Should we consider the cases of absolute value before or after differentiating?

66 Views Asked by At

Consider some function: $\DeclareMathOperator{argmin}{argmin}$

$$f(w) = \underset{w}{\argmin}\:(z-w)^2 + \lambda\gamma |w|\:\:\:\:\:\:\:\:\:where\:\lambda,\gamma\in\mathbb{R}^+$$

I'm quite confused when dealing with functions containing absolute value terms for the purposes of differentiation (w.r.t $w$). In some cases, I've seen people differentiate the function including the absolute value terms first; resulting in:

$${f}'(w) = 2(w-z) + \lambda\gamma\frac{w}{|w|}$$

and then considering the 2 cases where:

$$w\in\mathbb{R}^+$$ $$w\in\mathbb{R}^-$$

Meanwhile, I've also seen the case where people would consider the cases first (notice the inclusion of zero):

$$w\in\mathbb{R}^+_0$$ $$w\in\mathbb{R}^-_0$$

resulting in:

$${f}(w) = (z-w)^2 + \lambda\gamma w$$

$${f}(w) = (z-w)^2 - \lambda\gamma w$$

respectively, and then differentiating results in:

$${f}'(w) = 2(w-z) + \lambda\gamma$$

$${f}'(w) = 2(w-z) - \lambda\gamma$$

respectively.

For me there's a subtle difference in that the second way, $w=0$ can be a solution, whereas in the first way it would've been undefined. Sorry if I'm understanding this incorrectly but could someone shed some light on this issue please?

1

There are 1 best solutions below

0
On BEST ANSWER

It looks as though both methods lead to $f'(0)$ being undefined. The first case makes this a little more explicit, since you it has a division by zero, but the second case also can't define a derivative at that point. The derivatives of each of the pieces of your function give inconsistent results for $w=0$: $$\lim_{a\to0^-}f'(a)=2z-\lambda\gamma$$ $$\lim_{a\to0^+}f'(a)=2z+\lambda\gamma$$

If the left/right derivatives (or semi-derivatives) at a point aren't equal, then the function is not differentiable at that point.

This same issue arises if you just look at $g(x)=|x|$: no matter how you calculate it, splitting it into pieces before or after taking the derivative, its clear that the derivative is undefined at $x=0$, as the left and right derivatives aren't equal. Geometrically, you can see this from the sharp point at $x=0$ in a plot of $|x|$.