The classical Jensen integral inequality says:

Let $\mu$ be a probabilistic measure defined on some $\sigma$-algebra subsets of $\Omega$. If $f\colon \Omega \rightarrow \mathbb R$ be an integrable function with $f(x)\in (a,b)$ and $\phi\colon (a,b) \rightarrow \mathbb R$ a convex function then

$$ \phi\left(\int_{\Omega}f \mathrm d \mu\right) \leq \int_{\Omega} \phi \circ f\mathrm d \mu. $$

Are there generalizations of this result in the case when $f$ has values in a convex subset $D$ of $\mathbb R^n$ and $\phi$ is a a convex function on $D$ ?


Yes, for example Theorem 0.7 here. Note that we can push forward $\mu$ by $f$, thus reducing the inequality $$\phi\left(\int_{\Omega}f \,\mathrm d \mu\right) \leq \int_{\Omega} \phi \circ f \,\mathrm d \mu \tag1$$ to $$\phi\left(\int_{D}x \, \mathrm d \nu (x)\right) \leq \int_{D} \phi(x)\, \mathrm d \nu(x)\tag2$$ where $\nu=f_\#\mu$. That is, the value of $\phi$ at the baricenter of $(D,\nu)$ is at most the mean of $\phi$.

Anyway, the proof of (1) is similar to the standard Jensen's inequality. Let $\ell$ be a supporting function for $\phi$ at the baricenter $x_0=\int_{\Omega}f \,\mathrm d \mu$. Then $$\ell\left(\int_{\Omega}f \,\mathrm d \mu\right) =\int_{\Omega} \ell \circ f \,\mathrm d \mu \tag3$$ If we replace $\ell$ with $\phi$ in (3), the left side stays the same, while the other side can only go up. (Recall that $\ell(x_0)=\phi(x_0)$ and $\ell\le\phi$ everywhere). The existence of $\ell$ is a somewhat technical issue if $D$ is ugly and $\phi$ isn't continuous: see the notes linked above.

  • $\begingroup$ Many thanks for answer and reference. $\endgroup$ – Alex Jun 24 '13 at 18:12

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.