Let $f: (a,b) \to \Bbb R$ be a continuous function such that $$\forall x \in (a,b)\; \exists \epsilon_0 \forall \epsilon < \epsilon_0: f(x) \leq \frac{f(x+\epsilon)+f(x-\epsilon)}{2}. $$

Is $f$ necessarily convex?

By definition, $f$ is convex if $$\forall x,y \;\forall t \in [0,1]: f(tx+(1-t)y)\leq tf(x)+(1-t)f(y)$$

so if $f$ is convex it satisfies our property, as can be seen by taking $t=\frac 12, y=x'-\epsilon, x= x'+\epsilon$.

The problem with claiming the converse is tha t the convexity definition is global, while the property is local. On the other hand, the property is on a discrete second derivative and when the function is differentiable we do have a local description of convexity saying exactly the continuous analogue, namely $f'' \geq 0$ at all points.


I added the assumption that $f$ is continuous, as this post shows that in general $f$ may not be convex.


Assume the function is not convex. That is, there are $x<y\in (a,b), \lambda_{0} \in (0,1)$ s.t. $f(\lambda_0 y + (1-\lambda_0)x)>\lambda_0 f(y)+(1-\lambda_0)f(x)$. Equivalently, $f(\lambda_0 (y -x) + x) - \lambda_0 (f(y)-f(x)) - f(x) > 0$

Now, define $g:[0,1]\to \Bbb{R}$ by $g(\lambda)=f(\lambda (y -x) + x) - \lambda (f(y)-f(x)) - f(x)$. As $g$ is continuous with compact domain, there is $\lambda^*$ such that maximizes $g$. Hence $g(\lambda^*)\ge g(\lambda_0)>0$, and since $g(0)=g(1)=0, \lambda^* \in (0,1)$.

Now, let $\varepsilon>0$ small enough. We know $$ 0\le g(\lambda^*) - g(\lambda^*+\varepsilon) \\ = f(\lambda^* (y -x) + x) - \lambda^* (f(y)-f(x)) - f(x)\\ - f((\lambda^* +\varepsilon) (y -x) + x) + (\lambda^*+\varepsilon) (f(y)-f(x)) + f(x)\\ = f(\lambda^* (y -x) + x) - f((\lambda^* +\varepsilon) (y -x) + x) + \varepsilon (f(y)-f(x)) $$

and similarly,

$$ 0\le g(\lambda^*) - g(\lambda^*-\varepsilon) \\ = f(\lambda^* (y -x) + x) - f((\lambda^* -\varepsilon) (y -x) + x) - \varepsilon (f(y)-f(x)) $$ where at least one of the inequalities is strict. (*)

Hence $$ 0< f(\lambda^* (y -x) + x) - f((\lambda^* +\varepsilon) (y -x) + x) + \varepsilon (f(y)-f(x)) \\ + f(\lambda^* (y -x) + x) - f((\lambda^* -\varepsilon) (y -x) + x) - \varepsilon (f(y)-f(x)) \\ = 2f(\lambda^* (y -x) + x) - f((\lambda^* +\varepsilon) (y -x) + x) - f((\lambda^* -\varepsilon) (y -x) + x) $$

Now, if we denote $x^*=\lambda^* (y-x) + x$, and $\epsilon = \varepsilon (y - x)$, we can write the former inequality as $$ \frac{f(x^* - \epsilon) + f(x^* + \epsilon)}{2} < f(x^*) $$ which is true for all small enough $\epsilon$. Which is a contradiction.

To justify (*), as pointed out by @Ashkan:

First, let $M=\{\lambda^*\in[0,1]|g(\lambda^*)=\max_{[0,1]}g(\lambda)\}$. This set is nonempty, thus it has a supremum $\bar{\lambda}$. And this supremum is in $M$. To see this, take a sequence $(\lambda^*_n)_n \in M$ s.t. converges to $\bar{\lambda}$. By continuity of $g$, $g(\lambda^*_n)\to g(\bar{\lambda})$, and since $g(\lambda^*_n)=\max_{[0,1]}g(\lambda)$ for all $n$, we deduce $\bar{\lambda}\in M$.

With this, the justification of (*) is that we can take, w.l.o.g, $\lambda^*=\bar{\lambda}$. Hence, for all $\varepsilon>0$, $g(\lambda^*)>g(\lambda^* +\varepsilon)$.

  • 2
    $\begingroup$ you should better clarify this "where at least one of the inequalities is strict" $\endgroup$ – Red shoes Jul 1 '17 at 23:22
  • $\begingroup$ @Ashkan Done. Thanks for pointing this out. $\endgroup$ – Nate River Jul 2 '17 at 0:02
  • $\begingroup$ It is clear now! Very nice answer. +1 $\endgroup$ – Fimpellizieri Jul 2 '17 at 0:12
  • $\begingroup$ What a marvelous proof!! +1 And that addendum for showing that one of the inequalities is strict is great. $\endgroup$ – Paramanand Singh Jul 2 '17 at 1:09

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.