rLlZpnT02ZU
than the max over A of P theta minus P theta prime of A,
rLlZpnT02ZU
which means that for any A--
rLlZpnT02ZU
actually, let me write P theta hat and P theta star,
rLlZpnT02ZU
like we said, theta hat and theta star.
rLlZpnT02ZU
And so if I have a bound, say, on the total variation,
rLlZpnT02ZU
which is 0.01, that means that P theta hat--
rLlZpnT02ZU
every time I compute a probability on P theta hat,
rLlZpnT02ZU
it's basically in the interval P theta star of A,
rLlZpnT02ZU
the one that I really wanted to compute, plus or minus 0.01.
rLlZpnT02ZU
This has nothing to do with confidence interval.
rLlZpnT02ZU
This is just telling me how far I
rLlZpnT02ZU
am from the value of actually trying to compute.
rLlZpnT02ZU
And that's true for all A. And that's key.
rLlZpnT02ZU
That's where this max comes into play.
rLlZpnT02ZU
It just says, I want this bound to hold
rLlZpnT02ZU
for all possible A's at once.
rLlZpnT02ZU
So this is actually a very well-known distance
rLlZpnT02ZU
between probability measures.
rLlZpnT02ZU
It's the total variation distance.
rLlZpnT02ZU
It's extremely central to probabilistic analysis.
rLlZpnT02ZU
And it essentially tells you that every time--
rLlZpnT02ZU
if two probability distributions are close,
rLlZpnT02ZU
then it means that every time I compute a probability
rLlZpnT02ZU
under P theta but I really actually
rLlZpnT02ZU
have data from P theta prime, then
rLlZpnT02ZU
the error is no larger than the total variation.
rLlZpnT02ZU
OK.
rLlZpnT02ZU
So this is maybe not the most convenient way
rLlZpnT02ZU
of finding a distance.
rLlZpnT02ZU
I mean, how are you going--
rLlZpnT02ZU
in reality, how are you to compute this maximum
rLlZpnT02ZU
over all possible events?
rLlZpnT02ZU
I mean, it's just crazy, right?
rLlZpnT02ZU
There's an infinite number of them.
rLlZpnT02ZU
It's much larger than the number of intervals, for example,
rLlZpnT02ZU
so it's a bit annoying.
rLlZpnT02ZU
And so there's actually a way to compress it
rLlZpnT02ZU
by just looking at the basically function distance or vector
rLlZpnT02ZU
distance between probability mass functions or probability
rLlZpnT02ZU
density functions.
rLlZpnT02ZU
So I'm going to start with the discrete version
rLlZpnT02ZU
of the total variation.
rLlZpnT02ZU
So throughout this chapter, I will
rLlZpnT02ZU
make the difference between discrete random variables
rLlZpnT02ZU
and continuous random variables.
rLlZpnT02ZU
It really doesn't matter.
rLlZpnT02ZU
All it means is that when I talk about discrete,
rLlZpnT02ZU
I will talk about probability mass functions.
rLlZpnT02ZU
And when I talk about continuous,
rLlZpnT02ZU
I will talk about probability density functions.
rLlZpnT02ZU
When I talk about probability mass functions,
rLlZpnT02ZU
I talk about sums.
rLlZpnT02ZU
When I talk about probability density functions,
rLlZpnT02ZU
I talk about integrals.
rLlZpnT02ZU
But they're all the same thing, really.
rLlZpnT02ZU
So let's start with the probability mass function.
rLlZpnT02ZU
Everybody remembers what the probability mass
rLlZpnT02ZU
function of a discrete random variable is.
rLlZpnT02ZU
This is the function that tells me for each possible value
rLlZpnT02ZU
that it can take, the probability
rLlZpnT02ZU
that it takes this value.
rLlZpnT02ZU
So the Probability Mass Function, PMF,
rLlZpnT02ZU
is just the function for all x in the sample space
rLlZpnT02ZU
tells me the probability that my random variable is
rLlZpnT02ZU
equal to this little value.
rLlZpnT02ZU
And I will denote it by P sub theta of X.
rLlZpnT02ZU
So what I want is, of course, that the sum
rLlZpnT02ZU
of the probabilities is 1.
rLlZpnT02ZU
And I want them to be non-negative.
rLlZpnT02ZU
Actually, typically we will assume that they are positive.
rLlZpnT02ZU
Otherwise, we can just remove this x from the sample space.
rLlZpnT02ZU
And so then I have the total variation distance, I mean,
rLlZpnT02ZU
it's supposed to be the maximum overall sets of--
rLlZpnT02ZU
of subsets of E, such that the probability
rLlZpnT02ZU
of A minus probability of theta prime of A--
rLlZpnT02ZU
it's complicated, but really there's
rLlZpnT02ZU
this beautiful formula that tells me
rLlZpnT02ZU
that if I look at the total variation between P theta
rLlZpnT02ZU
and P theta prime, it's actually equal to just 1/2
rLlZpnT02ZU
of the sum for all X in E of the absolute difference between P
rLlZpnT02ZU
theta X and P theta prime of X.
rLlZpnT02ZU
So that's something you can compute.
rLlZpnT02ZU
If I give you two probability mass functions,
rLlZpnT02ZU
you can compute this immediately.
rLlZpnT02ZU
But if I give you just the densities
rLlZpnT02ZU
and the original distribution, the original definition
rLlZpnT02ZU
where you have to max over all possible events,
rLlZpnT02ZU
it's not clear you're going to be
rLlZpnT02ZU
able to do that very quickly.
rLlZpnT02ZU
So this is really the one you can work with.
rLlZpnT02ZU
But the other one is really telling you
rLlZpnT02ZU
what it is doing for you.
rLlZpnT02ZU
It's controlling the difference of probabilities
rLlZpnT02ZU
you can compute on any event.
rLlZpnT02ZU
But here, it's just telling you, well,