Suppose $\phi$ is a real function on $\mathbb{R}$ such that $$\phi \Bigl(\int_0^1f(x)\,dx\Bigr)\leq \int_0^1 \phi(f)\,dx$$ for every real bounded measurable function $f$. Prove that $\phi$ is then convex.

This one is in Rudin's Real and Complex Analysis, page 74, and, to be quite honest, I am unsure what my first step to solve this problem should even be. To help think of a general methodology, I attempted this similar question, from page 71 of the same book:

Assume $\phi$ is a continuous real valued function on $(a,b)$ such that $$\phi\Bigl(\frac{x+y}{2}\Bigr)\leq\frac{\phi(x)+\phi(y)}{2}$$ for all $x, y\in (a,b).$ Prove that $\phi$ is then convex (the conclusion does not follow if continuity is omitted from the hypothesis).

I take it that the latter question is just a special case of the former (assuming $\phi$ is bounded on $(a,b)$), so I attempted to solve the latter problem first. My attempt was to note that \begin{align*} \phi\big(\lambda x + (1-\lambda)y\big) &\leq \phi\Bigl(\frac{2\lambda x + 2(1-\lambda)y}{2}\Bigr) \\ &\leq \frac{1}{2}\Bigl(\phi\bigl(2\lambda x\bigr)+\phi \bigl(2(1-\lambda)y\bigr)\Bigr) \\ & \leq \phi\bigl(\lambda x\bigr)+\phi\bigl((1-\lambda)y\bigr) \end{align*} with the latter inequality justified simply by applying the hypothesis again twice to the previous expression.

From here, however, I am stuck; I cannot justify "popping out" the $\lambda$ and the $(1-\lambda)$ in the last inequality via some bound afforded by the hypothesis and/or the continuity of $\phi$.

  • 2
    $\begingroup$ You could check out this question. It has a lot of non-standard proofs in it and the question itself might give you a hint at the standard proof. $\endgroup$ – Milo Brandt Dec 26 '14 at 3:32
  • $\begingroup$ Meelo, thanks. Are you aware of any solution to the first part of my question, or would that depend on solving the second part of my question? $\endgroup$ – Darrin Dec 26 '14 at 3:36
  • 1
    $\begingroup$ Well, solving the second part would be good for the first part, though not strictly necessary - informally, if $f$ were a measure with value $\lambda$ for $\{x\}$ and $(1-\lambda)$ for $\{y\}$, then the integral will collapse to the definition of convexity. (Formally, you'd have to actually know about the formalisms of this topic, which I don't) $\endgroup$ – Milo Brandt Dec 26 '14 at 3:46

Consider $f$ such that $f(x) = x_1$ for $0\leq x \leq \alpha$ and $f(x) = x_2 $ for $\alpha < x \leq 1$. So the Jensen's inequality for integrals will be

$$\phi\left(\alpha x_1 + (1- \alpha)x_2 \right) \leq \alpha \phi(x_1) + (1-\alpha)\phi(x_2)$$

by changing values of $x_1$ and $x_2$, we have the result.


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.