If you want the true maximum and minimum, just track them with variables.
Depending on your input data, a probabilistic min/max may be "accurate enough". So if you only look at every number with 50% chance, then you probably have only a 50% of having the exact min or max. But maybe already a 75% chance of having at least the second largest/smallest etc. However, computing the random number to do an unbiased sampling is already more expensive than looking at all numbers for min/max. Skipping every second is dangerous: there might be an even/odd pattern in the data that screws you badly.