Sampling Theorems for Learning from Incomplete Measurements