This is more of a philosophical question.
Suppose that you are trying to choose a good password for a particular online service, say your bank's e-banking service. Now the bank has some restrictions on the passwords that it allows: you can only use (lowercase) letters and (decimal) digits, and the password can be up to 6 characters long. So what is a "good" way to select such a password?
One could do the following (call this scheme 1): select randomly 6 characters from the set [a-z0-9] with uniform probability. This scheme has the disadvantage that it may produce "weak" passwords such as those containing only letters or only digits (there is a 14.2%  chance of this happening).
So here's another idea (call this scheme 2): select randomly (and uniformly) one letter, one digit, and 4 characters from the set [a-z0-9]. Then use a random permutation of these 6 characters as the password. This guarantees that the password will have characters from both classes (letters and digits).
So the question is, which of the two scheme produces "better" passwords?
On the one hand, the second scheme produces passwords that will likely be more resilient to simple brute force attacks. On the other hand, the first scheme allows strictly more combinations than the second one: 2^31.0  vs 2^30.8 ; in fact the set of passwords of scheme 1 is a superset of those of scheme 2.
Note: I'm looking at this from the perspective of the party that generates the password and not the party that sets the password policy.
 (26^6+10^6)/36^6 = 0.142
 log2(26^6) = 31.0
 log2(36^6-26^6-10^6) = 30.8 i.e. all combinations of 6 characters excluding the ones with only letters (26^6) and the ones with only digits (another 10^6)