Outcomes of Frequencies calculations are usually quite sensitive to the geometry, so before computing the frequencies, one should make sure that the geometry is well converged at the level of the subsequent Frequencies calculation: the same model parameters and basis sets.
In all cases one should take care that the precision of Numerical Integration is adequate, preferably at least 5.0 (this is good advice anyway for a sound Frequencies calculation).
Doing one-point, rather than two-point differentiation will roughly save you half of the time needed to complete the calculation. Increasing the integration precision will work the other way. To obtain high-precision results using one-point differentiation requires for one thing that you use very small displacements (smaller than the defaults) and high accuracy of numerical integration. Recent studies [14, 108] suggest to use a) two-sided displacements, b) an integration precision of 6.0 (!!).
This may not always be feasible due to the high CPU costs, but it should at least stress the importance of accuracy in the computation of frequencies.
A computation of frequencies runs over discrete displacements of atomic coordinates. When using Cartesian displacement coordinates, the program applies symmetry to skip symmetry-equivalent displacements and thereby save CPU time. In the output and logfile you'll find in such a case that the 'frequency displacement counter' skips one or more values: the counter counts all possible displacements, while only the symmetry-unique ones are actually carried out.
Starting from ADF2005.01 symmetric displacements can be used. This speeds up the computation significantly and reduces the level of numerical noise in gradients by using the SMOOTH option.
