I want to show that if $f:\mathbb{E}\rightarrow\mathbb{R}$ is convex, and differentiable at $x$, then $\partial f(x) = \{ \nabla f(x) \} $ .

I understand that for a convex function, we have the following:

$$f(y) \le f(x)+\nabla f^T(y-x) \forall x,y \in dom \, f$$

and I know that the definition of the subdifferential of $f$ at $x$ is:

$$ \partial f(x) = \{ z | f(y) \ge f(x)+<z,y-x> \forall x,y \in dom\, f \} \}$$

So, I'm able to say that if $f$ is convex, then certainly $\nabla f(x) \in\partial f(x)$. But I do not understand how to show $\partial f(x) = \{ \nabla f(x) \} $?

(as a bit of a side note, I do not understand how we know that $f'(x)=\nabla f$? I know that it is differentiable at $x$, but that does not necessarily mean that the gradient exists right.....?

A function can be differentiable in all directions, and the gradient not exist, and we did not say that $f$ was continuous)


Define the function $$\phi(t):=f(x+t(y-x))$$ for $t\in [0,1]$. Since $f$ is convex implies $$\phi(t)\leqslant (1-t)f(x)+tf(y)=(1-t)\phi(0)+t\phi(1)$$ hence $\phi(t)$ is a convex function on $[0,1]$. Moreover differentiability of $f$ implies that $\phi$ is also differentiable. In particular we have $$\phi'(0)=\langle\nabla f(x),y-x\rangle$$ Convexity of $\phi$ yields $$\phi'(0)\leqslant \phi(1)-\phi(0)=f(y)-f(x)\Rightarrow\langle \nabla f(x),y-x\rangle\leqslant f(y)-f(x)$$ Therefore by definition $\nabla f(x)\in\partial f(x)$. Now suppose that $z\in \partial f(x)$ then $$f(y)\geqslant f(x)+\langle z,y-x\rangle$$ for all $y$. Let $y:=x+tw$ and define the function $$\psi(t):=f(y)-f(x)-\langle z,y-x\rangle=f(x+tw)-f(x)-\langle z,tw\rangle$$ This gives $$\psi'(t)=\langle\nabla f(x), w\rangle-\langle z,w\rangle$$ Since $z\in\partial f(x)$ then $\psi(t)\geqslant 0$ for all $t$. Also $\psi(0)=0$ hence at $t=0$ we have a minimum for $\psi$. This implies $\psi'(0)=0$ which is equivalent to saying $$\langle\nabla f(x), w\rangle=\langle z,w\rangle$$ Since the last equation holds for any arbitrary $w$ then $\nabla f(x)=z$. This proves that in fact $$\partial f(x)=\{\nabla f(x)\}$$

  • $\begingroup$ Thanks. It seems like the beginning part of the proof is the same as mine in the question, up until where you introduce $\psi(t)$, right? $\endgroup$ – makansij Apr 29 '18 at 22:31
  • $\begingroup$ Two questions: First I am not seeing how $\psi(t) \ge 0$? Secondly, I understand that $\psi’(0) =0$ but how do we know it holds for all $t$ and not just $t=0$? Is it because $\psi’(t)$ doesn’t depend on $t$? $\endgroup$ – makansij Apr 29 '18 at 22:32
  • $\begingroup$ It seems like your secret to showing uniqueness is that $y$ can be any point in the entire domain because $y=x+tw$? $\endgroup$ – makansij Apr 29 '18 at 22:39
  • 1
    $\begingroup$ First question: almost similar the first part although you need to justify how convexity interplays there. Second question: $\psi(t)\geqslant 0$ for all $t$ since we are letting $z$ be a subgradient of the convex function. Third question: Since $\psi(0)=0$ and $\psi$ is identically nonnegative it means that $t=0$ is a minimum. Then simply first order condition for a minimum. And also $y$ is arbitrary means you can take arbitrary $w$. $\endgroup$ – Arian Apr 30 '18 at 13:20
  • $\begingroup$ Thanks @Arian. It’s making more sense. When you say “justify how convexity interplays there”. Is saying I understand that for a convex function, we have the following: $f(y) \le f(x)+\nabla f^T(y-x) \forall x,y \in dom \, f$ not enough? If not, why not? Last thing: There must be some conditions on $t$ and $w$ in $y=x+tw$, right? $x+tw$ must still be in the domain of the function $f$, right? $\endgroup$ – makansij May 2 '18 at 23:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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