Difference between revisions of "Information Theory/Discrete Sources with Memory"

From LNTwww
Line 207: Line 207:
 
<br clear=all>
 
<br clear=all>
 
However, we do not consider the spectral properties of the AMI code here, but interpret this code information-theoretically:
 
However, we do not consider the spectral properties of the AMI code here, but interpret this code information-theoretically:
*Based on the symbol set size&nbsp; $M = 3$&nbsp; the decision content of the (ternary) encoded sequence is equal to&nbsp; $H_0 = \log_2 \ 3 ≈ 1.585 \hspace{0.05cm} \rm bit/symbol$.&nbsp; The first entropy approximation returns&nbsp; $H_1 = 1.5 \hspace{0.05cm} \rm bit/Symbol$, as shown in the following calculation:
+
*Based on the symbol set size&nbsp; $M = 3$&nbsp; the decision content of the (ternary) encoded sequence is equal to&nbsp; $H_0 = \log_2 \ 3 ≈ 1.585 \hspace{0.05cm} \rm bit/symbol$.&nbsp; The first entropy approximation returns&nbsp; $H_1 = 1.5 \hspace{0.05cm} \rm bit/symbol$, as shown in the following calculation:
 
    
 
    
 
:$$p_{\rm H} = p_{\rm L} = 1/2 \hspace{0.3cm}\Rightarrow \hspace{0.3cm}
 
:$$p_{\rm H} = p_{\rm L} = 1/2 \hspace{0.3cm}\Rightarrow \hspace{0.3cm}

Revision as of 12:17, 13 February 2023

A simple introductory example


$\text{Example 1:}$  At the  "beginning of the first chapter"  we have considered a memoryless message source with the symbol set  $\rm \{A, \ B, \ C, \ D\}$   ⇒   $M = 4$  .   An exemplary symbol sequence is shown again in the following figure as source  $\rm Q1$ .

With the symbol probabilities  $p_{\rm A} = 0.4 \hspace{0.05cm},\hspace{0.2cm}p_{\rm B} = 0.3 \hspace{0.05cm},\hspace{0.2cm}p_{\rm C} = 0.2 \hspace{0.05cm},\hspace{0.2cm} p_{\rm D} = 0.1\hspace{0.05cm}$  the entropy is

$$H \hspace{-0.05cm}= 0.4 \cdot {\rm log}_2\hspace{0.05cm}\frac {1}{0.4} + 0.3 \cdot {\rm log}_2\hspace{0.05cm}\frac {1}{0.3} + 0.2 \cdot {\rm log}_2\hspace{0.05cm}\frac {1}{0.2} + 0.1 \cdot {\rm log}_2\hspace{0.05cm}\frac {1}{0.1} \approx 1.84 \hspace{0.05cm}{\rm bit/symbol} \hspace{0.01cm}.$$

Due to the unequal symbol probabilities the entropy is smaller than  $H_0 = \log_2 M = 2 \hspace{0.05cm} \rm bit/symbol$.

Quaternary source without/with memory



The source  $\rm Q2$  is almost identical to the source  $\rm Q1$, except that each individual symbol is output not only once, but twice in a row:  $\rm A ⇒ AA$,  $\rm B ⇒ BB$,  and so on.

  • It is obvious that  $\rm Q2$  has a smaller entropy (uncertainty) than  $\rm Q1$.
  • Because of the simple repetition code,  the entropy
$$H = 1.84/2 = 0.92 \hspace{0.05cm} \rm bit/symbol$$
has only half the size, although the occurrence probabilities have not changed.


$\text{Conclusion:}$  This example shows:

  • The entropy of a source with memory is smaller than the entropy of a memoryless source with equal symbol probabilities.
  • The statistical bindings within the sequence  $〈 q_ν 〉$  have to be considered now,
  • namely the dependency of the symbol  $q_ν$  from the predecessor symbols  $q_{ν-1}$,  $q_{ν-2}$, ...


Entropy with respect to two-tuples


We continue to look at the source symbol sequence  $〈 q_1, \hspace{0.05cm} q_2,\hspace{0.05cm}\text{ ...} \hspace{0.05cm}, q_{ν-1}, \hspace{0.05cm}q_ν, \hspace{0.05cm}\hspace{0.05cm}q_{ν+1} .\hspace{0.05cm}\text{...} \hspace{0.05cm}〉$  and now consider the entropy of two successive source symbols.

