OK, you have asked a lot of questions in this vein, and they've mostly been answered, sometimes in great detail.
This is really not much different than the rest. You do seem to understand the idea of mean cycle size for such properties, which can only be understood when AES is modelled as a good pseudorandom permutation.
Your function $f,$ if it is a projection to a random subset that function is linear and partitions the space into different equal sized subsets. You already understand this from your comment.
If AES is pseudorandom, you cannot do better than the various time-memory tradeoffs, as expounded by Hellman, much later by Wiener (parallel collision search) etc.
If $f$ was nonlinear, it may not partition the input set $\{0,1\}^n$ to equal subsets, so then the analysis would no longer be correct, you might be stuck in a "small" subset, so the joint distribution of the subset sizes, namely
$$
X_a=\{ x \in \{0,1\}^{128}: f(x)=a\}
$$
would need to be analyzed as $a$ ranges over the codomain of $f$.
Finally, since AES can be well modelled as a good pseudorandom permutation it does not matter even one bit whether the key is zero, or it has mostly zero components. If it did, it would have come out at a first level with Sbox analysis, what if most input keys to most Sboxes were zero? Who cares, there is all the mixing steps, and constant addition which is composed with the galois field inverse in the Sbox