Version of 23/3/01

• REOPT keyword specifies that the eigenvector to be followed should be reoptimised in a BFGSTS search after the EF step and before the tangent space minimisation.
• bfgsts.f changed to allow for zero tangent space steps in the initial phase.
• FREEZE n1 n2 etc.: specifies that atoms n1, n2, etc. should be frozen. Only works for LBFGS minimization and BFGSTS if 3 or more non-linear atoms are frozen.
• ITINC=.TRUE. was set in the wrong place in lbfgs-- fixed. Fixed bug in BFGSSTEP. Periodic boundary conditions could cause glitches in LBFGS--fixed.
• Call to potential skipped in BFGSTS if no tangent space minimisation is to be performed. Trust ratio is not known, and steps of constant size are taken uphill until the curvature becomes negative.
• Iteration to convergence added in ORTHOG. XLBFGS affected, as well as other EF calculations.
• Changed BFGSTS so that the EF step has length PUSHOFF until there is a negative eigenvalue. Using large PUSHOFF steps when the curvature is positive requires that the eigenvector be converged more accurately, i.e. more steps needed in the variational BFGS/BFGS approach. Smaller PUSHOFF means that fewer such steps are required. However, using small PUSHOFF and a small number of steps (e.g. 2) for the variational eigenvector determination does not work with minimisation steps in the tangent space before we have a negative eigenvalue. SW works best with smaller PUSHOFF, small number of variational eigenvalue steps and no tangent space minimisation steps until we have a negative curvature--BLJ works best with a large PUSHOFF, accurate eigenvalue determination and a few tangent space minimisation steps.
• Fixed infinite loop in ORTHOG. Changed COORDS to be relative to centre of coordinates (or centre of mass).
• Added GAMESSUS and GAMESSUK keyword for GAMESS-US and GAMESS-UK.
• Changed BMIN and BEIG to put back previous coordinates if we are going to stop a minimisation without convergence. LBFGS doesn't know that we are going to stop, and changes the geometry!
• Added Dzugutov potential.
• Fixed bug in TRAD calculation.
• Removed estimation of inverse Hessian elements from Hessian.
• Changed lbfgs so that orthogonalisation occurs for the first step too.
• Fixed bug in fixed uphill step size for positive eigenvalue to make sure the steps are actually uphill!
• Added RADIUS keyword to define the radius of a spherical container,
• Added FIXD (using hard sphere moves) and RANSEED keywords.
• Fixed bug in HSMOVE for periodic boundary conditions distance rescaling.
• Fixed BFGSSTEP to work without reading a vector both with and without the NOHESS key. The vector is calculated as required.
• Added keywords INDEX and KEEPINDEX.
• Added keywords BSMIN and RKMIN for steepest-descent gradient only paths.
• Added FIXAFTER keyword and 'LC' atom type.
• Added GRADSQ keyword for optimisation of the modulus gradient.
• Added PV keyword for constant pressure optimisations.