Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
Drexel University or The Math Forum.
|
|
|
|
Re: Reconstruct determinant of the Hessian
Posted:
Dec 17, 2012 6:02 AM
|
|
karl <oudeis@nononet.com> writes:
>Hi all,
>
>in quasi Newton methods in optimization the Hessian of the target function is reconstructed.
>This means that one needs to store a matrix in each step.
>If one needs only the *determinant* of the Hessian, is
>there a trick to get it without storing the whole matrix?
>
>Thanks
>
>Karl
no!
clearly there is a recursive formula for the determinant, quite easy to
apply if you use the rank-one-multiplicative form of the update.
but how would you compute the next "s" and the next "y" needed for this
without computing them? and in order to compute them you need the matrix.
and the ''limited memory'' version also needs to store old vectors, and due
to its numerical instability works not well for longer memorys
sorry, no help
peter
|
|
|
|
