Roughly translated, this is saying that BigIntegers should not have leading zeroes in the representation. The word "minimal" is used because a BigInteger represented this way is represented with an array of minimum length (since an array that was identical except for leading zeroes would represent the same BigInteger, and therefore there would not be "exactly one representation for each BigInteger value").
Edited to add: of course, the value 0 can be represented with an array of length 0, thus ensuring that the requirement "most significant int must not be zero" is not violated.