In Information Theory, what is the lower bound that minimizes the value of $2^{l_x}$

56 Views Asked by At

Here is the question from my notes:

Suppose we wish to find a decipherable code that minimizes the expected value of $2^{l_x}$ for a probability distribution $p(x)$. Establish the lower bound $$E\big(2^{l_x}\big)\geq \bigg(\sum_x \sqrt{p(x)}\bigg)^2 \tag{1}$$ Hint: Use Cauchy-Schwarz

I am not really sure what to do in here but here is what I did: $E(l_x)=\sum_x p(x)l_x$ hence I say that $$E(2^{l_x})=\sum_x p(x)2^{l_x} \tag{2}$$ Hence $$\sum_x p(x)2^{l_x}\geq \bigg(\sum_x \sqrt{p(x)}\bigg)^2 \tag{3}$$ I take roots of both side and expand the summations: $$\sqrt{p(1)}+...\sqrt{p(n)}\leq\sqrt{p(1)2^{l_1}+...p(n)2^{l_n}}\tag{4}$$ The Cauchy Schwarz which I am supposed to use is I think something along those lines: $$|x_1y_1+...+x_ny_n|\leq\sqrt{x_1^2+...x_n^2}\sqrt{y_1^2+...y_n^2} \tag{5}$$ Any hints on what to do here would be greatly appreciated.

1

There are 1 best solutions below

0
On BEST ANSWER

Notice that we are speaking here of "decipherable codes", so we must put that restriction in play somehow. Fortunately we have Kraft-McMillan inequality.

So we must have $\sum_x 2^{-l_x}\le 1$.

Let's call $a_x = 2^{-l_x}$ and $b_x = p_x 2^{l_x}=p_x a_x^{-1} $

Then we want to bound $E=\sum b_x$ subject to $\sum a_x \le 1$

Now, for any non-negative $a_x,b_x$ , we can write Cauchy-Schwarz as $$\sum_x \sqrt{a_x b_x} \le \sqrt{ \sum_x a_x}\sqrt{ \sum_x b_x} \tag{1}$$

In our case this gives

$$\sum_x \sqrt{p_x} \le \sqrt{E} \sqrt{ \sum_x a_x} \le \sqrt{E} \tag{2}$$

which gives the desired bound.