In the EigenFollowing geometry optimization method, the geometry is changed on
each cycle; if the
ΔH_{f
} decreases, the cycle is completed. If it does
not drop, the step-size is reduced, and the
ΔH_{f
} recalculated. Only
when the
ΔH_{f
} decreases, compared to the previous cycle, is the current
cycle considered to be successful. During the calculation, the confidence
level or trust radius is continuously checked. If this becomes too small, the
calculation will be stopped. This can readily happen if (a) the geometry was
already almost optimized; (b) a reaction path or grid calculation is being
performed; (c) if the geometry is in internal coordinates and "big rings" are
involved; or (d) if the gradients are not correctly calculated (in a
complicated C.I., for example).

For cases (a) and (b), add `LET` and `
DDMIN=0`. In case (c) use
either mixed coordinates or entirely Cartesian coordinates. Case (d) is
difficult--if nothing else works, add `NOANCI`; this will always cause
the derivatives to be correctly calculated, but will also use a lot of time.

Adding `LET` and `DDMIN=0` is often very effective, particularly
when reaction paths are being calculated. The first geometry optimization might
take more cycles, but the resulting Hessian matrix is better tempered, and
subsequent steps are generally more efficient.