1

Let $\phi:\mathbb [0,\infty) \to [0,\infty)$ be a $C^2$ function, and let $c>0$ be a constant.

Suppose that for any $x_1,x_2>0, \alpha \in [0,1]$ satisfying $\alpha x_1 + (1- \alpha)x_2 =c$, we have $$ \phi(c)=\phi\left(\alpha x_1 + (1- \alpha)x_2 \right) \leq \alpha \phi(x_1) + (1-\alpha)\phi(x_2) \tag{1} $$

Is it true that $$ \phi(x) \ge \phi(c)+\phi'(c) (x-c) \tag{2} $$ for every $x$ in a neighbourhood $c$?

Does the answer change if we assume in addition that $\phi$ is strictly decreasing?

The classical proofs for "convexity of $\phi$ implies $\phi$ is above its tangents do not seem to adapt to this setup.

The converse claim holds:

Indeed,

$$ \phi(x_1) \ge \phi(c)+\phi'(c) (x_1-c) , \phi(x_2) \ge \phi(c)+\phi'(c) (x_2-c) $$ thus $$ \alpha \phi(x_1) + (1-\alpha)\phi(x_2) \ge \phi(c)+\phi'(c)(\alpha x_1 + (1- \alpha)x_2 -c)=\phi(c). $$

This formula here implies that $\phi''(c) \ge 0$.

Asaf Shachar
  • 25,111

2 Answers2

2

It's true. The most general formulation would be like this: Suppose your condition (1) is satisfied, and $$\phi'_{-}(c)=\lim_{h\to 0^-}\frac{\phi(c+h)-\phi(c)}{h}$$ exists. Then, $$\phi(x) \ge \phi(c)+\phi'_{-}(c) (x-c) \tag{3}$$ for $x>c.$ If $$\phi'_{+}(c)=\lim_{h\to 0^+}\frac{\phi(c+h)-\phi(c)}{h}$$ exists, then $$\phi(x) \ge \phi(c)+\phi'_{+}(c) (x-c) \tag{4}$$ for $0<x<c.$

Proof: Let $x>c,$ first. Then, for $h>0,$ we can choose $\alpha$ so that $$c=\alpha\,(c-h)+(1-\alpha)\,x$$ i.e. $$\alpha=\frac{x-c}{x-c+h},\quad 1-\alpha=\frac{h}{x-c+h}.$$ With $x_1=c-h$ and $x_2=x,$ your condition (1) becomes (after some algebra) $$\phi(x)\ge(x-c)\,\frac{\phi(c)-\phi(c-h)}{h}+\phi(c),$$ and $h\to0$ proves (3). The proof of (4) (with $x_1=x<c$ and $x_2=c+h>c$) is completely analogous and left to the reader. ;-)

  • You have a typo $\phi(c-h$ is $\phi(c-h)$ .Good day. (+1) – Miss and Mister cassoulet char Jun 26 '20 at 07:20
  • @Mister.Expandead Thank you! I've corrected it. –  Jun 26 '20 at 08:02
  • @ProfessorVector Thank you, this is an excellent answer. I wonder, do you know whether or not convexity at a point implies the existence of one-sided derivatives? I asked about it separately here: https://math.stackexchange.com/questions/3746986/does-convexity-at-a-point-imply-existence-of-one-sided-derivatives – Asaf Shachar Jul 06 '20 at 07:58
0

As you have already observed we have $\phi''(c) \geq 0$. Observe that if $\phi''(c) >0$ then $\phi' $ is increasing in some neighborhood of $c$ and the desired inequality follows immediately from Mean Value Theorem. For the general case apply this argument to $\phi (x) +\epsilon x^{2}$ and the take limit as $\epsilon \to 0$.

Kavi Rama Murthy
  • 311,013