Difference between revisions of "Aufgaben:Exercise 3.4Z: Equivalent Convolution Codes?"

From LNTwww
 
(2 intermediate revisions by the same user not shown)
Line 1: Line 1:
 
{{quiz-Header|Buchseite=Channel_Coding/Algebraic_and_Polynomial_Description}}
 
{{quiz-Header|Buchseite=Channel_Coding/Algebraic_and_Polynomial_Description}}
  
[[File:P_ID2666__KC_Z_3_4.png|right|frame|Non-systematic and <br>systematic convolutional encoder]]
+
[[File:EN_KC_Z_3_4.png|right|frame|Non-systematic and systematic convolutional encoder]]
The top illustration shows a convolutional encoder described by the following equations:
+
The top figure shows a convolutional encoder described by the following equations:
 
:$$x_i^{(1)} \hspace{-0.15cm} \ = \ \hspace{-0.15cm} u_{i}^{(1)} + u_{i-1}^{(1)}+ u_{i-1}^{(2)} \hspace{0.05cm},$$
 
:$$x_i^{(1)} \hspace{-0.15cm} \ = \ \hspace{-0.15cm} u_{i}^{(1)} + u_{i-1}^{(1)}+ u_{i-1}^{(2)} \hspace{0.05cm},$$
 
:$$x_i^{(2)} \hspace{-0.15cm} \ = \ \hspace{-0.15cm} u_{i}^{(2)} + u_{i-1}^{(2)} \hspace{0.05cm},$$
 
:$$x_i^{(2)} \hspace{-0.15cm} \ = \ \hspace{-0.15cm} u_{i}^{(2)} + u_{i-1}^{(2)} \hspace{0.05cm},$$
Line 8: Line 8:
  
 
We are looking for the transfer function matrices
 
We are looking for the transfer function matrices
* $\mathbf{G}(D)$&nbsp; of this non-systematic code, and  
+
* $\mathbf{G}(D)$&nbsp; of this non-systematic code,&nbsp; and  
 +
 
 
* $\mathbf{G}_{\rm sys}(D)$&nbsp; of the equivalent systematic code.
 
* $\mathbf{G}_{\rm sys}(D)$&nbsp; of the equivalent systematic code.
  
  
 
The matrix&nbsp; $\mathbf{G}_{\rm sys}(D)$&nbsp; is obtained in the following way:
 
The matrix&nbsp; $\mathbf{G}_{\rm sys}(D)$&nbsp; is obtained in the following way:
* One splits off from the&nbsp; $k &times; n$ matrix&nbsp; $\mathbf{G}(D)$&nbsp; in front a square matrix&nbsp; $\mathbf{T}(D)$&nbsp; with each&nbsp; $k$&nbsp; rows and columns. The remainder is denoted by&nbsp; $\mathbf{Q}(D)$.
+
* One splits off from the&nbsp; $k &times; n$&nbsp; matrix&nbsp; $\mathbf{G}(D)$&nbsp; in front a square matrix&nbsp; $\mathbf{T}(D)$&nbsp; with&nbsp; $k$&nbsp; rows and&nbsp; $k$&nbsp; columns.&nbsp; The remainder is denoted by&nbsp; $\mathbf{Q}(D)$.
* Then calculate the inverse matrix of&nbsp; $\mathbf{T}(D)$&nbsp; $\mathbf{T}^{-1}(D)$&nbsp; and from this calculate the matrix for the equivalent systematic code:
+
 
 +
* Calculate the inverse matrix&nbsp; $\mathbf{T}^{-1}(D)$&nbsp; of &nbsp; $\mathbf{T}(D)$.&nbsp; From this calculate the matrix for the equivalent systematic code:
 
:$${\boldsymbol{\rm G}}_{\rm sys}(D)= {\boldsymbol{\rm T}}^{-1}(D) \cdot {\boldsymbol{\rm G}}(D) \hspace{0.05cm}.$$
 
