koopman bug and fix

[euhruska]

Found a bug in the koopman implementation by analysing https://github.com/radical-collaboration/extasy-grlsd/issues/77 together with @fnueske and a suggestion how to solve it. Take line https://github.com/markovmodel/PyEMMA/blob/1a280481ed6cbbfc3fbceaacfa22136e26983678/pyemma/_ext/variational/solvers/direct.py#L82 with s = _np.array([ 1.90963508, 1.72242421, 1.18440072, -1.16845181]) leads to an m=4 in line https://github.com/markovmodel/PyEMMA/blob/1a280481ed6cbbfc3fbceaacfa22136e26983678/pyemma/_ext/variational/solvers/direct.py#L89 , while m=3 is expected (as for 3 positive eigenvalues, removing later the 4th)

Suggestion: increase the value 1e-16 to 1e-14 in line https://github.com/markovmodel/PyEMMA/blob/1a280481ed6cbbfc3fbceaacfa22136e26983678/pyemma/_ext/variational/solvers/direct.py#L84 This will lead to m=3, but not sure if that's the best fix as this might depend on the individual machine precision of the user

System: pyemma 2.5.4, python 2.7.14, numpy

[marscher]

solved, was a bug, caused by machine precision

Originally posted by @euhruska in https://github.com/radical-collaboration/extasy-grlsd/issues/77#issuecomment-421409376