Suppose $f: \mathbb{R}^n \rightarrow \mathbb{R}$ is a strongly convex function, i.e.,

$$ \langle \nabla f(y) - \nabla f(x) ,y-x \rangle \geq \theta \|y-x\|_2^2 \,\,\,\,\, \forall x,y \in \mathbb{R}^n $$


$$ z^T(\nabla^2 f(x) -\theta I)z\geq 0 \,\,\,\,\forall z \neq 0 $$

Also, suppose $\nabla f(x)$ is a Lipschitz with Lipschitz constant $L>0$, i.e., $ \|\nabla f(y)- \nabla f(x)\|_2 \leq L \|y-x\|_2$.

Show $\frac{1}{L}\|\nabla f(y)-\nabla f(x)\|_2^2 \leq \langle \nabla f(y) - \nabla f(x) ,y-x \rangle$

  • $\begingroup$ Your statement is definitely wrong. Just suppose $f(x)=\frac{1}{2}x^2$, then $L=1$ and the right side should be $(x-y)^2$. However, the left side is $\frac{1}{4}|y^4-x^4|$. It is clear that this inequality can't hold. $\endgroup$ – foxell May 6 '19 at 6:37
  • $\begingroup$ To know about the proof of some properties of strongly convex function, I recommend the book "Lecture on Convex Optimization" of Nesterov. $\endgroup$ – foxell May 6 '19 at 6:38
  • $\begingroup$ @foxell: Thank you for your careful reading. I had dropped $\nabla$ in the left hand side. By the way, your left hand side becomes $(y^2-x^2)$ . $\endgroup$ – Saeed May 6 '19 at 17:59

In order to get the conclusion we only need the convexity of $f(x)$ rather than the strong convexity. It is easy to show that if $\nabla f(x)$ is Lipschitz continuous with constant $L$ then we have:$$f(y)-f(x)-\langle\nabla f(x),y-x\rangle\leq\frac{L}{2}\| x-y\|^2$$ To show this inequality, we have:\begin{align} f(y)-f(x)-\langle\nabla f(x),y-x\rangle&=\int_{0}^{1}\langle \nabla f(x+t(y-x)),y-x \rangle dt-\langle\nabla f(x),y-x\rangle \\ &=\int_{0}^{1}\langle \nabla f(x+t(y-x))-\nabla f(x),y-x \rangle dt \\ &\leq\int_0^1\|\nabla f(x+t(y-x))-\nabla f(x)\|\|y-x\| dt\\ &\leq\int_0^1Lt\|y-x \|^2dt=\frac{L}{2}\|y-x\|^2 \end{align}

In order to get the result we need to consider the next function $$h(x)=f(x)-\langle \nabla f(y),x\rangle$$ It is clear that $h(x)$ is convex and also $\nabla h(x)$ is Lipschitz continuous with constant $L$, and hence $y\in\arg\min_{x}h(x)$ because $\nabla h(y)=0$. We have:$$ h(x-\frac{1}{L}\nabla h(x))\geq h(y) $$ Use the first inequality: $$ h(x-\frac{1}{L}\nabla h(x))\leq h(x)-\frac{1}{2L}\|\nabla h(x)\|^2 $$ We have: $$ h(y)\leq h(x)-\frac{1}{2L}\|\nabla h(x) \|^2 $$ $$f(y)+\langle\nabla f(y),x-y\rangle\leq f(x)-\frac{1}{2L}\| \nabla f(x)-\nabla f(y)\|^2 $$ $$f(y)-f(x)+\langle\nabla f(y),x-y\rangle\leq -\frac{1}{2L}\| \nabla f(x)-\nabla f(y)\|^2$$ Exchange $x,y$: $$f(x)-f(y)+\langle\nabla f(x),y-x\rangle\leq -\frac{1}{2L}\| \nabla f(x)-\nabla f(y)\|^2$$ Add them together then you can get the conclusion.

  • $\begingroup$ Could you show me how you get the first line? Also, why $h(x-\frac{1}{L}\nabla h(x)) \geq h(y)$? $\endgroup$ – Saeed May 7 '19 at 18:16
  • $\begingroup$ Ok, I have given the proof of the first line now. And the answer to your second question is that as I said $y$ is the global minimum of $h(x)$ so $h(y)\leq h(z),\forall z\in R^n$ $\endgroup$ – foxell May 8 '19 at 1:37
  • $\begingroup$ It was beautiful! $\endgroup$ – Saeed May 9 '19 at 0:53

By definition of strong convexity we have the following:

\begin{equation} f(x) \geq f(y) + <\nabla f(y), x-y> + \frac{\sigma}{2} \left\lVert x-y\right\rVert^2 \end{equation}

Similarly we have:

\begin{equation} f(y) \geq f(x) + <\nabla f(x), y-x> + \frac{\sigma}{2} \left\lVert x-y\right\rVert^2 \end{equation}

If you sum up these two inequalities and organize them you will end up with the following:

\begin{equation} <\nabla f(y) - \nabla f(x), y-x> \geq \sigma.\left\lVert x-y\right\rVert^2 \end{equation}

Which is what you have. We call this function $\sigma$-strongly convex. To give an example, "negative entropy" is a 1-strongly convex (also known as 1-SC) function wrt L1-norm which is used as regularization term for making a non-smooth function smooth in L1-setting (in learning with experts setup)--Please note that in L1-setting everything is a simplex.

Additionally, you can use "Hölder's inequality" to prove the general case shown below (not just norm 2):

\begin{equation} \left\lVert \nabla f(y) - \nabla f(x)\right\rVert_{\infty} \left\lVert x-y \right\rVert_{1} \geq <\nabla f(x) - \nabla f(y), x-y> \geq \sigma.\left\lVert x-y\right\rVert_{1}^2 \end{equation}

enter image description here

  • $\begingroup$ What I have is not what you have. My lower bound is in terms of $\nabla f(x)$ and $\nabla f(y)$ while yours is in terms of $\|x-y\|^2$. Please correct your derivation. $\endgroup$ – Saeed May 7 '19 at 18:13

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.