:$${\boldsymbol{\rm G}}_{\rm sys}(D)= {\boldsymbol{\rm T}}^{-1}(D) \cdot {\boldsymbol{\rm G}}(D) \hspace{0.05cm}.$$
* Since&nbsp; $\mathbf{T}^{&ndash;1}(D) \cdot \mathbf{T}(D)$&nbsp; yields the&nbsp; $k &times; k$ identity matrix&nbsp; $\mathbf{I}_k$&nbsp;, the transfer function matrix of the equivalent systematic code can be written in the desired form:
+
* Since &nbsp; $\mathbf{T}^{&ndash;1}(D) \cdot \mathbf{T}(D)$ &nbsp; yields the&nbsp; $k &times; k$&nbsp; identity matrix&nbsp; $\mathbf{I}_k$,&nbsp; the transfer function matrix of the equivalent systematic code can be written in the desired form:
 
:$${\boldsymbol{\rm G}}_{\rm sys}(D) = \big [ \hspace{0.05cm} {\boldsymbol{\rm I}}_k\hspace{0.05cm} ; \hspace{0.1cm} {\boldsymbol{\rm P}}(D) \hspace{0.05cm}\big ]  
 
:$${\boldsymbol{\rm G}}_{\rm sys}(D) = \big [ \hspace{0.05cm} {\boldsymbol{\rm I}}_k\hspace{0.05cm} ; \hspace{0.1cm} {\boldsymbol{\rm P}}(D) \hspace{0.05cm}\big ]  
 
\hspace{0.5cm}{\rm with}\hspace{0.5cm} {\boldsymbol{\rm P}}(D)= {\boldsymbol{\rm T}}^{-1}(D) \cdot {\boldsymbol{\rm Q}}(D) \hspace{0.05cm}.
 
\hspace{0.5cm}{\rm with}\hspace{0.5cm} {\boldsymbol{\rm P}}(D)= {\boldsymbol{\rm T}}^{-1}(D) \cdot {\boldsymbol{\rm Q}}(D) \hspace{0.05cm}.
 
\hspace{0.05cm}$$
 
\hspace{0.05cm}$$
  
The circuit below will certainly generate a systematic code with the same parameters&nbsp; $k$&nbsp; and&nbsp; $n$.  
+
*The circuit below will certainly generate a systematic code with the same parameters&nbsp; $k$&nbsp; and&nbsp; $n$.  
 +
 
  
In the subtask '''(5)''' it has to be clarified whether this is indeed the&nbsp; <i>equivalent systematic code</i>&nbsp;. That is, whether in fact for the two circuits exactly the same quantity&nbsp; $\{ \hspace{0.1cm} \underline{x} \hspace{0.1cm}\}$&nbsp; of code sequences results when all possible information sequences&nbsp; $\{ \hspace{0.1cm} \underline{u} \hspace{0.1cm} \}$&nbsp; is taken into account.
+
In subtask&nbsp; '''(5)'''&nbsp; it has to be clarified whether this is indeed the&nbsp; "equivalent systematic code".&nbsp; That is,&nbsp; whether in fact for the two circuits exactly the same quantity &nbsp; $\{ \hspace{0.1cm} \underline{x} \hspace{0.1cm}\}$ &nbsp; of code sequences results when all possible information sequences &nbsp; $\{ \hspace{0.1cm} \underline{u} \hspace{0.1cm} \}$ &nbsp; are taken into account.
  
