Hi Vadim,
1. Many experts advise to do multiple k-weight fitting to deal with correlated variables. Should one always use multiple k-weights, or is it better to switch to one kw value once the correlations are taken care of – to refine the remaining variables? Does it make any difference?
I wouldn't say that multiple k-weights "take care of correlations", as the correlations are generally unavoidable, and don't always drop by very much when using multiple k-weights either. As an example, I just tried kweight=2 v. kweight = 1,2,3 on a test case, and saw Correl(S02, sigma2) drop from 0.90 to 0.88, and Correl(DelR,E0) drop from 0.87 to 0.80. The correlations are smaller, but hardly gone. But I still recommend using multiple k-weights most of the time because it seems to improve the stability and reliability of the fit results (that is, it's harder for a small change in fit parameters to give a different result). That is, I use it unless the performance hit is unacceptable.
2. When modeling the Debye-Waller factors for multiple-scattering paths, is it possible to express them in terms of the sigma^2's of single-scattering paths that correspond to the atoms involved in the multiple scattering events; i.e. for a core—atomA—atomB—core path, can sigma^2 be obtained by some combination of core—atomA and core—atomB sigmas? It seems intuitively that they should be related, and also that the amplitudes of multiple scattering paths should be more sensitive to disorder. Does this make any sense?
In general, there is not a simple relation between the sigma2 for single- and multiple-scattering paths. As Anatoly points out, for some collinear paths there are some fairly simple approximations that seem to work. Other than that, it is often safe to assume that a sigma2 for a MS path will be larger than that of a SS path that is one of its legs. ;). There are other approaches to _calculating_ sigma2 for single- and multiple-scattering paths. From what I can tell, these tend to be computationally intensive, highly system-dependent, and difficult to reduce to a set of parameters that can be refined to match data. --Matt