quick check that `hmm.sample_by_observation_probabilities` isn't the culprit

trajectory_index_test_case.py

import numpy as np
import numpy.random as npr
import pyemma

# generate random trajectories, where each trajectory has a random length between 100 and 1000.
# since the trajectories have different lengths, we will by likely to see an error if `hmm` somehow
# scrambles the trajectory indices.
dtrajs_A = [npr.randint(0,100,x) for x in npr.randint(100,1000,8)]
dtrajs_B = [npr.randint(100,200,x) for x in npr.randint(100,1000,9)]
dtrajs_C = [npr.randint(200,300,x) for x in npr.randint(100,1000,10)]
dtrajs_D = [npr.randint(300,400,x) for x in npr.randint(100,1000,11)]

dtrajs = dtrajs_A + dtrajs_B + dtrajs_C + dtrajs_D

hmm = pyemma.msm.estimate_hidden_markov_model(dtrajs, nstates = 10, lag = 10, maxit = 1)

# draw many (traj_ind, frame_ind) pairs according to the MSM
indices = hmm.sample_by_observation_probabilities(10000)

# for each (traj_ind, frame_ind) pair, check that the frame_ind isn't out of bounds
for (traj_ind, frame_ind) in np.vstack(indices):
    assert(len(dtrajs[traj_ind]) > frame_ind)