The best lower bound for representing MAX_N is 2 hidden layers while the current upper bound is logarithmic in N

The best lower bound is 2, while the current upper bound is logarithmic in $N$.

http://arxiv.org/abs/2608.25221v1