Why is additivity assumption in additive models unrealistic? Is it more unrealistic than the assumption of Linearity? I have gone through the following questions regarding the assumptions of additive models. 
Assumptions of additive model
What does the additive assumption mean?
However, I do not quite get the intuitive understanding of why additive assumptions are regarded as unrealistic. 
And how unrealistic is it? (if it's quantifiable) 
Is it more unrealistic than the assumption of Linearity ?
Can anybody please demonstrate with a simple real-world example? 
 A: Suppose our inputs are $x\in\mathbb R^p$ and we are trying to model $y\in\mathbb R$ using a function $f : \mathbb R^p\to\mathbb R$. 
An additive model is of the form
$$
f(x) = \beta_0 + \sum_{i=1}^m\beta_ih_i(x_i)
$$
so we're saying that $f$, our potentially complicated function of a $p$-dimensional input, is built out of functions only of each coordinate separately. 
Note that if $h_i(x_i) = x_i$ then this reduces to a standard linear model, so additive models generalize linear models. 
This is restrictive because we're missing so many possible functions, like if $y = x_1x_2$ we won't be able to exactly capture that. Any function that needs to simultaneously consider multiple coordinates of $x$ at once won't be exactly model-able with an additive function. 
Another way to see why this is so restrictive is that once we fix our basis functions $\{h_1,\dots,h_m\}$ we are only considering functions in the span of this set, and that'll be at most an $m$-dimensional subspace of the probably infinite dimension function space that could contain the "true" function, so in that sense we have almost no chance of the true function being in our thin slice through this space that we're considering.
