Next: Appendix: Adaptive neural algorithms
Up: Appendix: Proofs
Previous: Proof of convergence of
Denote by
the result of applying
once the iteration step 2 in (26) on .
Let
be the eigenvalue decomposition of
.
Then we have
|
|
|
(35) |
|
|
|
(36) |
Note that due to the normalization, i.e. division of by
,
all the eigenvalues of
are in the interval [0,1].
Now, according to (35), for every eigenvalue
of
,
say ,
has a corresponding
eigenvalue
where h(.) is defined as:
|
(37) |
Thus, after k iterations, the eigenvalues of
are obtained as
)))),
where h is applied k times on the ,
which
are the eigenvalues of
for the original matrix
before the iterations.
Now, we have always
for
.
It is therefore clear that all the eigenvalues of
converge to 1, which means that
,
Q.E.D.
Moreover, it is not difficult to see
that the convergence is quadratic.
Next: Appendix: Adaptive neural algorithms
Up: Appendix: Proofs
Previous: Proof of convergence of
Aapo Hyvarinen
1999-04-23