To solve (5.15), define
as the element-wise inverse of
.
Then, (5.15) can be written
,
which means that
must be fully in the column space of , or that
|
(5.17) |
for some vector .
Therefore, to find the mean of the surrogate density,
we solve for the free variable such that
where
We can solve this by driving the square error to zero:
We can easily find derivative of
with respect
to :
where
is the diagonal matrix with diagonal elements
|
(5.22) |
We have found that if we use the negative-definite
Hessian approximation
|
(5.23) |
the resulting Newton-Raphson algorithm has excellent convergence properties
when starting with
.
Let
be the value of
at the solution to (5.18).
The algorithm to find
can be summarized as follows.
- Set iteration counter .
- To initialize, let
where is the vector of ones.
-
. Initially,
will be the
vector of ones.
-
- Compute derivative and Hessian according to (5.21),(5.22) and (5.23), then
update :
- Increment and go to step 3.
The algorithm is implemnted by software/lam_solve.m with dbound=0.
Although this method corresponds to classical
methods, it is based on a novel sampling argument
and can be extend to other manifolds as we will see.