  
Line 31: Line 34:
  
 
Hints:  
 
Hints:  
* This exercise belongs to the chapter&nbsp; [[Channel_Coding/Algebraic_and_Polynomial_Description| Algebraic and Polynomial Description]].
+
* This exercise belongs to the chapter&nbsp; [[Channel_Coding/Algebraic_and_Polynomial_Description| "Algebraic and Polynomial Description"]].
* Reference is made in particular to the pages&nbsp;
+
* Reference is made in particular to the sections&nbsp;
:: [[Channel_Coding/Algebraic_and_Polynomial_Description#Transfer_Function_Matrix|"Transfer Function Matrix"]]&nbsp; and&nbsp;
+
:* [[Channel_Coding/Algebraic_and_Polynomial_Description#Transfer_Function_Matrix|"Transfer Function Matrix"]]&nbsp; and&nbsp;
:: [[Channel_Coding/Algebraic_and_Polynomial_Description#Equivalent_systematic_convolutional_code|"Equivalent systematic convolutional code"]].
+
:* [[Channel_Coding/Algebraic_and_Polynomial_Description#Equivalent_systematic_convolutional_code|"Equivalent systematic convolutional code"]].
 
   
 
   
  
Line 56: Line 59:
 
- The third row of&nbsp; $\mathbf{G}(D)$&nbsp; is&nbsp; $(D, \, 1 + D, \, 1)$.
 
- The third row of&nbsp; $\mathbf{G}(D)$&nbsp; is&nbsp; $(D, \, 1 + D, \, 1)$.
  
{Specify&nbsp; $\mathbf{T}(D)$&nbsp; and&nbsp; $\mathbf{T}^{-1}(D)$&nbsp;. What is the determinant?
+
{Specify &nbsp; $\mathbf{T}(D)$&nbsp; and &nbsp; $\mathbf{T}^{-1}(D)$.&nbsp; What is the determinant?
 
|type="()"}
 
|type="()"}
 
- $\det {\mathbf{T}(D)} = 1$,
 
- $\det {\mathbf{T}(D)} = 1$,
Line 76: Line 79:
 
===Solution===
 
===Solution===
 
{{ML-Kopf}}
 
{{ML-Kopf}}
'''(1)'''&nbsp; Here $\underline{k = 2}$ and $\underline{n = 3}$ &nbsp;&#8658;&nbsp; Rate $\underline{R = 2/3}$.  
+
'''(1)'''&nbsp; Here&nbsp; $\underline{k = 2}$&nbsp; and&nbsp; $\underline{n = 3}$ &nbsp; &#8658; &nbsp; Rate&nbsp; $\underline{R = 2/3}$.  
*The memory order $\underline{m = 1}$ (number of memory elements per input).  
+
*The memory order&nbsp; $\underline{m = 1}$&nbsp; $($number of memory elements per input$)$.
*The influence length is equal to the sum of all memory elements &#8658; $\underline{\nu = 2}$.
+
 +
*The influence length is equal to the sum of all memory elements &nbsp; &#8658; &nbsp; $\underline{\nu = 2}$.
  
  
  
'''(2)'''&nbsp; The information bit $u_i^{(1)}$ affects only the first output $x_i^{(1)}$, while $u_i^{(2)}$ is used for $x_i^{(2)}$ and $x_i^{(3)}$.  
+
'''(2)'''&nbsp; The information bit&nbsp; $u_i^{(1)}$&nbsp; affects only the first output&nbsp; $x_i^{(1)}$,&nbsp; while&nbsp; $u_i^{(2)}$&nbsp; is used for&nbsp; $x_i^{(2)}$&nbsp; and&nbsp; $x_i^{(3)}$.  
*Thus, for the zeroth [[Channel_Coding/Algebraic_and_Polynomial_Description#Division_of_the_generator_matrix_into_partial_matrices| "partial matrix"]] is obtained:
+
*Thus,&nbsp; for the zeroth&nbsp; [[Channel_Coding/Algebraic_and_Polynomial_Description#Division_of_the_generator_matrix_into_partial_matrices| "partial matrix"]]&nbsp; is obtained:
 
:$${ \boldsymbol{\rm G}}_0 =  
 
:$${ \boldsymbol{\rm G}}_0 =  
 
\begin{pmatrix}
 
\begin{pmatrix}
Line 91: Line 95:
  
 
*The delayed inputs affect as follows:
 
