DAMP Instruction
DAMP damp [1] limse [15]
damp is usually left at the default value unless there is severe correlation,
e.g. when trying to refine a pseudo-centrosymmetric structure, or refining
with few data per parameter (e.g. from powder data). A value in the range
1-10000 might then be appropriate. The diagonal elements of the least-squares
matrix are multiplied by (1+damp/1000) before inversion; this is a version of
the Marquardt algorithm (J. Soc. Ind. Appl. Math., 11 (1963) 431-441). A
side-effect of damping is that the standard deviations of poorly determined
parameters will be artificially reduced; it is recommended that a final least-
squares cycle be performed with little or no damping in order to improve these
estimated standard deviations. Theoretically, damping only serves to improve
the convergence properties of the refinement, and can be gradually reduced as
the refinement converges; it should not influence the final parameter values.
However in practice damping also deals effectively with rounding error
problems in the (single-precision) least-squares matrix algebra, which can
present problems when the number of parameters is large and/or restraints are
used (especially when the latter have small esd's), and so it may not prove
possible to lift the damping entirely even for a well converged refinement.
If the maximum shift/esd (excluding the overall scale factor) is greater
than limse, all the shifts are scaled down by the same numerical factor so
that the maximum is equal to limse. If the maximum shift/esd is smaller than
limse no action is taken. This helps to prevent excessive shifts in the early
stages of refinement.