Difference between revisions of "Aufgaben:Exercise 2.7: C Programs "z1" and "z2""
From LNTwww
Line 53: | Line 53: | ||
===Solution=== | ===Solution=== | ||
{{ML-Kopf}} | {{ML-Kopf}} | ||
− | '''(1)''' After the first iteration of the loop $(m = 0)$ the variable $\text{sum = 0.2}$, at the next iteration $(m = 1)$ holds $\text{sum = 0.2 + 0.3 = 0.5}$. | + | '''(1)''' After the first iteration of the loop $(m = 0)$ the variable $\text{sum = 0.2}$, at the next iteration $(m = 1)$ holds $\text{sum = 0.2 + 0.3 = 0.5}$. |
− | *In both cases | + | *In both cases: $\text{sum}$ is less than $x = 0.75$. |
− | *Only | + | *Only at $m = 2$ the return condition is satisfied: $0.9 > x$. Thus $\underline{z1 = 2}$. |
− | '''(2)''' | + | '''(2)''' Correct are the <u>solutions 2 and 3</u>: |
− | * | + | *If one were to dispense with the auxiliary variable $x$ and write in line 8 instead "$\text{sum > random()}$" a new random value would be generated on each iteration of the loop, and $z1$ would then not have the desired properties. |
− | *$z1$ works according to the diagram on the page "Generation of | + | *$z1$ works according to the diagram on the page "Generation of multilevel random variables" in the theory section. There you find a faster implementation for the case of equal probabilities $(1/M)$. |
− | *In the first run $(m = 0)$ | + | *In this case: In the first run $(m = 0)$ the return condition is not satisfied due to the "Lesser/Equal" query; the output value is actually $z1 = 1$. |
− | '''(3)''' | + | '''(3)''' Correct are the <u>solutions 1, 3, and 4</u>: |
− | *It results in a binomially distributed random variable, with | + | *It results in a binomially distributed random variable, with $z2 \in \{0, 1, 2, 3, 4\}$. |
− | *For the calculation of the probability ${\rm Pr}(z2 = 0) = (1 -p)^{I}$ one needs the mathematical library. | + | *For the calculation of the probability ${\rm Pr}(z2 = 0) = (1 -p)^{I}$ one needs the mathematical library. |
− | *But exponentiation could also be realized by $I$& | + | *But exponentiation could also be realized by by multiplying $I$ times. |
− | '''(4)''' Because of line 6, | + | '''(4)''' Because of line 6, before the program loop $(i = 0)$ the field element $\text{p_array[0]}=(1 -p)^{I}$. |
*In the first iteration $(i = 1)$ the following value is entered: | *In the first iteration $(i = 1)$ the following value is entered: | ||
:$$\text{p_array[1]}=\frac{ p\cdot I}{ 1- p}\cdot\text{p_array[0]}= I\cdot p\cdot(1- p)^{ I- 1}={\rm Pr}(z2= 1) .$$ | :$$\text{p_array[1]}=\frac{ p\cdot I}{ 1- p}\cdot\text{p_array[0]}= I\cdot p\cdot(1- p)^{ I- 1}={\rm Pr}(z2= 1) .$$ | ||
− | *In the second iteration $(i = 2)$ the probability for the result "$z2=2$" is calculated: | + | *In the second iteration $(i = 2)$ the probability for the result "$z2=2$" is calculated: |
:$$\text{p_array[2]}=\frac{p\cdot (I- 1)}{ 2\cdot ( 1- p)}\cdot\text{p_array[1]}= \left({ I \atop { 2}}\right)\cdot p^{\rm 2}\cdot( 1- p)^{\rm 2}={\rm Pr}( z2 = 2) .$$ | :$$\text{p_array[2]}=\frac{p\cdot (I- 1)}{ 2\cdot ( 1- p)}\cdot\text{p_array[1]}= \left({ I \atop { 2}}\right)\cdot p^{\rm 2}\cdot( 1- p)^{\rm 2}={\rm Pr}( z2 = 2) .$$ | ||
Latest revision as of 14:33, 29 December 2021
The two C programs given here are suitable for generating discrete random variables:
- The function $z1$ generates an $M$–level random variable with the value set $\{0, 1$, ... , $M-1\}$. The associated probabilities are passed in the array $\text{p_array}$ with property "Float" The function $\text{random()}$ returns equally distributed float–random variables between $0$ and $1$.
- A second function $z2$ (source code see below) returns a special probability distribution specified by the two parameters $I$ and $p$. This is done using the function $z1$.
Hints:
- The exercise belongs to the chapter Generation of Discrete Random Variables.
- In particular, reference is made to the page Generation of multilevel random variables.
Questions
Solution
(1) After the first iteration of the loop $(m = 0)$ the variable $\text{sum = 0.2}$, at the next iteration $(m = 1)$ holds $\text{sum = 0.2 + 0.3 = 0.5}$.
- In both cases: $\text{sum}$ is less than $x = 0.75$.
- Only at $m = 2$ the return condition is satisfied: $0.9 > x$. Thus $\underline{z1 = 2}$.
(2) Correct are the solutions 2 and 3:
- If one were to dispense with the auxiliary variable $x$ and write in line 8 instead "$\text{sum > random()}$" a new random value would be generated on each iteration of the loop, and $z1$ would then not have the desired properties.
- $z1$ works according to the diagram on the page "Generation of multilevel random variables" in the theory section. There you find a faster implementation for the case of equal probabilities $(1/M)$.
- In this case: In the first run $(m = 0)$ the return condition is not satisfied due to the "Lesser/Equal" query; the output value is actually $z1 = 1$.
(3) Correct are the solutions 1, 3, and 4:
- It results in a binomially distributed random variable, with $z2 \in \{0, 1, 2, 3, 4\}$.
- For the calculation of the probability ${\rm Pr}(z2 = 0) = (1 -p)^{I}$ one needs the mathematical library.
- But exponentiation could also be realized by by multiplying $I$ times.
(4) Because of line 6, before the program loop $(i = 0)$ the field element $\text{p_array[0]}=(1 -p)^{I}$.
- In the first iteration $(i = 1)$ the following value is entered:
- $$\text{p_array[1]}=\frac{ p\cdot I}{ 1- p}\cdot\text{p_array[0]}= I\cdot p\cdot(1- p)^{ I- 1}={\rm Pr}(z2= 1) .$$
- In the second iteration $(i = 2)$ the probability for the result "$z2=2$" is calculated:
- $$\text{p_array[2]}=\frac{p\cdot (I- 1)}{ 2\cdot ( 1- p)}\cdot\text{p_array[1]}= \left({ I \atop { 2}}\right)\cdot p^{\rm 2}\cdot( 1- p)^{\rm 2}={\rm Pr}( z2 = 2) .$$
- For $I= 4$ and $p = 0.25$ we get the following numerical value ⇒ "$4$ over $2$" $=6$:
- $$\text{p_array[2]}={\rm Pr}( z 2=2)=6\cdot\frac{1}{16}\cdot\frac{9}{16} \hspace{0.15cm}\underline{=0.211}.$$