Dynamic allocation of memory has been implemented. Installation-defined workspace sizes are now only defaults which can be adjusted by input-keys. No need anymore to compile differently-sized program versions. It also improves in a general way the efficiency of memory usage.
Default (automatic) values for Numerical Integration precision have been increased AND now depend on the runtype (high default value for Frequencies, for instance).
Minor modifications in the Numerical Integration grid generator, leading to higher standard precision within the atomic core regions. This may result in slightly (but not essentially) different outcomes.
The electrostatic interaction energies, and their derivatives w.r.t. nuclear displacements, were computed from the fitted (i.e.: approximate) densities. An evaluation using the exact densities has been implemented. This generally leads to somewhat different results. The so-corrected inaccuracies in the old version are larger for atoms that had relatively inaccurate fit sets in their database files.
The SCF convergence criterion is relaxed during the intermediate cycles of optimizations. This reduces (usually) the total execution time. In some cases it may lead to using one or two geometry cycles more. This feature (a secondary SCF criterion) is controllable with the (extended) CONVERGENCE option to the SCF key. The (new) key FULLSCF can be used to suppress the feature altogether.
Optimization and TS search strategies have been modified to solve stability problems and improve robustness (geometry convergence sometimes failed after initial progress towards the minimum; TS searches very often failed to converge by randomly walking around the TS point).
Calculated energy gradients are adjusted to eliminate spurious components of rigid motions. This used to be treated in a slightly different way in Optimizations versus Frequency calculations. As a result, the computation of Frequencies in a previously optimized geometry could issue warnings that "GRADIENTS IN THE EQUILIBRIUM ARE NOT SMALL".
A new key DEPENDENCY has been introduced to handle numerical problems with very large basis sets that contain diffuse functions. It should not be necessary/useful to apply it to the standard basis sets provided with the package.
