I have an optimisation problem where I wish to find the fitted values $(\hat{y}_1, \hat{y}_2, \dots, \hat{y}_n)$ that minimise a pairwise 'distance' to observed values $(y_1, y_2, \dots, y_n)$. I would like to penalise negative distances more than positive distances, that is $\hat{y}_i - y = d$ for some $d > 0$, should be penalised less than if $\hat{y}_i - y = -d$. Ideally I would like positive distances to be penalised like an $\ell^2$ norm. Something like $\exp(-|\hat{y}_i - y|) + |\hat{y}_i - y|^2$ gives a crude example of what I am after, except this is not minimised at $d = 0$.
Can you please suggest some good objective functions? (In this context, good means having as many of the following properties as possible: 1) easily differentiable, 2) computationally easy, 3) convex, 4) continuous.)
EDIT: improving on my crude example, consider the function
$$g(x) = \lambda \exp(-x) + x^2$$
This has a minimum at $x = r$ such that:
$$g'(r) = -\lambda\exp(-r) + 2r = 0.$$
The minimum at $x = r$ is easily found by any root solver. Then, the objective function $$f(x) = \lambda\exp(-(x+r)) + (x+r)^2 - g(r)$$ satisfies all the desired properties with the caveat about the imprecision of calculating $r$.
Are there any other functions like the above?