Forte et al. [1991] used the method of Bayesian inference to invert for the radial viscosity profile which best fit the observed plate velocities in the degree range l = 1-8. Using a tomographic shear wave model SH425.2 [ Su and Dziewonski, 1991] as the driving force, they parameterized the mantle viscosity in five layers (0-100 km, 100-400 km, 400-670 km, 670-2600 km and 2600 km to the core mantle boundary). Their inversion produced a viscosity model with a low viscosity in the transition zone (i.e., 400-670 km) and high viscosity in the 100-400 km layer with a factor of 42 jump at 670 km. The plate-like divergence predicted by this model explains 48% of the variance in the observed plate divergence (in the range l = 1-15), while a model with an isoviscous mantle actually increases the variance. There is a significant trade-off between the top three layers (0-100 km and 400-670 km are correlated and 100-400 km is anti-correlated with the others), so a model with a low viscosity in the 100-400 km layer and higher viscosities in the 0-100 km layer and 400-670 km layers fits the data nearly as well.