Let $f:\mathbb R\to\mathbb R$ be a lower semicontinuous function, we define the Fréchet subdifferential of $f$ at $x\in\mathbb R$ by $$\partial^F f(x):=\left\{L\in\mathbb R: \liminf_{v\to0}\frac{f(x+v)-f(x)-Lv}{|v|}\geq0\right\}.$$ According to this paper, in the third page, authors claimed that this subdifferential satisfies a additive property, namely $$\partial^F(f+g)(x)\subset \partial^Ff(x)+\partial^Fg(x)\tag{$*$}$$ for any convex function $g:\mathbb R\to\mathbb R$ (here I choose the Banach space $\mathbb R$, then $g$ is automatically continuous and $\partial^F g(x)\neq\emptyset$ for all $x\in\mathbb R$.)

As usual, the summation in $(*)$ should be the Minkowski sum, hence in particular we have $\emptyset+S=\emptyset$ for all $S\subset\mathbb R$.

I'm trying to prove $(*)$, but here is a counterexample: Let $f(x)=-|x|, g(x)=|x|$ for all $x\in\mathbb R$, then $g$ is a convex function and $\partial^Fg(0)=[-1,1]$ and $\partial^F(f+g)(0)=\{0\}$; however, $\partial^Ff(0)=\emptyset$: if $L\in\partial^Ff(0)$, then $\liminf_{v\to0}\frac{-|v|-Lv}{|v|}\geq0$, i.e., $-1-|L|\geq0$, which is a contradiction. Now it seems that I've got a counterexample of $(*)$, but $(*)$ is used everywhere in literature, so what am I missing now?

In the same paper, at the end of the proof of Theorem 3.1, authors used an argument like this: if $f+h$ takes minimum at $\bar x$ with $h$ convex and $\partial^F h(\bar x)\neq\emptyset$, then $$0\in \partial^F(f+h)(\bar x)\subset \partial^Ff(\bar x)+\partial^Fh(\bar x);$$ combining this with $\partial^Fh(\bar x)\subset (-\delta,\delta)$ gives that $\partial^Ff(\bar x)\cap(-\delta,\delta)\neq\emptyset$. But my example above still indicates the failure of this argument.

Theorem 3.1 impiles the fact that $\operatorname{dom}(\partial^Ff)$ is dense in $\mathbb R$ if $f:\mathbb R\to\mathbb R$ is a lower semicontinuous function, which is a well-known result. So I must miss something here.

Thanks for any help!


1 Answer 1


In part (P3) of Definition 2.1 in the paper you linked, it is also required that $g$ be $\partial$-differentiable at $x$ (meaning that both $\partial g(x)$ and $\partial(-g)(x)$ are nonempty), which is not the case in your example (with $\partial=\partial^F$ and $g(\cdot)=|\cdot|$).

  • $\begingroup$ Thanks for your answer! I understand now. I didn't realize the stronger condition of $\partial$-differentiable, which is also the reason why the authors require $p=2>1$ in the proof of Theorem 3.1, since $|x|^2$ is smoother than $|x|$. However, I've seen in another paper, in the proof of Theorem 7, Rockafellar invoked the ordinary Ekeland's variational principle, in which case $h(x)=A|x|(A>0)$ and he still used the property $(*)$. Is there any other particular reason in Rockafellar's proof? $\endgroup$ Nov 26, 2023 at 2:33
  • $\begingroup$ @Fergns : I suggest you post this additional question separately. $\endgroup$ Nov 26, 2023 at 3:17

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.