Exercise 3.5: Kullback-Leibler Distance and Binomial Distribution

From LNTwww

Given probabilities

We assume here the  binomial distribution,  which is characterised by the parameters  $I$  and  $p$   
⇒   see the book  "Theory of Stochastic Signals":

  • Range of Values:
$$X = \{\hspace{0.05cm}0\hspace{0.05cm}, \hspace{0.15cm} 1\hspace{0.05cm},\hspace{0.15cm} 2\hspace{0.05cm},\hspace{0.15cm} \text{...}\hspace{0.1cm} ,\hspace{0.15cm} {\mu}\hspace{0.05cm}, \hspace{0.05cm}\text{...}\hspace{0.1cm} , \hspace{0.15cm} I\hspace{0.05cm}\}\hspace{0.05cm},$$
  • Probabilities:
$$P_X (X = \mu) = {I \choose \mu} \cdot p^{\mu} \cdot (1-p)^{I-\mu} \hspace{0.05cm},$$
  • Linear mean:
$$m_X = I \cdot p \hspace{0.05cm},$$
  • Variance:
$$\sigma_X^2 = I \cdot p \cdot (1-p)\hspace{0.05cm}.$$

In the part of the table highlighted in red, the probabilities  $P_X(X = \mu$)  of the binomial distribution under consideration are given.  In subtask  (1)  you are to determine the corresponding distribution parameters  $I$  and  $p$.


This given binomial distribution is to be approximated here by a  Poisson distribution  $Y$,  characterised by the rate  $\lambda$:

  • Range of values:
$$Y = \{\hspace{0.05cm}0\hspace{0.05cm}, \hspace{0.15cm} 1\hspace{0.05cm},\hspace{0.05cm} 2\hspace{0.05cm},\hspace{0.15cm} \text{...}\hspace{0.1cm} ,\hspace{0.15cm} {\mu}\hspace{0.05cm}, \hspace{0.05cm}\text{...}\hspace{0.1cm}\}\hspace{0.05cm},$$
  • Probabilites:
$$P_Y (Y = \mu) = \frac{\lambda^{\mu}}{\mu !} \cdot {\rm e}^{-\lambda} \hspace{0.05cm},$$
  • Expected values:
$$m_Y = \sigma_Y^2 = \lambda\hspace{0.05cm}.$$

In order to assess whether the probability mass function  $P_X(X)$  is sufficiently well approximated by  $P_Y(Y)$,  one can resort to the so-called  Kullback–Leibler distances  $\rm (KLD)$,  sometimes also called  "relative entropies"  in the literature.

Adapted to the present example, these are:

$$D(P_X \hspace{0.05cm}|| \hspace{0.05cm} P_Y) \hspace{0.15cm} = \hspace{0.15cm} {\rm E} \left [ {\rm log}_2 \hspace{0.1cm} \frac{P_X(X)}{P_Y(X)}\right ] \hspace{0.2cm}=\hspace{0.2cm} \sum_{\mu = 0}^{I} P_X(\mu) \cdot {\rm log}_2 \hspace{0.1cm} \frac{P_X(\mu)}{P_Y(\mu)} \hspace{0.05cm},$$
Enclosed results table
$$D(P_Y \hspace{0.05cm}|| \hspace{0.05cm} P_X) \hspace{0.15cm} = \hspace{0.15cm} {\rm E} \left [ {\rm log}_2 \hspace{0.1cm} \frac{P_Y(X)}{P_X(X)}\right ] \hspace{0.2cm}=\hspace{0.2cm} \sum_{\mu = 0}^{\infty} P_Y(\mu) \cdot {\rm log}_2 \hspace{0.1cm} \frac{P_Y(\mu)}{P_X(\mu)} \hspace{0.05cm}.$$

If  $\log_2$  is used, the pseudo–unit  "bit"  must be added to the numerical value.

In the adjacent table, the Kullback–Leibler–distance  $D(P_X \hspace{0.05cm}|| \hspace{0.05cm} P_Y)$  (in  "bit")  between the binomial PMF  $P_X(\cdot)$  and some Poisson approximations  $P_Y(\cdot)$  $($with five different rates $\lambda)$  is entered.

