7.3 Linear Least-Squares

It might make sense to skip this linear case, as the model functions we use for crystallography are anything but linear, but this case leads us into the case of non-linear least-squares minimization. A linear equation is one where each parameter contributes independently, so our model function, \(M(\mathbf {p},x)\) can be rewritten as \[M(\mathbf {p},x) = p_1 M_1(x) + p_2 M_2(x) + ... = \sum _k p_k M_k(x)\] where the \(M_k(x)\) terms are some function of \(x\) that we can evaluate.

We want to find the \(p_i\) values so that \[\chi ^2 = \sum _j w_j[y_{obs,j}-M(\mathbf {p},x_j)]^2\] is minimized. When that is true, then the derivative of \(\chi ^2\) with respect to each \(p_k\) value will be zero. That allows us to generate our system of \(m\) simultaneous equations (where k goes from 1 to \(m\) for each parameter value, \(p_k\)):

\[\frac {\partial \chi ^2}{\partial p_k} = -2\sum _j w_j[y_{obs,j}-M(\mathbf {p},x_j)]\frac {\partial M(\mathbf {p},x_j)}{\partial p_k} = 0\] Note that since \(M\) is a linear equation, \(\frac {\partial M(\mathbf {p},x_j)}{\partial p_k} = M_k(x_j) \) and substituting \(M(\mathbf {p},x_j) = \sum _i p_i M_i(x_j)\) we get: \[\sum _j w_j[y_{obs,j}-\{\sum _i p_i M_i(x_j)\}]M_k(x_j) = 0\] which can be rearranged as \[\sum _j w_jy_{obs,j}M_k(x_j) = \sum _i p_i \sum _j w_j M_i(x_j) M_k(x_j) \]

This can be cast as our previous equation, \[\sum _{k=1}^m A_{jk} p_k = b_j\] where \(b_j = \sum _j w_jy_{obs,j}M_k(x_j) \) and \(A_{jk} = \sum _j w_j M_i(x_j) M_k(x_j)\). Since we know the values of \(w_j\), \(y_{obs,j}\) and can compute \(M_k(x_j)\) we can solve for the \(p_k\) values by computing \(\mathbf {A}^{-1}\) as we did before in the previous section.