Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Let us consider a continuous function $f \colon \mathbb{R} \to \mathbb{R}$. Let us call $f$ weakly convex if $$ \int_{-\infty}^{+\infty}f(x)[\varphi(x+h)+\varphi(x-h)-2\varphi(x)]dx\geq 0 \tag{1} $$ for all $h \in \mathbb{R}$ and all $\varphi \in C_0^\infty(\mathbb{R})$ with $\varphi \geq 0$. I was told that $f$ is weakly convex if, and only if, $f$ is convex; although I can imagine that (1) is essentially the statement $f'' \geq 0$ in a weak sense, I cannot find a complete proof. Is this well-known? Is there any reference?

share|improve this question

2 Answers 2

up vote 2 down vote accepted

If you take a family of symmetric mollifiers $\rho_\sigma$ compactly supported in $B_\sigma(0)$, and let $f_\sigma = \rho_\sigma \ast f$ denote the convolution of $f$ with $\rho_\sigma$, then for any $\varphi\in C_c^\infty(\mathbb R)$ you have that $\int f_\sigma \varphi = \int f \varphi_\sigma$. In particular $$\int f_\sigma(x) [\varphi(x+h)+ \varphi(x-h) - 2\varphi(x) ]= \int f(x) [\varphi_\sigma(x+h)+ \varphi_\sigma(x-h) - 2\varphi_\sigma(x) ] \ge 0$$ for $\varphi \ge 0$. Hence after substitution and multiplication by $h^{-2}$ for $h\ne 0$, we get $$\int \frac{f_\sigma(x+h)+f_\sigma(x-h) - 2f_\sigma(x) }{h^2} \varphi(x) \ge 0$$ Letting $h\to 0$, it follows from this that $f_\sigma'' \ge 0$. So $f_\sigma$ is convex for all $\sigma > 0$. On the other hand, since $f$ is continuous, $f_\sigma \to f$ pointwise. So $f$ is convex, itself.

On the other hand, if $f$ is convex, then $f(x+h) - f(x) \ge f(x) - f(x-h)$, so $f(x+h) + f(x-h) -2 f(x) \ge 0$. From this the integral inequality follows.

share|improve this answer
    
Nice one. Just one Question. Is the limit of convex function (if it exists) a convex function? –  Tomás Oct 24 '12 at 18:24
    
@Tomás: Thank you. Yes, the pointwise limit of convex functions is itself convex. Just fix $\lambda \in [0,1]$ and $x,y$. Then you have $$f_\sigma((1-\lambda)x + \lambda y) \le (1-\lambda) f_\sigma(x) + \lambda f_\sigma(y)$$ for all $\sigma$. Now take the limit. –  Sam Oct 24 '12 at 18:39

By a change of variables (translation-invariance of Lebesgue measure) the given inequality can be equivalently rewritten as $$ \int [f(x+h)+f(x-h)-2f(x)]\varphi(x)\,dx \geq 0 \qquad \text{for all }0 \leq \varphi \in C^{\infty}_0(\mathbb{R})\text{ and all }h \gt 0. $$ If $f$ were not midpoint convex then there would be $x \in \mathbb{R}$ and $h \gt 0$ such that $f(x+h) + f(x-h) - 2f(x) \lt 0$. By continuity of $f$ this must hold in some small neighborhood $U$ of $x$, so taking any nonzero $\varphi \geq 0$ supported in $U$ would yield a contradiction to the assumed inequality.

Thus, $f$ is midpoint convex and hence convex because $f$ is continuous.


Edit: The converse direction should be clear: it follows from $f(x+h) + f(x-h) - 2f(x) \geq 0$ for convex $f$ and all $h \gt 0$ so that the integrand in the first paragraph is non-negative.

Finally, the argument above works without essential change for continuous $f \colon \mathbb{R}^n \to \mathbb{R}$.

share|improve this answer
1  
I like this solution much more than my own. (+1) Observing that midpoint convexity is equivalent to convexity for continuous functions really captures the main point here. –  Sam Oct 24 '12 at 18:57

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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