- solution of the CCS/CIS eigenvalue problem to generate
reasonable start vectors; the eigenvectors are converged in
this step only to a remaining residual norm <
`preopt`

- pre-optimization of the eigenvectors by a robust
modified Davidson algorithm (see ref. [10])
using the
`LINEAR CC RESPONSE SOLVER`until the norm of all residuals are below`preopt`

, combined with a DIIS extrapolation for roots assumed to be converged below the threshold`thrdiis`

. - solution of the nonlinear eigenvalue problem with a DIIS
algorithm using the
`DIIS CC RESPONSE SOLVER`until the norm of the residuals are below the required threshold`conv`

- almost degenerate roots in the same symmetry class
- complex roots (break down of the CC approximation close to conical intersections)
- large contributions from double excitations

`thrdiis`

(controlling the DIIS extrapolation in the
linear solver) should be set about one order of magnitude
smaller than the smallest difference between two eigenvalues
and `preopt`

(controlling the switch to the DIIS solver)
again about one order of magnitude smaller then `thrdiis`

.
Tighter thresholds or difficult situations can make it necessary
to increase the limit for the number of iterations `maxiter`

.

In rare cases complex roots might persist even with tight convergence
thresholds. This can happen for CC2 and CIS(D_{∞}) close
to conical intersections between two states of the same symmetry,
where CC response can fail due to its non-symmetric Jacobian.
In this case one can try to use instead the ADC(2) model.
But the nonlinear partitioned form of the eigenvalue problem used
in the `ricc2` program is not well suited to deal with such situations.

Large contributions from double excitations can not be monitored
in the output of the (quasi-) linear solver.
But it is possible to do in advance a CIS(D) calculation.
The CIS(D) results for the `||T2||`

diagnostic correlate usually
well with the CC2 results for this diagnostic.
Else the DIIS solver will print the `||T2||`

diagnostics
in each iteration if the print level is set > 3.
States with large double excitation contributions converge notoriously
slow (a consequence of the partitioned formulation used in the `ricc2` program).
However, the results obtained with second-order methods for
double excited states will anyway be poor. It is strongly recommended
to use in such situations a higher-level method.