Hi,

I have an applied statistics problem regarding invariant subspaces,
which I am unable to solve.  Perhaps someone can help me?
 
I am trying to minimise a function
 
  f = ln |U'*C*U| + trace( (U'*C*U)^(-1) (U'*S*U) )
 
with respect to the nxp matrix U (p<n), subject to the constraint
that U'*U=I.  C and S are both full-rank nxn matrices.
 
After applying some horrible matrix calculus using Lagrangian
multipliers, I am able to determine that a necessary condition on the
minimum is
 
  U*U'*C*U = C*U + S*U - C*U (U'*C*U)^(-1) U'*S*U.
 
Evidently only the subspace spanned by U is unique, since U*R also
satisfies the equations for any orthogonal R.  This observation also
makes it possible to assume either C or S diagonal without any loss of
generality.  I can also simultaneously diagonalise them (and even make
one of them identity), but only at the cost of complicating the
constraint.
 
Even with these insights, a method of solution has completely eluded me.
I can't even seem to solve it for the apparently simpler case where U
is a nx1 vector.
 
Any suggestions would be extremely, extremely welcome.
 
Thanks,
Fred

.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to