*The delayed inputs affect as follows:
** $u_{i&ndash;1}^{(1)}$ affects $x_i^{(1)}$,
+
:* $u_{i&ndash;1}^{(1)}$&nbsp; affects&nbsp; $x_i^{(1)}$,
** $u_{i&ndash;1}^{(2)}$ affects $x_i^{(1)}$ und $x_i^{(2)}$:
 
  
 +
:* $u_{i&ndash;1}^{(2)}$&nbsp; affects&nbsp; $x_i^{(1)}$&nbsp; and&nbsp; $x_i^{(2)}$:
  
*Thus the partial matrix $\mathbf{G}_1$ and the transfer function matrix $\mathbf{G}(D)$:
+
 
 +
*Thus,&nbsp; the partial matrix&nbsp; $\mathbf{G}_1$&nbsp; and the transfer function matrix&nbsp; $\mathbf{G}(D)$:
 
:$${ \boldsymbol{\rm G}}_1 =  
 
:$${ \boldsymbol{\rm G}}_1 =  
 
\begin{pmatrix}
 
\begin{pmatrix}
Line 107: Line 112:
 
\hspace{0.05cm}. $$
 
\hspace{0.05cm}. $$
  
*Therefore the <u>proposed solutions 1 and 3</u> are correct.  
+
*Therefore the&nbsp; <u>proposed solutions 1 and 3</u>&nbsp; are correct.
*Answer 2 cannot be correct, if only because no element with $D^2$ can occur in the transfer function matrix when $m = 1$.  
+
 +
*Answer 2 cannot be correct,&nbsp; because no element with&nbsp; $D^2$&nbsp; can occur in the transfer function matrix when&nbsp; $m = 1$.
 +
 
*$\mathbf{G}(D)$ is moreover a $2 &times; 3$ matrix; there is no third row.
 
*$\mathbf{G}(D)$ is moreover a $2 &times; 3$ matrix; there is no third row.
  
  
  
'''(3)'''&nbsp; Splitting $\mathbf{G}(D)$ gives the $2 &times; 2$ matrix.
+
'''(3)'''&nbsp; Splitting&nbsp; $\mathbf{G}(D)$&nbsp; gives the&nbsp; $2 &times; 2$&nbsp; matrix.
 
:$${ \boldsymbol{\rm T}}(D) =
 
:$${ \boldsymbol{\rm T}}(D) =
 
\begin{pmatrix}  
 
\begin{pmatrix}  
Line 125: Line 132:
 
\end{pmatrix} \hspace{0.05cm}. $$
 
\end{pmatrix} \hspace{0.05cm}. $$
  
*The correct solution is <u>solution 3</u>. For control:
+
*The correct solution is&nbsp; <u>solution 3</u>.&nbsp; For control:
 
:$${ \boldsymbol{\rm T}}(D) \cdot { \boldsymbol{\rm T}}^{-1}(D) \hspace{-0.15cm} \ = \ \hspace{-0.15cm}   
 
:$${ \boldsymbol{\rm T}}(D) \cdot { \boldsymbol{\rm T}}^{-1}(D) \hspace{-0.15cm} \ = \ \hspace{-0.15cm}   
 
\frac{1}{1+D^2} \cdot  
 
\frac{1}{1+D^2} \cdot  
Line 175: Line 182:
 
\end{pmatrix}\hspace{0.05cm}. $$
 
\end{pmatrix}\hspace{0.05cm}. $$
  
*The correct solution is therefore <u>proposals 1 and 3</u>.
+
*The correct solution is therefore the&nbsp; <u>proposals 1 and 3</u>.
  
