Re: [Cppqed-support] structure::InfiniteDetectedException when continuing a MCWF trajectory

[Raimar S. <rai...@ui...>]

Dear András,

thank you for adjusting and merging the fix.

On Saturday 21 April 2012 10:28:58 you wrote:

> Great! Thanks a lot for tracking this down. But this means that the first
> timestep after a continuation was always erroneous so far (that is, if the
> system time development had *Exact* part)? And this eventually produced an
> overflow when the trajectory was continued after long enough a time?

The stack when psi first gets corrupted is something like (left out some 
indirections):

quantumtrajectory::evolve ->
MCWF_Trajectory<RANK>::step ->
MCWF_Trajectory<RANK>::coherentTimeDevelopment ->
Evolved<A>::step ->
evolved::details::apply ->
gsl_odeiv_evolve_apply ->
MCWF_Trajectory<RANK>::derivs ->
Hamiltonian::addContribution ->
BinarySystem::addContribution ->
Hamiltonian<2,ONE_TIME>::addContribution

Here, after the call to addContribution of the binary Hamiltonian (the frees 
are ok in the first step, why?), the maximum absolute value in dpsidt is 
3.5e+130. I would think this is because the transformed non-hermitian 
Hamiltonian has some huge elements as we evaluate U(t) for large t. These 
large values carry on to psi as the odeiv stepper takes the huge derivatives 
and updates the state  (max is 4e+126). The next time derivs is called, the 
maximum absolute value of dpsidt is already on the order of 3e+257 and psi 
gets updated to something on the order of 3e+257. Then first dpsidt and 
afterwards psi overflows.
			
What I don't understand is why derivs gets called three times with different 
values of t (t=30, t=30.00014 and t=30.00021) before dpsidt overflows, but each 
time tIntPic0_=0. Shouldn't tIntPic0_ get updated?

And the question remains if continuation always was erroneous. I would argue 
that as long as the stepper makes it to the first update of tIntPic0_ without 
overflow, everything should be fine except if rounding errors skew the results. 
Physically large values of t should be legal, the problems arise if we cannot 
represent the transformation numerically accurate. 
 
Best regards
Raimar