While it is well known that uninformed random test generation often does a poor job at probing the parameter space, this has nothing to do with the CLT. Even hi

While it is well known that uninformed random test generation often does a poor job at probing the parameter space, this has nothing to do with the CLT. Even his first example with the bounded queue, one where the CLT does make predictions, he is incorrect to apply it. That's because what we care about is not the difference between the number of enqueue and dequeue operations (which has expected value 0), but the magnitude of the difference (which has expected value sqrt(N)). And what we care about is not the magnitude of the difference after completing the tests, but the greatest magnitude at any point along the random walk. This is governed by a more advanced theorem called the Law of the Iterated Logarithm, which tells us that the probability of finding the bug is quite a bit higher than the author realizes.