On Thu, Mar 18, 2004 at 10:50:28AM -0500, Maxim V. Lobanov wrote:
> > Your sample will probably broaden the lines (LX, LY, etc.) so much that
> > any attempt to vary the Gaussian
> > coefficients will yield nonsense.
> Just some remark (of course, I am not a great specialist):
> At least to my experience, there is always some Gaussian broadening from
> the sample as well, and (again it is my humble opinion only) at least U is
> better to allow to be refined.
> For example, Rietan manual states the problem in the following way:
> {
> U, V, and W tend to be highly correlated, with a result that various
> combinations of quite different values can lead to essentially the
> same variance, sigma^2. These three parameters, therefore, do not
> converge in a stable manner when refined simultaneously (Prince,
> 1993). In particular, refining P in addition to U, V, and W almost
> certainly affords a singular (non-positive definite) coefficient matrix.
This happens because parameters U, W and P in
sigma^2 = U*tan(th)^2 + V*tan(th) + W + P/cos(th)^2
multiply linearly dependent functions, as
tan(th)^2 = 1/cos(theta)^2 - 1
Therefore it makes no sense to refine U, W and P together as there
is an infinite number of U,W,P triplets that make up the same
sigma^2.
Pavol