  • All source symbols  $q_ν$  are taken from an alphabet with the symbol set size  $M$, so that for the combination  $(q_ν, \hspace{0.05cm}q_{ν+1})$  there are exactly  $M^2$  possible symbol pairs with the following $\text{combined probabilities}$:
$${\rm Pr}(q_{\nu}\cap q_{\nu+1})\le {\rm Pr}(q_{\nu}) \cdot {\rm Pr}( q_{\nu+1}) \hspace{0.05cm}.$$
  • From this, the  »compound entropy«  of two consecutive symbols can be computed:
$$H_2\hspace{0.05cm}' = \sum_{q_{\nu}\hspace{0.05cm} \in \hspace{0.05cm}\{ \hspace{0.05cm}q_{\mu}\hspace{0.01cm} \}} \ \sum_{q_{\nu+1}\hspace{0.05cm} \in \hspace{0.05cm}\{ \hspace{0.05cm} q_{\mu}\hspace{0.01cm} \}}\hspace{-0.1cm}{\rm Pr}(q_{\nu}\cap q_{\nu+1}) \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{{\rm Pr}(q_{\nu}\cap q_{\nu+1})} \hspace{0.4cm}({\rm unit\hspace{-0.1cm}: \hspace{0.1cm}bit/two–tuple}) \hspace{0.05cm}.$$
The index "2" symbolizes that the entropy thus calculated refers to  »two-tuples«.
  • To get the average information content per symbol,  $H_2\hspace{0.05cm}'$  has to be divided in half:
$$H_2 = {H_2\hspace{0.05cm}'}/{2} \hspace{0.5cm}({\rm unit\hspace{-0.1cm}: \hspace{0.1cm}bit/symbol}) \hspace{0.05cm}.$$
$$H_1 = \sum_{q_{\nu}\hspace{0.05cm} \in \hspace{0.05cm}\{ \hspace{0.05cm}q_{\mu}\hspace{0.01cm} \}} {\rm Pr}(q_{\nu}) \cdot {\rm log_2}\hspace{0.1cm}\frac {1}{{\rm Pr}(q_{\nu})} \hspace{0.5cm}({\rm unit\hspace{-0.1cm}: \hspace{0.1cm}bit/symbol}) \hspace{0.05cm}.$$
  • The index "1" is supposed to indicate that  $H_1$  considers only the symbol probabilities and not statistical bindings between symbols within the sequence.  With  $H_0 = \log_2 \ M$  the following size relation results:
$$H_0 \ge H_1 \ge H_2 \hspace{0.05cm}.$$
  • With statistical independence of the sequence elements   ⇒   $H_2 = H_1$.


The previous equations each indicate an  "ensemble mean value".   The probabilities required for the calculation of  $H_1$  and  $H_2$  can, however, also be calculated as time averages from a very long sequence or, more precisely, approximated by the corresponding  $\text{relative frequencies}$.

Let us now illustrate the calculation of entropy approximations  $H_1$  and  $H_2$  with three examples.

Ternary symbol sequence and formation of two-tuples

$\text{Example 2:}$  We will first look at the sequence  $〈 q_1$, ... , $q_{50} \rangle $  according to the graphic, where the sequence elements  $q_ν$  originate from the alphabet $\rm \{A, \ B, \ C \}$   ⇒   the symbol set size is  $M = 3$.

By time averaging over the  $50$  symbols one gets the symbol probabilities  $p_{\rm A} ≈ 0.5$,   $p_{\rm B} ≈ 0.3$  and  $p_{\rm C} ≈ 0.2$, with which one can calculate the first order entropy approximation:

$$H_1 = 0.5 \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{0.5} + 0.3 \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{0.3} + 0.2 \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{0.2} \approx \, 1.486 \,{\rm bit/symbol} \hspace{0.05cm}.$$

Due to the not equally probable symbols  $H_1 < H_0 = 1.585 \hspace{0.05cm}. \rm bit/symbol$.  As an approximation for the probabilities of two-tuples one gets from the above sequence:

$$\begin{align*}p_{\rm AA} \hspace{-0.1cm}& = \hspace{-0.1cm} 14/49\hspace{0.05cm}, \hspace{0.2cm}p_{\rm AB} = 8/49\hspace{0.05cm}, \hspace{0.2cm}p_{\rm AC} = 3/49\hspace{0.05cm}, \\\ p_{\rm BA} \hspace{-0.1cm}& = \hspace{0.07cm} 7/49\hspace{0.05cm}, \hspace{0.25cm}p_{\rm BB} = 2/49\hspace{0.05cm}, \hspace{0.2cm}p_{\rm BC} = 5/49\hspace{0.05cm}, \\\ p_{\rm CA} \hspace{-0.1cm}& = \hspace{0.07cm} 4/49\hspace{0.05cm}, \hspace{0.25cm}p_{\rm CB} = 5/49\hspace{0.05cm}, \hspace{0.2cm}p_{\rm CC} = 1/49\hspace{0.05cm}.\end{align*}$$

Please note that the  $50$  sequence elements can only be formed from  $49$  two-tuples  $(\rm AA$, ... , $\rm CC)$  which are marked in different colors in the graphic.

