Somewhere in the processing a division by the b-value is missing.
Note: the b-value in S_k = S_0 * exp[ -b * g_T * D * g ]
When compared to the Weighted LS method, the tensor values differ by a factor of the b-value ( 1000 in my case ), example: D_eigval = 0.001706 , D_wls = 1.7180
This leads to problems when computing the residual of the tensor fit. Here the exponent argument for the modelled signal [ -b * g_T * D_estimated * g ] contains again the b value and hence moves the whole exp[...] to almost zero if the values of the estimated tensor D_estimated are not scaled correctly.