Re: Implementation of the Schanda, Lescanne and Marion numeric dispersion model code in relax.Re: Implementation of the Schanda, Lescanne and Marion numeric dispersion model code in relax.P. S. Actually, I might take some of this text into the relax manual :)I have a few other implementation questions, but I'll send these in a different message for saner threads in the mailing list archives.Hi,The 2-site numeric dispersion models for CPMG data are now fully functional within relax! The following describes how these are implemented in relax (to be permanently archived at http://dir.gmane.org/gmane.science.nmr.relax.devel). I hope that you will check all of this by running relax and seeing for yourself. I also have a few questions below. Note that the code is in the relax_disp branch of the source code repository and that there is no official release at http://www.nmr-relax.com/download.html yet.
I have documented the numeric models now present in relax in Chapter 10 of the user manual. I have compiled the current manual for the branch and uploaded it to http://download.gna.org/relax/manual/relax_disp_manual.pdf. Please have a look and see what you think. The text is extremely minimal and needs to be expanded. If you have any suggestions or additional material (equations, figures, etc., and in LaTeX preferably), it would be happily accepted. For example equations for the matrices and the propagation used. The aim is to be a complete stand-alone relaxation dispersion reference for the users, pointing them to all the relevant literature at all possible points. Cheers!
As a summary, from the relax_disp.select_model user function description, the numeric models are:
'NS 2-site 3D': The reduced numerical solution for the 2-site Bloch-McConnell equations using 3D magnetisation vectors whereby the simplification R20A = R20B is assumed. Its parameters are {R20, ..., pA, dw, kex}.
'NS 2-site 3D full': The full numerical solution for the 2-site Bloch-McConnell equations using 3D magnetisation vectors. Its parameters are {R20A, R20B, ..., pA, dw, kex}.
'NS 2-site star': The reduced numerical solution for the 2-site Bloch-McConnell equations using complex conjugate matrices whereby the simplification R20A = R20B is assumed. It has the parameters {R20, ..., pA, dw, kex}.
'NS 2-site star full': The full numerical solution for the 2-site Bloch-McConnell equations using complex conjugate matrices with parameters {R20A, R20B, ..., pA, dw, kex}.
'NS 2-site expanded': The numerical solution for the 2-site Bloch-McConnell equations expanded using Maple by Nikolai Skrynnikov. It has the parameters {R20, ..., pA, dw, kex}.
I would be interested to hear if you have suggestions for saner model names.In the auto-analysis in relax - the blackbox analysis script to make the lives of users extremely easy - the *full models are not turned on by default. If curious, you can see auto-analysis script in the file auto_analyses/relax_disp.py. The auto-analysis script is normally hidden from the user. Also, please run the relax GUI with:
$ relax --guiand start the relaxation dispersion analysis to see how it is implemented. Note that the auto-analyses are used for the operation of all GUI analyses. Using scripts, the relax user is free to do anything they wish and implement any protocol they can imagine (if useful for other users, these can then be converted into an auto-analysis).
Part of the auto-analysis that I have implemented is the concept of using optimised parameters from a simpler or equivalent model to speed up optimisation by avoiding an expensive grid search. For the 'full' models above, I use the parameters of the simpler model and start with R20A = R20B = R20. Then, during optimisation, R20A and R20B diverge.
For the model equivalence, this is a bit different. I use the optimised parameter values from the analytic Carver and Richards model (CR72) as the starting parameters for the 'NS 2-site 3D', 'NS 2-site star' and 'NS 2-site expanded' models. This avoids the grid search and really speeds up the optimisation by orders of magnitude. It can be done because the CR72 results are often very similar to the numeric results. And even a failed CR72 solution is close enough to the minima to be used as a starting point for the numeric models.
Cheers,
Edward