  • The entropy approximation  $H_2$  should actually be equal to  $H_1$  since the given symbol sequence comes from a memoryless source.
  • Because of the short sequence length  $N = 50$  and the resulting statistical inaccuracy, however, a smaller value results:  
$$H_2 ≈ 1.39\hspace{0.05cm} \rm bit/symbol.$$


$\text{Example 3:}$  Now let us consider a  »memoryless  binary source«  with equally probable symbols, i.e. there would be  $p_{\rm A} = p_{\rm B} = 1/2$.  The first twenty subsequent elements are   $〈 q_ν 〉 =\rm BBAAABAABBBBBAAAABABAB$ ...

  • Because of the equally probable binary symbols   $H_1 = H_0 = 1 \hspace{0.05cm} \rm bit/symbol$.
  • The compound probability  $p_{\rm AB}$  of the combination  $\rm AB$  is equal to  $p_{\rm A} \cdot p_{\rm B} = 1/4$.  Also $p_{\rm AA} = p_{\rm BB} = p_{\rm BA} = 1/4$.
  • This gives for the second entropy approximation
$$H_2 = {1}/{2} \cdot \big [ {1}/{4} \cdot {\rm log}_2\hspace{0.1cm}4 + {1}/{4} \cdot {\rm log}_2\hspace{0.1cm}4 +{1}/{4} \cdot {\rm log}_2\hspace{0.1cm}4 +{1}/{4} \cdot {\rm log}_2\hspace{0.1cm}4 \big ] = 1 \,{\rm bit/symbol} = H_1 = H_0 \hspace{0.05cm}.$$

Note:   Due to the short length of the given sequence the probabilities are slightly different:  $p_{\rm AA} = 6/19$,  $p_{\rm BB} = 5/19$,  $p_{\rm AB} = p_{\rm BA} = 4/19$.


$\text{Example 4:}$  The third sequence considered here results from the binary sequence of  $\text{Example 3}$  by using a simple repeat code:

$$〈 q_ν 〉 =\rm BbBbAaAaAaBbAaAaBbBb \text{...} $$
  • The repeated symbols are marked by corresponding lower case letters.  But it still applies  $M=2$.
  • Because of the equally probable binary symbols, this also results in  $H_1 = H_0 = 1 \hspace{0.05cm} \rm bit/symbol$.
  • As shown in  "Exercise 1.3"  for the compound probabilities we obtain  $p_{\rm AA}=p_{\rm BB} = 3/8$  and  $p_{\rm AB}=p_{\rm BA} = 1/8$.  Hence
$$\begin{align*}H_2 ={1}/{2} \cdot \big [ 2 \cdot {3}/{8} \cdot {\rm log}_2\hspace{0.1cm} {8}/{3} + 2 \cdot {1}/{8} \cdot {\rm log}_2\hspace{0.1cm}8\big ] = {3}/{8} \cdot {\rm log}_2\hspace{0.1cm}8 - {3}/{8} \cdot{\rm log}_2\hspace{0.1cm}3 + {1}/{8} \cdot {\rm log}_2\hspace{0.1cm}8 \approx 0.906 \,{\rm bit/symbol} < H_1 = H_0 \hspace{0.05cm}.\end{align*}$$

A closer look at the task at hand leads to the following conclusion:

  • The entropy should actually be  $H = 0.5 \hspace{0.05cm} \rm bit/symbol$  (every second symbol does not provide new information)
  • The second entropy approximation  $H_2 = 0.906 \hspace{0.05cm} \rm bit/symbol$  is much larger than the entropy  $H$.
  • For the entropy determination, the second order approximation is not sufficient.  Rather, one must consider larger contiguous blocks of  $k > 2$  symbols.
  • In the following, such a block is referred to as  "$k$–tuple".


Generalization to $k$-tuple and boundary crossing


For abbreviation we write using the compound probability  $p_i^{(k)}$  for a  $k$–tuple in general:

$$H_k = \frac{1}{k} \cdot \sum_{i=1}^{M^k} p_i^{(k)} \cdot {\rm log}_2\hspace{0.1cm} \frac{1}{p_i^{(k)}} \hspace{0.5cm}({\rm unit\hspace{-0.1cm}: \hspace{0.1cm}bit/symbol}) \hspace{0.05cm}.$$
  • The control variable  $i$  stands for one of the  $M^k$  tuples. 
  • The previously calculated approximation  $H_2$  results with  $k = 2$.


$\text{Definition:}$  The  »entropy of a discrete source with memory«  has the following limit:

$$H = \lim_{k \rightarrow \infty }H_k \hspace{0.05cm}.$$

For the  »entropy approximations«  $H_k$  the following relations apply  $(H_0=\log_2 M)$:

$$H \le \text{...} \le H_k \le \text{...} \le H_2 \le H_1 \le H_0 \hspace{0.05cm}.$$


The computational effort will increase with increasing  $k$  except for a few special cases (see the following example) and depends on the symbol set size  $M$  of course:

  • For the calculation of  $H_{10}$  for a binary source  $(M = 2)$  one must average over  $2^{10} = 1024$  terms.
  • With each further increase of  $k$  by  $1$  the number of sum terms doubles.
  • In case of a quaternary source  $(M = 4)$  it must already be averaged over  $4^{10} = 1\hspace{0.08cm}048\hspace{0.08cm}576$  summation terms for the determination of  $H_{10}$.
  • Considering that each of these  $4^{10} =2^{20} >10^6$  $k$–tuple should occur in simulation/time averaging about  $100$  times (statistical guideline) to ensure sufficient simulation accuracy, it follows that the sequence length should be in this case greater than  $N = 10^8$.


