Vimal wrote:
> Hi,
>
> If in EM, I assume that p(x) is a known\complete pdf and q(x) being
> the pdf of incomplete data, then
>
> Can I say that:
> 1) For EM, E-step: \int p(x) log(q(x)) dx


No, I think you need tio improve your notation...

let p(x,y), q(x,y) be the true and modelled density of joint random
variables, where x is observed and y is unobserved .. then the
expectation step evaluates E (log(q(y|x))), where the expectation is
with respect to the density q(y|x) (ie. the conditional density of y
given x). IE.

  E-step: l*(x)= \int q(y|x))  log(q(y|x)) dy

Then the maximisation step is to maximise l*(x) (ie using the sample
value of x)... but this is notionally an approximation to the real
target of maximising the expectation of l*(x), where the expectation
is with respect to the marginal distribution of x.

  M-step : maximise \int p(x) l*(x) dx.

This would probably make more sense if further notation were
introduced relating to the parameters being estimated.


David Jones


.
.
=================================================================
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