Expectation of transformed mean of multiple random variables <= mean of the expectation of their transformation

43 Views Asked by At

I am trying to understand the following. If $M$ is the number of different random variables $X_1 ... X_M$, are there any conditions under which we can claim that:

$$E\left[f(\sum_{m=1}^M w_m X_m)\right] \leq \sum_{m=1}^M w_mE[f(X_m)]$$

So far, I think that it suffices for $f$ to be convex and for all $w_m$ to be non-negative with $\sum_{m=1}^M w_m = 1$. Because then, if we treat $E[f(.)]$ as itself a function, then this Jensen's formulation applies. However:

  • I am not sure this is a reasonable interpretation to make when applying Jensen inequality to random variables
  • if not, are there other conditions (for ex. bounds on the random variables or characteristics of $f$) that would make the inequality to hold?
1

There are 1 best solutions below

2
On

Jensen inequality can be relevant here, but not wrt your expectation (because Jensen is about taking the expectation inside-outside the function, which is not the case here).

If we instead leave the expectation aside, then, if $f$ is convex and $w_m \ge 0$, letting $W=\sum w_w$, and we can indeed apply Jensen (again, ignoring that $X_m$ are random variables) :

$$f \left(\frac{1}{W} \sum w_m X_m\right) \le \frac{1}{W} \sum w_m f(X_m)$$

In particular, if $W=1$:

$$f \left(\ \sum w_m X_m\right) \le \sum w_m f(X_m)$$

Taking expectations on both sides you get your inequality.