$GRADEX group  (optional, for RUNTYP=GRADEXTR)
 
   This group controls the gradient extremal following
algorithm.  The GEs leave stationary points parallel to
each of the normal modes of the hessian.  Sometimes a GE
leaving a minimum will find a transition state, and thus
provides us with a way of finding that saddle point.  GEs
have many unusual mathematical properties, and you should
be aware that they normally differ a great deal from IRCs.
 
   The search will always be performed in cartesian
coordinates, but internal coordinates along the way may
be printed by the usual specification of NZVAR and $ZMAT.
 
METHOD = algorithm selection.
         SR   A predictor-corrector method due to Sun
              and Ruedenberg (default).
         JJH  A method due to Jorgensen, Jensen and
              Helgaker.
 
NSTEP  = maximum number of predictor steps to take.
         (default=50)
 
DPRED  = the stepsize for the predictor step.
         (default = 0.10)
 
STPT   = a flag to indicate whether the initial geometry
         is considered a stationary point. If .TRUE.,
         the geometry will be perturbed by STSTEP along
         the IFOLOW normal mode.
         (default = .TRUE.)
 
STSTEP = the stepsize for jumping away from a stationary
         point. (default = 0.01)
 
IFOLOW = Mode selection option.  (default is 1)
         If STPT=.TRUE., the initial geometry will be
         perturbed by STSTEP along the IFOLOW normal mode.
         Note that IFOLOW can be positive or negative,
         depending on the direction the normal mode
         should be followed in. The positive direction
         is defined as the one where the largest component
         of the Hessian eigenvector is positive.
 
         If STPT=.FALSE. the sign of IFOLOW determines
         which direction the GE is followed in. A positive
         value will follow the GE in the uphill direction.
         The value of IFOLOW should be set to the Hessian
         mode which is parallel to the gradient to avoid
         miscellaneous warning messages.
 
GOFRST = a flag to indicate whether the algorithm should
         attempt to locate a stationary point.  If .TRUE.,
         a straight NR search is performed once the NR
         step length drops below SNRMAX.  10 NR step are
         othen allowed, a value which cannot be changed.
         (default = .TRUE.)
 
SNRMAX = upper limit for switching to straight NR search
         for stationary point location.
         (default = 0.10 or DPRED, whichever is smallest)
 
OPTTOL = gradient convergence tolerance, in Hartree/Bohr.
         Used for optimizing to a stationary point.
         Convergence of a geometry search requires the
         rms gradient to be less than OPTTOL.
         (default=0.0001)
 
HESS   = selection of the initial hessian matrix,
         when STPT=.TRUE.
       = READ causes the hessian to be read from a $HESS.
       = CALC causes the hessian to be computed. (default)
 
 
  ---- the next parameters apply only to METHOD=SR ----
 
DELCOR = the corrector step should be smaller than this
         value before the next predictor step is taken.
         (default = 0.001)
 
MYSTEP = maximum number of micro iteration allowed to
         bring the corrector step length below DELCOR.
         (default=20)
 
SNUMH  = stepsize used in the numerical differentiation
         of the Hessian to produce third derivatives.
         (default = 0.0001)
 
HSDFDB = flag to select determination of third derivatives.
         At the current geometry we need the gradient, the
         Hessian, and the partial third derivative matrix
         in the gradient direction.
 
         If .TRUE., the gradient is calculated at the
         current geometry, and two Hessians are calculated
         at SNUMH distance to each side in the gradient
         direction.  The Hessian at the geometry is formed
         as the average of the two displaced Hessians.
 
         If .FALSE., both the gradient and Hessian are
         calculated at the current geometry, and one
         additional Hessian is calculated at SNUMH in the
         gradient direction.
 
         The default double-sided differentiation produces
         a more accurate third derivative matrix, at the
         cost of an additional wave function and gradient.
         (default = .TRUE.)
 
           * * * * * * * * * * * * * * * * * * *
           See the 'further information' section
           for some help with GRADEXTR runs.
           * * * * * * * * * * * * * * * * * * *