The problem also seems to be related to the number of multinomial trials. Using 10 instead of 1e5 in sample = np.random.randint(0, 1e5, 10) produces much less bias. But already a value of 100 seems critical.
The problem also seems to be related to the number of multinomial trials. Using 10 instead of 1e5 in sample = np.random.randint(0, 1e5, 10) produces much less bias. But already a value of 100 seems critical.