Dear fellow EXAFSeteers,

I have a somewhat bizarre problem with Artemis fits. I am using Demeter 0.9.25 64bit on a Windows7 machine.

The Situation: I am trying to fit the first shell complexing ligands to aqueous metal complexes. I have sets of experiments at different temperatures but otherwise identical conditions and concentrations. When I fit all experiments separately, parameters such as coordination number and metal-ligand distance show systematic variation with temperature (as they should). Therefore, I would like to try fitting all these experiments together and define these parameters as functions of temperature, which would reduce the total number of free parameters relative to the experimental points and should also bring the uncertainties down a bit. To this end, I first made an Artemis project that contains several of the data sets (at different temperatures). Then I made individual parameters for each data set that are exactly as they are in the individual fits (parameter names are different, though) and duplicated the fourier transform parameters of the individual fits. Now each data set has its own set of parameters. No parameters are shared between data sets and there are no connections between data sets via "def" functions in the GDS or anything.

The Problem: When  start the fit, I get different results from the individual fits. Still mostly within the calculated error values (which are quite large at the moment), so not "significantly different" in the statistical sense of the words but still very different numbers. Whats more, if I add another data set with corresponding parameters (again no connection to the other data sets or parameters), the results for the parameters of the other data sets change again.

Am I right in assuming that this should not be the case or is there a connection which I am missing? Might this be a bug? Some fits react a bit sensitive to the choice of k- and R-space windows (but these were identical in individual and simultaneous fits). My fits are pushing the number of independent parameters a bit, according to the log-files (which is why I want to do the simultaneous fits with T-dependent parameters). Might this cause instability in the fit routine. In the simultaneous fits, the number of highly correlated parameters (>0.95) skyrockets (is ok in individual fits), which may be related to the next point. I suspect it may also have to do with the way Artemis decides that the fit is "done". If this is evaluated globally for all parameters somehow, then the observed behavior would not be surprising. But if it is evaluated on a per-parameter or per-data set basis, then it should not be like this.

Finally: Is my approach reasonable at all or is that a bad idea? Does someone else use this method? I have not seen it published but then again I am new to this and many publications are a bit reserved on the topic of fitting strategy.

Thanks a lot and sorry for the long read,

Anselm