Next: lsqlin, Previous: curvefit_stat, Up: Residual optimization [Index]
The backends for objf_type == "wls" (currently the only supported
type of objective function) compute covd (due to user request or
as a prerequisite for covp and corp) as a diagonal matrix
by assuming that the variances of data points are proportional to the
reciprocal of the squared weights and guessing the factor of
proportionality from the residuals. If covp is not defined
(e.g. because the Jacobian has no full rank), an attempt is made to
still compute its uniquely defined elements, if any. In corp,
interdependent parameters can cause elements of 1 or -1,
which in this case are not the real coefficients of correlation, but
rather indicate the direction of parameter interdependence. To be
consistent with this, an attempt is made (often not successful) to
identify parameter interdependence and mark it with elements of 1
or -1 in corp even if the respective elements of
covp can not be computed.