Does that mean that MLP are not universal function approximators? No.

Its a fact that MLP is capable of fitting arbritrary functions.

Does anything here deviate from the theory? No.