To all reviewers: on the limitation of problem setting: Gaussian inputs and non-overlapping filters

Our current result is for Gaussian inputs, and single hidden layer non-overlapping CNNs. "it would be much better to write down the proof of the claim in line 209... are required. We will add the derivation in line 209. "the proof indicates that η "where and how the assumptions in Theorem 4.3 are used and to explain what the conditions mean." Thank you for pointing out these issues.