As we have seen in Sec. , the inference
of the `unobserved' variables, based on the `observed' one,
for the problem represented graphically in the `Bayesian'
network of Fig. , consists
in evaluating `somehow'
|
|
|
(88) |
from which the most interesting probability distribution, at least
for the purpose of this paper,
can be obtained by marginalization
(see also Appendix A). Besides a normalization factor,
Eq. () is proportional to Eq. (),
hereafter indicated by `' for compactness, which can be
written making use of the chain rule obtained
following the bottom-up analysis of the graphical model of
Fig. :
in which
where
is the Kroneker delta (all other symbols belong to the
definitions of the binomial and the Beta distributions) and we have left
to define the prior distribution .
The distribution of interest is then obtained by summing up/integrating
where the limits of sums and integration will be written in
detail in the sequel.
As a first step we simplify the equation by summing
over
and and exploiting the Kroneker delta
terms () and ().
We can then replace
with
and with
with the obvious constraints
(i.e.
) and
.
The inferential distribution of interest
,
becomes then, besides constant factors and indicating
all the status of information on which the inference is based
as `', that is
,
where we have dropped all the terms not depending on the variables
summed up/integrated.
The two integrals appearing
in Eq. () are, in terms of the generic variable ,
of the form
d,
which defines the special function beta
B,
whose value
can be expressed
in terms of Gamma function as
B.
We get then