Equivalent Definition of Strong Convexity

841 Views Asked by At

Let $f \in \mathcal{C}^2(\mathbb{R}^n).$ Recall that we defined $f$ to be strongly convex if there exists $\beta > 0$ such that $\langle D^{2}f|_{x}y,y\rangle\ge\beta$ for every $x, y \in \mathbb{R}^{n}$ or equivalently if the function $g(x)=f(x)-\frac{\beta}{2}\|x\|^{2}$ is convex. Show that if there exists $\gamma>0$ such that $$f(tx+(1-t)y) \leq tf(x)+(1-t)f(y)-\gamma t(1-t)\|x-y\|^{2}\tag{*}$$ for all $x,y \in \mathbb{R}^{n}, t \in [0,1]$ then the function $g(x)=f(x)-\gamma\|x\|^2$ is convex.


I tried showing that the function $g(x)=f(x)-\gamma \|x\|^2$ is convex using the above assumption but did not strike any luck. Any hints on how to proceed for this problem are appreciated. Than you for you help.

1

There are 1 best solutions below

0
On

"Equivalently" is not really right here as the definition of strong convexity does not assume $f \in \mathcal{C}^2$. However, $f$ satisfies (*) iff $g$ is convex.

The proof: (*) is equivalent to $$ g(tx+(1-t)y)\le tg(x)+(1-t)g(y)+\gamma\text{Res} $$ where $$ \text{Res}=t\|x\|^2+(1-t)\|y\|^2-\|tx+(1-t)y\|^2-t(1-t)\|x-y\|^2. $$ It is straightforward calculation to show that $\text{Res}=0$ for $\ell^2$-norm. For example, the $\|x\|^2$ term is $$ t\|x\|^2-t^2\|x\|^2-t(1-t)\|x\|^2=(t-t^2-t+t^2)\|x\|^2=0. $$ I leave it for you to verify the rest.