If ints are 32 bit you would expect to be able to add 1 to 2^20, but that causes an overflow with this implementation. (low + high) / 2 is arguably worse because you turn the algorithm from working on any array into one that only works on arrays that are small enough not to cause the overflow.