  
'''(5)'''&nbsp; Correct <u>YES</u>. &nbsp; The lower circuit on the data sheet is identified by the equations. $x_i^{(1)} = u_i^{(1)}$ und $x_i^{(2)} = u_i^{(2)}$ sowie
+
'''(5)'''&nbsp; Correct is&nbsp; <u>YES</u>. &nbsp; The lower circuit on the data sheet is identified by the equations&nbsp; $x_i^{(1)} = u_i^{(1)}$,&nbsp; $x_i^{(2)} = u_i^{(2)}$,&nbsp; and
 
:$$x_i^{(3)}= x_{i-1}^{(3)} + u_i^{(2)} \hspace{0.15cm} \circ\!\!-\!\!\!-^{\hspace{-0.25cm}D}\!\!\!-\!\!\bullet\hspace{0.15cm}
 
:$$x_i^{(3)}= x_{i-1}^{(3)} + u_i^{(2)} \hspace{0.15cm} \circ\!\!-\!\!\!-^{\hspace{-0.25cm}D}\!\!\!-\!\!\bullet\hspace{0.15cm}
 
X^{(3)}(D)= X^{(3)}(D) \cdot D +U^{(2)}(D)$$
 
X^{(3)}(D)= X^{(3)}(D) \cdot D +U^{(2)}(D)$$
Line 184: Line 191:
 
\hspace{0.05cm}.$$
 
\hspace{0.05cm}.$$
  
This corresponds exactly to the last element of $\mathbf{G}_{\rm sys}(D)$ corresponding to subtask (4).
+
*This corresponds exactly to the last element of&nbsp; $\mathbf{G}_{\rm sys}(D)$&nbsp; from subtask&nbsp; '''(4)'''.
 
{{ML-Fuß}}
 
{{ML-Fuß}}
  
  
 
[[Category:Channel Coding: Exercises|^3.2 Polynomial Description^]]
 
[[Category:Channel Coding: Exercises|^3.2 Polynomial Description^]]

Latest revision as of 18:58, 10 November 2022

Non-systematic and systematic convolutional encoder

The top figure shows a convolutional encoder described by the following equations:

$$x_i^{(1)} \hspace{-0.15cm} \ = \ \hspace{-0.15cm} u_{i}^{(1)} + u_{i-1}^{(1)}+ u_{i-1}^{(2)} \hspace{0.05cm},$$
$$x_i^{(2)} \hspace{-0.15cm} \ = \ \hspace{-0.15cm} u_{i}^{(2)} + u_{i-1}^{(2)} \hspace{0.05cm},$$
$$x_i^{(3)} \hspace{-0.15cm} \ = \ \hspace{-0.15cm} u_{i}^{(2)}\hspace{0.05cm}.$$

We are looking for the transfer function matrices

  • $\mathbf{G}(D)$  of this non-systematic code,  and
  • $\mathbf{G}_{\rm sys}(D)$  of the equivalent systematic code.


The matrix  $\mathbf{G}_{\rm sys}(D)$  is obtained in the following way:

  • One splits off from the  $k × n$  matrix  $\mathbf{G}(D)$  in front a square matrix  $\mathbf{T}(D)$  with  $k$  rows and  $k$  columns.  The remainder is denoted by  $\mathbf{Q}(D)$.
  • Calculate the inverse matrix  $\mathbf{T}^{-1}(D)$  of   $\mathbf{T}(D)$.  From this calculate the matrix for the equivalent systematic code:
$${\boldsymbol{\rm G}}_{\rm sys}(D)= {\boldsymbol{\rm T}}^{-1}(D) \cdot {\boldsymbol{\rm G}}(D) \hspace{0.05cm}.$$
  • Since   $\mathbf{T}^{–1}(D) \cdot \mathbf{T}(D)$   yields the  $k × k$  identity matrix  $\mathbf{I}_k$,  the transfer function matrix of the equivalent systematic code can be written in the desired form:
$${\boldsymbol{\rm G}}_{\rm sys}(D) = \big [ \hspace{0.05cm} {\boldsymbol{\rm I}}_k\hspace{0.05cm} ; \hspace{0.1cm} {\boldsymbol{\rm P}}(D) \hspace{0.05cm}\big ] \hspace{0.5cm}{\rm with}\hspace{0.5cm} {\boldsymbol{\rm P}}(D)= {\boldsymbol{\rm T}}^{-1}(D) \cdot {\boldsymbol{\rm Q}}(D) \hspace{0.05cm}. \hspace{0.05cm}$$
  • The circuit below will certainly generate a systematic code with the same parameters  $k$  and  $n$.


