How about the approximation capability of a convolutional neural network? -

globally, deep cnn series of compositions of linear function , sigmoid function. how approximation capability?

given enough neurons, neural networks can approximate function.

here nice, visual proof.