$\text{Example 5:}$  We consider an alternating binary sequence   ⇒   $〈 q_ν 〉 =\rm ABABABAB$ ...  .  Accordingly it holds that  $H_0 = H_1 = 1 \hspace{0.1cm} \rm bit/symbol$.

In this special case, the  $H_k$ approximation must be determined independently from  $k$  by averaging only two compound probabilities:

  • $k = 2$:    $p_{\rm AB} = p_{\rm BA} = 1/2$     ⇒     $H_2 = 1/2 \hspace{0.2cm} \rm bit/symbol$,
  • $k = 3$:    $p_{\rm ABA} = p_{\rm BAB} = 1/2$     ⇒     $H_3 = 1/3 \hspace{0.2cm} \rm bit/symbol$,
  • $k = 4$:    $p_{\rm ABAB} = p_{\rm BABA} = 1/2$     ⇒     $H_4 = 1/4 \hspace{0.2cm} \rm bit/symbol$.


The entropy of this alternating binary sequence is therefore

$$H = \lim_{k \rightarrow \infty }{1}/{k} = 0 \hspace{0.05cm}.$$

The result was to be expected, since the considered sequence has only minimal information, which does not affect the entropy end value  $H$,  namely:
    "Does  $\rm A$  occur at the even or the odd times?"

You can see that  $H_k$  comes closer to this final value  $H = 0$  very slowly:  The twentieth entropy approximation still returns  $H_{20} = 0.05 \hspace{0.05cm} \rm bit/symbol$.


$\text{Summary of the results of the last sections:}$ 

  • Generally it applies to the  »entropy of a message source«:
$$H \le \text{...} \le H_3 \le H_2 \le H_1 \le H_0 \hspace{0.05cm}.$$
  • A  »redundancy-free source«   exists if all  $M$  symbols are equally probable and there are no statistical bindings within the sequence.
    For these hold  $(r$  denotes the relative redundancy $)$:
$$H = H_0 = H_1 = H_2 = H_3 = \text{...}\hspace{0.5cm} \Rightarrow \hspace{0.5cm} r = \frac{H - H_0}{H_0}= 0 \hspace{0.05cm}.$$
  • A  »memoryless source«   can be quite redundant  $(r> 0)$.  This redundancy then is solely due to the deviation of the symbol probabilities from the uniform distribution.  Here the following relations are valid:
$$H = H_1 = H_2 = H_3 = \text{...} \le H_0 \hspace{0.5cm}\Rightarrow \hspace{0.5cm}0 \le r = \frac{H_1 - H_0}{H_0}< 1 \hspace{0.05cm}.$$
  • The corresponding condition for a  »source with memory«  is
$$ H <\text{...} < H_3 < H_2 < H_1 \le H_0 \hspace{0.5cm}\Rightarrow \hspace{0.5cm} 0 < r = \frac{H_1 - H_0}{H_0}\le1 \hspace{0.05cm}.$$
  • If  $H_2 < H_1$, then  $H_3 < H_2$,   $H_4 < H_3$, etc.   ⇒   In the general equation, the  "≤" character must be replaced by the  "<" character.
  • If the symbols are equally probable, then again  $H_1 = H_0$,  while  $H_1 < H_0$  applies to symbols which are not equally probable.


The entropy of the AMI code


In chapter  "Symbol-wise Coding with Pseudo-Ternary Codes"  of the book  "Digital Signal Transmission", among other things, the AMI pseudo-ternary code is discussed.

Signals and symbol sequences for AMI code
  • This converts the binary sequence  $〈 q_ν 〉$  with  $q_ν ∈ \{ \rm L, \ H \}$  into the ternary sequence  $〈 c_ν 〉$  with  $q_ν ∈ \{ \rm M, \ N, \ P \}$.
  • The names of the source symbols stand for  $\rm L$(ow)  and  $\rm H$(igh)  and those of the code symbols for  $\rm M$(inus),  $\rm P$(lus),  and  $\rm N$(ull)   ⇒   "Zero".


The coding rule of the AMI code ("Alternate Mark Inversion") is:

  • Each binary symbol  $q_ν =\rm L$  is represented by the code symbol  $c_ν =\rm N$ .
  • In contrast,  $q_ν =\rm H$  alternates with  $c_ν =\rm P$  and  $c_ν =\rm M$  coded   ⇒   name "Alternate Mark Inversion".
  • This special encoding adds redundancy with the sole purpose of ensuring that the encoded sequence does not contain a DC component.