In subtask  (5)  it has to be clarified whether this is indeed the  "equivalent systematic code".  That is,  whether in fact for the two circuits exactly the same quantity   $\{ \hspace{0.1cm} \underline{x} \hspace{0.1cm}\}$   of code sequences results when all possible information sequences   $\{ \hspace{0.1cm} \underline{u} \hspace{0.1cm} \}$   are taken into account.




Hints:



Questions

1

What are the parameters of the encoder shown above?

$k \hspace{0.25cm} = \ $

$n \hspace{0.22cm} = \ $

$m \hspace{0.10cm} = \ $

$ν \hspace{0.28cm} = \ $

$R \hspace{0.18cm} = \ $

2

What is the form of the transfer function matrix  $\mathbf{G}(D)$?

The first row of  $\mathbf{G}(D)$  is  $(1 + D, \, 0, \, 0)$.
The first row of  $\mathbf{G}(D)$  is  $(1 + D^2, \, 0, \, D^2)$.
The second row of  $\mathbf{G}(D)$  is  $(D, \, 1 + D, \, 1)$.
The third row of  $\mathbf{G}(D)$  is  $(D, \, 1 + D, \, 1)$.

3

Specify   $\mathbf{T}(D)$  and   $\mathbf{T}^{-1}(D)$.  What is the determinant?

$\det {\mathbf{T}(D)} = 1$,
$\det {\mathbf{T}(D)} = D$,
$\det {\mathbf{T}(D)} = 1 + D^2$.

4

What is true for the equivalent systematic transfer function matrix?

The first row of  $\mathbf{G}_{\rm sys}(D)$  is  $(1, \, 0, \, 0)$.
The second row of  $\mathbf{G}_{\rm sys}(D)$  is  $(0, \, 1, \, 1 + D)$.
The second row of  $\mathbf{G}_{\rm sys}(D)$  is  $(0, \, 1, \, 1/(1 + D))$.

5

Are the two given circuits actually equivalent?

YES.
NO.


Solution

(1)  Here  $\underline{k = 2}$  and  $\underline{n = 3}$   ⇒   Rate  $\underline{R = 2/3}$.

  • The memory order  $\underline{m = 1}$  $($number of memory elements per input$)$.
  • The influence length is equal to the sum of all memory elements   ⇒   $\underline{\nu = 2}$.


(2)  The information bit  $u_i^{(1)}$  affects only the first output  $x_i^{(1)}$,  while  $u_i^{(2)}$  is used for  $x_i^{(2)}$  and  $x_i^{(3)}$.

$${ \boldsymbol{\rm G}}_0 = \begin{pmatrix} 1 & 0 & 0\\ 0 & 1 & 1 \end{pmatrix} \hspace{0.05cm}. $$
  • The delayed inputs affect as follows:
  • $u_{i–1}^{(1)}$  affects  $x_i^{(1)}$,
  • $u_{i–1}^{(2)}$  affects  $x_i^{(1)}$  and  $x_i^{(2)}$:


  • Thus,  the partial matrix  $\mathbf{G}_1$  and the transfer function matrix  $\mathbf{G}(D)$:
$${ \boldsymbol{\rm G}}_1 = \begin{pmatrix} 1 & 0 & 0\\ 1 & 1 & 0 \end{pmatrix} \hspace{0.3cm}\Rightarrow \hspace{0.3cm} { \boldsymbol{\rm G}}(D) = { \boldsymbol{\rm G}}_0 + { \boldsymbol{\rm G}}_1 \cdot D = \begin{pmatrix} 1+D & 0 & 0\\ D & 1+D & 1 \end{pmatrix} \hspace{0.05cm}. $$
  • Therefore the  proposed solutions 1 and 3  are correct.
  • Answer 2 cannot be correct,  because no element with  $D^2$  can occur in the transfer function matrix when  $m = 1$.
  • $\mathbf{G}(D)$ is moreover a $2 × 3$ matrix; there is no third row.


