Spellucci2009-07-25 22:50:33

In article <1102078227.oUG+AGoqz4rMvHBz6gPR9A@teranews>,

Martin Magnusson

g(z)= f(S*z), i.e. x=S*z

then

grad g(z) = S^T grad f(S*z) grad is a column

Hessian g(z) = S^T Hessian(f(S*z)) *S

hence your setting of the scaled Hessian is o.k. maybe you forgot to

scale the gradient properly?

but I cannot see why you scale such ..

one way to scale is :

the diagonal of Hessian g(z) should be the unit matrix and from your

data I cannot see a dramatic change here. but the eigenvalues of the Hessian

are here

1.0e+07 *

0.0013

0.0079

1.8872

so it is a little illconditioned and you might try to use a proper

unitary transformation plus scaling.

hth

peter

## Leave a Reply

You must be logged in to post a comment.