How do I show arithmetic mean is greater than or equal to geometric mean given non-negativity of KL divergence?

i.e. Given $D(p||q)=\sum_i p_i \log\left(\frac{p_i}{q_i}\right)\geq 0$, show that $\prod_{i=1}^n x_i^{a_i}\leq\sum_i a_ix_i,\;\forall a_i\geq 0\; s.t.\;\sum_i a_i=1$

I tried substituting $p_i=a_i$ and $q_i=x_i$ but I am only able to get G.M. on one side but not A.M. on the other.


Let's see. Taking logarithms on the target inequality, we get on the LHS $$L=\log \prod_{i=1}^n x_i^{a_i}=\sum a_i \log(x_i)$$

How can be related with the KL divergence? $a_i$ is (can be considered as) a probability function, but $x_i$ or $a_i y_i$ is not. Then let's normalize, define $Y= \sum a_i x_i$ and $y_i = a_i x_i /Y$ so that $y_i$ is a pf.

Then let's add the "desired" factors and see where it leads us to ...

$$L=\sum a_i \log\left(\frac{x_i}{Y} \frac{ a_i}{a_i} Y\right)=\sum a_i \log(y_i/a_i) + \sum a_i \log(Y) \tag 1 $$

The first term is close to what we need but we need to invert the fraction (hence change the log sign):

$$ \sum a_i \log(a_i/y_i) = \sum a_i \log(Y) -L \tag2$$

At last we can apply the KL divergence, the LHS is non-negative, then...

(I hope you can follow up from here.)


Problem: Assume that we know the following:

$\sum p_i\log \frac{p_i}{q_i} \ge 0$ for $p_i>0 , q_i>0, \forall i$ and $\sum p_i = 1, \sum q_i = 1 $.

Prove that $\prod x_i^{a_i} \le \sum a_i x_i$ for $x_i> 0, a_i > 0, \forall i$ and $\sum a_i = 1$.

Solution: By taking logarithm on both sides, the inequality to be proved is written as $$\sum\nolimits_i a_i\log x_i \le \log \sum\nolimits_i a_i x_i = (\sum\nolimits_i a_i)\log \sum\nolimits_i a_i x_i = \sum\nolimits_i a_i\log \sum\nolimits_j a_j x_j $$ or $$\sum\nolimits_i a_i \Big(\log \sum\nolimits_j a_j x_j - \log x_i\Big)\ge 0$$ or $$\sum\nolimits_i a_i \Big(\log a_i - \log \frac{a_ix_i}{\sum\nolimits_j a_j x_j}\Big)\ge 0.\tag{1}$$ Let $$p_i = a_i, \quad q_i = \frac{a_ix_i}{\sum_j a_j x_j}, \quad i=1, 2, \cdots, n.$$ We know that (1) holds.

We are done.


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.