Why are separable convolutional layers not recommended after layers having too few channels?

Since separable convolutional layers only have one spatial filter per input channel, you should avoid using them after layers that have too few channels, such as the input layer.

from "Géron, Aurélien. Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow (p. 475). O’Reilly Media. Kindle Edition. "