If the random sigma value fails, then try the next one (sigma+1),
and so on, instead of drawing a new random value. For small n,
this will avoid failures (for n=5, the probability of failure
with param=1 goes down from 1/150 to 0).