However, we do not consider the spectral properties of the AMI code here, but interpret this code information-theoretically:

  • Based on the symbol set size  $M = 3$  the decision content of the (ternary) encoded sequence is equal to  $H_0 = \log_2 \ 3 ≈ 1.585 \hspace{0.05cm} \rm bit/symbol$.  The first entropy approximation returns  $H_1 = 1.5 \hspace{0.05cm} \rm bit/symbol$, as shown in the following calculation:
$$p_{\rm H} = p_{\rm L} = 1/2 \hspace{0.3cm}\Rightarrow \hspace{0.3cm} p_{\rm N} = p_{\rm L} = 1/2\hspace{0.05cm},\hspace{0.2cm}p_{\rm M} = p_{\rm P}= p_{\rm H}/2 = 1/4\hspace{0.3cm} \Rightarrow \hspace{0.3cm} H_1 = 1/2 \cdot {\rm log}_2\hspace{0.1cm}2 + 2 \cdot 1/4 \cdot{\rm log}_2\hspace{0.1cm}4 = 1.5 \,{\rm bit/symbol} \hspace{0.05cm}.$$
  • Let's now look at two-tuples.  With the AMI code,  $\rm P$  cannot follow  $\rm P$  and  $\rm M$  cannot follow  $\rm M$.  The probability for  $\rm NN$  is equal to  $p_{\rm L} \cdot p_{\rm L} = 1/4$.  All other (six) two-tuples occur with the probability  $1/8$.  From this follows for the second entropy approximation:
$$H_2 = 1/2 \cdot \big [ 1/4 \cdot {\rm log_2}\hspace{0.1cm}4 + 6 \cdot 1/8 \cdot {\rm log_2}\hspace{0.1cm}8 \big ] = 1,375 \,{\rm bit/symbol} \hspace{0.05cm}.$$
  • For the further entropy approximations  $H_3$,  $H_4$, ...  and the actual entropy  $H$  will apply:
$$ H < \hspace{0.05cm}\text{...}\hspace{0.05cm} < H_5 < H_4 < H_3 < H_2 = 1.375 \,{\rm bit/symbol} \hspace{0.05cm}.$$
  • Exceptionally in this example we know the actual entropy  $H$  of the encoded sequence  $〈 c_ν 〉$:   Since no new information is added by the coder and no information is lost, it is the same entropy  $H = 1 \,{\rm bit/symbol} $  as the one of the redundancy-free binary sequence $〈 q_ν 〉$.


The  "Exercise 1.4"  shows the considerable effort required to calculate the entropy approximation  $H_3$.   Moreover,  $H_3$  still deviates significantly from the final value  $H = 1 \,{\rm bit/symbol} $.  A faster result is achieved if the AMI code is described by a Markov chain as explained in the next section.


Binary sources with Markov properties


Markov processes with  $M = 2$  states

Sequences with statistical bindings between the sequence elements (symbols) are often modeled by  $\text{Markov processes}$  where we limit ourselves here to first-order Markov processes.  First we consider a binary Markov process  $(M = 2)$  with the states (symbols)  $\rm A$  and  $\rm B$.

On the right, you can see the transition diagram for a first-order binary Markov process  however, only two of the four transfer probabilities given are freely selectable, for example

  • $p_{\rm {A\hspace{0.01cm}|\hspace{0.01cm}B}} = \rm Pr(A\hspace{0.01cm}|\hspace{0.01cm}B)$   ⇒   conditional probability that  $\rm A$  follows  $\rm B$ .
  • $p_{\rm {B\hspace{0.01cm}|\hspace{0.01cm}A}} = \rm Pr(B\hspace{0.01cm}|\hspace{0.01cm}A)$   ⇒   conditional probability that  $\rm B$  follows  $\rm A$ .


For the other two transition probabilities:  $p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}A} = 1- p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A}$  and   $p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}B} = 1- p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B} \hspace{0.05cm}.$

Due to the presupposed properties  $\text{Stationarity}$  and  $\text{Ergodicity}$  the following applies to the state or symbol probabilities:

$$p_{\rm A} = {\rm Pr}({\rm A}) = \frac{p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B}}{p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B} + p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A}} \hspace{0.05cm}, \hspace{0.5cm}p_{\rm B} = {\rm Pr}({\rm B}) = \frac{p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A}}{p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B} + p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A}} \hspace{0.05cm}.$$

These equations allow first information–theoretical statements about the Markov processes:

  • For  $p_{\rm {A\hspace{0.01cm}|\hspace{0.01cm}B}} = p_{\rm {B\hspace{0.01cm}|\hspace{0.01cm}A}}$  the symbols are equally likely   ⇒   $p_{\text{A}} = p_{\text{B}}= 0.5$.  The first entropy approximation returns  $H_1 = H_0 = 1 \hspace{0.05cm} \rm bit/symbol$, independent of the actual values of the (conditional) transition probabilities  $p_{\text{A|B}}$  and  $p_{\text{B|A}}$.
  • But the source entropy  $H$  as the limit value  $($for  $k \to \infty)$  of the  $\text{Entropy approximation of order  $k$}$   ⇒   $H_k$  depends very much on the actual values of  $p_{\text{A|B}}$  and  $p_{\text{B|A}}$  and not only on their quotients.  This is shown by the following example.


