The composition of multiple linear transformations is still a linear transformation; therefore, a multi-layer network without nonlinear activation functions has expressive power equivalent to a single-layer linear model.