#9 Failure of the 3-site dispersion models when pB and pC are zero.

[Edward d'Auvergne]

This was discovered by Luca Codutti in the ilsacripante relax fork on GitHub. All details were reported at https://github.com/nmr-relax/relax/pull/6. Further investigation shows that this is not only a 3-site MMQ model issue, but it also affects the 3-site R1rho model.

[Edward d'Auvergne]

This is specifically the issue from https://github.com/nmr-relax/relax/pull/6:

1st error in grid search
relax gives RelaxWarning error: invalid value encountered in true_divide during grid search.
nan data give inf values during simplex minimization which break the minimization.

The trace back brings you to the ns_mmq_3site.py function, and it turns out that the error is straightforward:
pB_pC is defined in relax/lib/dispersion/ns_mmq_3site.py as

pB_pC = pB + pC k_BA = pA * kex_AB / pA_pB k_AB = pB * kex_AB / pA_pB k_CB = pB * kex_BC / pB_pC k_BC = pC * kex_BC / pB_pC

As pC is defined as 1 -pA -pB, you get a divide by 0 error 3 lines later if pA is set to 1 and pB to 0 in the grid search. For this reason I added a pB_lower and a pA_upper function in /specific_analyses/relax_disp/parameter_object.py which specifically sets pB_lower to 0.01 if 3site exchange models are used and pA_upper to 0.99. This is a workaround, if you guys have a better solution I would be glad.

[Edward d'Auvergne]

This was fixed in commit 141031.