  • The respective entropy  $H(Y)$, which also depends on the rate  $\lambda$,  is given in the first row.
  • The columns for  $\lambda = 1$  are to be completed in subtasks  (3)  and (4) .
  • In subtask  (6)  these results are to be interpreted.



Hints:

$$A\hspace{0.05cm}' = 0.4096 \cdot {\rm lg} \hspace{0.1cm} \frac{0.4096}{0.3679} + 0.2048 \cdot {\rm lg} \hspace{0.1cm} \frac{0.2048}{0.1839} + 0.0512 \cdot {\rm lg} \hspace{0.1cm} \frac{0.0512}{0.0613} + 0.0064 \cdot {\rm lg} \hspace{0.1cm} \frac{0.0064}{0.0153} + 0.0003 \cdot {\rm lg} \hspace{0.1cm} \frac{0.0003}{0.0031} \hspace{0.05cm},$$
$$B\hspace{0.05cm}' = 0.1839 \cdot {\rm lg} \hspace{0.1cm} (0.1839) + 0.0613 \cdot {\rm lg} \hspace{0.1cm} (0.0613) + 0.0153 \cdot {\rm lg} \hspace{0.1cm} (0.0153) + 0.0031 \cdot {\rm lg} \hspace{0.1cm} (0.0031) + 0.0005 \cdot {\rm lg} \hspace{0.1cm} (0.0005) + 0.0001 \cdot {\rm lg} \hspace{0.1cm} (0.0001)$$
$$\Rightarrow \hspace{0.3cm} A\hspace{0.05cm}' \hspace{0.15cm} \underline {= 0.021944} \hspace{0.05cm},\hspace{0.5cm} B\hspace{0.05cm}' \hspace{0.15cm} \underline {= -0.24717} \hspace{0.05cm}.$$


Questions

1

What are the characteristics of the present binomial distribution?   Hint:  Enter (at most) one decimal place.

$I \hspace{0.47cm} = \ $

$p \hspace{0.47cm} = \ $

$m_x \ = \ $

$\sigma^2_x \hspace{0.25cm} = \ $

2

Which Kullback–Leibler distance should be used here?

Neither of the two distances is applicable.
$D(P_X \hspace{0.05cm}|| \hspace{0.05cm} P_Y)$  is more suitable
$D(P_Y \hspace{0.05cm}|| \hspace{0.05cm} P_X)$  is more suitable
Both Kullback–Leibler distances are applicable.

3

Calculate the appropriate Kullback–Leibler distance  $($abbreviated here as  $D)$  for  $\lambda = 1$.  Consider the auxiliary quantity  $A\hspace{0.05cm}'$.

$D \ = \ $

$\ \rm bit$

4

Calculate the entropy  $H(Y)$  of the Poisson approximation with rate  $\lambda = 1$.  Consider the auxiliary quantity  $B\hspace{0.05cm}'$.

$H(Y) \ = \ $

$\ \rm bit$

5

Which of the following statements are true?

In the  $H(Y)$  calculation, all terms have the same sign.
In the   $D(P_X \hspace{0.05cm}|| \hspace{0.05cm} P_Y)$  calculation all terms have the same sign.

6

How do you interpret the completed results table?

According Kullback–Leibler distance, you should choose  $\lambda = 1$.
$\lambda = 1$  guarantees the best approximation  $H(Y) ≈ H(X)$.


Solution

(1)  With the binomial distribution, all probabilities are  ${\rm Pr}(X > I) = 0$   ⇒   $\underline{I = 5}$.  Thus, for the probability that  $X =I = 5$,  we get:

$${\rm Pr} (X = 5) = {5 \choose 5} \cdot p^{5} = p^{5} \approx 0.0003 \hspace{0.05cm}.$$

Thus one obtains for

  • the characteristic probability:   $p= (0.0003)^{1/5} = 0.1974 \hspace{0.15cm} \underline {\approx 0.2}\hspace{0.05cm},$
  • the linear mean (expected value):   $m_X = I \cdot p \hspace{0.15cm} \underline {= 1}\hspace{0.05cm},$
  • the variance:   $\sigma_X^2 = I \cdot p \cdot (1-p) \hspace{0.15cm} \underline {= 0.8}\hspace{0.05cm}.$