(3)  Splitting  $\mathbf{G}(D)$  gives the  $2 × 2$  matrix.

$${ \boldsymbol{\rm T}}(D) = \begin{pmatrix} 1+D & 0 \\ D & 1+D \end{pmatrix} \hspace{0.3cm}\Rightarrow \hspace{0.3cm}{\rm det}\hspace{0.1cm}{ \boldsymbol{\rm T}}(D) = (1+D) \cdot (1+D) = 1+D^2 $$
$$\Rightarrow \hspace{0.3cm}{ \boldsymbol{\rm T}}^{-1}(D) = \frac{1}{1+D^2} \cdot \begin{pmatrix} 1+D & 0 \\ D & 1+D \end{pmatrix} \hspace{0.05cm}. $$
  • The correct solution is  solution 3.  For control:
$${ \boldsymbol{\rm T}}(D) \cdot { \boldsymbol{\rm T}}^{-1}(D) \hspace{-0.15cm} \ = \ \hspace{-0.15cm} \frac{1}{1+D^2} \cdot \begin{pmatrix} 1+D & 0 \\ D & 1+D \end{pmatrix} \cdot \begin{pmatrix} 1+D & 0 \\ D & 1+D \end{pmatrix} =$$
$$ \ = \ \hspace{-0.15cm} ... \hspace{0.1cm}= \frac{1}{1+D^2} \cdot \begin{pmatrix} 1+D^2 & 0 \\ 0 & 1+D^2 \end{pmatrix} = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}\hspace{0.05cm}. $$


(4)  According to the data sheet applies:

$${ \boldsymbol{\rm P}}(D) \hspace{-0.15cm} \ = \ \hspace{-0.15cm} { \boldsymbol{\rm T}}^{-1}(D) \cdot { \boldsymbol{\rm Q}}(D) = \frac{1}{1+D^2} \cdot \begin{pmatrix} 1+D & 0 \\ D & 1+D \end{pmatrix} \cdot \begin{pmatrix} 0 \\ 1 \end{pmatrix} =$$
$$\ = \ \hspace{-0.15cm} \frac{1}{1+D^2} \cdot \begin{pmatrix} (1+D)\cdot 0 + 0 \cdot 1 \\ D\cdot 0 + (1+D)\cdot 1 \end{pmatrix} = \frac{1}{1+D^2} \cdot \begin{pmatrix} 0 \\ 1+D \end{pmatrix} = \begin{pmatrix} 0 \\ 1/(1+D) \end{pmatrix} $$
$$\Rightarrow \hspace{0.3cm} {\boldsymbol{\rm G}}_{\rm sys}(D) = \begin{pmatrix} 1 & 0 & 0\\ 0 & 1 & 1/(1+D) \end{pmatrix}\hspace{0.05cm}. $$
  • The correct solution is therefore the  proposals 1 and 3.


(5)  Correct is  YES.   The lower circuit on the data sheet is identified by the equations  $x_i^{(1)} = u_i^{(1)}$,  $x_i^{(2)} = u_i^{(2)}$,  and

$$x_i^{(3)}= x_{i-1}^{(3)} + u_i^{(2)} \hspace{0.15cm} \circ\!\!-\!\!\!-^{\hspace{-0.25cm}D}\!\!\!-\!\!\bullet\hspace{0.15cm} X^{(3)}(D)= X^{(3)}(D) \cdot D +U^{(2)}(D)$$
$$\Rightarrow \hspace{0.3cm} G(D) = \frac {X^{(3)}(D)}{U^{(2)}(D)} = \frac {1}{1+D} \hspace{0.05cm}.$$
  • This corresponds exactly to the last element of  $\mathbf{G}_{\rm sys}(D)$  from subtask  (4).