$\text{Example 6:}$  We consider three binary symmetric Markov sources that are represented by the numerical values of the symmetric transition probabilities  $p_{\rm {A\hspace{0.01cm}\vert\hspace{0.01cm}B} } = p_{\rm {B\hspace{0.01cm}\vert\hspace{0.01cm}A} }$.  For the symbol probabilities the following applies:  $p_{\rm A} = p_{\rm B}= 0.5$, and the other transition probabilities have the values

Three examples of binary Markov sources with  $p_{\rm {A\hspace{0.01cm}\vert\hspace{0.01cm}B} } = p_{\rm {B\hspace{0.01cm}\vert\hspace{0.01cm}A} }$ 
$$p_{\rm {A\hspace{0.01cm}\vert\hspace{0.01cm}A} } = 1 - p_{\rm {B\hspace{0.01cm}\vert\hspace{0.01cm}A} } = p_{\rm {B\hspace{0.01cm}\vert\hspace{0.01cm}B} }.$$
  • The middle (blue) sequence with  $p_{\rm {A\hspace{0.01cm}\vert\hspace{0.01cm}B} } = p_{\rm {B\hspace{0.01cm}\vert\hspace{0.01cm}A} } = 0.5$  has the entropy  $H ≈ 1 \hspace{0.1cm} \rm bit/symbol$.  That means:   In this special case there are no statistical bindings within the sequence.
  • The left (red) sequence with  $p_{\rm {A\hspace{0.01cm}\vert\hspace{0.01cm}B} } = p_{\rm {B\hspace{0.01cm}\vert\hspace{0.01cm}A} } = 0.2$  has less changes between  $\rm A$  and  $\rm B$.  Due to statistical dependencies between neighboring symbols the entropy is now smaller:  $H ≈ 0.72 \hspace{0.1cm} \rm bit/symbol$.
  • The right (green) symbol sequence with  $p_{\rm {A\hspace{0.01cm}\vert\hspace{0.01cm}B} } = p_{\rm {B\hspace{0.01cm}\vert\hspace{0.01cm}A} } = 0.8$  has the exact same entropy  $H ≈ 0.72 \hspace{0.1cm} \rm bit/symbol$  as the red sequence.  Here you can see many regions with alternating symbols  $($... $\rm ABABAB$ ... $)$.


This example is worth noting:

  • If you had not used the Markov properties of the red and green sequences, you would have reached the respective result  $H ≈ 0.72 \hspace{0.1cm} \rm bit/symbol$  only after lengthy calculations.
  • The following sections show that for a source with Markov properties the final value  $H$  can be determined from the entropy approximations  $H_1$  and  $H_2$  alone.   Likewise, all entropy approximations  $H_1$  and  $H_2$  can also be calculated from  $H_k$  for  $k$–tuples in a simple manner   ⇒   $H_3$,  $H_4$,  $H_5$, ...   $H_{100}$, ...


Simplified entropy calculation for Markov sources


Markov processes with  $M = 2$  states

We continue to assume the first-order symmetric binary Markov source.  As in the previous section, we use the following nomenclature for

  • the transition probabilities   ⇒   $p_{\rm {A\hspace{0.01cm}|\hspace{0.01cm}B}}$,   $p_{\rm {B\hspace{0.01cm}|\hspace{0.01cm}A}}$,  $p_{\rm {A\hspace{0.01cm}|\hspace{0.01cm}A}}= 1- p_{\rm {B\hspace{0.01cm}|\hspace{0.01cm}A}}$,   $p_{\rm {B\hspace{0.01cm}|\hspace{0.01cm}B}} = 1 - p_{\rm {A\hspace{0.01cm}|\hspace{0.01cm}B}}$,  
  • the ergodic probabilities   ⇒   $p_{\text{A}}$  and  $p_{\text{B}}$,
  • the compound probabilities, for example   ⇒   $p_{\text{AB}} = p_{\text{A}} - p_{\rm {B\hspace{0.01cm}|\hspace{0.01cm}A}}$.


