In article <[EMAIL PROTECTED]>,
John Hicken <[EMAIL PROTECTED]> wrote:
>*** post for FREE via your newsreader at post.newsfeed.com ***

>I'm doing a module on Probability Theory, and I'm not sure I 'get' the
>concept of conditional probability of a random variable over a sigma field.
>I think I understand the mathematics of the situation, but I don't really
>have an intuitive grasp of the situation.

>By a sigma field I mean a set B, of subsets of a set A, containing the empty
>set, and closed under taking complements, and countable intersections.

>According to the definition I've been given, for a random variable Y on
>probability space (A,F,P), with E[|Y|] < infinity and D a sub sigma field of
>F.  A version of E[Y|D] is any function from A to the real numbers that is D
>measurable, and whose integral over each member of D is the same as that of
>Y, the integrals being over P.  I say all that so we know exactly what we
>are talking about.

>What seems to be happening is that the expectation over a sigma field D, is
>sort of sampling the random variable Y, so it acts in the same way over the
>sets in D.

This is incorrect.

>But perhaps someone could tell me if there is a better way of understanding
>intuitively what this means.

Since the argument is by means of the Radon-Nikodym
Theorem, we need to go back to that to get the
understanding.  BTW, one can define conditional
expectation with respect to a measure which is not
a probability measure, and this can be useful.

At any rate, suppose we have a finite measure space (A, D, P),
and we have a (signed) measure Q on D, and for convenience we
will assume that |Q(S)|/P(S)| is bounded for all S in D.  We
do not need such strong assumptions, but the R-N theorem can
be reduced to this, and the intuition remains valid.

Now for any partition Z of D into a finite (countable can be
used as well) family of sets of positive P measure, define
r_Z(x) = Q(S)/P(S) for that S which contains x.  Take a
sequence of partitions Z_n which spreads out the r's; under
the conditions given, making \int (r_Z_n)^2 dP go to its
least upper bound will work, but otherwise a maximal spreading
out is needed, and any will do.  Then r_Z_n will converge in
measure, and the limit is the Radon- Nikodym derivative, which
for your problem is the conditional expectation.

So the intuition is to take conditional expectations on finite
partitions, and pass to the limit using a sequence of partitions
which use as much of the information as they can.  It is the
natural extension from discrete partitions.

-- 
This address is for information only.  I do not claim that these views
are those of the Statistics Department or of Purdue University.
Herman Rubin, Department of Statistics, Purdue University
[EMAIL PROTECTED]         Phone: (765)494-6054   FAX: (765)494-0558
.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to