(2)  Proposed solution 2 is correct:

  • Using  $D(P_Y \hspace{0.05cm}|| \hspace{0.05cm} P_X)$  would always result in an infinite value regardless of  $λ$  since for  $\mu ≥ 6$ :
$$P_X (X = \mu) = 0 \hspace{0.05cm},\hspace{0.3cm}P_Y (Y = \mu) \ne 0 \hspace{0.05cm}.$$
  • Even though the probabilities  $P_Y (Y = \mu)$  become very small for large  $μ$  they are still  "infinitely larger"  than  $P_X (X = \mu)$.



(3)  We use the first Kullback–Leibler distance:

$$D = D(P_X \hspace{0.05cm}|| \hspace{0.05cm} P_Y) =\hspace{0.2cm} \sum_{\mu = 0}^{5} P_X(\mu) \cdot {\rm log}_2 \hspace{0.1cm} \frac{P_X(\mu)}{P_Y(\mu)} \hspace{0.05cm}.$$
  • Using the logarithm of base ten  $(\lg)$,  we obtain for the Poisson approximation with  $\lambda = 1$:
$$D \hspace{0.05cm}' = 0.3277 \cdot {\rm lg} \hspace{0.1cm} \frac{0.3277}{0.3679} + A \hspace{0.05cm}' = -0.016468 + 0.021944 = 0.005476 \hspace{0.05cm}.$$
  • After converting to the logarithm of base two  $(\log_2)$ , we finally obtain:
$$D = D(P_X \hspace{0.05cm}|| \hspace{0.05cm} P_Y) = \frac{0.005476}{{\rm lg} \hspace{0.1cm}(2)} \hspace{0.15cm} \underline {\approx 0.0182\ {\rm (bit)}}\hspace{0.05cm}.$$


(4)  Using the logarithm of base ten, the entropy of the Poisson approximation  $(\lambda = 1)$:

$$H\hspace{0.05cm}'(Y) = -{\rm E} \left [{\rm lg} \hspace{0.1cm} {P_Y(Y)} \right ] = -2 \cdot 0.3679 \cdot {\rm lg} \hspace{0.1cm} (0.3679) - B\hspace{0.05cm}' = 0.31954 + 0.24717 = 0.56126.$$
  • Converting to  "bit"  gives the result we are looking for:
$$H(Y) = \frac{0.56126}{{\rm lg} \hspace{0.1cm}(2)} \hspace{0.15cm} \underline {= 1.864\ {\rm (bit)}} \hspace{0.05cm}.$$


(5)  Correct is statement 1.  In the numerical calculation of the Kullback–Leibler distance

  • the contribution of the  $μ$–th term is positive,  if  $P_Y(\mu) > P_X(\mu)$,
  • the contribution of the  $μ$–th term is negative,  if  $P_Y(\mu) < P_X(\mu)$.


Kullback–Leibler distance and entropy


(6)  Proposed solution 1 is correct:

  • It can also be seen from the graph that  $D(P_X \hspace{0.05cm}|| \hspace{0.05cm} P_Y) =0.0182$  bit  is not undercut by any  $λ$–value other than  $λ = 1$  (green crosses).
  • Furthermore, one can see from this plot that a better entropy approximation is achieved with  $λ = 0.9$  than with  $λ = 1$  (blue circles):
$$H(Y) = 1.795\ {\rm bit} \hspace{0.15cm}\approx \hspace{0.15cm} H(X) = 1.793\ {\rm bit}\hspace{0.05cm}.$$
The second proposed solution is therefore wrong.
  • With  $λ = 1$ , the  linear means  of the two random variables coincide:
$$m_X = m_Y= 1.$$
  • With  $λ = 0.9$ the  second moments  agree:
$$m_X + \sigma_X^2 = m_Y + \sigma_Y^2= 1.8.$$

Whether this statement is relevant, we leave undecided. 

Because:   Due to the steady increase of   $H(Y)$  with increasing  $λ$ , it is clear that for some  $λ$–value   $H(Y) = H(X)$  must indeed hold.