We now compute the  $\text{Entropy of a two-tuple}$  (with the unit "bit/two-tuple"):

$$H_2\hspace{0.05cm}' = p_{\rm A} \cdot p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}A} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{ p_{\rm A} \cdot p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}A}} + p_{\rm A} \cdot p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{ p_{\rm A} \cdot p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A}} + p_{\rm B} \cdot p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{ p_{\rm B} \cdot p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B}} + p_{\rm B} \cdot p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}B} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{ p_{\rm B} \cdot p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}B}} \hspace{0.05cm}.$$

If one now replaces the logarithms of the products by corresponding sums of logarithms, one gets the result  $H_2\hspace{0.05cm}' = H_1 + H_{\text{M}}$  with

$$H_1 = p_{\rm A} \cdot (p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}A} + p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A})\cdot {\rm log}_2\hspace{0.1cm}\frac {1}{p_{\rm A}} + p_{\rm B} \cdot (p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B} + p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}B})\cdot {\rm log}_2\hspace{0.1cm}\frac {1}{p_{\rm B}} = p_{\rm A} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{p_{\rm A}} + p_{\rm B} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{p_{\rm B}} = H_{\rm bin} (p_{\rm A})= H_{\rm bin} (p_{\rm B}) \hspace{0.05cm},$$
$$H_{\rm M}= p_{\rm A} \cdot p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}A} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{ p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}A}} + p_{\rm A} \cdot p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{ p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}A}} + p_{\rm B} \cdot p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{ p_{\rm A\hspace{0.01cm}|\hspace{0.01cm}B}} + p_{\rm B} \cdot p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}B} \cdot {\rm log}_2\hspace{0.1cm}\frac {1}{ p_{\rm B\hspace{0.01cm}|\hspace{0.01cm}B}} \hspace{0.05cm}.$$

$\text{Conclusion:}$  Thus the  »second entropy approximation«  (with the unit "bit/symbol")  is:

$$H_2 = {1}/{2} \cdot {H_2\hspace{0.05cm}'} = {1}/{2} \cdot \big [ H_{\rm 1} + H_{\rm M} \big] \hspace{0.05cm}.$$


It is to be noted:

  • The first summand  $H_1$   ⇒   first entropy approximation depends only on the symbol probabilities.
  • For a symmetrical Markov process   ⇒   $p_{\rm {A\hspace{0.01cm}|\hspace{0.01cm}B}} = p_{\rm {B\hspace{0.01cm}|\hspace{0.01cm}A}} $   ⇒   $p_{\text{A}} = p_{\text{B}} = 1/2$   the result for this first summand is  $H_1 = 1 \hspace{0.1cm} \rm bit/symbol$.
  • The second summand  $H_{\text{M}}$  must be calculated according to the second of the two upper equations.
  • For a symmetrical Markov process you get  $H_{\text{M}} = H_{\text{bin}}(p_{\rm {A\hspace{0.01cm}|\hspace{0.01cm}B}})$.


Now, this result is extended to the  $k$–th entropy approximation.  Here, the advantage of Markov sources over other sources is taken advantage of, that the entropy calculation for  $k$–tuples is very simple.  For each Markov source, the following applies

$$H_k = {1}/{k} \cdot \big [ H_{\rm 1} + (k-1) \cdot H_{\rm M}\big ] \hspace{0.3cm} \Rightarrow \hspace{0.3cm} H_2 = {1}/{2} \cdot \big [ H_{\rm 1} + H_{\rm M} \big ]\hspace{0.05cm}, \hspace{0.3cm} H_3 ={1}/{3} \cdot \big [ H_{\rm 1} + 2 \cdot H_{\rm M}\big ] \hspace{0.05cm},\hspace{0.3cm} H_4 = {1}/{4} \cdot \big [ H_{\rm 1} + 3 \cdot H_{\rm M}\big ] \hspace{0.05cm},\hspace{0.15cm}{\rm usw.}$$

$\text{Conclusion:}$  With the boundary condition for  $k \to \infty$, one obtains the actual source entropy

$$H = \lim_{k \rightarrow \infty} H_k = H_{\rm M} \hspace{0.05cm}.$$

From this simple result important insights for the entropy calculation follow:

  • For Markov sources it is sufficient to determine the entropy approximations  $H_1$  and  $H_2$.  Thus, the entropy of a Markov source is
$$H = 2 \cdot H_2 - H_{\rm 1} \hspace{0.05cm}.$$
  • Through  $H_1$  and  $H_2$  all further entropy approximations  $H_k$  are also fixed for  $k \ge 3$ :
$$H_k = \frac{2-k}{k} \cdot H_{\rm 1} + \frac{2\cdot (k-1)}{k} \cdot H_{\rm 2} \hspace{0.05cm}.$$


However, these approximations are not very important.  Mostly only the limit value  $H$  is of interest.  For sources without Markov properties the approximations  $H_k$  are calculated only to be able to estimate this limit value, i.e. the actual entropy.


Notes:

  • In the  "Exercise 1.5"  the above equations are applied to the more general case of an asymmetric binary source.
  • All equations in this section also apply to non-binary Markov sources  $(M > 2)$ as shown in the next section.


Non-binary Markov sources


Ternary and quaternary first order Markov source

The following equations apply to each Markov source regardless of the symbol set size  $M$:

$$H = 2 \cdot H_2 - H_{\rm 1} \hspace{0.05cm},$$
$$H_k = {1}/{k} \cdot \big [ H_{\rm 1} + (k-1) \cdot H_{\rm M}\big ] \hspace{0.05cm},$$
$$ \lim_{k \rightarrow \infty} H_k = H \hspace{0.05cm}.$$

These equations allow the simple calculation of the entropy  $H$  from the approximations  $H_1$  and  $H_2$.

We now look at the transition diagrams sketched on the right:

  • A ternary Markov source  $\rm MQ3$  $(M = 3$,  blue coloring$)$,  and
  • a quaternary Markov source  $\rm MQ4$  $(M = 4$,  red color$)$.


In  "Exercise 1.6"  the entropy approximations  $H_k$  and the total entropy  $H$  are calculated as the limit of  $H_k$  for  $k \to \infty$ .   The results are shown in the following figure.  All entropies specified there have the unit "bit/symbol".

Entropies for  $\rm MQ3$,  $\rm MQ4$  and the  $\rm AMI$  code

These results can be interpreted as follows:

  • For the ternary markov source  $\rm MQ3$  the entropy approximations of  $H_1 = 1.500$  above  $H_2 = 1.375$  up to the limit  $H = 1.250$  continuously decreasing.  Because of  $M = 3$  ⇒   $H_0 = 1.585$.
  • For the quaternary Markov source  $\rm MQ4$  one receives  $H_0 = H_1 = 2.000$  (since four equally probable states) and  $H_2 = 1.5$.   From the  $H_1$-  and  $H_2$-values all entropy approximations  $H_k$  and the final value  $H = 1.000$  can be calculated.
  • The two models  $\rm MQ3$  and  $\rm MQ4$  were created during the attempt to calculate the  $\text{AMI code}$  to be described information–theoretically by Markov sources.  The symbols  $\rm M$,  $\rm N$  and  $\rm P$  stand for  "minus",  "zero"  and  "plus".
  • The entropy approximations  $H_1$,  $H_2$  and  $H_3$  of the AMI code (green markers) were calculated in the  "Exercise 1.4".  On the calculation of  $H_4$,  $H_5$, ... had to be omitted for reasons of effort.  But the final value of  $H_k$  for  $k \to \infty$   ⇒   $H = 1.000$  is known.
  • You can see that the Markov model  $\rm MQ3$  for  $H_0 = 1.585$,  $H_1 = 1.500$  and  $H_2 = 1.375$  yields exactly the same numerical values as the AMI code.   On the other hand  $H_3$  ⇒   $(1.333$  instead of  $1.292)$  and especially the final value  $H$  ⇒   $(1.250$  compared to  $1.000)$.
  • The model  $\rm MQ4$  $(M = 4)$  differs from the AMI code  $(M = 3)$  with respect to the decision content  $H_0$  and also with respect to all entropy approximations  $H_k$.   Nevertheless, $\rm MQ4$  is the more suitable model for the AMI code, since the final value  $H = 1,000$  is the same.
  • The model  $\rm MQ3$  yields entropy values that are too large, since the sequences  $\rm PNP$  and  $\rm MNM$  are possible here, which cannot occur in the AMI code.   Already with the approximation  $H_3$  the difference is slightly noticeable, in the final value  $H$  even clearly  $(1.25$  instead of  $1)$.


In the model  $\rm MQ4$  the state  "Zero"  was split into two states  $\rm N$  and  $\rm O$  (see upper right figure in this section):

  • Here applies to the state  $\rm N$:   The current binary symbol  $\rm L$  is encoded with the amplitude value  $0$  (zero), as per the AMI rule.  The next occurring  $\rm H$ symbol, on the other hand, is displayed as  $-1$  (minus), because the last symol  $\rm H$  was encoded as  $+1$  (plus).
  • Also with the state  $\rm O$  the current binary symbol  $\rm L$  is represented with the ternary value  $0$.  In contrast to the state  $\rm N$  however, the next occurring  $\rm H$ symbol is now encoded as  $+1$  (plus) since the last  $\rm H$ symbol was encoded as  $-1$  (minus).


$\text{Conclusion:}$ 

  • The  $\rm MQ4$ output sequence actually follows the rules of the AMI code and has entropy  $H = 1.000 \hspace{0.15cm} \rm bit/symbol$.
  • But because of the new state  $\rm O$  now  $H_0 = 2.000 \hspace{0.15cm} \rm bit/symbol$  $($against  $1.585 \hspace{0.15cm} \rm bit/symbol)$  is clearly too large.
  • Also all  $H_k$ approximations are larger than with the AMI code.  First for  $k \to \infty$  the model  $\rm MQ4$  and the AMI code match exactly:   $H = 1.000 \hspace{0.15cm} \rm bit/symbol$.


Exercises for the chapter


Exercise 1.3: Entropy Approximations

Exercise 1.4: Entropy Approximations for the AMI Code

Exercise 1.4Z: Entropy of the AMI Code

Exercise 1.5: Binary Markov Source

Exercise 1.5Z: Symmetrical Markov Source

Exercise 1.6: Non-Binary Markov Sources

Exercise 1.6Z:Ternary Markov Source